spin-torque driven macrospin dynamics subject to thermal noise
TRANSCRIPT
Spin-Torque Driven Macrospin Dynamics subject to
Thermal Noise
by
Daniele Pinna
A dissertation submitted in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
Department of Physics
New York University
January 2015
Andrew D. Kent
Daniel L. Stein
Dedication
To the most spectacular Bird on this side of the Holocene boundary.
iii
Acknowledgements
This doctoral work would not have been possible without the support of my two
advisers: Dan Stein and Andy Kent. They have supplied me with both the directive
and incentive to explore the problems discussed in this thesis. I am indebted to
them for the freedom and the space they have given me to take my research in an
autonomous direction. Witnessing my professional development sifted and molded
by their experienced oversight could not have been more fortuitous for me.
I also thank the Physics department of New York University for accepting
and supporting me throughout these 5.5 years of studentship. The experience
accumulated through both my teaching and research assistanships wouldn’t have
gone by so smoothly without the many people handling all sorts of bureaucracy
and background logistics.
Any journey must come about with significant personal growth. In this respect,
I acknowledge the many graduate colleagues that enabled me to call this Physics
department a “home”. I particularly thank Roberto Gobbetti, Sven Kreiss, and
Colm Kelleher for allowing me to bounce ideas around and discussing the most
random topics that our minds could cook up.
More scientifically speaking, I’d also like to thank the many people that have
defined the Kent Lab throughout my stay. They have supplied me with crucial
context, without which this research work would have never made any sense to me
to begin with.
I would also like to acknowledge friends and family for the immensity of their
love and support throughout these years. It feels as if Graduate school could not
have taken place without your presence allowing me to grow into a more mature
molt of myself. Thank you for existing and being who you are.
iv
Last but not least, I am greatful to the National Science Foundation and IARPA
for providing the funding to support my Ph.D. research. My thesis research was
supported by NSF-DMR-1006575, NSF-DMR-1309202 and by IARPA contract
W911NF14-C-0089.
v
Abstract
This thesis considers the general Landau-Lifshitz-Gilbert theory underlying the
magnetization dynamics of a macrospin magnet subject to spin-torque effects and
thermal fluctuations, as a function of the spin-polarization angle. The macrospin
has biaxial magnetic anisotropy, typical of thin film magnetic elements, with an
easy axis in the film plane and a hard axis out of the plane. We will argue that
when magnetic diffusion due to spin-torque and thermal noise effects happen on a
timescale that is much larger than the conservative precessional timescale due to
material anisotropies, it becomes possible to explore steady-state dynamics per-
turbatively by analytically averaging the magnetization dynamics over constant
energy orbits. This affords the possibility to simplify the magnetization dynamics
to a 1D stochastic differential equation governing the evolution of the macropsin’s
anisotropy energy. Current induced steady-state motions are then found to appear
whenever the magnetization settles onto a stable constant energy trajectory where
a balance of spin-torque and damping effects is achieved: with the remaining gyro-
magnetic motion due to anisotropy fields driving precessions. After averaging, all
the relevant dynamical scenarios are found to depend on the ratio between hard
and easy axis anisotropies. We derive analytically the range of currents for which
in-plane and out-of-plane limit cycles exist and discuss the regimes in which the
constant energy orbit averaging technique is applicable. We find that there is a
critical angle of the spin-polarization necessary for the occurrence of such states
and predict a hysteretic response to applied current. This model can be tested
in experiments on orthogonal spin-transfer devices, which consist of both an in-
plane and out-of-plane magnetized spin-polarizers, effectively leading to an angle
between the easy and spin-polarization axes.
vi
Contents
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiv
List of Appendices xxv
Introduction 1
1 General Formalism 7
1.1 Magnetic Energies . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.1 Exchange Energy . . . . . . . . . . . . . . . . . . . . . . . . 11
1.1.2 Dipolar Energy . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.1.3 Magnetocrystalline/Shape Anisotropy Energy . . . . . . . . 14
1.1.4 Magnetostatic and Zeeman Energy . . . . . . . . . . . . . . 16
1.2 Macrospin Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2.1 The Landau-Lifshitz (LL) equation . . . . . . . . . . . . . . 18
1.2.2 Gilbert Damping . . . . . . . . . . . . . . . . . . . . . . . . 20
vii
1.2.3 Slonczewski Spin-Torque . . . . . . . . . . . . . . . . . . . . 21
1.2.4 Thermal Effects . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.3 The Stochastic Landau-Lifshitz-Gilbert Slonczewski Equation (sLLGS) 26
2 Stochastic Calculus 32
2.1 Gaussian Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2 Stochastic Differential Equations . . . . . . . . . . . . . . . . . . . 33
2.3 Interpretation of Stochastic Integrals . . . . . . . . . . . . . . . . . 35
2.4 The Fokker-Planck Equation . . . . . . . . . . . . . . . . . . . . . . 39
2.5 Limits of Stochastic Modeling . . . . . . . . . . . . . . . . . . . . . 41
2.5.1 Gradient Systems . . . . . . . . . . . . . . . . . . . . . . . . 42
2.5.2 Non-Gradient Systems . . . . . . . . . . . . . . . . . . . . . 43
2.6 Stochastic Macrospin Dynamics: Reprise . . . . . . . . . . . . . . . 44
3 Numerical Methods 48
3.1 Euler-Maruyama Scheme . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2 Heun Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3 Strong and Weak Convergence of a Numerical Scheme . . . . . . . . 53
3.4 GPU: The CUDA Environment . . . . . . . . . . . . . . . . . . . . 56
4 Uniaxial Macrospin Model 63
4.1 Collinear Spin-Torque Model . . . . . . . . . . . . . . . . . . . . . . 67
4.1.1 Collinear High Current Regime . . . . . . . . . . . . . . . . 67
4.2 Tilted Spin-Torque Model . . . . . . . . . . . . . . . . . . . . . . . 69
4.3 Thermally Activated Regime . . . . . . . . . . . . . . . . . . . . . . 72
4.4 Switching Time Probability Curves . . . . . . . . . . . . . . . . . . 80
viii
5 Constant Energy Orbit-Averaged Dynamics 85
5.1 Constant Energy Orbit-Averaged (CEOA) dynamics . . . . . . . . . 87
5.2 IP Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.2.1 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . 96
5.2.2 D > D0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.2.3 D < D0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.2.4 CEOA Breakdown . . . . . . . . . . . . . . . . . . . . . . . 104
5.2.5 Thermally Activated Switching . . . . . . . . . . . . . . . . 111
5.3 OOP Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.3.1 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . 116
5.3.2 CEOA Breakdown . . . . . . . . . . . . . . . . . . . . . . . 118
5.3.3 Thermal Stability, Precession Linewidth, Phase, Amplitude
and Power Fluctuations . . . . . . . . . . . . . . . . . . . . 122
5.4 Experimental Verification of OOP orbits . . . . . . . . . . . . . . . 133
5.4.1 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.4.2 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6 Structure of Escape Trajectories 145
6.1 Friedlin-Wentzell (FW) theory . . . . . . . . . . . . . . . . . . . . . 146
6.2 Macrospin Escape Paths . . . . . . . . . . . . . . . . . . . . . . . . 153
Conclusion 158
Appendices 163
Bibliography 172
ix
List of Figures
1 A schematic of the origin of the spin transfer torque. Layers A,B
and C are non-magnetic layers. . . . . . . . . . . . . . . . . . . . . 2
2 Illustration of a spin valve/MTJ device. . . . . . . . . . . . . . . . . 4
1.1 Constant energy trajectories for D = 10. ε < 0 trajectories are
shown in red whereas ε > 0 trajectories are shown in blue. Notice
how two distinct basins exist for positive and negative energy tra-
jectories. The separatrix (corresponding to ε = 0) separating the
different basins is shown in black. . . . . . . . . . . . . . . . . . . . 19
1.2 Relaxation trajectory of the magnetization for a free layer under the
effects of Gilbert damping and the conservative LL torque [65]. . . . 21
1.3 Relaxation trajectory for the magnetization of a free layer under
the effects of Gilbert damping, the conservative LL torque, and
stochastic thermal noise [65]. . . . . . . . . . . . . . . . . . . . . . . 25
x
1.4 A typical trilayer spin-valve consisting of a free magnetic layer sand-
wiched between an in plane magnetized reference layer and out-of-
plane magnetized polarizer layer. The net spin-torque acting on the
free layer will generally appear tilted away from the easy-axis of
the magnet. The advantage of such devices lies in the ability to in-
stantly torque the magnetic free layer without the need of a thermal
incubation time to destabilize the magnetization initially. . . . . . . 29
1.5 Easy-easy x and hard-axis z magnetic anisotropy directions are
shown along with spin-polarization direction np. The spin-polariation
is tilted by an angle ω with respect the magnetic easy axis. . . . . . 30
3.1 Floating-Point Operations per Second for the CPU and GPU [131]. 57
3.2 Memory Bandwidth for the CPU and GPU [131]. . . . . . . . . . . 57
3.3 A schematic view of a CUDA streaming multiprocessor with 8 scalar
processor cores [132]. . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.1 Histogram distribution of mx after letting the magnetic system relax
to thermal equilibrium (103 natural time units). The overlayed red
dashed line is the theoretical equilibrium Boltzmann distribution.
In the inset we show a semilog-plot of the probability vs. m2x de-
pendency. As expected, the data scales linearly with slope equal to
ξ ≡ K/kBT = 80: the ratio between total anisotropy and thermal
energy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
xi
4.2 Current amplitude vs. mean switching time. Blue line shows the
fit of the ballistic limit to the numerical data (in blue crosses). Red
line shows the improvement obtained by including diffusion gradient
terms. Times are shown in units of (T · s) where T stands for Tesla:
real time is obtained upon division by µ0HK . . . . . . . . . . . . . . 70
4.3 mx: green > 0, red < 0 for applied current I = 5. The plane dissect-
ing the sphere is perpendicular to the uniaxial anistropy axis. Its
intersection with the sphere selects the regions with highest uniaxial
anisotropy energy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.4 Mean switching time behavior for various angular tilts and cur-
rents larger than the critical current obtained by numerically solving
(1.29). Each set of data is rescaled by its critical current such that
all data plotted has Ic = 1. Angular tilts are shown in the legend in
units of π/36 such that the smallest angular tilt is 0 and the largest
is π/4. Times are shown in units of (T ·s) where T stands for Tesla:
real time is obtained upon division by µ0HK . . . . . . . . . . . . . . 73
xii
4.5 Mean switching time behavior in the sub-critical low current regime
obtained by numerically solving the full macrospin dynamics. Times
are shown in units of (T · s) where T stands for Tesla: real time
is obtained upon division by µ0HK . The red and green line are
fits to the data with the functional form 〈τ〉 = C exp(−ξ(1 − I)µ),
where µ is the debated exponent (either 1 or 2) and C is deduced
numerically. The red curve fits the numerical data asymptotically
better the green curve. The difference between the red line and
(4.11) is that our theoretical prediction includes a current dependent
prefactor which was not fitted numerically. The differences between
numerical data and (4.10) is due to numerical inaccuracies out to
such long time regimes. The differences between (4.10) and (4.11),
on the other hand, quantify the reach of the crossover regime. . . . 76
4.6 Mean switching time behavior in the sub-critical low current regime
obtained by numerically solving the macrospin dynamics. Various
polarizer tilts are compared by rescaling all data by the appropriate
critical current value. Times are shown in units of (T · s) where T
stands for Tesla: real time is obtained upon division by µ0HK . . . . 78
xiii
4.7 Influence of precessional orbits on transient switching as seen from
the switching time probability curve in the supercritical current
regime. The case shown is that of an angular tilt of π/3 subject
to a current intensity of 2.0 times the critical current. Data was
gathered by numerically solving (1.29). The non-monotonicity in
the probability curve shows the existence of transient switching.
Times are shown in units of (T · s) where T stands for Tesla: real
time is obtained upon division by µ0HK . . . . . . . . . . . . . . . . 83
4.8 Spin-torque induced switching time probability curves for various
angular configurations of polarizer tilt (a sample normalized current
of 10 was used) obtained by numerically solving (1.29). A log-log
y-axis is used following (4.17) to make the tails of the probability
distributions visible. Times are shown in units of (T · s) where T
stands for Tesla: real time is obtained upon division by µ0HK . . . . 84
5.1 Orbital frequencies plotted as a function of ε for differentD. To com-
pare the results, the positive portion of ε axis has been rescaled by
D. Frequency is expressed in units of (GHz/T). Physical frequency
is obtained upon multiplying by µ0HK . The sharp minimum in the
frequency is a result of the precessional period diverging at ε = 0. . 90
5.2 Critical currents versus the ratio of the hard and easy axis anisotropies
D. The blue curve is I1C and the red curve is I0
C . For D < D0, cur-
rents greater than I1C lead to deterministic switching (labeled DS).
For D > D0 currents between I0C and I1
C lead to limit cycles (LC).
Limit cycles can also occur for currents just below and approxi-
mately equal to I1C for D < D0, as shown in Figure 5.5 . . . . . . . 99
xiv
5.3 Three sample regimes of deterministic energy flow ε as a function
of energy for D > D0. Coloring is included to better distinguish
the various curves. (blue thick dashed curve) I < I1C : Subcrit-
ical regime, thermal noise must oppose a positive energy flow to
achieve switching; (green dash-dotted curve) I1C < I < I0
C : Limit
cycle regime; and (red dotted curve) I > I0C : Supercritical regime,
negative flow leads to deterministic switching. . . . . . . . . . . . . 100
5.4 Three sample regimes of deterministic energy flow ε as a function
of energy for D < D0. Coloring is included to better distinguish
the various curves. (blue thick dashed curve) I < I0C : Subcritical
regime, thermal noise must oppose a positive energy flow to achieve
switching; (green dash-dotted curve) I0C < I < I1
C : Crossover
regime, switching is still achieved via thermal activation but the un-
stable equlibirum has now shifted; and (red dotted curve) I > I1C :
Supercritical regime, negative flow leads to deterministic switching. 102
5.5 Energy flow for a D = 4 macrospin and applied current I = 2.82 <
I0C . Circles and squares respectively represent stable and unstable
equilibria. For these parameters (D = 4 and I = 2.82), two stable
equilibria of the zero temperature dynamics coexist. . . . . . . . . . 103
xv
5.6 a) Mean switching time versus current for D = 4, α = 0.04 and
ξ = 80 at different θ angular tilts with φ = 0 kept fixed. All
currents have been rescaled by 1/ cos θ. Times are shown in units of
(s ·T ) where T stands for Tesla: real time is obtained upon division
by µ0HK . For visual guidance, the critical currents I0C , I1
C and
limit current IM have been included. In a regime where the CEOA
technique is applicable, the switching data from the various angular
configurations should all fall on top of each other. b) Double y-axis
plot of max[T (ε)|∂t|ε||] and the percent deviation of data from (a))
as a function of normalized current. In the current range where the
deterministic flow achieves its minimum, the deviation of the data
does also. As the critical current I0C is approached, deviation spikes
are observed analogously to what can be inferred by the theory. . . 108
5.7 a) Mean switching time versus current for D = 50, α = 0.04 and
ξ = 80 at different θ angular tilts with φ = 0 kept fixed. All
currents have been rescaled by 1/ cos θ. Times are shown in units
of (s · T ) where T stands for Tesla: real time is obtained upon
division by µ0HK . For visual guidance, the critical currents I0C , I1
C
and limit current IM have been included. b) Double y-axis plot
of max[T (ε)|∂t|ε||] and the percent deviation of data from (a)) as
a function of normalized current. In the current range where the
deterministic flow achieves its minimum, the deviation of the data
does also. As the critical current I0C is approached, deviation spikes
are observed analogously to what can be inferred by the theory. . . 109
xvi
5.8 a) Mean switching time versus current for D = 7, α = 0.04 and
ξ = 80 at different θ angular tilts with φ = 0 kept fixed. All
currents have been rescaled by nz = 1/ cos θ. Times are shown in
units of (s ·T ) where T stands for Tesla: real time is obtained upon
division by µ0HK . For visual guidance, the critical currents I0C , I1
C
and limit current IM have been included. b) Double y-axis plot
of max[T (ε)|∂t|ε||] and the percent deviation of data from (a)) as
a function of normalized current. In the current range where the
deterministic flow achieves its minimum, the deviation of the data
does also. As the critical current I0C is approached, deviation spikes
are observed analogously to what can be inferred by the theory. . . 110
xvii
5.9 a) Scaling dependence of mean switching time as a function of ap-
plied current I for models with varying D < D0. ξ is the energy
barrier height and I1C the critical current threshold for deterministic
switching. b) Fit of (5.45) to the form (1 − I/I1C)β. Dashed lines
represent continuation of analytical results outside the technique’s
regime of validity. Fitting exponent β is plotted as a function of
applied current for models with varying D. In the limit of small
D the exponent approaches the constant value β = 2 consistent
with previous uniaxial macrospin results [136,155,159]. For D > 0,
the exponent β depends nonlinearly on the applied current inten-
sity. Only for values D ∼ D0 do we notice that in the limit of
small applied currents, β → 1 as suggested by similar energy diffu-
sion studies from the literature [139, 154]. For intermediate values
D0 > D > 0 the low current limit of β can be obtained analytically
(5.47). On the other hand, in the limit I → I1C the exponent β can
be shown to diverge for all non-zero values of D. . . . . . . . . . . . 113
xviii
5.10 Three regimes of deterministic energy flow ε as a function of energy
for D = 10. (blue-dashed) I < IOOP: Subcritical regime. Energy
flows from positive to negative energy basins due to dynamics being
globally dissipative (overdamped). (red-dashdotted) I > Imax: Su-
percritical regime. Energy flows towards limiting stable value ε = D
due to dynamics being overdriven by applied current. (green-dotted)
IOOP < I < Imax: Oscillator regime. Energy flow will stabilize at
a fixed point corresponding to a precessing oscillator state. In this
regime, the fixed point represents a constant energy trajectory where
spin-torque and damping effects balance. . . . . . . . . . . . . . . . 119
5.11 Steady-state ensemble energy as a function of dimensionless ap-
plied current I (rescaled by IOOP = (2/π)√D + 1) for a model with
D = 10, ξ = 80 and α = 0.04. Red line shows an analytic fit to
numerical data within the current limits defined by the theory (for
reference Imax/IOOP ≈ 4.97). Insets show density plots in spheri-
cal coordinates of 5120 numerical trajectories for a sample with a
2.56ωC tilt between easy and spin-polarization axes, driven by a cur-
rent of I/IOOP = 4 (top), and I/IOOP = 15 (bottom). The dotted
line denotes the conservative trajectory. . . . . . . . . . . . . . . . . 121
5.12 Standard deviation of the energy distribution plotted as a function of
dimensionless applied current I (rescaled by IOOP = (2/π)√D + 1)
for D = 10, ξ = 80 and α = 0.04. The solid blue line shows
the theoretical prediction (5.64) calculated within the current limits
defined by the theory (for reference Imax/IOOP ≈ 4.97). . . . . . . . 124
xix
5.13 Inverse quality factor (5.66) vs. applied current for D = 10 set at
room temperature (ξ = 80). Red dashed line denotes the upper
bound of the validly of the CEOA formalism: Imax/IOOP ≈ 4.97 for
the parameters chosen. . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.14 Switching probability vs. spin-current pulse length for a macrospin
model with D = 10, ω = 2.12ωC) driven by a spin-current intensity
of I = 2.75 IOOP in the absence of thermal noise. Times are shown
in units of (s · T ) where T stands for Tesla: real time is obtained
upon division by µ0HK . Before the current pulse is switched on,
the magnetic ensemble is taken to be antiparallel to the easy-axis of
the magnetic film. Switching probability is defined as the ensemble
fraction that relaxes into a parallel configuration upon switching
the current pulse off. The right-hand vertical axis plots the evo-
lution of the average 〈mz〉 component. In the absence of thermal
noise the oscillator remains coherent at all times and its periodic
motion is clearly seen. Due to the deterministic nature of the zero-
temperature dynamics, the macrospin will deterministically switch
either into the parallel or antiparallel state at all times. . . . . . . . 130
xx
5.15 Switching probability vs. spin-current pulse length for a macrospin
model with D = 30, ω = 3ωC driven by a spin-current intensity
of I = 5 IOOP in the presence of thermal noise corresponding to
ξ = 80 (left) and ξ = 1200 (right). Times are shown in units
of (s · T ) where T stands for Tesla: real time is obtained upon
division by µ0HK . Before the current pulse is switched on, the
magnetic ensemble is taken to be antiparallel to the easy-axis of the
magnetic film. Switching probability is defined as the fraction of the
ensemble that relaxes into a parallel configuration upon switching
the current pulse off. For long pulse times the switching probability
converges to a value indicating that the phase of the OOP precession
has decohered. The red dashed lines are a qualitative graphical
representation of the decoherence time. . . . . . . . . . . . . . . . . 131
5.16 Log-log plot of ensemble decoherence time vs. energy barrier height
to thermal energy ratio ξ for a macrospin model with D = 30,
ω = 3ωC driven by a spin-current intensity of I = 1.5 Iswitch. Times
are shown in units of (s · T ) where T stands for Tesla: real time is
obtained upon division by µ0HK . Linear regression (solid lines) of
data points demonstrates a transition between a phase noise domi-
nated regime τdec ∝ 1/T below a certain critical inverse temperature
ξ < ξC . Above ξ > ξC (T < TC), both amplitude and phase noise
contribute to ensemble decoherence. . . . . . . . . . . . . . . . . . . 132
xxi
5.17 Resistance versus in-plane applied field hysteresis loops. The major
loop (black curve) shows the switching of both the free and reference
layers. The minor loop (blue curve) shows the response of just the
free layer. The loop is centered at 41 mT due to dipolar interactions
between the reference and free layer. Inset: Schematic of the spin-
valve’s layer stack showing the out-of-plane magnetized polarizing
layer (OP), in-plane magnetized free layer (FL) and reference layer
(RL). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.18 Differential resistance versus current at various easy axis applied
fields starting from the P state. The magnitude of the current |Idc|
is increased (black curves) and then decreased (red curves). (a) The
black curve shows switching from P to AP at 1.5 mA and also -1.2
mA, i.e. the switching occurs for both polarities of the current. At
larger positive and negative current the resistance change is inter-
mediate of that of the P to AP transitions. On reducing the current
there is a transition from the intermediate resistance (IR) state into
an AP state. (b) At 40 mT switching from P to AP only occurs
for positive polarity current and on reducing the current there is an
IR to AP state transition for |Idc| ≤ 1 mA. (c) At 42 mT switching
from P to AP again only occurs for positive polarity current. How-
ever, on reducing the applied field the transition is from IR to P for
|Idc| ≤ 1 mA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
xxii
5.19 Current swept state diagram of an OST spin valve device showing
the threshold currents for switching as a function of applied easy
axis field. IP−APC and IAP−IRC are labeled by solid and open blue
symbols. IAP−PC and IP−IRC are labeled by solid and open red sym-
bols. The green curves indicate the IIR−APC (crosses) and IIR−PC
(dashes), showing the bistability range of the IR states. . . . . . . . 138
5.20 (a) Representative FL minor hysteresis loops measured with a slowly
swept field at several fixed currents. The scale bar shows ∆RAP−P =
0.1 Ω, the resistance difference between the AP and P states. (b)
State diagram constructed from dV/dI|H hysteresis loops. The
color represents ∆R, the resistance difference between field increas-
ing and field decreasing measurements. The central zone (orange
color) corresponds to the AP/P bistable zone. Black dashed curves
trace the boundaries between P, AP and IR states. . . . . . . . . . 139
5.21 Simulations of an ensemble of 5000 macrospins represented as a
state-diagram with current increasing (a) and current decreasing
(b). The three relevant states AP, P and IR, are color coded as
red, green and blue respectively. Each data point is represented by
a RGB color that is determined by the proportion of the ensemble
populating each corresponding state. Currents are shown in units
of switching positive current at zero field and room temperature.
Applied fields are shown in units of anisotropy field. The parameters
used in the simulation are described in the main text. . . . . . . . . 144
6.1 Diagram of momentum ellipse parametrized by γ. γ0 and π + γ0
correspond to instanton and anti-instanton solutions respectively. . 148
xxiii
6.2 Contour plot of the norm of the drift field taken from the Maier-
Stein model [177]. On the left, for α = 3, two global minima are
present at the unstable (0, 0) and stable (1, 0) equilibria respectively
along with a saddle along the x-axis. On the right, for α = 5, the
previous saddle has become a local maxima due to the appearance
of two new local minima off the x-axis. . . . . . . . . . . . . . . . . 152
6.3 Escape trajectory for a uniaxial macrospin model with j = 0.3 and
ω = 0. Blue line shows the least-action result of numerical integra-
tion of FW dynamics. Dashed red line is the analytical result shown
in equation (6.28). . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.4 Escape trajectory for a uniaxial macrospin model with j = 0.3 and
ω = 0.2π. Blue line shows the least-action result of numerical in-
tegration of FW dynamics. Dashed red line is the analytical result
shown in equation (6.28). . . . . . . . . . . . . . . . . . . . . . . . . 157
xxiv
List of Appendices
Appendix A Constant Energy Orbit Averaging of sLLGS Equation . . . . . 163
Appendix B Orbit averaging of a Stratonovich Equation . . . . . . . . . . . 169
xxv
Introduction
Over the past three decades, the study of spintronics has set the stage for emer-
gent technologies capable of exploiting the intrinsic “spin” of the electron, and its
associated magnetic properties, to operate, probe and manipulate novel solid-state
devices. The first major acknowledgement of this success came with the award-
ing of the 2007 Nobel prize in Physics to Albert Fert [1] and Peter Grunberg [2]
for their discovery of the giant magnetoresistance effect (GMR). In the GMR ef-
fect, the resistance of a magnetic multilayer is found to depend strongly on the
relative magnetization orientations of the various layers composing it. When the
magnetizations are parallel the measured resistance is small, and when the magne-
tizations are antiparallel the resistance is higher. This influence of magnetization
on current flow later led Slonczewski [3] and Berger [4] to suggest that there may
also be an influence of the current on the magnetization. The practical implica-
tions of this line of thoughthas led to great research interest in what are known
as spin-transfer phenomena, where spin-polarized currents are used to manipulate
the magnetic moments of ferromagnetic structures without the need for external
magnetic fields. Particularly, sweeping advances capable of employing the electron
“spin” to store data in switcheable magnetic states have proceeded to define the
backbone of today’s non-volatile data storage industry.
Spin-transfer torque (STT), occurs when a current of spin-polarized electrons
travels through a ferromagnetic material and deposits angular momentum via an
exchange interaction with the ferromagnet’s macroscopic magnetization (see Fig-
ure 1). Due to conservation of the total angular momentum in this interaction, the
net change in the angular momentum of the current flowing through the ferromag-
netic layer must be equal and opposite to the net change in angular momentum of
1
Figure 1: A schematic of the origin of the spin transfer torque. Layers A,B and Care non-magnetic layers.
the ferromagnetic layer itself along with any potential excitations of the underlying
lattice. Moreover, since angular momentum is proportional to magnetization, with
a conversion factor given by the gyromagnetic ratio γ, the absorption of angular
momentum results in a change in the magnetization direction of the ferromagnet.
Since its experimental confirmation, the study of STT has grown into a thriving
field of both experimental [5–10] and theoretical research [11–17]. Over the last
decade STT has shown tremendous practical potential as a data storage technology
via its ability to quickly read and write information into high density, non-volatile
memory structures by reversing the orientation of magnetic bits. It has also shown
promise in its ability to drive magnetic spin-torque oscillators (STO) at GHz fre-
quencies with narrow linewidths, which has a variety of uses [18–25], as well as
its ability to move magnetic domain walls, which may have uses in magnetic data
storage and logic devices [26–32].
From a practical standpoint, the typical material structures consist of a sandwich-
2
like layering (called a stack) of different nanometer thick magnetic metals separated
by a non-magnetic layer made of either a metallic conductor (spin valve) or a thin
insulator through which electrons can tunnel quantum mechanically (magnetic
tunnel junction). The various layers can generally differ in the coercive fields they
exhibit, with one layer being fabricated to possess a particularly weak coercive field
leading it to be “softer” in terms of its preferred magnetic orientation. This layer
is known as the free layer. The ability to tune, switch and control the magnetic
orientation of the free layer lies at the basis of all spin-transfer technologies. The
other layers in the stack, are built to retain a fixed magnetic orientation. If on one
hand they serve to spin-polarize the charge current traversing through the device,
on the other, they allow for the experimental determination of the free layer’s rela-
tive orientation via the GMR effect. As a result, fixed layers are also often referred
to as polarizer or reference layers.
Figure 2 presents an illustration of a spin valve/MTJ device. When electrons
are made to flow through the device from the bottom towards the top, the electrons
become spin-polarized by the reference layer. When this transmitted spin-current
encounters the free layer it induces a spin-torque on the free layer which tries
to align the free and fixed layer magnetizations. Alternatively, by applying the
current in the opposite direction (from top towards the bottom), the direction of
the spin-torque can be reversed. In the latter case, electrons which have a net
spin-polarization opposite to the direction of the fixed layer’s magnetization, will
be reflected off the fixed layer’s interface inducing a spin-torque on the free layer.
The resulting effect tries to anti-align the free and fixed layer magnetizations [16].
The spin-torque also allows for a number of dynamical regimes.
For data storage applications, the free layer serves as the fundamental unit
3
Figure 2: Illustration of a spin valve/MTJ device.
of memory, the magnetic bit, and operation involves using spin-torque to switch
the orientation of the free layer from one metastable state to another. This use of
STT was confirmed shortly after its theoretical prediction [33,34] with much of the
early research focusing on understanding the effects of magnetic anisotropy [35,36],
temperature [37, 38], and spin-current strength [39, 40] on the switching dynam-
ics. Another important application involves using spin-torque effects to generate
large-angle steady-state magnetic excitations in the free layer [41, 42]. Persistent
magnetization oscillations have been predicted to be used as wide-band tunable
RF oscillators [43] operating in the GHz to THz frequency range. To these ends, it
is of importance to understand the physics of current induced magnetic excitations
in the presence of noise with attention to the parameters that determine the tun-
ability of quality factors in these systems. Overall, whereas much of the research
into these phenomena has focussed on experimental and micromagnetic simula-
tions [44], a simple theoretical model for describing the richness of the observed
free layer dynamics has been lacking.
4
This thesis is organized as follows. In Chapter 1, we will outline the macrospin
model used to describe the free layer dynamics throughout this thesis. The model
approximates the free layer as an evolving, single domain, magnetic spin subject to
anisotropic, spin-torque and thermal effects. As this results in the simplest possible
reduction of the true microscopic degrees of freedom existing in reality, it forms
the basis for spin-torque driven magnetic phenomena. A proper understanding of
its dynamical regimes will represent a platform upon which the true complexity
of experimental phenomena can be quantified and tested. Due to the stochastic
nature of the model’s dynamics and the non-conservative quality of its dynamical
flow, Chapter 2 will then summarize a few basic results regarding the stochastic
modeling of physical systems. One of the main successes of the work underlying this
thesis has been the development of an efficient and massively parallelized code for
the numerical simulation of the dynamics derived in Chapter 1. It’s peculiarity lies
in the use of computer graphics cards (rather than CPU and computational cluters)
to perform all numerical integrations. This allowed study of the ensemble evolution
of the stochastic macrospin dynamics by focusing on the underlying Langevin
behavior rather than solving a partial differential equation for the evolution of the
probability density of magnetization states over time. Chapter 3 will outline both
the numerical methods used and the hardware employed for their execution.
Numerical methods in hand, the macrospin dynamics will be analyzed in depth
throughout Chapters 4 and 5. Chapter 4 will focus on the simplest anisotropic
scenario: the uniaxial macropin model. In this chapter we will derive the statis-
tical properties underlying thermally activated switching phenomena. The main
result of this chapter will be the resolution of a long standing debate regarding the
scaling of mean switching time with applied current. In particular we show that
5
the orientation of the polarized current relative to that of the macrospin’s prefer-
ential magnetization axis does not significantly influence the thermally activated
dynamics. In Chapter 5, we will develop an averaging theory based on time-scale
separation to significantly reduce the complexity of the macrospin’s dynamics. We
will derive a one-dimensional stochastic differential equation describing the evo-
lution and diffusion of the macrospin’s energy over its energy landscape. After
initially verifying that the results presented in Chapter 4 can be accurately repro-
duced with this technique, we will proceed to study the various dynamical regimes
accessible by a macrospin presenting a more general biaxial anisotropy energy
landscape. The main result of this chapter will consist in a surprisingly simple
set of physical conditions relating the spin valve’s spin-torque efficiency and the
magnetic anisotropy strengths to predict whether a macrospin should be expected
to switch or behave like a spin-torque oscillator. We argue that these details are
of importance to both researchers and industry professionals alike in establish-
ing guidelines for the development of practical devices. Finally, in Chapter 6, we
will briefly touch upon the properties of thermally driven dynamical systems sub-
ject to non-gradient flow. We do so to explore the structure and topology of the
macrospin’s random transitions between equilibrium orientations. We will argue
that the spin-torque’s non-gradient nature cannot be expected to alter the most
probable escape paths significantly.
6
Chapter 1
General Formalism
The precise derivation of the term magnet, which has now become the most
common one, is difficult to ascertain. Lucretius (99-55 BC) says it was called
“magnet” from the place from which it was obtained in the native hills of the
Magnesians. However, Pliny the Elder (23-79 AD) relates a more colorful legend,
as copied from the poet Nicander (second century B.C.), that the shepherd Magnes,
while guarding his flock on the slopes of Mount Ida, suddenly found the iron crook
of his staff clinging to a stone, which has become known after him as the ‘Magnes
stone’, or magnet. [45]
What is certain, and well known, is that the first quantitative investigations of
electric and magnetic phenomena were performed by Charles-Augustin de Coulomb
(1736-1806) [46]. Their unified theoretical description was then obtained by James
Clerk Maxwell (1831-1879) [47] whereby the dynamical behavior of electric and
magnetic fields were characterized on a macroscopic length scale. However, on an
atomic length scale, quantum theory must be employed for the proper microscopic
description of the physical properties of matter to emerge.
7
The investigation of magnetic processes in the fine ferromagnetic particles we
will be interested in is, however, on a somewhat intermediate level. The size of the
particles is on the order of nanometers or micrometers leading to the necessary con-
sideration of magnetic domain formation, for which Maxwell’s equations will not
be a sufficient description. For one, effects which originate from the atomic struc-
ture of solids have to also be take into account. Magnetocrystalline anisotropy, for
example, is caused by the crystal lattice: the periodical positions of the atoms com-
posing the solid. Furthermore, the exchange interaction between spin momentum
of electrons is a typical quantum mechanical effect.
However, employing full quantum mechanical models to study the properties of
fine ferromagnetic particles is impossible with today’s available computing powers.
One is forced to ‘neglect’ quantum mechanics, ignoring the atomic nature of mat-
ter, and consider a semi-classical approximation in a continuous medium. Such a
theory, was first introduced by Landau and Lifshitz [48] in an attempt to study the
structure of the domain wall between two anti parallel magnetic domains. William
Fuller Brown contributed several works to the discipline and is responsible for
naming the theory micromagnetics [49]. He wanted to emphasize the fact that this
theory should be capable of describing the details of the walls which separate mag-
netic domains as opposed to domain theory which instead considers the domains,
neglecting the walls in between.
In outlining the elements of the micromagnetic theory, this chapter will ulti-
mately attempt to channel the reader’s interest towards the behavior of the finer
magnetic particles, whose low nanometer size scale can be approximated further
as a single magnetic domain known as a Stoner-Wohlfarth model. The derived
physical principles governing the behavior of such a system will allow us to avoid
8
the full spatially extended detail formalized by the micromagnetic theory. This
‘reduced’ theory will, however, retain very rich dynamical properties which have
found fertile applications in describing recent experimental developments in the
field of spin-transfer phenomena.
1.1 Magnetic Energies
To attempt a thermodynamics theory of spontaneous magnetism, we must as-
sume that the magnetic moment M of a material together with any applied external
field H and temperature T , provide a complete characterization of the states of our
magnetic system. It is then known that the thermodynamic potential controlling
spontaneous transformations under fixed H and T is the Gibbs free energy G. If
H and µoM (with µ0 the permeability of free space) are conjugate work variables,
denoting by F the Helmholtz free energy, one has that
G(M,H, T ) = F − µ0M ·H, (1.1)
and thermodynamics equilibrium is reached when G attains its globally minimum
value [50]. In our magnetic system, the internal degrees of freedom, which give
rise to spontaneous transformations, can be represented by the magnetic moment
M itself. It will be assumed to vary during internal microscopic processes on a
characteristic relaxation time which is much shorter that the time scale over which
M varies significantly due to thermodynamic forces driving the system towards
global equilibrium. This implies that the system relaxes by passing through a
sequence of non equilibrium states, each characterized by a well-defined value of
M.
9
The equation of state of the system for the conjugate work variables H and M
will then be given by
H =1
µ0
(∂F
∂M
)T
, µ0M = −(∂G
∂H
)T
. (1.2)
So far we have not considered the dependence of the magnetization on spatial
coordinates. Magnetic materials, can exhibit very complex magnetization patterns.
We, thus, subdivide our ferromagnetic body into elementary volumes large enough
to contain a statistically significant number of atoms and yet small enough with
respect to the typical length scale over which the magnetization varies significantly.
The assumption which led to the definition of the Gibbs free energy implies, then,
that the relaxation time over which individual volumes reach thermal equilibrium
with respect to the given value of the magnetization M(r) is much shorter than
the time over which the system as a whole approaches equilibrium through time
changes of M(r).
There are five significant contributions to the Gibbs free energy of a ferromag-
netic body: the exchange energy, the dipolar energy, the magnetocrystalline/shape
anisotropy energy, the magnetostatic energy, and the Zeeman energy due to the
interaction with the external field H. [51] We omit the magnetoelastic energy,
which arises from magnetostriction for two reasons. When a ferromagnet is mag-
netized, it shrinks (or expands) along the direction of magnetization resulting in a
volume change and, with it, the saturation magnetization defined as the magnetic
moment per unit volume. In micromagnetics, however, it is a basic assumption
that the saturation magnetization remains constant. Secondly, a large part of the
10
magnetostrictive effects can be expressed in the same mathematical form as the
magnetocrystalline anisotropy. If the anisotropy constants are extracted from ex-
periment, as they often are, all magnetostrictive effects are virtually included for
free, thus allowing us to ignore its effects explicitly in our treatment.
1.1.1 Exchange Energy
The Coulomb interaction between electrons with overlapping orbits of neigh-
boring atoms can be very strong on short length scales. In ferromagnets (antifer-
romagnets), the minimum Coulomb energy will be achieved when their orbits are
anti-symmetric (symmetric) causing their spins to align (anti-align) due to Pauli’s
exclusion principle. This interaction can have a field strength as strong as 100
Tesla for neighboring spins. It’s effects are typically captured by introducing the
Heisenberg Hamiltonian of the exchange energy, which is usually written in the
form
Hex = −M∑i,j=1
JijSi · Sj, (1.3)
where Jij is the exchange integral (taken to be positive for ferromagnets and
negative for antiferromagnets), which can be calculated using quantum mechan-
ics [52,53]. It decreases rapidly with increasing distance between atoms with spin
operator S, allowing the sum to be taken only over nearest neighbors. Writing J for
Jij and replacing the spin operators by classical vectors, we rewrite the exchange
energy as
Eex = −JS2∑i,j|i 6=j
cosφi,j. (1.4)
11
Further developing the cosine into its Taylor series expansion for small φi,j and
noting that, in this limit |φi,j| ' |mi −mj| ' |(r · ∇)m| with ri the displacement
vector between neighboring lattice points (where we define from now on m =
M/MS), the exchange energy can be written as
Eex = JS2∑NN
φ2i,j = JS2
∑i
∑ri
[(r · ∇)m]2 . (1.5)
Changing the summation over i to an integral over the entire ferromagnetic
body, we finally get:
Eex =
∫V
A[(∇mx)
2 + (∇my)2 + (∇mz)
2] d3r. (1.6)
The exchange constant appearing is A = JS2c/a, where a is the distance between
nearest neighbors and c = 1, 2, 4 for a simple cubic, body centered cubic and
face centered cubic crystal structure. The typical length scale over which the
exchange energy acts, known as the exchange length, is intimately related to the
exchange constant: lex =√
2A/µ0MS. Within the range of the exchange length,
the exchange interaction dominates over magnetic body and spin directions can
be expected to not change significantly. For common magnetic materials such as
Co, Fe, Ni and Permalloy, the exchange lengths are all several nanometers in size.
Assuming that the size of the magnetic element we wish to model is on the order
of the exchange length scale, we can safely assume that the element will behave as
a single coherent domain with Eex = 0.
12
1.1.2 Dipolar Energy
As the exchange interaction weakens with the size of the magnetic sample, it
begins to compete with an opposing effect: the dipolar interaction. At large dis-
tances, atomic spins see themselves as magnetic dipoles. Contrary to the exchange,
the dipolar interaction will attempt to make spins align or anti-align with each
other depending on their relative position. The minimum energy configuration of
the magnetic texture will be such that a balance of short distance alignment and
long distance anti-alignment takes place. The magnetic sample can be expected
to form domains, where within each domain the spins are all oriented in the same
direction while the spin directions in different domains may be different.
Without burdening ourselves with a long derivation, we limit ourselves to stat-
ing the continuum representation of the dipolar energy, namely [54]:
Edipole =
∫V
1
2Hd(r) ·m(r)d3r (1.7)
where Hd is known as the demagnetization field. It is the sum of the dipolar
interaction between the magnetic moment at position r and all other positions r′.
It will necessarily depend on the shape and surface of the magnetic material:
Hd(r) = ∇(∫
V
∇ ·m(r)
|r− r′|d3r′ −
∮S
n(r′) ·m(r′)
|r− r′|d2r′
). (1.8)
The same consideration reserved for the exchange interaction applies for the
dipolar interaction also. If we wish to study nanometer size magnetic elements with
strong exchange interactions, the system is expected to behave monodomain-like
and dipolar effects can be disregarded.
13
1.1.3 Magnetocrystalline/Shape Anisotropy Energy
The Heisenberg “exchange” Hamiltonian is completely isotropic and its energy
levels do not depend on the direction in space in which the crystal element is magne-
tized. In the absence of other interaction terms, its magnetization direction would
not have a preferred orientation. Real magnets, however, are not isotropic, and
exhibit preferential magnetic orientations. The most common type of anisotropy
is the magnetocrystalline anisotropy, which is caused by the spin-orbit interaction
of the electrons. The structure of electron orbitals is, in fact, closely linked to the
crystallographic structure making the net magnetic moment prefer to align along
well-defined crystallographic axes. As such, a magnetic material will be magnetized
more easily along certain spatial directions than others. The magnetocrystalline
anisotropy is usually small compared to the exchange energy. The preferred di-
rection of magnetization, however, will be entirely determined by this anisotropy
along with any corrections due to the element’s shape: shape anisotropy.
Even though the spin-orbit interactions leading to the anisotropy can be eval-
uated from first principles [55], it is typically easier to consider it as arising phe-
nomenologically from the crystal symmetries (via power series expansions) and
extracting the relevant coefficients from experiment. In hexagonal crystals, the
anisotropy energy is a function of only one parameter, the angle between the mag-
netization and the preferential axis: known as the c-axis or easy-axis. Experiments
show that it is typically symmetric with respect to the basal plane of crystal (lead-
ing to vanishing odd power contributions of cos θ in the power series expansion
which are not allowed by time reversal invariance). To second order, we can then
write the anisotropy energy as
14
Eani =
∫V
[−K1 cos2 θ(r′) +K2 cos4 θ(r′)
]d3r′
=
∫V
[−K1m
2x(r′) +K2m
4x(r′)]
d3r′, (1.9)
where the easy-axis is chosen to be parallel to the x-axis. It is known from experi-
ment, that terms of higher order, and in most cases even K2, are small. If K1 < 0,
then the easy axis is in the basal plane of the crystal whereas if K1 > 0 the easy-
axis will be out of the plane of the crystal. The actual shape of the sample will
further influence the anisotropy energy due to dipolar interaction between spins
preferring that the magnetization reside in the crystal plane. This is known as
shape anisotropy. In some cases, the net effects of shape and magnetocrystalline
anisotropies can be written in a mathematical form as
Eanitot =
∫V
[N1m
2x(r) +N2m
2y(r) +N3m
2z(r)
]d3r. (1.10)
If we further suppose that the sample behave like a monodomain m(r) ≡ m
for r in the material, and that its magnetization is normalized |m| = 1, the total
anisotropy can be rewritten as:
Eanitot = K(Dm2z −m2
x), (1.11)
where K > 0 is the easy-axis (uniaxial) total anisotropy energy, and we further take
the convention of referring to the z-axis (the hard-axis) as that least energetically
favorable for the magnetization orientation. D is then the dimensionless ratio of
hard- and easy-axis anisotropies.
15
1.1.4 Magnetostatic and Zeeman Energy
Lastly, a magnetic potential will be induced whenever a magnetic body is in the
presence of an externally applied field H. This will attempt to align the sample’s
magnetization along it. Its contribution is known as the Zeeman energy, and can
be written as
EZ = −µ0MS
∫V
m ·H d3r. (1.12)
For a strongly magnetized body, even in the absence of external fields, a magneto-
static energy is generated by the presence, within the magnetic body, of an internal
field HD (known as the demagnetization field) which acts in the opposite direction
to the overall magnetization of the body itself. A strong magnetic body will typi-
cally develop magnetic domain structures in order to minimize this magnetostatic
energy. It’s form is derived from Maxwell’s equations and results in:
ED = −µ0MS
2
∫V
m ·H d3r, (1.13)
where for a monodomain magnet, this contribution is included in the shape anisotropy.
1.2 Macrospin Dynamics
Due to the linearity of Maxwell’s equations, the superposition principle allows
for the addition of all the terms contributing to the Gibbs free energy G. A
full micromagnetic treatment will generally employ all those discussed thus far.
However, as hinted in the first part of this chapter, this work will be interested in
modeling a Stoner-Wohlfarth monodomain magnetic body with magnetization M
16
of constant magnitude MS. The body is assumed to have a size lm along the y
direction, and size a in both the x and z directions. The total volume of the object
is then V = a2lm. The energy landscape experienced by M is generally described
by only two of the terms described: the Zeeman energy due to any externally
applied field Hext, and the magnetocrystalline/shape anisotropy energy with easy-
and hard-axes chosen along the x and z directions respectively. We take advantage
of the sample’s constant magnetization to define a normalized unit magnetization
vector m = M/MS.
The total energy landscape of the macrospin can then be written as:
E(m) = KPm2z −Km2
x −MSVm ·Hext, (1.14)
where KP = µ0M2SV and K = (1/2)µ0MSV HK are the hard- and easy-axis
anisotropy energies respectively with µ0HK the Stoner-Wohlfarth switching field
(in units of Tesla) fixing the coercivity of the magnetic model. In the spirit of the
previous sections, we rescale the magnetic energy by K, allowing us to simplify
the energy landscape of the model into
ε = E(m)/K =[Dm2
z −m2x − 2h ·m
], (1.15)
where we define h = Hext/HK and D ≡ KP/K = MS/HK as the ratio of the
anisotropies. Such an energy landscape, in the absence of applied fields, generally
selects stable magnetic configurations parallel and anti-parallel to x.
17
1.2.1 The Landau-Lifshitz (LL) equation
The torque ~τ acting on a physical system will relate directly to the rate of
change in angular momentum L of the system by Newton’s second law: ∂tL = ~τ .
The torque acting on a magnetic moment M subject to a magnetic field H is in
turn ~τ = M ×H. Since the magnetic moment of a point particle is linked to its
angular momentum by the gyromagnetic ratio1 M = −γL, we must have
ΓLL =dm
dt= −γm×H (1.16)
as the equation of motion for the element’s magnetic moment (known as the
Landau-Lifshitz (LL) equation).
If we are interested in the dynamical properties and time evolution of the
macrospin’s magnetization, we must consider its precession in an effective magnetic
field generated by (1.15). The effective interaction field Heff is then given by
Heff = − 1
µ0MSV∇mE(m) = −HK [Dmzz−mxx− h] . (1.17)
The dynamics describe an undamped precession of the magnetization vector m
about the effective field’s direction with Larmor frequency ωL = γµ0|Heff |. As
Heff is the gradient of the energy, the motion of the magnetization described by
(1.16) preserves the energy of the macrospin. As a result, under the effect of the
LL torque, the magnetization m travels along closed contours of constant energy
(see Figure 1.1).
1γ = gµB
~ ' 1.76× 1011 rads·T with g ' 2 is the Lande factor and µB the Bohr magneton.
18
Figure 1.1: Constant energy trajectories for D = 10. ε < 0 trajectories are shownin red whereas ε > 0 trajectories are shown in blue. Notice how two distinct basinsexist for positive and negative energy trajectories. The separatrix (correspondingto ε = 0) separating the different basins is shown in black.
19
1.2.2 Gilbert Damping
From experiments, however, it is known that changes in the magnetization
decay in finite time. For instance, in hysteresis curve measurements, applying a
sufficiently strong field to a magnetic material causes the magnetization to saturate
along the direction of the field. This is in stark contrast to the simple precessional
motion described by the LL torque (1.16) and indicates that dissipative effects con-
tribute to the magnetization dynamics. The basic mechanism driving the damping
is due to electron scattering [56, 57], although other mechanisms have also been
considered in the literature. These range from magnon-induced currents [58], to
magnon-magnon scattering [59], spin-orbit coupling [60], and spin-shot noise [61].
The simplest phenomenological way to incorporate these effects into the magneti-
zation dynamics was suggested by Gilbert [62]:
ΓD =α
MS
m× m (1.18)
with dimensionless Gilbert damping parameter α. This Gilbert damping (GD)
torque term in the equation of motion leads to the decay of the LL precessions and
aligns the magnetization along the effective magnetic field. The strength of the
Gilbert dissipation is proportional to the phenomenological dimensionless damping
constant α. In modern nanomagnetic devices its value is can be as small as α =
0.01 [33, 63, 64], allowing for hundreds of precessional cycles of the magnetization
about the effective field prior to equilibration (see Figure 1.2).
The resulting dynamics m = ΓLL + ΓD, arising from both damping and pre-
cession about an effective field, is known as the Landau-Lifshitz-Gilbert (LLG)
equation. They are sufficient to describe the evolution of a magnetization subject
20
Figure 1.2: Relaxation trajectory of the magnetization for a free layer under theeffects of Gilbert damping and the conservative LL torque [65].
to applied fields in real world scenarios. It is however not yet suited to capturing
effects due to thermal noise and spin-transfer torques. We now proceed to outline
both.
1.2.3 Slonczewski Spin-Torque
In 1996 both Slonczewski [3] and Berger [4] proposed a novel way of manipu-
lating the magnetization of nanoscale magnetic objects by passing a spin-polarized
current through the magnetic layer. The resulting effect has become known as
spin-torque (ST), or spin-transfer torque (STT), due to the spin-angular momen-
tum that is transferred from the spin-polarized current to the magnetic moment.
A simple explanation for this effect is as follows. Spin-polarized electrons entering
the free layer find themselves either aligned (with the amplitude ∝ cos(θ/2)), or
anti aligned (with the amplitude ∝ sin(θ/2)) with the free layer magnetization di-
rection, where θ is the angle between the current’s polarization and the macrospin’s
magnetization axes. At the interface, two main quantum processes contribute to
21
transferring ~ angular momentum from the itinerant electron to the macroscopic
magnetization of the free layer [66]. The first is the result of spin-dependent reflec-
tion/transmission due to differing energy bands of the ”spin-up” and ”spin-down”
conducting electrons across the interface. The other is generated by classical de-
phasing of spins in the ferromagnet due to their differing spin-dependent preces-
sional frequency around the local magnetization texture. When summed over all
Fermi surface electrons, these processes reduce the transverse component of the
transmitted and reflected spin currents to nearly zero for most systems of interest
within a short span of the ferromagnet/spacer layer interface. It follows that, to a
good approximation, the torque on the magnetization is proportional to the trans-
verse component of the incoming spin current and that maximal effects should be
most prominent in thin magnetic layers.
The corresponding non-conservative term in the macroscopic equation of mo-
tion takes the form
ΓS = −γj [m× (m× np)] , (1.19)
where spin-torque effects are assumed to be brought about by a flow of current
polarized in the direction np proportional to the spin-angular momentum deposited
per unit time j = (~/2e)ηJ/µ0MSHKd, with η = (J↑ − J↓)/(J↑ + J↓) the spin-
polarization factor of incident current J and thickness d of the magnetic free layer.
Here the spin-current is represented by an effective magnetic moment j = jnp
whose orientation and magnitude can depend on the actual state of the macrospin’s
magnetization. For the purpose of this work, though, we will assume it to be a
static quantity that acts on the magnetization dynamics. In passing, we note that
the self-induced magnetic field of the current is ignored here as the dimension a,
22
characterizing the thickness of the magnetic system, is considered to be smaller
than 1000A and spin-current effects are expected to dominate over any current
induced magnetic fields at this scale.
The form of the spin-torque effects thus state that the amount of angular mo-
mentum absorbed by the macrospin is proportional to the orthogonal component
of the spin-polarization axis np. As we will soon show, this exchange of angular
momentum can act to pump (subtract) energy into (from) the system depending
on the polarization of the spin-current in relation to the easy-axis x of the free
layer.
An additional “field like” torque may result from non-equilibrium spin-accumulation
in the free layer
ΓFST = γσj (m× np) , (1.20)
which causes the free layer magnetization to precess about the direction of the spin-
current polarization np, and σ represents the relative strength of the “field like”
torque compared to the STT. This spin-accumulation arises from the fact that the
transverse components of the spin-current persist with a characteristic relaxation
length λS, which may vary from only a few angstroms to several nanometers [67].
While some experimental [68] and theoretical [69] literature has suggested the
strength of this additional “field like” torque can be comparable in strength to the
STT given by (1.19), most have found σ 1 [70–74]; thus it will not be considered
in the remainder of this thesis.
23
1.2.4 Thermal Effects
The deterministic description of the previous sections is incomplete, especially
for small enough magnetic domains, due to the absence of thermal noise. Noise
must accompany the presence of Gilbert damping for the fluctuation-dissipation
theorem (FDT) to be satisfied. This was first realized by W.F. Brown [75]. Ther-
mal effects are included by considering uncorrelated fluctuations in the effective
interaction field: Heff → Heff + Hth. We model the stochastic contribution Hth by
specifying its correlation properties, namely:
〈Hth〉 = 0
〈Hth,i(t)Hth,k(t′)〉 = 2Cδi,kδ(t− t′). (1.21)
Thermal noise manifests itself as random fluctuation of the magnetization away
from the deterministic trajectory (see Figure 1.3). At low energies it acts to push
the magnetization away from the easy-axis, giving the free layer positive average
initial energy and an average initial deviation from the easy-axis. This effect will
be critical in some of the switching dynamics discussed.
We also note that current flow is a source of shot noise, which at low frequencies
acts like a white-noise source in much the same way as thermal noise [76]: j →
j+δjS, where δjS(t) is an isotropic Gaussian random component of the spin-current
with
〈δjS,i(t)δjS,k(t′)〉 = 2C(θ)δi,kδ(t− t′), (1.22)
24
Figure 1.3: Relaxation trajectory for the magnetization of a free layer under theeffects of Gilbert damping, the conservative LL torque, and stochastic thermalnoise [65].
where the noise correlator is dependent on the angle between the spin-polarization
axis np and the magnetization m. The noise correlator C(θ) is calculated in [14]
as
C(θ) =~jSF(θ)
2MSVcoth
(eΦ
2kBT
)(1.23)
where Φ is the voltage bias across the magnetic layer and jSF(θ) is the spin-flip
current given by
jSF(θ) =~Φ
4eVMS
[GP sin2
(θ
2
)+GAP cos2
(θ
2
)], (1.24)
where GP and GAP are the conductances for the electrons with spin parallel and
anti parallel to the free layer respectively. Since the parallel conductance is al-
ways larger than the anti-parallel conductance (GP > GAP ), the spin-shot noise is
strongest when the fixed and free layers are anti-aligned. It is therefore interesting
to understand when this additional source of noise plays a role.
25
For a magnetic layer coupled to unpolarized leads, the current induced noise
on the magnetization dynamics was found to be ΓL/ΓR(1+ΓL/ΓR)2
Φ [77], where Φ is the
voltage drop across the magnetic layer, while ΓL/ΓR is a dimensionless ratio char-
acterizing the coupling strength of the magnetic layer to the left (L) and right (R)
leads. Thus the noise is maximal (Φ/4) for perfectly symmetrical couplings, and
is smaller in the limit of highly asymmetric contacts. This basic behavior, and the
order of magnitude of the effect, is not likely to be modified by polarized leads. We
argue that the temperatures at which experiments have been performed, current
noise effects are not important. For an all metallic device, such as a spin-valve
nanopillar, the couplings are nearly symmetrical and, at the critical current, a
typical voltage drop across the magnetic layer is less than 10 mV or, equivalently,
1 K. For a magnetic tunnel junction device Φ can be ∼ 1 V. However, in this case
the coupling is asymmetric. One lead (L) forms a magnetic tunnel junction with
the nanomagnet, while the other (R) a metallic contact. This gives ΓR/ΓL > 104
and a relevant energy ∼ 1 K, again far lower than room temperature. Therefore,
since experiments are typically performed at room temperature where T = 300K,
where thermal noise dominates, only thermal noise is considered hereafter.
1.3 The Stochastic Landau-Lifshitz-Gilbert Slon-
czewski Equation (sLLGS)
The contribution to the magnetization dynamics brought about by thermal
noise and spin-torque will significantly alter the magnetization dynamics. The
original LLG equation m = ΓLL + ΓD used to describe the dynamics will be
modified into:
26
m =− γm× (Heff + Hth) + α (m× m)
− γjm× (m× np) , (1.25)
where the first term proportional to the gyromagnetic ratio γ is conservative and
precessional in nature, the second is dissipative and attributed to damping with
strength α, while the third represents a non-conservative forcing due to spin-torque
effects brought about by a flow of current polarized in the direction np. Having
assumed that the system magnetization is constant in magnitude will then imply
that m will evolve over the surface of a unit sphere and satisfy the condition that
m ·m = 0. This notion can be used to solve for m and rewrite the Landau-Lifshitz
dynamics in the more convenient form:
m =− γ′m× (Heff + Hth)− αγ′m× [m× (Heff + Hth)]
− γ′jm× (m× np) + γ′αjm× np, (1.26)
where γ′ = γ/(1 +α2) is known as the Gilbert ratio, Heff is given by (1.17) and we
have essentially rewritten the dynamics by solving explicitly for m. In passing, we
note that the manipulation just performed has given rise to the “field like” torque
γ′αjm× np, which acts like an applied external field pointed along the direction of
the spin-current polarization. Due to the magnitude of the leading factor α, this
torque is typically small and can be ignored for most purposes.
The macrospin’s geometry is fully determined by two angles: ω the angle be-
tween the spin-polarization np and x easy-axis and the azimuthal angle ψ char-
acterizing the extent to which np, x and z are coplanar (see Fig. 1). A tilted
27
spin-polarization axis allows modeling a spin-torque that results from more than
one “polarizing” layer in a spin-valve (or MTJ) stack or, more generally, a free
layer that has an easy-axis tilted relative to the spin-polarization axis. This is par-
ticularly relevant to experiments employing a perpendicular polarizer layer with an
in-plane magnetized spin-valve, consisting of a free and reference layer [63, 78–83]
(see Fig.1.4). The two fixed layer contributions lead to a net spin-torque which
can be formally thought to arise from a tilted spin polarizer [84–87]. In this case,
the effective spin-polarization will be tilted with respect to the easy-axis of the
free layer. Without loss of generality, taking for simplicity the ‘reference’ and ‘po-
larizer’ magnetization to be aligned with the x- and z- axes respectively (collinear
with the anisotropy axes), the net spin polarization axis can be written as:
j = jnp =j√
η2ref + η2
pol
(ηref x + ηpolz) , (1.27)
where ηref and ηpol are the spin polarization factors of the two magnetic layers
taken separately. The effective tilt angle ω can then be written in terms of the
ratio of the spin-torque efficiencies ω = atan(ηpol/ηref). The net polarization factor
appearing in (1.19) will then be η = 1.
For the main body of this work, externally applied magnetic fields will not be
considered. Considerations regarding their effects will be given in the conclusion.
With this in mind, we write the deterministic drift terms appearing in (1.26):
28
Figure 1.4: A typical trilayer spin-valve consisting of a free magnetic layer sand-wiched between an in plane magnetized reference layer and out-of-plane magnetizedpolarizer layer. The net spin-torque acting on the free layer will generally appeartilted away from the easy-axis of the magnet. The advantage of such devices liesin the ability to instantly torque the magnetic free layer without the need of athermal incubation time to destabilize the magnetization initially.
29
Figure 1.5: Easy-easy x and hard-axis z magnetic anisotropy directions are shownalong with spin-polarization direction np. The spin-polariation is tilted by an angleω with respect the magnetic easy axis.
mx = Dmzmy
+ α[(Inx +mx)(1−m2
x) + Inzmx(√
1− n2mz − nmy) +Dmxm2z
]my = −(D + 1)mxmz
+ α[mymz(Inz
√1− n2 +Dmz)−mxmy(Inx +mx) + Inzn(1−m2
y)]
mz = mxmy
− α[(Inz√
1− n2 +Dmz)(1−m2z) +mxmz(Inx +mx) + Inznmzmy
].
(1.28)
where nx = cosω, nz = sinω and n = sinψ. One of these equations is neces-
sarily redundant due to the |m| = 1 restraint.
In general, however, taking into account all stochastic contributions due to
thermal noise, the full magnetization dynamics will read:
30
mi = Ai(m) +Bik(m) Hth,k, (1.29)
where the expressions for the drift vector A(m) and diffusion matrix B(m) terms,
written in vectorial form, are
A(m) = m× heff − αm× (m× heff)
− αIm× (m× np)− α2Im× np, (1.30)
Bik(m) =√C[−εijkmj − α(mimk − δik)]. (1.31)
Where we define I = j/(αµ0MSHK), heff = Heff/HK and introduced the natu-
ral timescale τ = γ′µ0HKt. We will refer to this set of three, coupled, stochas-
tic differential equations describing the magnetization dynamics as the stochastic
Landau-Lifshitz-Gilbert-Slonczewski (sLLGS) equation. The notation chosen for
expressing the stochastic contribution ‘Hth,k’ implies that noise is to be inter-
preted in the sense of Stratonovich calculus [88] whose meaning will be presented
and justified in the following chapter.
31
Chapter 2
Stochastic Calculus
As we have seen in the previous chapter, the effect of thermal fluctuations can
be introduced in the macrospin dynamics by adding a random fluctuation field
to the effective magnetic field. A typical magnetization trajectory can then be
obtained by integrating the equations of motion. In this chapter we discuss the
mathematical framework under which one can model numerically a differential
equation with stochastic contributions.
2.1 Gaussian Noise
When one considers the effect of temperature on a dynamical system, it is as-
sumed that the resulting perturbations are of very high frequency. “Very high”
means that the timescale of the fluctuations is much smaller than the typical
timescale of the deterministic portion of the dynamics. In our case that the fre-
quency is expected to be well above the typical precession frequency of the mag-
netization vector. Thus, the fluctuating field used to simulate thermal activation
is chosen to be represented by a stochastic process. Furthermore, it is assumed
32
to be Gaussian white noise because the fluctuations emerge from the interaction
of the magnetization with a large number of independent microscopic degrees of
freedom, themselves subject to equivalent stochastic processes: phonons, conduct-
ing electrons, nuclear spins, etc. [89] As a result of the central limit theorem, the
fluctuation field is Gaussian distributed.
A stochastic process η(t) is called Gaussian white noise [90], if its two time
covariance is delta function correlated:
〈η(t)〉 = 0 (2.1)
〈η(t)η(t′)〉 = 2Cδ(t− t′) (2.2)
The Fourier transform of the covariance function can be shown to be indepen-
dent of frequency
F (ω) =
∫ds〈η(t)η(t+ s)〉 exp(iωs)
= 2C
∫dsδ(s) exp(iωs)
= 2C (2.3)
due to the absence of a correlation time.
2.2 Stochastic Differential Equations
Let us consider, for simplicity, a 1D stochastic differential equation (also called
Langevin equation or SDE) with additive noise:
33
x(t) = A(x(t), t) + η(t). (2.4)
The first term appearing in the equation is the deterministic drift term and its
effect on the dynamics of x(t) are perturbed by the noisy diffusion term η(t) which
is a Gaussian random variable.
The increase dx of the dynamical variable over an infinitesimal time step will
then be (to first order)
dx(t) = A(x(t), t
)dt+ dW (t) (2.5)
where we define
dW (t) =
∫ t+dt
t
dt′η(t′). (2.6)
Interpreting the above integral as the limit of a sum, dW will also be a Gaussian
random variable due to it being a sum of Gaussian random variables. Thus we must
have 〈dW (t)〉 = 0 and, employing the covariance properties of η(t), the variance
of dW will be
〈dW (t)2〉 =
∫ t+dt
t
dt1
∫ t+dt
t
dt2〈η(t1)η(t2)〉
=
∫ t+dt
t
dt1
∫ t+dt
t
dt22Cδ(t1 − t2)
= 2Cdt. (2.7)
If intervals [t, t+dt] and [t′, t′+dt] were to not overlap, we must have 〈dW (t)dW (t′)〉 =
0.
A first basic result of stochastic calculus should now be apparent to the reader.
34
The second moment of dW (t) is linear in dt. On the other hand, dW (t) is only of
order√
dt, which can be stated clearly by writing
dW (t) =√
2Cη(t)√
dt. (2.8)
In mathematics, dW (t) is known as a Wiener process and, by slight abuse of nota-
tion, stochastic differential equations such as (2.5) are often conveniently written
in the following form:
x(t) = A(x(t), t
)+ W (t). (2.9)
2.3 Interpretation of Stochastic Integrals
Let us now complicate the matter further by considering a scenario with mul-
tiplicative noise as opposed to simple additive noise:
x(t) = A(x(t), t
)+B
(x(t), t
)η(t), (2.10)
where from now on, for notational clarity, we assume that the variance of η(t) has
been absorbed inside B(x(t), t
)such that 〈η(t)η(t′)〉 = δ(t − t′). The increment
dx over a short time interval dt will then be:
dx(t) =
∫ t+dt
t
dt′A(x(t′), t′
)+
∫ t+dt
t
dt′B(x(t′), t′
)η(t′). (2.11)
If B(x(t), t) and η(t) were both continuous, the second integral term could be
shown to equal B(x(t), t)η(t)dt in the limit of small dt according to the first inte-
gral mean-value theorem. The noise η(t), however, has been chosen by definition
35
to represent a stochastic process and is not continuous. It is therefore not clear at
what time ti ∈ [t, t+ dt] the dynamical variable x(ti), appearing in the multiplica-
tive factor B(x(ti), t), is to be computed. Choosing a convention for computing∫ t+dt
tdt′B(x(t′), t′)η(t′) amounts to selecting a specific type of calculus that one
wishes to work in, resulting in a stochastic integral which can generally be written
as
∫ t+dt
t
dt′B(x(t′), t′)η(t′) = B(x(t) + αdx(t), t
) ∫ t+dt
t
dt′η(t′), (2.12)
with α ∈ [0, 1]. The integral of η(t) can now again be interpreted as defining a
Wiener process dW (t) allowing us to write (2.11) implicitly as:
dx = A(x+ αdx, t
)dt+B
(x+ αdx, t
)dW (t), (2.13)
where from now one we drop the explicit temporal dependence of x(t) for notational
convenience. Expanding the rhs to first order in dx(t) will then result in
dx = A(x, t)dt+ α∂xA(x, t)dtdx
+B(x, t)dW (t) + α∂xB(x, t)dxdW (t)
+O(dx2) (2.14)
which, upon substituting (2.13) back in to obtain an expansion in powers of dt one
finds
36
dx = A(x, t)dt+B(x, t)η(t)√
dt
+ αB(x, t)∂xB(x, t)η2(t)dt
+O(dt3/2)
=[A(x, t) + αB(x, t)∂xB(x, t)η2(t)
]dt+B(x, t)η(t)
√dt, (2.15)
where we have once again used (2.8) to express the Wiener process in terms of the
time interval explicitely. The above equation shows how the particular choice of α
will generally result in an additional drift term, which contains both α and η2(t).
The latter can be replaced by 1 for terms up to the order of dt. The increment dx
must then satisfy the two following noise averaged statistics:
〈dx〉 = [A(x, t) + αB(x, t)∂xB(x, t)] dt (2.16)
〈dx2〉 = B2(x, t)dt (2.17)
In particular, setting α = 0, we get
dx = A(x, t)dt+B(x, t)η(t)√
dt, (2.18)
known as the Ito interpretation of the stochastic differential equation. This is often
indicated by writing (2.10) in the form
x(t) = A(x, t) +B(x, t) · W. (2.19)
Setting α = 1/2, we get
37
dx =
[A(x, t) +
1
2B(x, t)∂xB(x, t)
]dt+B(x, t)η(t)
√dt, (2.20)
known as the Stratonovich interpretation of the stochastic differential equation.
This is often indicated by writing (2.10) in the form
x(t) = A(x, t) +B(x, t) W. (2.21)
We observe at once that Ito and Stratonovich interpretations differ from each
other only through the additional noise induced drift term:
1
2B(x, t)∂xB(x, t). (2.22)
This implies that one is free to switch from one calculus convention to the
other simply by taking into account the corresponding diffusion drift term. As
an example, to convert a Stratonovich SDE into its corresponding Ito form, one
simply writes:
x = A(x, t) +B(x, t) W
= A(x, t) +1
2B(x, t)∂xB(x, t) +B(x, t) · W. (2.23)
Due to the different drift terms, the two interpretations can be expected to yield
different dynamical properties. In the mathematics community, the Ito calculus
is most commonly used due to its conceptual simplicity arising from the property
that noise increments∫ t+dt
tdsη(s) and x(t) are statistically independent as implied
by (2.12): 〈B(x, t)η(t)〉 = 0 [91]. On the other hand, in the physics community,
38
the Stratonovich interpretation is preferred due to its deeper physical origin. Since
the noise term in (2.10) models, in a coarse-grained sense, the effect of microscopic
degrees of freedom that have finite (albeit short) correlation times τ such that
〈η(t)η(t′)〉 ∝ 2Cτ exp[−|t− t′|/τ ], (2.24)
this term should be physcally interpreted as the limit in which these correlation
times go to zero. The Wong-Zakai theorem [92] then says that in the zero cor-
relation time limit, the colored noise becomes white noise and we obtain the
Stratonovich interpretation of a SDE driven by white noise.
2.4 The Fokker-Planck Equation
The net effect of a dynamic driven by (2.10) will consist in a deterministic flow
due to A(x)1 accompanied by a random diffusion in configuration space due to
stochastic contributions. Since no two trajectories can be expected to be identical,
it is often of interest to know the probability density of finding the dynamical sys-
tem in a particular dynamical state at a given moment in time: ρ(x, t). Assuming
that the system is initially identified with a precise initial condition x(t0) = x0,
one must then have ρ(x0, t0) = δ(x − x0). We now proceed to derive an equation
governing the evolution of ρ(x, t) as a result of an underlying stochastic dynamic
driving the evolution of x(t).
We start by considering the continuity equation for ρ(x, t) expressed in integral
form
1From now on we will omit the explicit time dependence unless necessary.
39
ρ(x, t+ dt) =
∫dx′ρ(x, t+ dt|x′t)ρ(x′, t), (2.25)
where ρ(x, t+ dt|x′t) is the conditional probability distribution of x at time t+ dt
given that it was x′ at time t. It is defined by:
ρ(x, t+ dt|x′t) = 〈δ [x(t+ dt)− x]〉x′t, (2.26)
where the average is over the random noise η(t) and x(t + dt) − x(t) = dx(t) is
determined, to first order, by (2.16). Taylor expanding the conditional probability
around x0 yields:
ρ(x, t+ dt|x′t) = δ(x− x0)− 〈dx(t)〉∂xδ(x− x0)
+1
2〈dx2(t)〉∂2
xδ(x− x0) + ... (2.27)
Using then equation (2.25), we obtain
∂tρ(x, t) = ρ(x, t+ dt)− ρ(x, t)
= ∂x [−A(x)− αB(x)∂xB(x)] ρ(x, t) +1
2∂2x
[B2(x)ρ(x, t)
]= −∂x
[A(x) + (α− 1)B(x)∂xB(x)− 1
2B2(x)∂x
]ρ(x, t), (2.28)
where we show the effects of the calculus α explicitly. This is generally known as
the Fokker-Planck (FP) equation although many mathematicians refer to it also
as the Kolmogorov forward equation [93].
The above procedure can be repeated identically (albeit with a bit more alge-
40
bra) for a multi dimensional SDE:
x = A(x) + B(x) · W, (2.29)
where the dynamical state and deterministic drift are now represented by vectors,
and the stochastic effects are due to a stochastic vector W and state dependent
variance matrix B which need not be square. In writing down the above equation
only we have used “·” to signify simply matrix-vector multiplication instead of a
particular choice of stochastic calculus. The corresponding FP equation for the
evolution of the probability density can then be written as:
∂tρ(x, t) = −∇x ·
G ·[G−1 ·
(A + (α− 1)
(B · ∇x
)· BT
)−∇x
]ρ(x, t)
,
(2.30)
where we have defined G ≡ B · BTas the product between B and its transpose. It
is relevant to note that the FP equation can be written as a standard continuity
equation ∂tρ(x, t) = ∇x · J(x, t) where J is the probability current:
J(x, t) = G ·[G−1 ·
(A + (α− 1)
(B · ∇x
)· BT
)−∇x
]ρ(x, t). (2.31)
2.5 Limits of Stochastic Modeling
The work discussed up to this point is premised on the notion that true physical
systems are expected to be approximated by a SDE similar to (2.10) or (2.29). It
is not immediately apparent though, whether the deterministic drift A(x) should
41
be identical to the physical force field driving the dynamics in the absence of
noise. This is particularly important for a model employing the Stratonovich
interpretation as we have shown that it leads to the appearance of an extra drift-
diffusion term to the deterministic dynamics. Often, for example, one may have
physical intuition as to how the noise may be affecting the system without having
a clear idea as to what the deterministic forces are [94,95]. In this section we will
consider two separate scenarios that may arise when modeling physical systems
with SDEs.
2.5.1 Gradient Systems
Gradient systems correspond to physical phenomena for which a well defined
energy landscape U(x) is known. Equilibrium statistical mechanics then imposes
that, at long times, the physical system’s statistical behavior should be described
by the equilibrium Boltzmann distribution ρ(x, t → ∞) = ρB(x) ∝ exp[−βU(x)]
with β = 1/kBT the inverse temperature. Another property of gradient systems
at equilibrium is that they must obey detailed balance [96]:
ρ(xf tf |x0t0)ρB(x0) = ρ(x0tf |xf t0)ρB(xf ), (2.32)
implying that all thermally induced transitions are just as likely as their reverse
process. The result of this is that, at equilibrium, the probability current J must
vanish identically. Plugging then the equilibrium Boltzmann distribution into
(2.31), drift and stochastic contributions must be such to satisfy:
− ∇xU(x)
kbT= G
−1 ·[A + (α− 1)
(B · ∇x
)· BT
]. (2.33)
42
The scenario is very simple to work out for systems with additive noise G ≡
B = 1, resulting in the drift term satisfying usual gradient dynamics
A = −∇xU/kBT. (2.34)
If, however, the same system is to be modeled with multiplicative noise, equilibrium
conditions are satisfied by the following conditions:
A = A− (α− 1)(B · ∇x
)· BT
(2.35)
G−1 · A = −∇xU/kBT, (2.36)
Implying that the specific form of the stochastic contributions must be known, or
at least suggested, in advance for the system of equations to be solved.
2.5.2 Non-Gradient Systems
Contrary to gradient systems, non-gradient systems do not admit an equi-
librium Boltzmann distribution ρB(x) ∝ exp[−βU(x)] in terms of a precise en-
ergy landscape. An equilibrium distribution can, however, still be mathematically
shown to exist in terms of an effective action or quasi-potential S(x) such that
ρeq(x) ∝ exp[−βS(x)], where the equilibrium stastical properties will not satisfy
detailed balance anymore. Violations of detailed balance will, in turn result in a
non-vanishing equilibrium probability current J which may be observed experimen-
tally in the form of Brownian vortices [97,98]. In this case there is no unambiguous
way to associate the drift A, effective action S(x) or stochastic variance B of the
43
Langevin equation to the experimentally observed probability current. There is
simply no physical intuition to guide the choice of an effective potential such that
limt→∞ρ ∝ exp[−S(x)/kT ] unless the knowledge of both the explicit form of noise
driving the system as well as the drift field are given.
The quasi-potential S(x) may be non-differentiable in certain regions of con-
figuration space leading to highly non-trivial, thermal dynamical behavior. A
typical outcome of this scenario is the appearance of caustics and soliton-like es-
cape trajectories [99–103]. This is of particular importance since knowledge of the
system’s quasi-potential is necessary to estimate escape rates within the framework
of Kramer’s transition-state theory [90,104].
2.6 Stochastic Macrospin Dynamics: Reprise
In developing the theoretical model for the macrospin’s dynamics, we derived
expressions for the drift vector A(m) and diffusion matrix B(m) terms appearing
in
mi = Ai(m) +Bik(m) Hth,k. (2.37)
Written in vectorial form they read:
A(m) = m× heff − αm× (m× heff)
− αIm× (m× np)− α2Im× np, (2.38)
Bik(m) =√C[−εijkmj − α(mimk − δik)], (2.39)
44
where now it should be clear what we mean by Stratonovich multiplicative noise
‘Hth,k’.
The precise noise correlation C is set by the fluctuation-dissipation theorem. In
fact, in the absence of applied currents, the macrospin model is a gradient system
expected to attain thermal equilibrium satisfying a Boltzmann distribution set by
the magnetization energy:
ρB(m) ∝ exp
(−E(m)
kBT
)(2.40)
An associated Fokker-Planck equation (first constructed by Brown [75]) de-
scribing the evolution of the probability density ρ(m, t) on the m-sphere can be
written, as already explained, in the form of a continuity equation for ρ(m, t)
∂ρ(m, t)
∂t= −∇ · J(m, t) (2.41)
with
J(m, t) ≡ Aρ(m, t)−∇m
[2CB · BT
ρ(m, t)]
(2.42)
the probability current (not to be confused with the spin-angular momentum de-
posited by the spin-polarized current discussed in (1.16))2. The divergence and
2The additional drift-diffusion term arising from the Stratonovich convention can be shownto vanish identically. One has (using Einstein summation):
(B · ∇x
)· B
T= Bik
∂Bjk∂mj
= −αmk [εijkmj + α(δik −mimk)]
= −α2(mim
2k − δikmk
)= 0, (2.43)
since m2k = |m| = 1
45
gradient are defined with respect to the magnetization components on the unit m-
sphere. Writing out the FP equation explicitly (in the absence of applied current)
leads to: 3
∂ρ
∂t= −∇m ·
[−m× heff − αm× (m× heff )
+ 2C(1 + α2)m× (m×∇m)]ρ. (2.45)
Plugging (2.40) into the FP equation and noting that:
∇m · (m× heff ) = 0 (2.46)
∇mρ = KkBT
heff ρ, (2.47)
Boltzmann equilibrium is achieved at long times (∂tρ→ 0) with diffusion constant
given by:
C =αkBT
2K(1 + α2)=
α
2(1 + α2)ξ, (2.48)
with ξ ≡ K/kBT the energy barrier height divided by the thermal energy. This
general relationship between variance of thermal effect and damping is known as
the Fluctuation-Dissipation theorem, originally formulated by Nyquist [105], and
later proven by Callen and Welton [106].
3 Given a Langevin equation in the form dX = A(t,X)dt+ B(t,X) · dW(t) the FP equationfor the probability distribution of state variable X, written in component form, is:
∂P
∂t= − ∂
∂Xi
[(Ai −DBjk
∂Bik∂Xj
−DBikBjk∂
∂Xj
)P
](2.44)
46
The validity of the stochastic modeling of the macrospin dynamics is thus
based on our ability to reproduce the Boltzmann equilibrium properties in the
absence of applied currents. The incorporation of non-vanishing current effects
alters an initially gradient-like physical system into a non-gradient one. As already
mentioned, one of the main results of this thesis is the development of a theory
capable of capturing steady-state limit cycle dynamics which imply that a non-
vanishing probability current exists in the macrospin dynamics at long times.
47
Chapter 3
Numerical Methods
The translation of a numerical integration scheme valid for deterministic dif-
ferential equations does not necessarily yield a proper algorithm in the stochastic
case. Depending on the selected deterministic scheme, its unconditional transla-
tion might converge to an Ito solution, to a Stratonovich solution, or to neither of
them. Even when a scheme is shown to converge, usually the order of convergence
is lower than that of its corresponding deterministic counterpart. This has to be
considered when choosing the size of the discretization time step.
We note that the macrospin dynamics introduced, rely on the saturation mag-
netization |m| = 1 remaining constant. This is a very important assumption when
modeling such a system. It in fact allows for the phase space of the dynamics
to be 2D due to the magnetization being constrained on the surface of a sphere
along with being an inertia-less physical system (the differential equations are first
order in time). By the Poincare-Bendixson theorem [107, 108], the deterministic
portion of the evolution will not exhibit any chaotic behavior. This fact is of great
importance when choosing to construct a numerical integration scheme capable of
48
capturing the long time behavior of the sLLGS equation.
The numerical solution of the Fokker-Planck equation (2.30) is a challenging
problem. Being a partial differential equation, the Fokker–Planck equation can be
solved analytically only in special cases. Various approaches have been explored
in the literature for obtaining numerical solutions. Suzuki’s scaling theory [109]
and normal mode analyses [110] have both proved useful for obtaining approx-
imate solutions. However, scaling theory is accurate only to a few percent for
intermediate times (i.e. those between the initial and equilibrium states) in the
case of a bistable system [111] and normal mode analyses may suffer from slow
convergence for general problems. A cumulant moment method has been used
succesfully by Desai and Zwanzig [112] for a nonlinear self-consistent dynamic
mean-field theory model [113]. The slow convergence of the cumulant hierarchy
was later observed in a study of a transient bimodality carried out by Brey, Casado
and Morillo [114]. Path-integral methods have been utilized by a number of au-
thors [115–117]. Wehner and Wolfer [118] have presented a practical formalism
that numerically evaluates the path integrals involving Onsager-Machlup func-
tionals and reduces errors to a few percent. Monte Carlo techniques [119] are
useful for providing information about certain properties of the system in terms
of the moments of the stochastic processes without the need for direct reference
to the probability density distribution. In the case where the entire distribution
function is required, direct approaches, such as those based on an eigenfunction
expansion [120, 121] or finite-difference methods [122–124], are frequently used.
The eigenfunction expansion method is applicable to a general class of linear prob-
lems. Through this approach, various spectral methods can be used to provide
extremely accurate solutions of the Fokker-Planck equation, albeit at the cost of
49
extensive computational time and resources. Furthermore, the finite-difference
method is known to lead often to stiff systems of ordinary differential equations.
All the methods listed typically require advanced parallel computation techniques
running on large computational clusters.
In this work, we will follow a different approach focusing on the direct numer-
ical integration of the underlying Langevin SDE (1.29). Stochastic simulation has
gained acceptance due to its straightforward implementation and robustness with
respect to different sorts of problems. The continuous increase of the efficiency of
available computer hardware has been acting in favor of stochastic simulation, mak-
ing it increasingly more popular. The recent evolution of computer architectures
towards multiprocessor and multicore platforms also resulted in improved perfor-
mance of stochastic simulation. Let us note that in the case of a low-dimensional
system, stochastic simulation often uses ensemble averaging to obtain the values
of observables, which in turn is an example of a so-called “embarrassingly parallel
problem” and it can directly benefit from a parallel architecture. In other cases,
mostly where a large number of interacting subsystems are investigated, the imple-
mentation of the problem on a parallel architecture is less trivial, but still possible.
The recent emergence of general-purpose computing techniques on graphics pro-
cessing units (GPUs) has been a breakthrough in computational science. Current
state of the art GPUs are now capable of performing computations at a rate of
about 5 TFLOPS (Trillion FLOating Point operations per Second) per single sili-
con chip. It must be stressed that 1 TFLOPS, as a performance level, was achieved
in 1996 only throught he use of huge and expensive supercomputers such as the
ASCI Red Supercomputer (which had a peak performance of 1.8 TFLOPS [125]).
The numerical simulations of SDEs can easily benefit from the parallel GPU ar-
50
chitecture. This however requires careful redesign of the algorithms along with a
particular consideration of their convergence properties.
In this chapter we present a practical introduction to solving SDEs numerically
with particular focus to applications involving NVIDIA GPUs using Compute Uni-
fied Device Architecture (CUDA) [126]. In the first two Sections we outline two
practical stochastic integration schemes and discuss their convergence properties
in Section 3. We will then conclude with a section on GPUs and outline code de-
ployment using the CUDA architecture. Unless otherwise stated, we shall consider
a 1D Ito process1 x = x(t), t0 ≤ t ≤ T satisfying a scalar stochastic differential
equation with multiplicative noise
dx(t) = A(x(t), t
)dt+B
(x(t), t
)dW (t) (3.1)
with the initial value x(t0) = x0.
3.1 Euler-Maruyama Scheme
For a given discretization t0 = τ0 < τ1 < · · · < τn < · · · < τN = T of the
time interval [t0, T ], a Euler-Maruyama [127] approximation is a continuous time
stochastic process y = y(t), t0 ≤ t ≤ T satisfying the iterative scheme
yn+1 = yn + A(yn, τn
)∆n +B
(yn, τn
)∆Wn, (3.2)
for n = 0, 1, 2, ..., N−1 with initial value y0 = x0, where yn = y(τn), ∆n = τn+1−τn
denotes the time discretization interval and ∆Wn = Wτn+1−τn is the increment of
the stochastic process. If the diffusion coefficient B ≡ 0, the stochastic iteration
1All schemes presented can be straightforwardly extended to the multidimensional case.
51
scheme (3.2) reduces to the deterministic Euler scheme for the ordinary differential
equation x = A(x, t). The random increments ∆Wn are independent Gaussian
random variables with zero mean 〈∆Wn〉 = 0 and variance 〈(∆Wn)2〉 = 2C(τn+1−
τn).
For the integration of the Langevin equation (1.29) with constant step size ∆t
the Euler-Maruyama scheme results in
mi = mi(t) + Ai(m, t
)∆t+Bik
(m, t
)∆Wk (3.3)
with 〈∆Wk〉 = 0 and 〈∆Wk∆Wl〉 = 2Cδkl∆t. C is the variance of the thermal
field as derived in (2.48) through the Fluctuation-Dissipation theorem.
In the context of Stratonovich stochatic calculus, the deterministic drift has to
be augmented by a noise induced drift-diffusion term (2.22) which gives
mi(t+ ∆t) = mi(t) +
[Ai(m, t
)+ 2C
1
2Bjk
∂Bik
∂mj
]∆t+Bik
(m, t
)∆Wk. (3.4)
3.2 Heun Scheme
The improved Euler-Maruyama or Heun scheme [89] is an example of a predictor-
corrector method. The predictor is given by a simple Euler-Maruyama type inte-
gration. If we consider the Langevin equation (1.29), the predictor is
mi = mi(t) + A(m, t
)∆t+Bik
(m, t
)∆Wk, (3.5)
52
where ∆t is again the discretization time step and ∆Wk are Gaussian random
numbers, whose first two moments are 〈∆Wk〉 = 0 and 〈∆Wk∆Wl〉 = 2Cδkl∆t.
The Heun scheme is then given by
mi(t+ ∆t) = mi(t) +1
2
[Ai(m, t+ ∆t
)+ Ai
(m, t
)]∆t
+1
2
[Bik
(m, t+ ∆t
)+Bik
(m, t
)]∆Wk. (3.6)
3.3 Strong and Weak Convergence of a Numeri-
cal Scheme
The definition of convergence is similar to the concept for ordinary differential
equation solvers, aside from the differences caused by the fact that a solution to a
SDE is a stochastic process, and each computed trajectory is only one realization
of that process. Each computed solution path y(t), using the Euler-Maruyama
scheme for example, gives a random value at T , so that y(T ) is a random variable
as well. The difference between the values of the Ito process and its numerical
approximation at time T , e(T ) = x(T ) − y(T ), is therefore a random variable.
We will say that a time discrete approximation y∆t with maximum step size ∆t
converges strongly to x at time T if
lim∆t→0|〈x(T )− y∆t(T )〉| = 0. (3.7)
We say that a solver converges strongly with order γ > 0 if the error in the moments
is γth order in the step size, or
53
|〈x(T )− y∆t(T )〉| = O (∆tγ) (3.8)
for sufficiently small step size ∆t. Although the Euler method for ordinary differ-
ential equations has order 1, the strong order for the stochastic Euler-Maruyama
scheme for stochastic differential equations is only 1/2 [128].
Strong convergence allows accurate approximations to be computed on an in-
dividual basis. For some applications, such detailed pathwise information may be
required. In other cases, however, the goal is to ascertain the probability distribu-
tion of the solution x(T ), and single realizations are not of primary interest.
Weak solvers seek to fill this need. They can be simpler than corresponding
strong methods, since their goal is to replicate the probability distribution only.
In this respect, we will say that a discrete time approximation y∆t with maximum
step size ∆t converges weakly to x at time T if
lim∆t→0|〈f(x(T )
)− f
(y∆t(T )
)〉| = 0 (3.9)
for all polynomials f(x). According to this definition, all moments converge as
∆t → 0. If the stochastic contributions to the Langevin equation are zero and
the initial value is deterministic, the definition agrees with the strong convergence
definition, and the usual ordinary differential equation definition.
Weakly convergent methods can also be assigned an order of convergence. We
say that a solver converges weakly with order γ > 0 if the error in the moments is
γth order in the step size, or
|〈f (x(T ))− f(y∆t(T )
)〉| = O (∆tγ) (3.10)
54
for sufficiently small step size ∆t.
In general, the rates of weak and strong convergence do not agree. Unlike
the case of ordinary differential equations, where the Euler method has order 1,
the Euler-Maruyama method for SDEs has strong order 1/2. However, Euler-
Maruyama is guaranteed to converge weakly with order 1. If the drift and diffusion
coefficients are almost constant, the Euler-Maruyama scheme gives good numerical
results. In practice, however, this is rarely the case and then the results can become
very poor. Therefore, the use of higher integration schemes is recommended.
The stochastic Heun scheme, on the other hand, converges with strong order
1 and weak order 2 to the solution of the general system of Langevin equations.
There are two main reasons for the choice of the Heun scheme for the numerical
integration of the stochastic sLLGS equation. The first, is that the Heun scheme
can be shown to converge towards Stratonovich solutions of the stochastic differ-
ential equations without alterations to the deterministic drift term. Secondly, the
deterministic part of the differential equations is integrated with a weak second
order accuracy in ∆t, which renders the Heun scheme numerically more stable
than Euler type schemes. In the remainder of this work, we will be interested
solely in weak order results (averages, variances, mean first passage times, etc.)
and, as such, the Heun scheme will prove sufficient for these tasks. Specifically,
we will numerically solve the dynamical macrospin equation (1.29) employing the
fluctuation-dissipation theorem result (2.48) to compute the strength of the ran-
dom kicks perturbing the system. Unless stated otherwise, statistics are gathered
from an ensemble of 5120 events with a natural integration stepsize of ∆τ = 0.01.2
For concreteness, we set the Landau damping constant α = 0.04. Different barrier
2Corresponding to a physical step size ∆t = γ′HK∆τ ' 10−12 − 10−11s depending on thevalue of HK .
55
heights were explored although the main results are shown for a barrier height of
ξ = 80 (unless stated otherwise) which, for experimentally relevant results, corre-
sponds to room temperature [81, 129,130]. To explore the simulations out to long
time regimes, events were simulated in parallel on NVidia Tesla C2050, M2070 and
Quadro FX 580 graphics cards.
3.4 GPU: The CUDA Environment
Driven by the insatiable market demand for realtime, high-definition 3D graph-
ics, the programmable Graphic Processor Unit or GPU has evolved into a highly
parallel, multithreaded, manycore processor with tremendous computational horse-
power and very high memory bandwidth, as illustrated by Figures 3.1 and 3.2.
The reason behind the discrepancy in floating-point capability between the CPU
and the GPU is that the GPU is specialized for compute-intensive, highly parallel
computation and therefore designed such that more transistors are devoted to data
processing rather than data caching and flow control.
CUDA (Compute Unified Device Architecture) is the name of a general purpose
parallel computing architecture of modern NVIDIA GPUs. The name CUDA is
commonly used in a wider context as a means of referring not only to the hardware
architecture of the GPU, but also to the software components used to program that
hardware. In this sense, the CUDA environment also includes the NVIDIA CUDA
compiler and the system drivers and libraries for the graphics adapter.
From the hardware standpoint, CUDA is implemented by organizing the GPU
around the concept of a streaming multiprocessor (SM). A modern NVIDIA GPU
contains tens of multiprocessors. A multiprocessor consists of 8 scalar processors
56
Figure 3.1: Floating-Point Operations per Second for the CPU and GPU [131].
Figure 3.2: Memory Bandwidth for the CPU and GPU [131].
57
Figure 3.3: A schematic view of a CUDA streaming multiprocessor with 8 scalarprocessor cores [132].
58
(SPs), each capable of executing an independent thread (see Figure 3.3). The
multiprocessors have four types of on-chip memory:
• a set of 32-bit registers (local, one set per scalar processor),
• a limited amount of shared memory (48 kB for devices having Compute
Capability3 2.0, shared between all SPs in a MP),
• a constant cache (shared between SPs, read-only),
• a texture cache (shared between SPs, read-only).
The amount of on-chip memory is very limited in comparison to the total global
memory available on a graphics device (a few kilobytes vs hundreds of megabytes).
Its advantage lies in the access time, which is two orders of magnitude lower than
the global memory access time.
The CUDA programming model is based upon the concept of a kernel. A
kernel is a function that is executed multiple times in parallel, each instance run-
ning in a separate thread. The threads are organized into one-, two- or three-
dimensional blocks, which in turn are organized into one- or two-dimensional grids.
The blocks are completely independent of each other and can be executed in any
order. Threads within a block however are guaranteed to be run on a single mul-
tiprocessor. This makes it possible for them to synchronize and share information
efficiently using the on-chip memory of the SM.
In a device having Compute Capability 2.0 or higher, each multiprocessor is
capable of concurrently executing at least 1536 active threads [131]. In practice,
3The compute capability of a device is represented by a version number, also sometimes calledits “SM version”. This version number identifies the features supported by the GPU hardwareand is used by applications at runtime to determine which hardware features and/or instructionsare available on the hardware being accessed.
59
the number of concurrent threads per SM is also limited by the amount of shared
memory and it thus often does not reach the maximum allowed value. The CUDA
environment also includes a software stack. For all CUDA versions greater than
v2.1 (the latest version is v5.2), it consists of a hardware driver, system libraries
implementing the CUDA API, a CUDA C compiler and two higher level mathe-
matical libraries (CUBLAS and CUFFT). CUDA C is a simple extension of the C
programming language, which includes several new keywords and expressions that
make it possible to distinguish between host (i.e. CPU) and GPU functions and
data.
For the macrospin model described by (1.29), we use a single CUDA kernel,
which is responsible for advancing the system by a predefined number of timesteps
of size ∆t. We exploit the parallelization capabilities if the GPU by allowing each
path to be calculated in a separate thread. For CUDA devices, it makes sense to
keep the number of threads as large as possible. This enables the CUDA scheduler
to better utilize the available computational power by executing threads when
other ones are waiting for global memory transfers to be completed [131]. It also
ensures that the code will execute efficiently on new GPUs, which, by the Moore’s
law, are expected to be capable of simultaneously executing exponentially larger
numbers of threads.
Depending on the precise graphics card being used, initial testing was per-
formed to establish the optimal amount of macrospins that the kernel can accept
without losing performance speed. For the case of the Tesla C2050 (448 cores) it
was found that requesting more than 5120 independent realizations via one single
kernel call typically overflowed the GPU memory banks.
The majority of the investigations were typically interested in long time be-
60
havior of the macrospin dynamics. As such, to increase the efficiency of the code,
multiple timesteps were calculated in a single kernel invocation. The results of the
intermediate integration steps did not need to be copied to the host (CPU) mem-
ory. This makes it possible to limit the number of global memory accesses in the
CUDA threads thus, noticeably improving code performance. In fact, for typical
simulations exploring macrospin behavior out to the microsecond timescale (∼ 105
integration time steps), the time required for accessing the global memory (gener-
ally at kernel call and, successively, at kernel exit) was found to be comparable to
the actual time taken by the GPU to evolve the system numerically.
When the CUDA kernel is called, initial magnetizations configurations are
loaded from the global system memory and cached into local variables. All cal-
culations are then performed using these variables and at the end of the kernel
execution, their values are written back to the global memory where they can be
read and written to file. Often an initial kernal call was made with random initial
magnetizations and then evolved for 105 time steps to allow the ensemble elements
to thermalize.
Each path is associated with its own state of the random number generator
(RNG), guaranteeing independence of the noise terms between different threads.
The initial RNG seeds for each thread are chosen randomly using a standard
integer random number generator available on the host system. Since CUDA does
not provide any random number generation routines by default, to generate the
large number of necessary random numbers, we chose a proven combination [133] of
the three-component combined Tausworthe “taus88” [134] and the 32-bit “Quick
and Dirty” LCG [135]. The hybrid generator provides an overall period of around
2121 which is over 60 orders of magnitude larger than the total number of random
61
numbers needed for the production of this thesis. In our kernel, two pairs of
uniform variates are generated per time step and then transformed into Gaussian
variates using the Box–Muller transform. As stated in the previous section, the
integration is performed using a weak order 2 Heun scheme, which uses three of
the four Gaussian variates for a single time step (discarding the fourth).
62
Chapter 4
Uniaxial Macrospin Model
The role of spin-torque can be considered from an energy landscape point of
view. This chapter, we will be dedicated to studyng a uniaxial model in which an
axially symmetric shape anisotropy leads to the presence of a hard-plane. As we
will see, this allows for a detailed exact treatment of the stochastic magnetization
dynamics which will serve as both the starting point for our theoretical develop-
ment as well as a benchmark for the stochastic numerical schemes employed. The
choice of a uniaxial sample corresponds to setting D = 0 or vanishing hard-axis
anisotropy energy (1.11). Such a system is typically realized by using a circu-
lar magnetic domain which suppresses the shape anisotropy of the magnet. The
magnetic energy of the monodomain is:
E(m) = −Km2x. (4.1)
In general, the change in the macrospin’s energy over time can be obtained
after some straightforward algebra starting from (1.26) and is found to be1:
1In deriving we neglect contributions due to the thermal field Hth
63
1
µ0MsV HK
E = − [αm× heff − I(αnp −m× np)] · (m× heff) . (4.2)
This expression shows how current pumps energy into the system. In the absence
of current, the damping dissipates energy and, as one would expect, the dynamical
flow is toward the minimum energy configuration (mx = 1 or mx = −1). The
sign preceding the current term allows the expression to become positive in certain
regions of magnetic configuration space. Furthermore, by averaging over constant
energy trajectories, one can construct an equivalent dynamical flow equation in
energy space. We will develop this approach in the next chapter and use it to
prove the appearance of stable limit cycles by considering which constant energy
trajectories lead to a canceling of the flow E = 0 in (4.2).
Starting from an initially stable magnetic state, spin-torque effects will tend
to drive the magnetization toward the current’s polarization axis (np). Once the
current is turned off, the projection of the magnetization vector along the uni-
axial anisotropy axis (mx) will determine which stable energy state (parallel or
anti-parallel to the x-axis) the magnetic system will relax to as long as the energy
rati ξ ≡ K/kBT is large enough to prevent random fluctuations from acciden-
tally reversing the magnetization on experimentally relevant timescales. As such,
switching dynamics are best studied by studying the x component of equation
(1.29), namely:
mx = α[(I cos(ω) +mx)(1−m2
x) + I sin(ω) cos(ψ)mx(mz − tan(ψ)my)]
+ α2I sin(ω)my +
√α
ξ(1−m2
x) W, (4.3)
where ω and ψ were defined in reference to Figure 1.4, and the multiple stochastic
64
contributions in (1.29) have been compounded by employing general rules for the
addition of random variables (see Appendix B for details). W is, as before, a stan-
dard mean zero, variance 1, Wiener process, and its prefactor explicitly expresses
the strength of the compounded stochastic effects. Whereas, in general, (4.3) is
not useful as it explicitly depends on the dynamics of both the mz and my compo-
nents of the magnetization, it is a convenient dynamical tool in the limit of small
ω where the dynamical equations for mx decouple completely from the other mag-
netization components thus reducing the complexity of (1.29) to a straightforward
1-D problem.
A crucial test of our numerical scheme will be to succesfully reproduce the
thermal equilibrium properties of the uniaxial model. We solve (1.29) directly by
initially assigning to the elements in our ensemble a random magnetization with
negative mx component and allowing the system to relax in the absence of applied
currents. We assume ξ large enough2 such that, before current affects the system,
thermalization will have only been achieved within the antiparallel energy well with
no states having had time to thermally switch to the parallel orientation on their
own. A typical histogram of thermalized magnetic orientations resulting from this
exercise is shown in Figure 4.1 and confirms proper functioning of the numerical
stochastic model by accurately reproducing the expected Boltzmann equilibrium
distribution. Once the system is properly thermalized, we can proceed to turn on
a current and allow the system to evolve for a fixed amount of time. Once this
time has passed, we can let the system relax again at zero current and study what
fraction of the ensemble has switched due to the thermally assisted spin-torque
effects.
2The assumption is valid as long as the mean first passage time for thermally diffusing acrossenergy wells is much larger than the experimental timescale.
65
Figure 4.1: Histogram distribution of mx after letting the magnetic system relax tothermal equilibrium (103 natural time units). The overlayed red dashed line is thetheoretical equilibrium Boltzmann distribution. In the inset we show a semilog-plot of the probability vs. m2
x dependency. As expected, the data scales linearlywith slope equal to ξ ≡ K/kBT = 80: the ratio between total anisotropy andthermal energy.
66
4.1 Collinear Spin-Torque Model
Having derived the necessary expressions for our uniaxial macrospin model’s
dynamics, it is useful to consider the following simplification. Let us take the
uniaxial anisotropy and spin-current axes to be collinear, namely, np ≡ x (or
ω = 0). In such a scenario, the stochastic LLG equation simplifies significantly. In
particular, (4.3) reduces to the simplified form:
mx = α(I +mx)(1−m2x) +
√α
ξ(1−m2
x) W. (4.4)
In this symmetric scenario, as anticipated, magnetization dynamics have been suc-
cesfully reduced to a straightforward 1-D problem. For a general value of I < 1,
the evolution of mx has two local minima and a saddle. The two stable configura-
tions are at mx = −1 and mx = 1, while the saddle is located at mx = −I. For
currents I > Ic = 1 there is only one stable minimum. Above the critical current,
spin torque pushes all magnetic configurations toward the mx = 1 state. This
regime is particularly important not just for its simplicity but also for its similar-
ity to the pure field switching model. The collinear spin-torque model is, in fact,
mathematically identical to a field switching model with applied field of intensity
I applied parallel to the uniaxial anisotropy axis of the magnetic system [109].
4.1.1 Collinear High Current Regime
In the high current regime I Ic we expect the deterministic dynamics to
dominate over thermal effects. We refer to this interchangeably also as ballistic
evolution. The determistic (drift) contribution of (4.4) can then be solved ana-
lytically given an initial configuration mx = −m0. The switching time τswitch will
67
simply be the time taken to get from some mx = −m0 < 0 to mx = 0 and reads:
τswitch(m0) =1
α
∫ 0
−m0
dm
(I +m)(1−m2)
=1
2α(I2 − 1)
I log
[1 +m0
1−m0
]− log
[1−m2
0
]− 2 log
[I
I −m0
]. (4.5)
Since the magnetic states are considered to be in thermal equilibrium before the
current is turned on, one should average the above result over the equilibrium
Boltzmann distribution in the starting well to obtain the average switching time
〈τswitch〉B. Such an initial distribution will be:
ρB(mx) =
√ξ exp[−ξ]F [√ξ]
exp[ξm2x], (4.6)
where F [x] = exp(−x2)∫ x
0exp(y2)dy is Dawson’s integral [158]. This expression
can be used to compute the average switching time numerically.
As the intensity of spin-currents becomes closer to Ic, thermal effects increas-
ingly contribute. Moreover, diffusion gradients add to the deterministic drift, which
can be shown explicitly by writing (4.4) in its equivalent Ito form. Doing so leads
us to a first correction of the ballistic dynamics due to thermal influences. The
x-component behavior then reads:
mx = α(I +mx)(1−m2x)−
α
2ξmx +
√α
ξ(1−m2
x)W (4.7)
The first term on the right hand side is still the ballistic flow that we have just
discussed. The second term is the diffusion-drift term discussed in Chapter 2. The
68
contribution of such a term generates a net motion away from the stable minima of
the ballistic equations as one expects to see under the influence of thermal effects.
Again, we can solve the drift dominated flow analytically to compute the switching
time. Including the effects due to the diffusion-drift term, we find:
τswitch(m0) =1
α
∫ 0
−m0
dm
(I +m)(1−m2)− (m/2ξ)
=1
α
∑j
log [1 + (m0/wj)]
3w2j + 2Iwj − (1− 1
2ξ)
(4.8)
Where the wj are the three zeros of the cubic equation w3+Iw2−(1− 12ξ
)w−I = 0.
As before, the average switching 〈τswitch〉B time will simply be given by averaging
numerically over the Boltzmann distribution ρB. In Figure 4.2, fits are shown
comparing theory to simulation data. As expected both expressions coincide in
the limit of very large currents (I Ic).
4.2 Tilted Spin-Torque Model
In the high current regime (I Ic), where np = x (i.e. the polarizer tilt is
aligned with the x-axis), the ballistic equation for mx was shown to decouple from
the other components, and the dynamics became one dimensional and determin-
istic. For the more general case where the uniaxial anisotropy axis may have any
tilt with respect to the x-axis, such a critical current is not as intuitively defined.
Unlike the collinear limit, a critical current, above which all magnetic states per-
ceive a net flow towards an increasing global mx > 0, does not exist. One can in
fact plot mx over the unit sphere to see what regions allow for an increasing and
69
Figure 4.2: Current amplitude vs. mean switching time. Blue line shows the fitof the ballistic limit to the numerical data (in blue crosses). Red line shows theimprovement obtained by including diffusion gradient terms. Times are shown inunits of (T · s) where T stands for Tesla: real time is obtained upon division byµ0HK .
70
Figure 4.3: mx: green > 0, red < 0 for applied current I = 5. The plane dissectingthe sphere is perpendicular to the uniaxial anistropy axis. Its intersection with thesphere selects the regions with highest uniaxial anisotropy energy.
decreasing projection as the current is changed. An example of this is shown in
Figure 4.3.
Unfortunately, regions characterizing mx < 0 flow can be shown to persist at
all currents. The approach is refined by requiring that on average, over constant-
energy precessional trajectories, the flow is toward the positive uniaxial anisotropy
axis: 〈mx〉 > 0 [35]. Such trajectories are found by solving the flow equations
with I = α = 0. Solutions correspond to circular librations about the uniaxial
anisotropy axis. The critical current is then redefined to be the minimum current
71
at which 〈mx〉 > 0 at all possible precessional energies. This is easily done and
results in:
I ≥ maxε
[−ε
cosω
]=
1
cos(ω)= Ic, (4.9)
thus allowing for a direct comparison of dynamical switching results between dif-
ferent angular configurations with polarizer tilt ω. In our discussion of (4.3) we
mentioned how in the general ω 6= 0 tilted polarizer case there is no way to reduce
the dimensionality of the full dynamical equations. In such a state, precessional
trajectories might allow for a magnetization state to temporarily transit through a
mx > 0, “switched” configuration, even though it might spend the majority of its
orbit in a mx < 0, “unswitched” configuration. This allows for a much richer mean
switching time behavior, especially for currents greater than the critical current,
as shown in Figure 4.4, and discussed more in depth later.
4.3 Thermally Activated Regime
For currents I < Ic, switching relies on thermal effects to stochastically push
the magnetization from one energy minimum to the other. It is of interest to un-
derstand how switching probabilities and switching times depend on temperature
and applied current. This is easily done in stochastic systems with gradient flow.
In such cases an energy landscape exists and a steady state probability distribution
can be constructed via Kramer’s theory for noise-induced escape from a potential
well [90] to derive approximate low-noise switching probabilities.
Unfortunately, though, spin-torque effects introduce a non-gradient term, and
the resulting LLG equation does not admit an energy landscape in the presence of
72
Figure 4.4: Mean switching time behavior for various angular tilts and currentslarger than the critical current obtained by numerically solving (1.29). Each set ofdata is rescaled by its critical current such that all data plotted has Ic = 1. Angulartilts are shown in the legend in units of π/36 such that the smallest angular tilt is0 and the largest is π/4. Times are shown in units of (T · s) where T stands forTesla: real time is obtained upon division by µ0HK .
73
applied currents. The collinear simplification, however, is an exception. As already
described, in the absence of a tilted polarizer the dynamics become effectively
one dimensional since the mx component decouples from the other magnetization
components. Consider then (4.4): because it is decoupled from the other degrees of
freedom, we can construct a corresponding one-component Focker-Planck equation.
The evolution in time of the distribution of mx is then:
∂tρ(mx, t) = L[ρ](mx, t),
where
L[f ] = −α∂mx[(mx + I)(1−m2
x)−1
2ξ(1−m2
x)∂mx
]f.
For high energy barriers and low currents, the switching events from one basin
to the other are expected to be rare. The probability of a double reversal should
be even smaller. We therefore model the magnetization reversal as a mean first
passage time (MFPT) problem with absorbing boundaries at the mx = −I saddle
point. The MFPT will then be given by the solution of the adjoint equation
(L†〈τ〉(mx) = −1) [90]:
α
2ξexp(−ξ(mx + I)2)∂mx
[(1−m2
x) exp(ξ(mx + I)2)]∂mxτ(mx) = −1
subject to the boundary condition 〈τ〉(0) = 0. This can be solved to give:
〈τ〉(mx) =2ξ
α
∫ 0
mx
duexp(−ξ(u+ I)2)
1− u2
∫ u
−1
ds exp(ξ(s+ I)2). (4.10)
The rightmost integral can be computed explicitly. Retaining only dominant terms,
74
the final integral can be computed by a saddlepoint approximation to give:
〈τ〉 ' 2
√π
α
exp(ξ(1− I)2)F (√ξ(1− I))
1− I2. (4.11)
Such a square exponential dependence has been previously derived by Taniguchi
and Imamura [136, 137] as well as Butler et al. [138], although a τ ∝ exp(ξ(1 −
I)) dependence, proposed elsewhere in the literature [27, 72, 139], has also been
successfully used to fit experimental data [129]3.
To decide between these experimental dependences, we fit the scaling behaviors
in Figure 4.5. The square exponential dependence fits the data better, confirming
analytical results. Furthermore, comparison of the asymptotic expression (4.11) to
the full theoretical prediction obtained by solving (4.10) numerically demonstrates
that even for mean switching times of the order 10−6 T · s, asymptoticity still is
not fully achieved.
All that remains is to consider the effects of angular tilt on the switching
properties in the thermally activated regime. Insight into this problem can be
obtained by invoking (4.3) again. For small values of α, the term in square brackets
is of leading order over the second ballistic term depending on my. This allows us,
in the small α regime, to neglect the second ballistic term altogether.
We now concentrate on the behavior of the term in square brackets. For sub-
critial currents, switching will depend on thermal activation for the most part.
We expect an initially anti-parallel configuration to not diffuse very far away from
its local energy minima. It will remain that way until a strong enough thermal
kick manages to drive it over the energy barrier. Because of this, the second term
3It is important to note that experiments at fixed temperature are by necessity performedover a limited range of I/Ic and thus cannot truly distinguish an exponent of 1 or 2
75
Figure 4.5: Mean switching time behavior in the sub-critical low current regimeobtained by numerically solving the full macrospin dynamics. Times are shownin units of (T · s) where T stands for Tesla: real time is obtained upon divisionby µ0HK . The red and green line are fits to the data with the functional form〈τ〉 = C exp(−ξ(1 − I)µ), where µ is the debated exponent (either 1 or 2) andC is deduced numerically. The red curve fits the numerical data asymptoticallybetter the green curve. The difference between the red line and (4.11) is thatour theoretical prediction includes a current dependent prefactor which was notfitted numerically. The differences between numerical data and (4.10) is due tonumerical inaccuracies out to such long time regimes. The differences between(4.10) and (4.11), on the other hand, quantify the reach of the crossover regime.
76
appearing in the square brackets will generally be close to zero as the macrospin
awaits thermal switching. To make the statement more precise, one can imagine
the magnetic state precessing many times before receiving a thermal kick strong
enough to take it over the saddle. The second term can then be averaged over a
constant energy trajectory where the average 〈mz〉 will vanish identically. Hence,
in the subcritical regime, (4.3) can be rewritten in the following approximate form:
mx ' α(I cos(ω) +mx)(1−m2x) +
√α
ξ(1−m2
x) W. (4.12)
This, is reminiscent of the 1D dynamics discussed in relation to the collinear limit,
and shown explicitly in (4.4). The only difference between the two is the substitu-
tion I → I cos(ω). In other words, the thermally activated dynamics are identical
for all angular tilts up to a rescaling by the critical current. We then expect that
the mean switching time dependences remain functionally identical to the collinear
case for all polarizer tilts. We have confirmed this by comparison with data from
our simulations, and the results are shown in Figure 4.6. As predicted, all mean
switching time data from different polarizer tilts collapse on top of each other after
a rescaling by each tilt’s critical current.
As already hinted in the introduction to the collinear model, a strong mathe-
matical analogy exists with the field switching model studied in the literature [140–
144]. By considering a macrospin model with external magnetic field applied
collinear to the uniaxial anisotropy axis one obtains a dynamical equation for
the magnetization vector analogous to (4.4) with the field strength in place of
the applied current. In fact, one can think of writing an effective energy land-
scape U(mx) = −K(m2x − 2Imx), in terms of which, the equilibrium Boltzmann
distribution has precisely the same form as that given by Brown [75]. The ther-
77
Figure 4.6: Mean switching time behavior in the sub-critical low current regimeobtained by numerically solving the macrospin dynamics. Various polarizer tiltsare compared by rescaling all data by the appropriate critical current value. Timesare shown in units of (T · s) where T stands for Tesla: real time is obtained upondivision by µ0HK .
78
mally activated behavior discussed in the literature also reproduces an exponent
2 scaling dependence such as what we have shown here. One must understand,
though, that spin torque effects are generally non conservative and it is only in
the collinear scenario that they may be interpreted in this way. Upon introduc-
ing an angular tilt between uniaxial and spin-polarization axes, the analogy with
the field switching model will break down (see (4.3)). It is interesting, though,
to quantify the crossover between the spin torque and field switched macrospin
model. Coffey [140, 141] has already discussed the effect of angular tilt between
anisotropy and applied switching field axes. We introduced noise in the macrospin
model by considering a random applied magnetic field. In (1.29), we showed the
full form of the dynamical equations. To write the dynamical equations for the
field switched model, we need to suppress current effects and simply introduce a
term identical to the stochastic contribution with the exception that now the ap-
plied field will not be random but fixed at a specific angular separation from the
uniaxial anisotropy axis. Writing the dynamical equations for the field switched
model is straightforward:
mx = α[(h cos(ω) +mx)(1−m2
x) + h sin(ω)(my − αmxmz)]
+
√α
ξ(1−m2
x) W,
where instead of an applied current I, the dynamics depend on an applied field
with strength h. Comparison with (4.3) shows how, in general, the two evolution
equations are very different from each other. In the thermally activated regime,
however, where one is able to average over constant energy trajectories due to their
timescales being much smaller than those required for actual diffusion or magnetic
79
torque (h hcrit), the second term in square brackets will be averaged away and
one is left with a stochastic evolution equation identical to (22). All thermally
activated switching will then again be functionally identical for all angular tilts up
to a trivial rescaling of the applied field.
In comparing our scaling relationships between current and mean switching
time with the previous literature, a subtle issue must be addressed. Results ob-
tained by Apalkov and Visscher [139] rely upon an initial averaging of the dy-
namics in energy space over constant energy trajectories (limit for small damping)
and only subsequently applying weak noise methods to extrapolate switching time
dependences. The small damping and weak noise limits are singular and the or-
der in which they are taken is important. Both limits radically alter the form of
the equations: whereas both limits suppress thermal effects, the first also severely
restrains the deterministic evolution of the magnetic system. Our approach con-
siders the weak noise limit and, only in discussing the effects of an angular tilt
between polarization and easy axes do we employ the small damping averaging
technique to obtain functional forms for the mean switching time. The switching
time data shown seems to justify, in this particular case, an interchangeability be-
tween these two limits. More generally, however, one should not expect the two
limits to commute.
4.4 Switching Time Probability Curves
Up until now, we have analyzed the main properties of spin-torque induced
switching dynamics by concentrating solely on the mean switching times. In exper-
iments, though, one generally constructs full probability curves. The probability
80
that a given magnetic particle has a switching time τswitch ≤ τ can be explicitly
written as:
P [τswitch ≤ τ ] =
∫ m(τ)
0
dxρB(x)
= exp[−ξ(1−m(τ)2)]F [√ξm(τ)]
F [√ξ]
, (4.13)
where m(τ) is the initial magnitization that is switched deterministically in time τ .
Once one has evaluated m(τ), the probability curve follows. Ideally, in the ballistic
regime, one would like to invert the ballistic equations. Unfortunately, though, the
solutions of such ballistic equations are generally transcendental and cannot be
inverted analytically. Even in the simpler collinear case, as can be seen from
equations (4.5) and (4.8), no analytical inversion is possible. One must instead
compute the inversion numerically4. Nonetheless, one can construct appropriate
analytical approximations by inverting the dominant terms in the expressions. In
the case of (4.5), for example, one has that for currents much larger than the
critical current:
τ(m) ' I
2α(I2 − 1)log[
1 +m
1−m] (4.14)
which can be inverted to give:
m(τ) = tanh[ατI2 − 1
I]. (4.15)
Plugging into expression (4.13) for the τ probability curve, one has:
4Easily achieved thanks to the monotonicity of their form.
81
P [τswitch ≤ τ ] = exp
[− ξ
cosh2[ατ I2−1I
]
]F [√ξ tanh[ατ I
2−1I
]]
F [√ξ]
, (4.16)
This expression can be truncated to a simpler form by noting that the leading
exponential term dominates over the ratio of Dawson functions. Furthermore, if
one considers the limit of large values for τ (or, analogously, I 1), the ‘cosh’ can
be also approximated by its leading exponential term. We are finally left with:
P [τswitch ≤ τ ] ' exp
[− ξ
cosh2[ατ I2−1I
]
]
∼ exp
[−4ξ exp[−2ατ
I2 − 1
I]
], (4.17)
which is very similar in form to what has already been reported and used for fitting
experimental data in the literature [35, 129].
In the low current regime, one constructs probability curves by considering the
mean switching time and modeling a purely thermal reversal as a decay process
with rate given by equation (4.11). The fraction of switched states then vary in
time as:
P (mx > 0) = 1− exp(−t/〈τ〉). (4.18)
Upon introducing a tilt polarizer, precessional effects can be seen directly on the
switching probability curves in the super-critical regime. One expects that in the
initial phases of switching, the fraction of switched states is sensitive to the time
at which the current is turned off. One may accidentally turn off the current
82
Figure 4.7: Influence of precessional orbits on transient switching as seen fromthe switching time probability curve in the supercritical current regime. The caseshown is that of an angular tilt of π/3 subject to a current intensity of 2.0 timesthe critical current. Data was gathered by numerically solving (1.29). The non-monotonicity in the probability curve shows the existence of transient switching.Times are shown in units of (T · s) where T stands for Tesla: real time is obtainedupon division by µ0HK .
during a moment of transient passage through the switched region along the pre-
cessional orbit. This was checked and verified from our numerical simulations (see
Figure 4.7). More generally, effects similar to the “waviness” seen in the mean
switching time curves can be seen in the probability curves as well, as the angle of
the polarizer tilt is allowed to vary (see Figure 4.8). Only a numerical solution of
the LLG equation can bring such subtleties to light.
83
Figure 4.8: Spin-torque induced switching time probability curves for various an-gular configurations of polarizer tilt (a sample normalized current of 10 was used)obtained by numerically solving (1.29). A log-log y-axis is used following (4.17) tomake the tails of the probability distributions visible. Times are shown in units of(T · s) where T stands for Tesla: real time is obtained upon division by µ0HK .
84
Chapter 5
Constant Energy Orbit-Averaged
Dynamics
Within the macrospin picture, current-induced steady-state motions appear
when the magnetization settles into a stable oscillatory trajectory that balances
the spin-torque and damping [145]. This oscillatory behavior will be characterized
by magnetization precession at a frequency associated with the element’s magnetic
anisotropy energy, which can tuned, for example, by modifying the element’s shape
(i.e. magnetic shape anisotropy) or magnetocrystalline anisotropies. Thermal noise
can, however, alter the frequency and amplitude of the motion as well as change the
conditions under which steady-state precession occurs. As a result, it is important
to know both how an applied current will influence the amplitude and frequency of
a stable magnetic oscillation and how thermal noise will perturb this configuration
by inducing amplitude and phase noise.
If amplitude and phase diffusion due to spin-torque and thermal noise effects
occur on a timescale much larger than that of magnetization precession, it becomes
85
possible to analyze the steady-state dynamics perturbatively [12]. In this case, the
magnetization dynamics will consist of a fast gyromagnetic precession whose am-
plitude slowly changes over time due to spin-torque and thermal effects. This
has successfully been used to study the dynamical and thermal stability of nano-
magnets subject to spin-polarized currents [146]. This separation of dynamical
timescales falls under the framework of multiscale analysis, which can be applied
in various ways.
Three different approaches have been proposed in the literature in the context
of spin-transfer. Apalkov and Visscher [139] employed an effective Fokker-Planck
(FP) equation, which described the diffusion of a macrospin’s energy under the
influence of both spin-transfer torque and thermal noise. This has been used to in-
terpret results on studies of thermally activated magnetic switching [129,147]. Kim,
Slavin and Tiberkevich [12, 148] studied the Landau-Lifshitz-Gilbert-Slonczewski
(LLGS) equation by noting its analogy to the van der Pol oscillator equation [149].
This resulted in an elegant treatment of the leading nonlinear effects governing the
oscillatory equilibrium steady-state dynamics of the spin-wave eigenmodes. The
approach [9] has had success in explaining the experimentally observed dependence
of the oscillator’s output power on bias current for spin-valves and magnetic tun-
nel junctions [8, 73, 150–152], as well as providing a framework for the extension
of multiscaling methods to spatially extended magnetic systems in which multiple
coupled spin-wave modes may be excited [153].
Finally, macrospin dynamics subject to thermal noise have been modeled using
a stochastic Langevin equation for the time evolution of the macrospin energy by
Newhall and Vanden-Eijnden [154] and in previous work by the Author [155]. This
allows for a reduction in complexity of the sLLGS equations to a 1D stochastic
86
differential equation. In this Chapter, stochastic energy space dynamics will be
developed and used to describe the full nonlinear dependence of mean switching
time on applied currents [156,157] for biaxial macrospin models (log τ ∝ (1−I)β(I))
as an analytic continuation of the uniaxial macrospin model. Recently, Dunn
and Kamenev have extended this approach to propose AC current-driven resonant
switching [65].
Furthermore, recent research on spin-torque oscillators has focused on the ex-
citation of stable in-plane (IP) and out-of-plane (OOP) precessions about the easy
and hard magnetic anisotropy axes of thin film nanomagnets with biaxial mag-
netic anisotropy. In deriving our reduced constant energy orbit-averaged (CEAO)
theory, we present a treatment of these precessional dynamics valid over a wide
range of parameters. We will focus particularly on the OOP dynamics and show
the conditions under which precessional motion about the hard axis occurs. The
oscillator behavior we find is reminiscent of that observed in experiments on a
spin-valve where spin-torque effects are due to the influence of both a perpendic-
ularly magnetized polarizer and in-plane magnetized reference layer [42] like the
one discussed in Chapter 2. The precessional dynamics are found to be stable at
room temperature and, as a result, have great potential for the development of
spin-torque nano-oscillators.
5.1 Constant Energy Orbit-Averaged (CEOA) dy-
namics
In the absence of damping and thermal noise, the dynamics (1.29) preserve the
macrospin’s energy which, expressed in dimensionless form, reads:
87
ε =E(m)
K= Dm2
z −m2x, (5.1)
The conservative trajectories come in two different types. For −1 < ε < 0 the
magnetization gyrates around the easy axis x and is said to be precessing “in-
plane” (IP). For 0 < ε < D, the magnetization precesses about the hard axis z
and is said to be precessing “out-of-plane” (OOP) (refer to Figure (1.1)). The
evolution of such trajectories can be described analytically by solving the sLLGS
equation in the absence of noise, damping and spin-transfer torque [162]:
m0x = −Dm0
zm0y
m0y = (D + 1)m0
zm0x
m0z = −m0
ym0x (5.2)
For IP trajectories [156] one has
m0x(t) = ±
√D − εD + 1
dn[√
D − εt, k2IP
](5.3)
m0y(t) =
√1 + ε sn
[√D − εt, k2
IP
](5.4)
m0z(t) =
√1 + ε
D + 1cn[√
D − εt, k2IP
], (5.5)
where k2IP ≡ D 1+ε
D−ε and sn[·], dn[·], cn[·] are Jacobi elliptic functions [158]. The
period of these trajectories as a function of energy can be expressed as a complete
88
elliptic integral of the first kind:
T (ε) =4√D − ε
∫ 1
0
dx√(1− x2)(1− k2
IPx2)
=4√D − ε
K(k2IP). (5.6)
The amplitudes of an orbit’s precession, projected onto the z-y plane, are1
Az(ε) =
√1 + ε
D + 1(5.7)
Ay(ε) =√
1 + ε (5.8)
Analogously, for OOP trajectories
m0x(t) =
√D − εD + 1
cn[√
D(1 + ε)t, k2OOP
](5.9)
m0y(t) =
√D − εD
sn[√
D(1 + ε)t, k2OOP
](5.10)
m0z(t) = ±
√1 + ε
D + 1dn[√
D(1 + ε)t, k2OOP
], (5.11)
with k2OOP ≡ D−ε
D(1+ε). The projected precession amplitudes in the x-y plane are:
Ay(ε) =
√D − εD
(5.12)
Ax(ε) =
√D − εD + 1
, (5.13)
and the period of the librations is given by:
1precession is around the x-axis
89
Figure 5.1: Orbital frequencies plotted as a function of ε for different D. To com-pare the results, the positive portion of ε axis has been rescaled by D. Frequencyis expressed in units of (GHz/T). Physical frequency is obtained upon multiplyingby µ0HK . The sharp minimum in the frequency is a result of the precessionalperiod diverging at ε = 0.
T (ε) =4√
D(1 + ε)
∫ 1
0
dx√(1− x2)(1− k2
OOPx2)
=4√
D(1 + ε)K(k2
OOP) (5.14)
A sample of these trajectories for positive and negative energies is shown in
Figure 1.1, and orbital frequency as a function of energy is plotted in Figure 5.1.
The unit magnetic sphere can be separated into four distinct basins, two corre-
sponding to ε < 0 dynamics and the others two to ε > 0. For large values of D
the ε > 0 OPP basin can lead to larger oscillatory resistance signals than ε < 0 IP
basin due to the larger precessional amplitudes (5.13).
Upon introducing the contributions of spin-torque, damping and thermal noise,
a macrospin’s dynamical evolution will deviate from a constant energy trajec-
90
tory. Applied currents can reorient the magnetization by pumping energy into the
magnetic system. We may then ask how the constant energy trajectories will be
perturbed. This can be expressed mathematically by computing how the magneti-
zation energy changes as a result of sLLGS evolution. Taking the time derivative
of (5.1), we write 2:
ε = 2 [Dmzmz −mxmx] (5.15)
as the dynamical evolution equation for the macrospin’s energy. Expressing the
time derivatives of the magnetization components in terms of the full stochastic
sLLGS dynamics by using (1.29), one obtains a stochastic evolution equation of
the form
ε = f(m) + g(m) W. (5.16)
We now consider qualitatively how the macrospin dynamics change if the
timescale for energy pumping/sinking, due to the collective effects of damping,
spin-torque and thermal noise, is much larger than the precessional period of the
conservative dynamics. In such a scenario, the full stochastic sLLGS dynamics
might be expected to follow constant energy trajectories fairly closely, with the
macrospin drifting slowly from one constant energy trajectory to the other. Av-
eraging the right hand side (RHS) of (5.16) over constant energy trajectories will
then lead to a single stochastic differential constant-energy orbit-averaged (CEOA)
equation for the evolution of the macrospin’s energy
2The chain rule for stochastic variables is unchanged if the multiplicative noise follows theStratonovich convention.
91
〈∂tε〉 = f(ε) + g(ε) Wε. (5.17)
This approach is justified when the energy drift over the period of a single conser-
vative orbit T (ε)ε is sufficiently small. As will be shown in detail in the next two
sections, due to the different structure of the constant energy orbits in the ε < 0
and ε > 0 basins, the averaging procedure must be performed separately in the
two basins using solutions (5.3) and (5.9) respectively.
Under such assumptions, thermal noise will influence the dynamics in two dis-
tinct ways. The first is by nudging the magnetization onto different energy orbits
(effectively diffusing over the macrospin’s energy landscape). The second, is by
perturbing the precessional phase of the magnetization along a given constant en-
ergy orbit. As such, (5.17) must be supplemented by an equation describing the
stochastic evolution of the dynamical phase. This can be written down by not-
ing that noise must influence energy and phase diffusion identically because it is
isotropic:
〈∂tχ〉 =2π
T (ε)+ g(ε) Wχ, (5.18)
where T (ε) is the period of the orbit at energy ε. We will distinguish between the
two independent noise terms Wε and Wχ by the fact that they act in orthogonal
directions: respectively away and along the constant energy orbit. Whereas (5.17)
will be shown to not depend explicitly on the phase χ, (5.18) will however depend
explicitly on the energy ε.
92
5.2 IP Dynamics
We proceed first by consider damping, applied current and thermal noise ef-
fects on an IP ε < 0 orbit. We must first average (5.16) over conservative negative
energy trajectories (5.3) (see Appendix A for more information). The qualita-
tive properties of such trajectories are better understood by parametrizing them
geometrically as follows. The energy landscape (5.1) determines the geometrical
conditions these trajectories must satisfy. For ε < 0 we have
Dεs2 − 1
εq2 = −1 (5.19)
s2 + q2 + p2 = 1. (5.20)
These are satisfied by the parametrization:
m0x(w) = ±
√|ε| cosh(s)
m0y(w) = ±
√D + |ε|D
√1− γ2
− cosh2(s)
m0z(w) =
√|ε|D
sinh(s),
γ2− =
|ε|(D + 1)
D + |ε|, (5.21)
with limits −acosh(1/γ−) < s < acosh(1/γ−).
We are now ready to average (5.15) over constant energy orbits. The averaging
procedure is simplified by noting that we are interested in switching behavior
starting from the ε < 0 basin and that in that basin the constant energy trajectories
are symmetric with respect to the x axis since they necessarily precess about the
macrospin’s easy axis. As such, the majority of terms obtained by inserting (1.28)
93
into (5.15) will average to zero. The remaining nonzero terms lead to the averaged
energy equation:
〈∂tε〉 = −2α[I cos(ω)(1 + ε)〈mx〉+ (D + 1)〈m2
x〉+ ε(D − ε)]
+ 2
√α(D + 1)
ξ
√〈m2
x〉+D − εD + 1
ε W (5.22)
where 〈·〉 implies averaging over constant energy orbits as described. Construc-
tion of the stochastic term requires averaging the variance of (5.15) following the
rules of additivity for Gaussian random variables (see Appendix B). It is seen
from (5.22) that the applied current factors into this equation only in the form
I cos(ω), where cos(ω) is the cosine of the angular tilt between easy and spin po-
larization axes. This leads to a more general form of the tilt scaling result obtained
in the previous chapter [159], namely that scalings between switching times and
applied current will be functionally identical independent of the angle of the in-
coming spin-polarized current. The only caveat here is that every current must
be rescaled by multiplying it by cos(ω). Analogously, the azimuthal tilt parame-
ter ψ does not appear at all in the energy-averaged equation, implying that only
coplanar setups between polarizer, easy and hard axes need to be considered in
what follows. With this understanding, the geometrical tilts will be absorbed into
I ≡ I cosω.
The averages appearing in (5.22) can be computed analytically (see Appendix
A) giving:
〈mx〉 = ± 2π
T (ε)√D + 1
〈m2x〉(ε) =
4
T (ε)√D + 1
√D
D + 1− γ2−η1 (5.23)
94
with
T (ε) = 4
√D + 1− γ2
−
D(D + 1)K(1− γ2
−) = 4
√D + 1− γ2
−
D(D + 1)η0(γ−) (5.24)
η0(ε) =
∫ 1
0
dx√(γ2− + (1− γ2
−)x2)(1− x2)=
K(1− 1γ2−
)
γ−= K(1− γ2
−) (5.25)
η1(ε) =
∫ 1
0
dx
√γ2− + (1− γ2
−)x2
1− x2= γ−E(1− 1
γ2−
) = E(1− γ2−) (5.26)
γ−(ε) =|ε|(D + 1)
D + |ε|, (5.27)
where K(x) and E(x) are elliptic functions of the first and second kind respec-
tively (we have simplified the results by employing a non-trivial identity proved
in Appendix A.3). The ± appearing in (5.23) represents whether the dynamics
are taking place in the parallel or antiparallel IP basin respectively. The energy
equation representing the dynamics starting in the antiparallel negative energy
well (0 < |ε| < 1,mx < 0) then reads (expressed in terms of γ−3):
∂t|ε|(γ−) =8α
T (γ−)
(D(D + 1)1/3
D + 1− γ2−
)3/2
×[η1(γ−)− γ2
−η0(γ−) +1− γ2
−
D
(η1(γ−)− πI
2
√D + 1− γ2
−
D
)]
+ 4
√α
ξT (γ−)
(D(D + 1)
D + 1− γ2−
)1/4√η1(γ−)− Dγ2
−
D + 1− γ2−η0(γ−) W.
(5.28)
3We can allow ourselves the freedom to switch between expressions involving γ− and ε. γ2− =|ε|(D+1)D+|ε| is a monotonically increasing function of |ε| with the convenient property that |ε| = 0→γ− = 0 and |ε| = 1→ γ− = 1. As such limits written in terms of γ− and ε are equivalent.
95
In outlining the steps above, we have succeeded in reducing the multidimen-
sional complexity of the full magnetization dynamics to a one-dimensional stochas-
tic differential equation whose properties we now proceed to study. The relation-
ship to the previous chapter [156,159] on uniaxial macrospins can be rederived by
considering equations (5.2) and (5.28) in the limit D → 0. Doing so leads to the
much simplified energy diffusion equation [155]
∂t|ε| = 2α√|ε|(1− |ε|)(
√|ε| − I) + 2
√α
ξ|ε|(1− |ε|) W, (5.29)
with a stable energy minimum at ε = −1 and saddle point at ε = −I2. This equa-
tion also shows that, for currents I > Ic = max|ε|∈[0,1]
√|ε| = 1, the flow becomes
negative for all values of |ε|, so that in this regime all states will deterministically
switch. More specifically, substituting |ε| = m2x and ∂t|ε| = 2mxmx one rederives
(4.12) exactly.
5.2.1 Stability Analysis
In the absence of applied currents, ∂t|ε| > 0 as determined by (5.28) and ε
flows toward its minimum value of −1 (the stable fixed point of its dynamics).
To understand switching one must therefore investigate under what circumstances
ε may become greater than zero, thus implying a transition from the red to the
green trajectories discussed in Figure 1.1. To this aim, it suffices to understand the
behavior of the energy flow at the stable fixed point of the zero current dynamics
(|ε| = γ = 1) and at the energy threshold for switching (|ε| = γ = 0). These
96
limiting flows are found by computing the integrals η0(γ) and η1(γ). One finds:
limγ−→1
η0(γ−) ∼ π2
+ π8
DD+1−γ2−
(1− γ2−) + o((1− γ2
−)2) (5.30)
limγ−→1
η1(γ−) ∼ π2− π
8D
D+1−γ2−(1− γ2
−) + o((1− γ2−)2), (5.31)
and
limγ−→0+
η0(γ−) ∼ log(γ−) + o(log(γ−)γ2−) (5.32)
limγ−→0+
η1(γ−) ∼ 1 + o(γ−). (5.33)
Using these results in (5.28), one finds that in the absence of thermal noise (i.e.,
the zero temperature limit), the limiting flows are:
limγ−→0+
∂t|ε| =8α
T
√D
D + 1
[D + 1− πI
2
√D + 1
D
](5.34)
limγ−→1
∂t|ε| =4πα
T
√D + 1
D(1− γ2
−)[D + 2− 2I
]. (5.35)
As expected, both eqs. (5.34) and (5.35) show how the qualitative behavior of
the deterministic energy flow can be tuned by the value of applied current. The
stability of the zero current stable fixed point (|ε| = γ− = 1) can in fact be be
rendered unstable once (5.35) changes sign at the applied current value:
I |ε|=1c ≡ I1
C =D + 2
2. (5.36)
Analogously, at |ε| = γ− = 0 the flow is either positive or negative depending on
the applied current. Using (5.34), we find that the sign of the flow switches from
97
positive to negative at
I |ε|=0c ≡ I0
C =2
π
√D(D + 1). (5.37)
Identical expressions for the critical currents have recently been derived by Taniguchi [147]
via different means. These two critical stability thresholds are identical at a critical
value of D:
D0 = 2π2 − 4
16− π2
[1 + 2
√4 + 2π2
π2 − 4
]' 5.09. (5.38)
The value of D relative to D0, as well as the applied current, will select between
two qualitatively different dynamical regimes. We now proceed to explore both.
5.2.2 D > D0
When the ratio between hard and easy axis anisotropies is greater than the
threshold value D0, I0C > I1
C (see Figure 5.2). Unless the applied current is greater
than I0C , deterministic switching of the monodomain cannot be achieved. This
differs from conclusions drawn by previous work [35] where the critical current
for deterministic switching was assumed to be that which renders the stable fixed
point (|ε| = γ− = 1) unstable.
For small currents such that I < I1C , on the other hand, the deterministic flow
pushes the energy toward the zero current stable fixed point. In such a scenario,
switching can only occur via thermal activation over the |ε| = γ− = 0 effective
barrier.
Finally, for currents I0C > I > I1
C , the zero current stable fixed point has
been rendered unstable while the flow at the switching energy threshold remains
98
Figure 5.2: Critical currents versus the ratio of the hard and easy axis anisotropiesD. The blue curve is I1
C and the red curve is I0C . For D < D0, currents greater
than I1C lead to deterministic switching (labeled DS). For D > D0 currents between
I0C and I1
C lead to limit cycles (LC). Limit cycles can also occur for currents justbelow and approximately equal to I1
C for D < D0, as shown in Figure 5.5
positive. The monodomain will move (deterministically) from the |ε| = γ− = 1
fixed point, but switching over the |ε| = γ− = 0 must still take place via thermal
activation. This implies that a new stable energy equilibrium exists for some value
0 < γS < 1. The precessional macrospin dynamics will then manifest itself in the
form of stable limit cycles. This has been observed both experimentally [41] and
numerically [154]. These three dynamical scenarios are displayed in Figure 5.3
where the deterministic energy flow is plotted as a function of energy.
99
Figure 5.3: Three sample regimes of deterministic energy flow ε as a function ofenergy for D > D0. Coloring is included to better distinguish the various curves.(blue thick dashed curve) I < I1
C : Subcritical regime, thermal noise must oppose apositive energy flow to achieve switching; (green dash-dotted curve) I1
C < I < I0C :
Limit cycle regime; and (red dotted curve) I > I0C : Supercritical regime, negative
flow leads to deterministic switching.
100
5.2.3 D < D0
When the ratio between hard and easy axis anisotropies is smaller than the
threshold value D0, we have I1C > I0
C (see Figure 5.2). The threshold critical
current that switches all magnetic states deterministically is now I1C [35]. Above
this value of the applied current, switching will occur independently of thermal
noise effects.
For applied currents such that I0C < I < I1
C , a saddle point emerges for some
value 0 < γU < 1. For switching to occur, thermal activation must move the
monodomain energy past this saddle point. As the current is lowered, γU becomes
progressively smaller until the limiting value γU = 0. At this point, switching
requires thermal activation throughout the whole energy range. These three dy-
namical scenarios are displayed in Figure 5.4 where the deterministic energy flow
is plotted as a function of energy.
It is interesting to note that the condition D < D0 can also result in the
appearance of limit cycles. Limit cycle regimes in fact can be seen for very small
applied ranges of current less than I0C (Figure 5.5).
The uniaxial macrospin model is a particular case of a D < D0 model (D = 0).
By taking the limit D → 0, we find I0C → 0 and I1
C → 1, as discussed in the
previous section. [159] For all values of the applied current strictly between 0
and 1, uniaxial macrospin switching must take place via thermal activation over
an effective energy barrier.
101
Figure 5.4: Three sample regimes of deterministic energy flow ε as a function ofenergy for D < D0. Coloring is included to better distinguish the various curves.(blue thick dashed curve) I < I0
C : Subcritical regime, thermal noise must oppose apositive energy flow to achieve switching; (green dash-dotted curve) I0
C < I < I1C :
Crossover regime, switching is still achieved via thermal activation but the unstableequlibirum has now shifted; and (red dotted curve) I > I1
C : Supercritical regime,negative flow leads to deterministic switching.
102
Figure 5.5: Energy flow for a D = 4 macrospin and applied current I = 2.82 < I0C .
Circles and squares respectively represent stable and unstable equilibria. For theseparameters (D = 4 and I = 2.82), two stable equilibria of the zero temperaturedynamics coexist.
103
5.2.4 CEOA Breakdown
In this section we will determine under what conditions the CEOA approach is
valid. The fundamental assumption is that the deterministic precession timescale
of the constant energy orbits be small compared to the timescale over which signif-
icant energy diffusion due to noise and spin-torque transfer takes place. We now
quantify this condition and obtain precise limits in terms of the applied current
intensity I and the anisotropy ratio D. For our approximations to be valid, the
averaged energy flow (T (ε)∂t|ε|) over any given orbit must be small compared to
the maximum allowable energy diffusion range (0 < |ε| < 1):
εmax T (ε)|∂t|ε|| = γ−
max T (γ−)|∂t|ε|| 1, (5.39)
where we continue to use the variable γ−(ε) = |ε|(D + 1)/(D + |ε|). To proceed,
we consider the deterministic drift and random components of the energy flow
separately. Noise averaging (5.28) will give the contribution of the drift to the
energy flow. The intensity of the orbit averaged drift then reads:
T (γ−)|〈∂t|ε|〉| =8α
(D(D + 1)1/3
D + 1− γ2−
)3/2
×∣∣∣∣∣η1(γ−)− γ2−η0(γ−) +
1− γ2−
D
(η1(γ−)− πI
2
√D + 1− γ2
−
D
)∣∣∣∣∣ ,(5.40)
where angular brackets denote noise averaging. This expression can be shown to
be finite for all values of γ−, I and D. Furthermore, for applied currents greater
than I0C , (5.40) has a maximum at γ− = 0; its value in this limit is given by (5.34).
104
Enforcing the conditions for validity discussed above results in an upper bound for
the current:
I (1 +1
8α√D
)I0C ≡ IM . (5.41)
In the limitD → 0, this expression converges to the correct uniaxial limit inequality
I (4πα)−1, discussed previously in [155].
When switching occurs deterministically (and I > I0C), this bound is the sole
limit to the validity of the CEOA approximation. This is because random con-
tributions to the energy flow dynamics have zero mean. However, in scenarios
where switching occurs due to thermal activation, one must consider the standard
deviation of random contributions to the energy flow. These determine the rate
of the rare events that drive the system over its confining energy barrier. The
energy flow’s standard deviation is√〈[∂t|ε| − 〈∂t|ε|〉]2〉. The orbit-averaged rate
of random events driving noise-induced energy diffusion can then be written as:
T (γ)√〈[∂t|ε| − 〈∂t|ε|〉]2〉 = 8
√α
ξη0(γ)
(η1(γ)− Dγ2
D + 1− γ2η0(γ)
), (5.42)
which can be shown to be both a monotonically decreasing function of γ− and
logarithmically divergent for γ− → 0. The latter is due to the limiting behavior
of η0(γ−) (see (5.32)) for all D > 0. The physical origin of this behavior lies in
the divergence of the constant energy precessional period (5.24) as the zero energy
switching threshold is reached (a fact also pointed out by Taniguchi et al. [147]).
The relevant maxima of (5.42) will be located at the energy flow’s saddle point
γU . Unfortunately, solving for γU while imposing that the RHS of (5.42) be much
less than 1 results in a set of transcendental equations which can only be solved
numerically. However, we can set a precise bound by requiring that the saddle
105
point never be at γU = 0 (where (5.42) diverges for all D > 0). The applied
current at which such a saddle point occurs has already been derived in (5.37),
leading us to the lower bound inequality
I I0C . (5.43)
as the validity condition for CEOA in thermally activated scenarios. The reason
for limiting ourselves to applied currents greater than I0C in deriving the upper
bound condition (5.41) is now justified. Condition (5.43) is to be considered valid
at all finite temperatures.
As a result of (5.43), there are only two scenarios in which thermally activated
switching can be analyzed and understood using CEOA dynamics. The first is in
the limit D → 0 (the uniaxial macrospin limit) where one has γ(ε) → 1 indpen-
dently of ε. Divergences in the intensity of random contributions are then avoided
for all ε and (5.42) takes the simple form 4π√
(α/ξ)(1− ε). At ε = 0 it converges
to the finite value 4π√α/ξ (which may still be large, however, depending on the
ratio α/ξ). The second scenario is for models with D < D0 and applied current
values I > I0C . In these, the saddle point to be reached via thermal activation
shifts to nonzero values γU 6= 0 where (5.42) does not diverge and the validity
condition may be satisfied.
The above discussion demonstrates that thermally activated switching between
dynamical limit cycle equilibria cannot be modeled with our technique. Thermal
switching from a limit cycle, in fact, must proceed via noise diffusion up to a
saddle point of the type γU = 0 where noise contributions diverge. Nonetheless, it
should be noted that this does not necessarily invalidate the conditions for their
106
existence discussed in the previous section. In fact, relaxation to a limit cycle is
a drift-driven process for which noise related contributions average to zero. Our
analysis merely shows that, once the macrospin has relaxed to its limit cycle state,
further dynamics must proceed via thermal activation for which CEOA does not
provide a suitable description.
In Figure 5.6, the dynamical behavior for D = 4 is displayed. Here D < D0
and I0C < I1
C . Fig 5.6a shows the dependence between rescaled current and mean
switching time for different polar angular tilts θ and fixed azimuthal angle φ. For
reference, the critical currents I0C , I1
C and IM are shown. For large applied currents,
deviations between the different sets of data are clearly visible. This discrepancy
decreases as the current is lowered.
However, for I < I0C , the procedure is expected to fail due to the previously
discussed divergence of the orbital period. To show this, we calculate the maximum
deviation of the mean switching time between all the angular data points at each
value of the rescaled current (see Figure 5.6b). As a further test, we have also
numerically computed the LHS of (5.39) and compared it to the angular deviations.
The critical currents I0C , I1
C and the limit current IM are again shown for reference.
The theory is in satisfactory agreement with numerical simulations as can be seen
from both the good alignment of the deviation minimum with the minimum of the
maximum flow and the rapid spike in deviation as I0C is approached.
This discussion implies that the CEOA technique applies best to small D
macrospin models, and becomes increasingly inaccurate for larger D. Figures 5.7
and 5.8 show the same analysis of a macrospin with D = 50 and one with D = 7.
Fig. 5.7a indicates that the technique is not applicable for sufficiently high values
of D. Nonetheless, theory and numerical results approach each other near the
107
Figure 5.6: a) Mean switching time versus current for D = 4, α = 0.04 and ξ = 80at different θ angular tilts with φ = 0 kept fixed. All currents have been rescaledby 1/ cos θ. Times are shown in units of (s · T ) where T stands for Tesla: realtime is obtained upon division by µ0HK . For visual guidance, the critical currentsI0C , I1
C and limit current IM have been included. In a regime where the CEOAtechnique is applicable, the switching data from the various angular configurationsshould all fall on top of each other. b) Double y-axis plot of max[T (ε)|∂t|ε||] andthe percent deviation of data from (a)) as a function of normalized current. In thecurrent range where the deterministic flow achieves its minimum, the deviation ofthe data does also. As the critical current I0
C is approached, deviation spikes areobserved analogously to what can be inferred by the theory.
108
Figure 5.7: a) Mean switching time versus current for D = 50, α = 0.04 andξ = 80 at different θ angular tilts with φ = 0 kept fixed. All currents have beenrescaled by 1/ cos θ. Times are shown in units of (s · T ) where T stands for Tesla:real time is obtained upon division by µ0HK . For visual guidance, the criticalcurrents I0
C , I1C and limit current IM have been included. b) Double y-axis plot
of max[T (ε)|∂t|ε||] and the percent deviation of data from (a)) as a function ofnormalized current. In the current range where the deterministic flow achievesits minimum, the deviation of the data does also. As the critical current I0
C isapproached, deviation spikes are observed analogously to what can be inferred bythe theory.
minimum of the deterministic flow, as seen in Fig. 5.7b. As in Fig. 5.6, as currents
drop below I0C , deviations in the data rapidly spike.
The analysis and its comparison to numerical data indicate that the CEOA
technique is best suited for studying macrospin dynamics in the crossover regime
separating ballistic from fully activated thermal switching. Furthermore, the sta-
bility analysis of the energy dynamics (5.28) manages to capture the emergence of
limit cycles even though it fails to describe thermally activated processes proceed-
ing from them.
109
Figure 5.8: a) Mean switching time versus current for D = 7, α = 0.04 and ξ = 80at different θ angular tilts with φ = 0 kept fixed. All currents have been rescaledby nz = 1/ cos θ. Times are shown in units of (s · T ) where T stands for Tesla:real time is obtained upon division by µ0HK . For visual guidance, the criticalcurrents I0
C , I1C and limit current IM have been included. b) Double y-axis plot
of max[T (ε)|∂t|ε||] and the percent deviation of data from (a)) as a function ofnormalized current. In the current range where the deterministic flow achievesits minimum, the deviation of the data does also. As the critical current I0
C isapproached, deviation spikes are observed analogously to what can be inferred bythe theory.
110
5.2.5 Thermally Activated Switching
Scenarios in which thermally activated switching can be studied using the ap-
proach described above are generally rather limited. Nonetheless, in the previous
section, the technique was shown to be applicable in models with D < D0 for
currents I0C < I < I1
C . Starting from (5.28) one could in principle construct a
Fokker-Planck equation describing energy diffusion and then attempt to solve an
appropriate mean first passage time problem numerically. Instead, we will simplify
the thermal activation problem by deriving the exponential scaling dependence be-
tween mean switching time and current, using an analytical tool described in [155].
In that paper, it was shown how, starting from energy diffusion dynamics analo-
gous to (5.28), the exponential scaling dependence can be written in terms of an
integral from the initial stable state to the saddle using a Friedlin-Wentzell [172]
type formulation:
log(〈τ〉) ∝ 2
∫ 1
|εU |
f(ε′)
g2(ε′)dε′ (5.44)
where f(ε) and g(ε) are, respectively, the drift and noise terms of (5.28), now
expressed in terms of the energy ε. This integral can be computed numerically
once the saddle point of the energy flow has been identified (e.g. via a Newton
algorithm). The integrand can be further expanded for D >> 1 and approximated
around the stable point |ε| = 1. This gives, to first order in 1− |ε|,
log(〈τ〉) ∝ ξ
∫ 1
|εU |
(1− (I/I0
C)
√D
D + ε′
(1− ε′
η1(ε′)− ε′η0(ε′)
))dε′
' ξ
∫ 1
|εU |
(1− (I/I0
C)1− ε′
1− ε′ζ(D)
)dε′ (5.45)
111
where ζ(D) is given by
ζ(D) =I1C
I0C
=π
4
(D + 2√D(D + 1)
). (5.46)
The approximated integral can now be written in closed form in terms of hyper-
geometric functions. The resulting expression is accurate to within 2% for values
D > 0.1.
In Fig. 10 we plot the dependence of the mean switching time as a function of
applied current, and compare it to log(τ) ∝ ξ(1 − I/IC)β. In the limit D → 0,
the uniaxial switching exponent β = 2 is recovered. For larger values of D, the
exponent β depends non-linearly on the applied current I. In the limit I → 0 the
limiting value of β can be derived analytically:
limI→0
β(I) = limI→0
log(1− q(D)
I0CI)
log(1− I/I1C)
= q(D)I1C
I0C
, (5.47)
with
q(D) ≡∫ 1
0
√D
D + x
(1− x
η1(x)− xη0(x)
)dx. (5.48)
The same calculation can be performed in the limit I → I1C and used to show
that the exponent β diverges for all non-zero values of D. The mean switching
time is nonetheless well behaved, as can be seen in Fig. 5.9a. This differs from the
results obtained by Taniguchi [147], in which the limiting value of the exponent β is
roughly 2.2, as the applied current approaches the critical threshold for determinisic
switching. Unfortunately, the numerical verification of β in regimes where CEOA
is applicable is a complicated matter. To do so, one would need to know the form
of the exponential prefactor in the scaling which, in turn, is also expected to be
112
Figure 5.9: a) Scaling dependence of mean switching time as a function of appliedcurrent I for models with varying D < D0. ξ is the energy barrier height and I1
C
the critical current threshold for deterministic switching. b) Fit of (5.45) to theform (1− I/I1
C)β. Dashed lines represent continuation of analytical results outsidethe technique’s regime of validity. Fitting exponent β is plotted as a function ofapplied current for models with varying D. In the limit of small D the exponentapproaches the constant value β = 2 consistent with previous uniaxial macrospinresults [136, 155, 159]. For D > 0, the exponent β depends nonlinearly on theapplied current intensity. Only for values D ∼ D0 do we notice that in the limit ofsmall applied currents, β → 1 as suggested by similar energy diffusion studies fromthe literature [139,154]. For intermediate values D0 > D > 0 the low current limitof β can be obtained analytically (5.47). On the other hand, in the limit I → I1
C
the exponent β can be shown to diverge for all non-zero values of D.
non-linearly dependent on the applied current as already discussed in relation to
the simpler uniaxial macrospin model in the previous chapter [159].
5.3 OOP Dynamics
In this approach, we now consider damping, applied current and thermal noise
effects on an OOP ε > 0 orbit. First, we average (5.16) over conservative positive
energy trajectories (5.9) (see Appendix A). Due to the symmetry of such trajecto-
ries, most terms average to zero with the remaining nonzero terms leading to the
113
CEOA equation:
〈∂tε〉 = 2α[I(D − ε)(sinω cos2 ψ)〈mz〉 −D(D + 1)〈m2
z〉+ ε(1 + ε)]
+ h(ε)
+
√2αD(D + 1)
ξ
√〈m2
z〉 −ε(1 + ε)
D(D + 1)· Wε, (5.49)
where angular brackets 〈·〉 denote averaging over a constant-energy trajectory with
energy ε. The second drift term (following the square brackets) h(ε) is a result of
transforming (1.29) into its Ito representation before performing the average over
orbits (see Appendix B). As a result, the multiplicative noise terms appearing in
the averaged energy equation above are now interpreted in the Ito sense4.
We note that, as has been found for negative (IP) CEOA states [156], the
dynamics as a function of applied current for different spin-polarization tilts are
identical, the current is simply rescaled by sinω cos2 ψ (refer to Fig. 1.5). This al-
lows us to numerically verify the CEOA approach by checking that the macrospin’s
evolution over some (properly rescaled) applied current is exactly identical for dif-
ferent tilts of the spin-polarization axis.
The associated equation describing the stochastic evolution of the dynamical
phase can at once be written down also:
〈∂tχ〉 =2π
T (ε)+
√2αD(D + 1)
ξ
√〈m2
z〉 −ε(1 + ε)
D(D + 1)· Wχ, (5.50)
where, again, T (ε) is the period of the orbit at energy ε. The structure of the
phase dynamics will become important when we discuss different aspects of phase
noise in later sections.
4As usual, we distinguish equations written in Ito vs. Stratonovich form by writing themultiplicative noise as ‘· ˙Wε,φ’
114
To compute the averages 〈mz〉 and 〈m2z〉 explicitly, we note that the positive
energy trajectories can be geometrically parametrized as follows:
m0x(s) =
√ε sinh(s),
m0y(s) = ±
√1 + ε
√1− γ2 cosh2(s)
m0z(s) = ±
√ε
Dcosh(s)
γ2+ =
ε(D + 1)
D(ε+ 1), (5.51)
where the parameter s ranges from −acosh(1/γ) < s < acosh(1/γ). Upon com-
puting the averages explicitly (Appendix A), the CEOA equations for the positive
energy dynamics (0 < ε < D), expressed in terms of γ+, read 5
∂tε(γ+) =πα
η0(γ+)
D(D + 1)
[D(1− γ2+) + 1]3/2
×±I(1− γ2
+)− 2
π
√D(1− γ2
+) + 1
[η1(γ+)−
γ2+
(D(1− γ2+) + 1)
η0(γ+)
]+ h(ε)+√
2α
ξ
D(D + 1)
D(1− γ2+) + 1
(η1(γ+)
η0(γ+)− γ2
+
D(1− γ2+) + 1
)· Wε (5.52)
∂tχ(γ+) =π
2η0(γ+)
√D(D + 1)
D(1− γ2+) + 1
+√2α
ξ
D(D + 1)
D(1− γ2+) + 1
(η1(γ+)
η0(γ+)− γ2
+
D(1− γ2+) + 1
)· Wχ (5.53)
where η0(γ+) = K[1− γ2+] and η1(γ+) = E[1− γ2
+] are again expressed in terms of
5We can allow ourselves the freedom to switch between expressions involving γ+ and ε. γ2+ =ε(D+1)D(ε+1) is a monotonically increasing function of ε with the convenient property that ε = 0 →γ+ = 0 and ε = D → γ+ = 1. As such limits written in terms of γ+ and ε are equivalent.
115
complete elliptic integrals of the first and second kind. For notational simplicity,
the geometrical tilts have been absorbed into I ≡ I sinω cos2 ψ.6 It is important
to note the applied current acts either to positively or negatively dampen the
dynamics depending on which ε > 0 basin the magnetization is in (see Figure 1.1).
The second drift term h(ε) appearing in (5.52), is the drift correction due to our
change to Ito calculus. As discussed in Appendix B, the extra drift term results in a
negligible correction. The following analysis will, analogously to the IP case, hence
ignore its second order effects although they can be reintroduced straightforwardly
if higher quantitative accuracy is desired.7
In following the outlined procedure, we have again reduced the complexity of
the magnetization dynamics to a one-dimensional stochastic differential equation,
whose properties we will now show to be analytically tractable.
5.3.1 Stability Analysis
As seen from (5.52), in the absence of applied currents, the deterministic drift
portion (first term on the RHS) of the energy diffusion dynamics is globally neg-
ative, ∂tε < 0. The energy ε flows from positive to negative energy basins toward
its minimum value of −1. This is consistent with our physical notion of the ε > 0
basins being energetically unfavorable. Upon introducing an applied current, the
behavior remains unchanged as long as no tilt is present between easy and spin-
polarization axes (ω = 0). If a nonzero tilt is introduced into the system, the
symmetry of the two positive energy basins is broken. In particular, due to the
dependence on ±I (everything else inside the curly brackets is always negative),
6Note that in contrast to the in-plane precessional dynamics discussed in [156], the current isrescaled by sinω as opposed to cosω
7The Ito drift-diffusion correction becomes relevant for dynamics close to the ε = 0 separatrix.
116
a critical current will exist, corresponding to a fixed point in the energy dynamics
appearing in the positive z, ε > 0 basin. The presence of a fixed point in the energy
dynamics can then be expected to correspond to a stable precessional (limit cycle)
state of the magnetization dynamics. The dynamics in the negative z basin, on the
other hand, will continue to be globally dissipative. Physically this is explained by
the fact that the tilt ω biases the magnetic evolution away from one basin in favor
of the other.
The critical current at which a fixed point appears can be obtained by studying
the behavior of the energy dynamics in the limit ε = γ+ → 0. Requiring that
limε→0
T(ε)ε ∝ −2√D + 1 + πI = 0, (5.54)
we obtain
IOOP =2
π
√D + 1, (5.55)
as the current at which a stable fixed point appears at ε = 0. Increasing I further
will shift the fixed point to higher energies. Qualitatively, this will result in an
increase of frequency and decrease of amplitude of the limit cycle oscillations. The
maximum possible energy obtainable by the oscillator is ε = D. This is achieved
when 8
Imax = D +1
2. (5.56)
Increasing the current beyond Imax simply overdrives the magnetization. As we
will see later, the CEOA approximation breaks down beyond this point and stable
8One analogously seeks a null net drift of the energy dynamics at ε = D: limε→D
T(ε)ε = 0
117
oscillations disappear. Fig. 5.10 shows a sample of the drift field due to (5.52) for
I < IOOP, IOOP < I < Imax, and I > Imax. IOOP and Imax represent the lower
and upper threshold currents for the appearance of steady-state precessions in the
stable OOP basin due to the nonlinear character of the magnetization dynamics.
Comparing with the CEOA treatment of magnetic switching [156] discussed
in the previous section, we note that Iswitch9, the critical current for switching,
equals√DIOOP. As such, the minimal currents sustaining stable OOP precessional
states are generally smaller than the critical switching current. This results in the
prediction of a hysteretic dependence of IPOOP transitions on applied current,
which has been observed recently in experiment [130]. In detail, since Iswitch =
Iswitch cosω and IOOP = IOOP sinω cos2 ψ, one can see that the relation between
direct critical switching current and threshold current for sustainment of OOP
precessions is
IOOP =Iswitch√
D tanω cos2 ψ. (5.57)
5.3.2 CEOA Breakdown
For our approximations to be valid, the averaged energy flow (T (ε)|∂tε|) over
any given orbit must be small compared to the maximum allowable energy varia-
tions (0 < ε < D):
εmax T (ε)|∂tε| D. (5.58)
9We limit ourselves to considering samples with D > D0. As discussed in the section on IPdynamics, the switching current for such models is I0C = 2
π
√D(D + 1). From now on we will
refer to I0C as Iswitch in the text. We remind the reader that the tilde stands to signify a currentrescaled by cos(ω) as relevant to the IP dynamics: Iswitch = I0C = I0C cos(ω).
118
Figure 5.10: Three regimes of deterministic energy flow ε as a function of energy forD = 10. (blue-dashed) I < IOOP: Subcritical regime. Energy flows from positiveto negative energy basins due to dynamics being globally dissipative (overdamped).(red-dashdotted) I > Imax: Supercritical regime. Energy flows towards limitingstable value ε = D due to dynamics being overdriven by applied current. (green-dotted) IOOP < I < Imax: Oscillator regime. Energy flow will stabilize at a fixedpoint corresponding to a precessing oscillator state. In this regime, the fixed pointrepresents a constant energy trajectory where spin-torque and damping effectsbalance.
119
The analysis proceeds identically to the previous section on IP dynamics [147,156]
so we simply state the results for OOP dynamics. For CEOA to be applicable one
must have IOOP . I . Imax.
In Figure 5.11 we show a comparison between theory and numerical results by
plotting average energy 〈ε〉 as a function of applied current. Ensembles consisting
of 5120 macrospins were initialized antiparallel to the easy-axis and allowed to
relax subject to a steady applied current. Upon varying the angular tilt ω between
easy and spin-polarization axes, we notice that the data follow our theory down to
a minimum critical angle ωC . For angular tilts less than ωC , stable positive energy
steady states cease to be accessible regardless of the applied current. The origin
of this angular cutoff is geometrical in nature and corresponds to the necessity
for the spin-polarization axis to be pointing inside the positive energy basin. The
condition for this to happen can be seen from (5.1) by solving for the separatrix
of the energy basins. One obtains
ωC =π
2− arctan(
√D), (5.59)
which is in excellent agreement with numerical data. This geometrical intuition
can be seen from theory by determining the tilt for which the threshold current
for OOP precessions equals that for direct switching. Starting with (5.57), and
setting ψ = 0 for convenience, leads to (5.59).
For large currents I > Imax, numerical results seem to indicate a steady drop in
ensemble energy as the applied current is increased. In fact, contrary to the CEOA
description, the macrospin’s magnetization ceases to precess around the hard axis
and instead settles into a magnetic configuration where all static torques balance
and spin-torque effects compete with the magnetic anisotropies.
120
Figure 5.11: Steady-state ensemble energy as a function of dimensionless appliedcurrent I (rescaled by IOOP = (2/π)
√D + 1) for a model with D = 10, ξ = 80 and
α = 0.04. Red line shows an analytic fit to numerical data within the current limitsdefined by the theory (for reference Imax/IOOP ≈ 4.97). Insets show density plots inspherical coordinates of 5120 numerical trajectories for a sample with a 2.56ωC tiltbetween easy and spin-polarization axes, driven by a current of I/IOOP = 4 (top),and I/IOOP = 15 (bottom). The dotted line denotes the conservative trajectory.
121
5.3.3 Thermal Stability, Precession Linewidth, Phase, Am-
plitude and Power Fluctuations
So far, we have provided an analytical approach that enables the study of the
properties of OOP dynamics. Once the strength of the applied current I has
been chosen, and provided that the angular tilt of the spin-polarization vector is
sufficient (ω > ωC), the average energy ε0 = 〈ε〉 of the equilibrium steady state tra-
jectory can be obtained by solving for the fixed point of the energy dynamics (5.52)
for OOP and (5.28) for IP dynamics respectively. Due to the dependence of the
precessional period T(ε) on the energy of the orbit, the expected precessional fre-
quency can be inferred.
Thermal noise will, however, perturb the magnetization about the fixed point,
resulting in fluctuations of the macrospin’s energy around its average ε0 value and
diffusion of its phase χ along the relevant constant energy orbit. These deviations
are believed to be the source of the oscillator’s experimentally measured frequency,
linewidth and phase decoherence. We will now proceed to derive an estimate for
such linewidths.
The general stochastic energy evolution equation (5.52,5.28) can be written
concisely as
∂tε = f(ε, I) + h(ε) + g(ε) · Wε, (5.60)
where f(ε, I), h(ε) and g(ε) are, respectively, the deterministic drift, Ito drift-
diffusion correction and multiplicative noise. For definiteness, we will concentrate
on OOP steady-state precessions in this section. The methods and techniques can
however be extended to the study of steady-state IP precessional dynamics also.
122
Following Ref. [22], one can use the stochastic energy evolution equation to
compute the mean time one must wait to observe a thermal excitation out of an
OOP trajectory. The asymptotic dependence of such a mean escape time is then
log(〈τjump〉) ∝ 2
∫ ε0(I)
0
dxf(x, I)
g2(x)
= ξ
(ε0 −
I
IOOP
∫ ε0
0
dxD − x√
1 + x(Dη1(x)− xη0(x))
), (5.61)
where ε0 ≡ 〈ε〉 is the usual solution of the fixed point equation (dependent on I).
Due to the dependence of the equilibrium oscillator energy on the applied current
ε0(I), the thermal stability of the OOP precessional states will depend non-linearly
on the applied current I.
The Fokker-Planck (FP) equation is:
∂tρ = ∂ε
[f(ε, I)ρ− 1
2g2(ε)∂ερ
], (5.62)
whose solution describes the full evolution of the energy distribution ρ(ε, t) as
a function of time (Appendix B). At equilibrium (∂tρ = 0), the saddle point
approximation can be used to determine a steady state distribution
ρeq(ε) ∝ exp
[2
∫ ε
0
dxf(x, I)
g2(x)
]' exp
[f ′(ε0, I)
g2(ε0)(ε− ε0)2
], (5.63)
that is valid as long as I > IOOP. We can then write an expression for the amplitude
noise by computing the variance of the energy in an equilibrium OOP distribution:
〈(ε− ε0)2〉 ' g2(ε0)
2|f ′(ε0, I)|. (5.64)
123
Figure 5.12: Standard deviation of the energy distribution plotted as a func-tion of dimensionless applied current I (rescaled by IOOP = (2/π)
√D + 1) for
D = 10, ξ = 80 and α = 0.04. The solid blue line shows the theoretical predic-tion (5.64) calculated within the current limits defined by the theory (for referenceImax/IOOP ≈ 4.97).
In Figure 5.12 we compare the theoretical approximation resulting from (5.64)
with the equilibrium energy variance extracted from our numerical simulations.
Whereas the variance does not appear to rescale trivially with the spin-polarizer
tilt, all tilts seem to show a variance versus applied current curve that peaks
within the same general region predicted by our rough estimate. For currents
I ' IOOP, Imax the approximation breaks down due to failure of the CEOA ap-
proximation.
Using (5.63), we see that all energy-dependent stationary characteristics 〈Q〉 =∫Q(ε)ρeq(ε) of the oscillator can be computed via distribution averaging. However,
we employ our saddle point estimate (5.64) to study thermal fluctuations. The
relative fluctuation of a quantity Q(ε) at equilibrium will be given by δQ/Q =
124
(Q′(ε)/Q(ε))|ε=ε0√〈(ε− ε0)2〉.
As a first example, the experimentally observed oscillator power depends on
the square of the oscillator’s precession amplitude along the in-plane direction.
Having chosen a coordinate system with the reference magnetic layer aligned in-
plane, power fluctuations are directly proportional to fluctuations in the precession
amplitude of the oscillator as projected along the in-plane axial direction. From
our previously derived expression of the oscillation amplitude along the in-plane
direction (5.13), one has:
δP
P=δA2
x
A2x
'√〈(ε− ε0)2〉. (5.65)
Analogously, denoting the oscillation frequency by ν(ε) = 2π/T (ε), one finds
for the precession linewidth quality factor Q dependence on amplitude noise:
1
Q=δν
ν' T′(ε0)
T(ε0)
√〈(ε− ε0)2〉. (5.66)
Fig. 5.13 shows how the quality factor is a monotonically increasing function of
applied current. Overall, increasing the driving current reduces the linewidth of
the oscillator in line with classical oscillator theory which predicts a linewidth
scaling dependent on the ratio of the thermal and oscillator energy (kBT/ε). In
practice, however, at currents high enough for the breakdown of the macrospin
model, micromagnetic effects due to Oersted fields are expected to complicate the
physical picture in non-trivial ways.
One may proceed further and ask whether the CEOA formalism is capable
of shedding light on the phase noise and, more generally, the phase decoherence
driving a magnetic system. The assumption that “sufficiently weak” noise drives
125
Figure 5.13: Inverse quality factor (5.66) vs. applied current for D = 10 set atroom temperature (ξ = 80). Red dashed line denotes the upper bound of thevalidly of the CEOA formalism: Imax/IOOP ≈ 4.97 for the parameters chosen.
diffusion from one energy orbit to another does not impose any limit on how
strong the noise driving the phase of the actual constant energy oscillation can
be. Both phase noise due to thermal diffusion along a given constant energy
orbit and amplitude noise can drive phase decoherence in a magnetic system. As
such, the relative intensity of both effects must be determined to understand phase
decoherence.
To do so, we consider how energy fluctuations about the ε0 equilibrium fixed
point influence the phase dynamics described in (5.52). Let ε(t) ≡ ε0 + δε(t) and
expand (5.60) in powers of δε. Denoting F (ε) ≡ f(ε)+h(ε), the resultant stochastic
differential equation can be formally integrated to give:
δε(t) = eF′(ε0)t
[c+ g(ε0)
∫ t
0
dt′e−F′(ε0)t′ · Wε
], (5.67)
126
where primes represent differentiation with respect to energy (F ′(ε0) ≡ ∂εF |ε=ε0),
and c is an (unimportant) initial condition. |F ′| represents the relaxation rate of
amplitude fluctuations to the ε0 baseline. Given the explicit dependence of the
phase χ on the energy evolution, such energy fluctuations are expected to play a
crucial role in the thermally driven phase dynamics.
Expanding the phase dynamics about ε0 to lowest order, we have:
∂tχ =2π
T (ε0)− 2πT ′(ε0)
T 2(ε0)δε(t) + g(ε0) · Wχ. (5.68)
Substituting (5.67) into (5.68) and recalling that Wε and Wχ are uncorrelated
stochastic processes, the expected phase variance at equilibrium can be evaluated
to give (we suppress the dependence on ε0):
〈∆χ2〉(t) =g2
[1 +
(2πT ′
F ′T 2
)2]|t|
+g2
2F ′
(2πT ′
F ′T 2
)2 [4(
1− eF ′|t|)−(
1− e2F ′|t|)]. (5.69)
which closely resembles the more general prediction from oscillator theory. [146,
161] Since the power spectrum can be written as a Fourier transformation of the
correlation function 〈exp[i(χ(t) − χ(t))]〉 ≈ exp[i〈χ(t) − χ(t)〉] exp[−〈∆χ2〉(t)/2],
the linewidth can be predicted [160] by inspecting (5.69).
The temporal dependence of the phase variance is responsible for the decoher-
ence of the magnetic ensemble over time. We interpret the decoherence time τdec as
the timescale necessary for the ensemble to homogeneously distribute itself along a
given constant energy orbit similarly to what is shown in Fig. 5. We quantify τdec
127
by asking on what timescale the width of the phase distribution begins to encom-
pass the entire constant energy orbit: 〈∆χ2〉(τdec) = 4π2. Although the temporal
dependence is generally quite complicated, two limiting regimes can be explored.
For low enough temperatures, the phase decoherence time τdec will be larger than
the relaxation timescale of the amplitude fluctuations τdec 1/|F ′|. Decoherence
can then be expected to mostly take place due to the differences in orbital evo-
lution at the different energies explored by the amplitude fluctuations. This will
eventually lead the spin ensemble to decohere and thermalize to a homogenous
distribution of phases relative to the referential ε0 orbit. The dominant amplitude
fluctuations driving such a low temperature regime result in a linear dependence
of the phase variance.
〈∆χ2(t)〉 ≈ g2
[1 +
(2πT ′
F ′T 2
)2]|t|. (5.70)
Due to the dependence of the multiplicative noise term in (5.52) on temperature
(g(ε) ∝√T ), the decoherence time τdec ∝ T−1/2 ∝
√ξ can be predicted to depend
on the inverse square root of temperature. Furthermore, a linear dependence on
time will imply a Lorentzian power spectrum with linewidth ∆νL = (g2/2π)(1+µ2)
(µ = 2πT ′/F ′T 2).
In a high temperature limit, pure phase noise will compete with the amplitude
noise effects by decohering the ensemble on a timescale smaller than the amplitude
fluctuation relaxation rate τdec 1/|F ′|. The exponential contributions in (5.69)
cease to be negligible and the approximate temporal dependence of the phase
variance can be written to second order in time as:
128
〈∆χ2(t)〉 ≈ g2
[|t|+ 2
(2πT ′
F ′T 2
)2
|F ′||t|2]. (5.71)
If (2πT ′/√|F ′|T 2)2 1 (typically the case when ε0 D), the term linear in
time can be dropped altogether resulting in a purely quadratic dependence of the
phase variance on time. In such a scenario, the decoherence time can be expected
to scale linearly with the inverse temperature τdec ∝ T−1 ∝ ξ. A phase variance
scaling quadratically in time will in turn lead to a gaussian power spectrum with
linewidth ∆νL =√
2gµ2F ′/2π.
We explore these predictions by studying switching probability curves of a
macrospin ensemble at varying temperatures for applied current intensities and
effective spin-polarization axial tilt consistent with an OOP precessional behav-
ior. Upon switching the current off, the phase of the oscillator will select the
macrospin’s relaxation outcome (either parallel or antiparallel to the easy axis of
the magnetic film) with high probability. In the absence of thermal noise, a cur-
rent pulse of fixed duration will lead to either a parallel or antiparallel relaxed
state after the pulse terminates (see Fig. 5.14) with absolute certainty. At nonzero
temperatures, however, oscillator ensemble phase decoherence is expected due to
thermal noise. As a result, long spin-current pulse times will lead to equally likely
parallel (antiparallel) relaxation due to ensemble thermalization along the OOP
constant energy orbit. In Fig. 5.15 we find good qualitative agreement between
such an understanding of phase decoherence behavior and numerical simulations.
The equilbrium probability bias for higher P switching is due to some of the states
thermally equilibrating into the IP energy basin before the current pulse is switched
off.
129
Figure 5.14: Switching probability vs. spin-current pulse length for a macrospinmodel with D = 10, ω = 2.12ωC) driven by a spin-current intensity of I =2.75 IOOP in the absence of thermal noise. Times are shown in units of (s · T )where T stands for Tesla: real time is obtained upon division by µ0HK . Before thecurrent pulse is switched on, the magnetic ensemble is taken to be antiparallel tothe easy-axis of the magnetic film. Switching probability is defined as the ensemblefraction that relaxes into a parallel configuration upon switching the current pulseoff. The right-hand vertical axis plots the evolution of the average 〈mz〉 component.In the absence of thermal noise the oscillator remains coherent at all times andits periodic motion is clearly seen. Due to the deterministic nature of the zero-temperature dynamics, the macrospin will deterministically switch either into theparallel or antiparallel state at all times.
130
Figure 5.15: Switching probability vs. spin-current pulse length for a macrospinmodel with D = 30, ω = 3ωC driven by a spin-current intensity of I = 5 IOOP
in the presence of thermal noise corresponding to ξ = 80 (left) and ξ = 1200(right). Times are shown in units of (s · T ) where T stands for Tesla: real timeis obtained upon division by µ0HK . Before the current pulse is switched on, themagnetic ensemble is taken to be antiparallel to the easy-axis of the magnetic film.Switching probability is defined as the fraction of the ensemble that relaxes into aparallel configuration upon switching the current pulse off. For long pulse timesthe switching probability converges to a value indicating that the phase of theOOP precession has decohered. The red dashed lines are a qualitative graphicalrepresentation of the decoherence time.
131
Figure 5.16: Log-log plot of ensemble decoherence time vs. energy barrier heightto thermal energy ratio ξ for a macrospin model with D = 30, ω = 3ωC drivenby a spin-current intensity of I = 1.5 Iswitch. Times are shown in units of (s · T )where T stands for Tesla: real time is obtained upon division by µ0HK . Linearregression (solid lines) of data points demonstrates a transition between a phasenoise dominated regime τdec ∝ 1/T below a certain critical inverse temperatureξ < ξC . Above ξ > ξC (T < TC), both amplitude and phase noise contribute toensemble decoherence.
The switching probability curves can be employed to numerically extract the
decoherence time at different temperatures. Fig. 5.16 shows a log-log plot of τdec
on ξ for a D = 30 model with a ω = 3ωC tilt, driven by a I = 1.5 Iswitch applied
current. Linear regression to numerical data shows an inverse proportionality
τdec ∝ 1/T ∝ ξ between decoherence time and temperature for temperatures
larger than a certain critical temperature. For T < TC , however, both amplitude
and phase noise seem to contribute to ensemble decoherence thus not allowing us
to probe the pure amplitude noise decoherence mechanism previously discussed.
132
Overall, the theoretical techniques developed in this chapter allow for a very de-
tailed characterization of the most relevant dynamical regimes that the macrospin
can be subject to via spin-torque induced excitations. The constant energy aver-
aged theory has allowed for very precise predictions to be made regarding when
a magnetic element can be expected to behave as a switching element or a spin-
torque nano-oscillator. In the next section we will briefly explore experimental
evidence for the OOP precessionary orbits.
5.4 Experimental Verification of OOP orbits
A polarizing layer that is magnetized perpendicularly to the free layer can signif-
icantly improve write speed and energy efficiency of spin-transfer torque magnetic
random access memories [81–83,168–171] by providing a large initial spin-transfer
torque. An orthogonal spin-torque OST device, of this kind (refer to Fig. 1.4)
was shown to also function as a microwave oscillator in the previous section, since
the polarizer can produce precessional magnetization dynamics, with the free layer
precessing OOP [42, 78, 79]. This precessional motion can also be used for ultra-
fast magnetization switching. For instance, sub-nanosecond switching has been
observed to be bipolar and to induce magnetization precession [81, 82, 170]. Fur-
ther, the effect of applied fields on these curretn-induced switching thresholds has
not yet been reported or considered in model studies. In this section we report
on the experimental observation of bipolar switching in OST spin valve devices as
discussed in [130]. Hysteretic transitions into intermediate resistance (IR) states
at large current are found to exist, with the IR state persisting to currents less
than the threshold currents for P to AP and AP to P switching.
133
5.4.1 Experiment
The layer stack studied consists of a perpendicularly magnetized spin-polarizing
layer, a non-magnetic metallic spacer layer, a free magnetic layer followed by an-
other non-magnetic spacer layer and a reference magnetic layer as illustrated in
the inset of Figure 5.17. The polarizer consists of a Co/Pd and Co/Ni multilayer,
with the Co/Ni multilayer closest to the free layer (FL), providing a highly spin-
polarized current [81]. The FL is a 3 nm thick CoFeB layer. The full layer stack is
6.2 [Co/Pd][Co/Ni]/10 Cu/3 CoFeB/10 Cu/12 CoFeB, with the layer thicknesses
indicated in nanometers. The stack was patterned into nanopillar devices with var-
ious shapes and sizes using e-beam lithography and ion-milling. Here we present
results on 50 nm × 100 nm devices in the shape of an ellipse. The magnetic easy
axis of the free layer is int he film plane along the long axis of the ellipse due to
magnetic shape anisotropy. Shape anisotropy also sets the magnetization direction
of the 12 nm thick CoFeB reference layer (RL).
Figure 5.17 shows measurements of the differential resistance (dV/dI) as a
function of applied field alogn the easy axis. The measurements are made with a
lock-in amplifier using an ac current of 200µA at a frequency of 473 Hz. A field
sweep from -200 mT to 200 mT (major hysteresis loop) shows steps in resistance
of 0.1 Ω indicative of switching of the FL from P to AP relative to the RL. The
coercive field of the RL is about 150 mT. A minor loop (-50 mT to 140 mT) shows
the switching of only the FL. the change in resistance between P and AP states is
∆RAP−P = 0.1 Ω. The coercive field for P to AP FL transitions is H+C = 59 mT
and the coercive field for P to AP FL transitions is H−C = 23 mT. The minor loop
is centered at H0 = (H+C + H−C )/2 = 41 mT due to dipolar coupling between the
Fla nd RL. Thus an external field of H0 corresponds (on average) to zero effective
134
Figure 5.17: Resistance versus in-plane applied field hysteresis loops. The majorloop (black curve) shows the switching of both the free and reference layers. Theminor loop (blue curve) shows the response of just the free layer. The loop iscentered at 41 mT due to dipolar interactions between the reference and free layer.Inset: Schematic of the spin-valve’s layer stack showing the out-of-plane magne-tized polarizing layer (OP), in-plane magnetized free layer (FL) and reference layer(RL).
field applied to the FL.
Current induced switching was characterized by measuring the differential resis-
tance as a function of current for a series of easy axis applied fields. The magnetic
state (P or AP) is first set by applying a large magentic field 200 mT and then a
lower field (larger than 25 mT for the P state, less than 58 mT for the AP state).
Then the current Idc was slowly ramped (' 0.1mA/s) from 0 to ±5 mA and then
ramped back to 0 mA, with dV/dI versus I recorded at each measuring field. Pos-
itive current corresponds to electron flow from polarizing to the reference layer
(from bottom to the top of the layer stack represented in the inset of Figure 5.17).
In this case the spin-torque associated with the RL favors an AP state for positive
currents. Representative measurement results starting from the P state are shown
135
in Figure 5.18. Similar results were found in measurements starting from the AP
state, which are discussed below.
In Figure 5.18(a), as the magnitude of the current is increased (black curves),
there is a discrete increase in differential resistance of ∆RAP−P = 0.1 Ω, associated
with a P to Ap transition. This is seen to occur for both polarities of the current.
On further increasing the current there is a change in resistance of about half of
∆RAP−P , i.e. a transition into an intermediate resistance state (IR). On decreasing
the current (red curves) the resistance eventually returns to that of the device’s AP
state. Near zero effective applied field (H ' H0), P to AP switching is only seen at
positive current polarity (Figure 5.18(b) and (c)). Whereas, for negative current,
only P to IR transitions occur as the current is increased. When the magnitude
of the current is decreased, there is a transition from an IR state to an AP state
for applied magnetic fields smaller than H0 (= 41 mT) and to a P state for fields
larger than H0.
This seemingly complex switching behavior can be summarized by plotting
the threshold currents for switching between resistance states in a current-applied
field state diagram (Figure 5.19). Each symbol in this diagram corresponds in a
discrete change in the resistance. The solid symbols represent resistance changes
of ∆RAP−P corresponding to transitions between P and AP states. They form
a diamond-shaped central zone within which both P and AP states are possible.
When the current is greater than 2.9 mA or is less than -1.4 mA, the step change
in resistance is less than ∆RAP−P . The boundaries are labeled by open symbols
and correspond to P and AP to IR transitions. These boundaries meet and join
the P to AP transition boundaries. Further, they define two triangular zones that
encompass IR states at high current magnitudes (both for positive and negative
136
Figure 5.18: Differential resistance versus current at various easy axis applied fieldsstarting from the P state. The magnitude of the current |Idc| is increased (blackcurves) and then decreased (red curves). (a) The black curve shows switching fromP to AP at 1.5 mA and also -1.2 mA, i.e. the switching occurs for both polaritiesof the current. At larger positive and negative current the resistance change isintermediate of that of the P to AP transitions. On reducing the current thereis a transition from the intermediate resistance (IR) state into an AP state. (b)At 40 mT switching from P to AP only occurs for positive polarity current andon reducing the current there is an IR to AP state transition for |Idc| ≤ 1 mA.(c) At 42 mT switching from P to AP again only occurs for positive polaritycurrent. However, on reducing the applied field the transition is from IR to P for|Idc| ≤ 1 mA.
137
Figure 5.19: Current swept state diagram of an OST spin valve device showing thethreshold currents for switching as a function of applied easy axis field. IP−APC andIAP−IRC are labeled by solid and open blue symbols. IAP−PC and IP−IRC are labeledby solid and open red symbols. The green curves indicate the IIR−APC (crosses)and IIR−PC (dashes), showing the bistability range of the IR states.
current polarities). As the current is swept back to zero, two parabolic shaped
curves (green) show the IR to P or AP transitions.
The general features of the state diagram of an OST spin valve device are
the following: (1) For magnetic fields near the FL coercive fields (H+C and H−C ),
current induced switching is bipolar. For fields close to but less than H+C , AP to
P transitions occur for both positive and negative currents and for fields near but
greater than H−C , P to AP transitions occur for both current polarities. (2) Near
H0, the center of the FL’s hysteresis loop, the switching occurs for only one current
polarity, positive current for the P to AP transition and negative current for the
AP to P current. (3) At large currents, transitions into an IR state are observed
138
Figure 5.20: (a) Representative FL minor hysteresis loops measured with a slowlyswept field at several fixed currents. The scale bar shows ∆RAP−P = 0.1 Ω, theresistance difference between the AP and P states. (b) State diagram constructedfrom dV/dI|H hysteresis loops. The color represents ∆R, the resistance differencebetween field increasing and field decreasing measurements. The central zone (or-ange color) corresponds to the AP/P bistable zone. Black dashed curves trace theboundaries between P, AP and IR states.
and this state persists even as the current is reduced well below the threshold
current for P/AP transitions. These features were seen in all ten 50 nm × 100 nm
ellipse devices that were studied.
The device states can also be determined by measuring the differential resis-
tance as a function of field at constant current. This is shown in Figure 5.20.
Figure 5.20(a) shows representative hysteresis loops at several currents. At zero
dc current the coercive field (HC = (H+C − H−C )/2) of the FL is largest and the
coercive field decreases as the current magnitude increases. For currents greater
than 2.9 mA or less than -1.4 mA, a plateau at a resistance between that of the
P and AP state resistances is seen, with the field range of the plateau increasing
with current magnitude.
A field swept state diagram is constructed as follows. The resistance measured
with decreasing field is subtracted from the resistance measured with increasing
139
field. The resistance difference ∆R is then plotted on a color scale versus current
and field (Figure 5.20(b)). ∆R is nonzero only in field ranges in which the device
response is hysteretic. The boundaries between zero and and non-zero ∆R regimes
are the boundaries between the P, AP and IR states. Thus the same general switch-
ing characteristics are observed in both current and field swept measurements.
5.4.2 Modeling
The device characteristics just discussed can be understood by employing the
biaxial macrospin model outlined in the first chapter of this thesis, with energy
landscape U(m) = K(Dm2z − m2
x). Spin-torque contributions due to both the
polarizer (magnetized out-of-the film plane, along z) and RL (magnetized in the
film plane along x) can be described in terms of effective spin-polarization direction
that is tilted with respect to the plane:
ΓS = Im× (m× nS),
nS =ηR
1− λRmx
x +ηP
1− λPmz
z. (5.72)
Here ηR,P and λR,P are the spin polarizations and spin-torque asymmetry pa-
rameters for the RL and polarizer layer respectively. I = (~/2e)I/(µ0M2s V ) is a
normalized applied current. Therefore, the combined spin-torque acting on the FL
magnetization can be written as:
ΓS = I
√1 + tan2(ωeff)
1− λRmx
m× (m× n) (5.73)
140
where the orientation of the effective spin-polarization axis n is tilted with respect
to the in-plane (IP) direction by an angle ωeff with
tan(ωeff) =ηRηP
(1− λRmx
1− λPmz
)= tan(ω)
1− λRmx
1− λPmz
. (5.74)
Naturally, in the case of ηP = 0 (no out-of-plane (OOP) polarizer), ΓS will reduce
to the conventional collinear spin-torque expression and n will lie in plane.
A qualitative understanding of central zone of the state-diagram can be seen
from the form of the spin-torque in (5.72). The torque associated with the reference
layer is initially collinear with the damping torque. It thus leads to switching
via the antidamping mechanism, typical of spin-transfer devices with collinear
magnetizations. However, the spin-transfer torque from the polarizer (∝ m ×
(m× z)) is equivalent to an effective field in the direction m× z, which is initially
in the direction of the FL’s medium axis y. Such a field reduces the FL’s easy axis
coercive field (for both current polarities) as is the case in the Stoner-Wohlfarth
model with a medium axis magnetic field. In the Stoner-Wohlfarth model the result
is an astroid shaped switching boundary, which resembles the diamond shaped
bistable central zone of the state diagram.
More quantitatively, the spin-torque asymmetry parameters λR,P lead to torques
that depend on the magnetization state of the FL. For example, different current
magnitudes are typically necessary for AP to P and P to AP switching [3]. Gen-
erally, larger currents are needed for P to AP switching at the same field, as seen
in the state diagram (see Figure 5.19). For simplicity in our model, we consider
λP = 0, as mz is typically small during the switching process. Therefore, λR ac-
counts for the main asymmetries we observe in spin-torque switching. We study
the switching by simulating an ensemble of 5000 macrospins under the influence of
141
spin-torque, thermal noise and an applied field. The ensemble is initially taken to
be thermalized in one of its (two possible) equilibrium easy-axis states before a cur-
rent is applied. After a current is applied the ensemble reaches a new steady-state
configuration over the course of a microsecond. The simulation is then repeated at
a higher current in incremental steps using the steady-state ensemble of the previ-
ous current as the initial condition. Upon reaching the limit of our current range,
we incrementally reduce the current to reproduce the procedure in the current
swept experiments.
Simulations results are plotted in Figure 5.21 both for current ramped-up (a),
and current ramped-down (b) cases, with parameters determined as follows. D is
governed by magnetic shape anisotropy and is calculated based on the FL’s shape to
be 17. The spin-torque asymmetry is taken to reproduce the measured ratio of pos-
itive I+C to negative I−C switching currents at effective zero field I+
C /I−C = 2, giving
λR = 0.5. K = 12µ0MsHKV is estimated to be 3.3×10−19 J (i.e. ξ = K/kBT = 80,
with T = 300 K), taking µ0Ms = 1.5T and µ0HK = 35 mT. Simulations were run
with spin-torque ratios ηP/ηR = 0, 0.24, 0.51, 0.68 and damping α = 0.04. The
results are shown as a colormap where red and green represent in-plane AP/P
configurations and blue corresponds to IR states. Figure 5.21 shows results for
ηP/ηR = 0.68. (Results for the other spin-torque ratios studies are shown in the
Supplementary Materials of [130].) Depending on the proportions in which the
ensemble is partitioned between the three available states, bistability regions arise
and are represented by a superposition of the colors, an example being the dark
yellow (green+red) P/AP bistability region at the center of the state diagram. The
magnetic field is normalized to HK and current is normalized by the positive criti-
cal current I+C at zero temperature. The normalized critical current (I+
C /IC ,I−C /IC)
142
and coercive field (H+C /IC ,H−C /IC) are smaller than 1 because of thermal fluctua-
tions.
The simulation captures the main switching features observed in the experi-
ment both for current ramping up and down. First, we observe a distorted di-
amond shaped central P/AP bistable central zone (Figure 5.21(a)), which shows
bipolar switching near the layer’s coercive field. The distortion (e.g. the lower
switching current for AP to P transitions) is associated with the non-zero spin-
torque asymmetry parameter λR. Second, there are transitions into an out-of-plane
precessional state which we associate with the IR state we observe experimentally.
We find that the threshold current for P/AP to IR transitions is higher when
the current is increasing than the IR to P/AP transitions when the current is
decreasing (Figure 5.21(b)), as observed in experiment. The vertical boundaries
in Figure 5.21(b) are artifacts that result from not having applied large enough
currents to realize an IR state when the current was increasing. In this case, when
decreasing the current, the ensemble appears to not transition out of its initial
AP or P configuration. Deviations between the experiment and macrospin model
appear at large currents. For example, the curvature of the AP to IR switching
boundary is positive in the simulation but negative in the experimental data. This
indicates that more sophisticated models, such as micromagnetic models, of the
magnetization dynamics may be needed to explain the large current dynamics.
The hysteretic transitions to the IR state can thus be understood under the frame-
work of the analytic theory presented in this thesis, which examined the influence
of the spin-torque ratio (ηP/ηR) on the magnetization dynamics.
143
Figure 5.21: Simulations of an ensemble of 5000 macrospins represented as a state-diagram with current increasing (a) and current decreasing (b). The three relevantstates AP, P and IR, are color coded as red, green and blue respectively. Eachdata point is represented by a RGB color that is determined by the proportion ofthe ensemble populating each corresponding state. Currents are shown in unitsof switching positive current at zero field and room temperature. Applied fieldsare shown in units of anisotropy field. The parameters used in the simulation aredescribed in the main text.
144
Chapter 6
Structure of Escape Trajectories
In this final chapter, work we will attempt to construct a general overview
of the effective energy barriers characterizing thermally driven systems. Since
gradient systems exhibit energy barriers which can be derived directly from the
energy landscape of the physical system, we will be particularly interested in non-
gradient systems (see the final section of Chapter 2). Let us then consider an
n-dimensional Langevin equation of motion with multiplicative noise understood
in the Ito sense.
x = F(x) +√εH(x) · W (6.1)
where F(x) and H(x) are the model’s given drift field and space-dependent vari-
ance, with ε the noise strength. At any given moment in time, the stochastic
contribution results from the addition of multiple gaussian processes with different
variances. These can, in turn, all be condensed into one single gaussian process
˙W with effective variance given by square summation:∑
j HijWj = gi˙W where
g2i =
∑j H
2ij.
145
6.1 Friedlin-Wentzell (FW) theory
Since the probability of witnessing a specific dynamical trajectory depends on
the realization of the noise ˙W in the system, the assumed gaussianity of the Wiener
process allows us to state that, in the limit of weak noise ε, the probability of the
system evolving from an initial position xA to a final state xB will be given by:
PA→B ∝ exp
[−1
2
∫ B
A
W 2dt
]= exp
[− 1
2ε
∫ B
A
(∑i
(xi − Fi)2
g2i
)dt
]. (6.2)
In the limit of small noise ε, the most likely path between xA and xB will
then be the one that, in maximizing the probability of realization, minimizes the
effective Lagrangian LFW =∑
i(xi−Fi)2
2g2iwhere the suffix FW stands for Friedlin
and Wentzell, who systematically defined this approach [172]. We will call the
minimizing trajectory the most probable escape path (MPEP) and the action cor-
responding to its transition between the equilibrium and unstable fixed points of
F will represent the effective energy barrier of the escape.
Denoting by G the diagonal matrix with elements Gii = g2i and the associ-
ated drift vector with elements fi = Fi/g2i , the Friedlin-Wentzell Lagrangian and
associated Hamiltonian can then be conveniently written as:
LFW =1
2(x− F) · G−1 · (x− F) (6.3)
HFW =1
2
[(p + f) · G · (p + f)− f · G · f
], (6.4)
where p is the canonical momentum with elements pi = ∂xiLFW . Upon inspection
146
of (6.3), it is at once apparent that escape dynamics are equivalent to those of
an electrically charged particle moving on a Riemannian manifold with positive
definite metric G under the action of a vector potential A = −f and electric scalar
potential Φ = −f · G · f = −|f |2G
,1 whose equations of motion can be obtained at
once using Hamilton’s equations:
x = G · p + F = G · (p + f) (6.5)
p = −∇xHFW , (6.6)
where optimal escape trajectories are those for which the Hamiltonian vanishes
identically. Setting HFW = 0, for any given state vector x, the set of momenta
satisfying the zero energy condition defines an n-dimensional ellipse in momentum
space centered in p = −f , having axes ai =√
f · G · f/g2i .
Limiting ourselves to a 2-dimensional system for simplicity2, the momentum
ellipse can be conveniently parametrized as:
px(γ) =|f |Ggx
cos γ − fx (6.7)
py(γ) =|f |Ggy
sin γ − fy. (6.8)
For any given position in configuration space, this momentum ellipse defines all
possible least-action motions accessible by an escape trajectory traveling through x.
1Given a Riemannian metric Q, we will denot the inner product of two vectors a and b as〈a,b〉Q. Analogously, the inner product on Q of a vector with itself will be written 〈a,a〉Q ≡ |a|2Q.
2The procedure being presented can be straightforwardly generalized to higher dimensions byconstructing the parametrization of a n-dimensional ellipse.
147
Figure 6.1: Diagram of momentum ellipse parametrized by γ. γ0 and π + γ0
correspond to instanton and anti-instanton solutions respectively.
Anti-instanton trajectories (p = 0) characterizing noiseless relaxation correspond
to γ = γ0 ≡ −atan( fyfx
gygx
) and reverse-drift motion (p = −2f) correponds to
γ = π + γ0. Figure 6.1 shows a typical diagram of such a momentum ellipse.
Since γ parametrizes momentum at each point in configuration space, we can
substitute (6.7) into the Hamiltonian equations (6.5) obtaining the following set
of equations satisfied by x solely as a function of x and γ:
x = gx|f |G cos γ (6.9)
y = gy|f |G sin γ. (6.10)
The importance of γ as characterizing the direction of escape is at once apparent.
148
In fact, constructing y/x one obtains that the slope of the escape trajectory is
∂xy = (gy/gx) tan γ. In presence of solely additive, or isotropic noise, G ≡ 1,
(6.9) shows that thermally driven and deterministic dynamics will evolve at the
same rate, |x| = |f |. This is trivial for instanton x = −f and anti-instanton
x = f trajectories, and could have also been obtained directly by writing down the
effective Lorentz dynamics
x = ∇|f |2 − x× (∇× f) (6.11)
and noting that upon multiplying by x one obtains ∂t|x|2 = ∂t|f |2 which again
implies that the dynamical speed is equal to the norm of the drift field. This notion
has been employed in the literature [173, 174] to construct an efficient numerical
scheme (the String method) capable of computing transition pathways, free energy
barriers and transition rates in complex systems. The notion that the thermal
escape speed is identical to the modulus of the drift field at each point in the
configuration space of the system further allows us to write the Lorentz equation
of motion in the general form:
(S + A) · x = 2∇|f |, (6.12)
where S and A are skew-symmetric and antisymmetric matrices respectively. Dy-
namical equations of this form have been extensively studied in recent litera-
ture [175, 176] where the effective potential 2|f | appearing on the right side of
(6.12) is argued to be a suitable Lyapunov function for the dynamics of x.
This result can be argued more generally, even for systems with multiplicative
noise by inspection of the Friedlin-Wentzel Lagrangian:
149
LFW =1
2|x− F|2
G−1 =
1
2
[|x|2
G−1 + |F|2
G−1 − 2x · G−1 · F
]= |f |2
G− x · f
= |f |2G
(1− cos Ψ) (6.13)
where we have employed the following identities:
F = G · f (6.14)
x · G−1 · x = |f |2G, (6.15)
and defined
Ψ ≡ acos
(x · f|f |2
G
)(6.16)
as the angle between the instantaneous escape velocity x and the field orientation
f at x. We find that, as long as Ψ does not vary too much over the course of the
escape trajectory, the effective action U(x) =∫ x
x0LFWdt will be dominated by the
behavior of |f |G which is in agreement with the statement that the norm of the
drift field coarsely captures the structure of the system’s action.
Often, non-gradient dynamics depending on a tunable set of parameters can
exhibit radically different thermally driven behavior as soon as a critical parameter
threshold is crossed [100,103]. This is reminiscent of phase transitions and has been
explored extensively in the literature. Let us then look at a sample system under
the lens developed up to now. We consider the following 2-dimensional stochastic
150
model taken from Maier and Stein [177,178]:
x = x(1− x2 − αy2) +√εW (6.17)
y = −y(1 + x2) +√εW, (6.18)
where noise is isotropic and the drift field is non-gradient (∇ × f 6= 0) for all
α > 1 (the model’s tunable parameter). For dynamics constrained on the x ≥ 0
half-plane, the point xS = (1, 0) is a stable fixed point of the dynamics whereas
xU = (0, 0) is an unstable saddle point. One is typically interested in the most
probable escape path (MPEP) leading from xS to xU . The system is known to
follow the x-axis along a soliton trajectory that counters the flow for all α < 4. At
α = 4, the MPEP bifurcates into two trajectories symmetric with respect to the
x-axis, with all other unsuccessful escape trajectories forming caustics with focus
point along the x-axis.
A simple analysis of the norm of the drift field unveils that in transitioning
across the critical threshold αC = 4, the structure of the extrema of f changes
sharply. If for α < 4, the norm exhibits two global minima at the fixed points
along with a saddle along the x-axis, for α > 4 two symmetric local minima
appear off the x-axis transforming the previous saddle into a local maxima. These
local minima are responsible for lowering the effective action of escape trajectories
transiting off the x-axis (see Figures 6.2).
151
Figure 6.2: Contour plot of the norm of the drift field taken from the Maier-Steinmodel [177]. On the left, for α = 3, two global minima are present at the unstable(0, 0) and stable (1, 0) equilibria respectively along with a saddle along the x-axis.On the right, for α = 5, the previous saddle has become a local maxima due tothe appearance of two new local minima off the x-axis.
152
6.2 Macrospin Escape Paths
The results of the previous section briefly allow us to make an important ob-
servation regarding the MPEPs expected to be followed by a macrospin subject to
spin-torques and thermal noise. For any non-zero applied current, the drift dynam-
ics (1.28) are non-gradient due to the non-conservative character of spin transfer
torque effects. One might then wonder whether the structure of the escape dy-
namics might be expected to change drastically as the dynamical parameters are
altered.
In the collinear uniaxial limit, the Hamiltonian dynamics arising from the min-
imization of the Friedlin-Wentzell action can be solved analytically. The F-W
Lagrangian for the thermally activated dynamics of the macrospin, expressed in
spherical coordinates (θ,φ), are:
LFW =1
2
[(θ − fθ)2 + sin2(θ)(φ− fφ)2
]. (6.19)
Passing to the Hamiltonian formalism we have:
HFW =1
2
[p2θ +
p2φ
sin2(θ)
]+ pθfθ + pφfφ, (6.20)
where pθ and pφ are the conjugate momenta of θ and φ.
Hamilton’s equations are then
153
θ = pθ + fθ (6.21)
φ =pφ
sin2(θ)pφ + fφ (6.22)
pθ =cos(θ)
sin3(θ)p2φ − pθ∂θfθ − pφ∂θfφ (6.23)
pφ = −pθ∂φfθ − pφ∂φfφ, (6.24)
where, up to this point, the setup is completely general and extensible to models
of any complexity (anisotropy ratio D 6= 0, positive tilts between axes, etc...). For
a uniaxial macrospin with collinear easy-/spin-current axes, the drift vector field
of the θ,φ dynamics can be obtained by rewriting (1.28) in spherical coordinates.
One then obtains
fθ = −α(j + cos(θ)) sin(θ) (6.25)
fφ = − cos(θ), (6.26)
where, in addition, j is the normalized spin-torque intensity.
We immediately notice that φ is a cyclic variable for this model and, as a result,
we expect pφ = c constant. Imposing our interest in a zero energy trajectory
starting from the hamiltonian HFW we can at once solve for pθ:
pθ = −fθ
[1±
√1− c
(2fφfθ
+αc
ξfθ sin2(θ)
)]. (6.27)
This solution is physically valid as long as the radicand is greater than or equal
154
to zero for all values of θ in the range of interest. This necessarily implies that
pφ = c = 0 which, in turn, is expected to remain constant for the whole switching
dynamics. The azimuthal coordinate φ can hence be expected to evolve following
the deterministic drift field even during thermal escape: φ = fφ. As a consequence,
pθ = −fθ[1 ± 1]. This is consistent with our knowledge of the gradient character
of the dynamics. In fact, pθ = 0 corresponds to standard relaxation to equilibrium
and pθ = −2fθ corresponds to an against-gradient instanton escape of the θ coor-
dinate. It is interesting to note that in the latter case, the precessional dynamics
of φ remain unaltered and, as such, escape from the stable well does not merely
correspond to a reflection in time.
Having been able to solve for the general momenta in terms of the spherical
coordinates, we can analytically derive the relation between φ and θ along the
escape trjectories. Dividing φ by θ from Hamilton’s equations, one has:
φ(θ) =1
α
∫ θ
π/2
cos(θ)
(j + cos(θ)) sin(θ)
=1
α(1− j2)log [tan(θ/2)]− j log
[−sin(θ/2) cos(θ/2)
j + cos(θ)
]− j log(2j)
+ const (6.28)
where the integration constant depends on the choice of the initial φ value at
the singular θ = π pole. Figure 6.3 compares this analytical result to the escape
trajectories obtained by numerically solving Hamilton’s equtions using a standard
shooting method. Various initial orientations of the escape trajectory at the θ = π
pole are simulated and, for each, the FW action is computed. Upon correcting for
the additive constant at the end of equation (6.28) we obtain very good matching
155
Figure 6.3: Escape trajectory for a uniaxial macrospin model with j = 0.3 andω = 0. Blue line shows the least-action result of numerical integration of FWdynamics. Dashed red line is the analytical result shown in equation (6.28).
between theory and numerics.
The inclusion of a small tilt ω between easy- and spin polarization axes does
not alter the escape dynamics. It simply rescales the current magnitude appearing
in 6.28 by the tilt-dependent critical switching current (j → j cosω) fits the data
very well (see Figure 6.4).
Upon passing to a biaxial macrospin model, it is not possible to solve the FW
dynamics analytically and no straightforward way exists to modify our collinear
uniaxial escape trajectory in a satisfactory way. We can however look at the norm
of the macrospin drift field and check wither its topological structure is stable
as the main parameters of the theory (α and j) are tuned. Without diving into
any calculations, we limit ourselves to state that the smallness of the damping
parameter α precludes any drastic alterations to the set of maxima and minima
156
Figure 6.4: Escape trajectory for a uniaxial macrospin model with j = 0.3 andω = 0.2π. Blue line shows the least-action result of numerical integration of FWdynamics. Dashed red line is the analytical result shown in equation (6.28).
appearing in the norm of the macrospin’s drift. The contribution of the precessional
terms will in fact dominate for all α ' 10 and lower. This should not surprise the
reader as this property of the macrospin dynamics is precisely what allowed the
development of the CEOA theory.
157
Conclusion
This thesis work has constructed and studied the theory underlying the dynam-
ics of a biaxial macrospin subject to thermal noise and the presence of spin-torque
due to both a perpendicularly magnetized polarizer and an in-plane magnetized
reference layer. Their combined spin-torque effects leads to an effective tilt ω
between the easy- and spin-polarization axes. The full stochastic magnetization
dynamics were solved employing the parallelization properties of graphics process-
ing units (GPUs) to repeatedly integrate the macrospin’s Langevin dynamics in an
effort to reconstruct its non-equilibrium ensemble properties. The technique has
allowed us to explore thermally activated behavior out to microsecond timescales.
We have also presented a theory capable of reducing the complexity of the
3D macrospin dynamics, under the action of both spin-torque and thermal noise,
to a 1D stochastic differential equation in the energy space of the macrospin. This
was achieved by averaging the sLLGS dynamics over constant energy trajectories
that were derived analytically.
Under such an approximation, the resulting theory predicts that the geometries
involved influence the respective dynamics in very precise ways. Particularly, we
found that the angular tilt between spin polarization and the easy-axes factors
158
into the dynamics only as a trivial rescaling of the applied current. Similarly, the
relative orientation of the hard axis is predicted to play no role in the switch-
ing dynamics under conditions where the constant-energy orbit averaging (CEOA)
assumption is a valid. We employed the theory to study the macrospin’s behav-
ior both within in-plane (IP) and out-of-plane (OOP) energetic basins. In both
scenarios, the main parameter characterizing the different dynamical regimes was
found to be the ratio D between the hard and easy axis anisotropies.
For IP dynamics, two critical currents I1C = (D + 2)/(2 cosω) and I0
C =
(2/π)√D(D + 1)/ cosω were found to exist: the relative magnitude of these crit-
ical currents depending nonlinearly on D. For D > D0 ' 5.09, we showed that
stable limit cycle magnetization precessions appear in well-defined current ranges;
transitions between these stable limit cycles proceeds through thermal activation.
When D < D0, limit cycles generally do not appear and the switching dynamics
become qualitatively similar to those of a uniaxial macrospin model.
Friedlin-Wentzell theory was employed to analytically study the exponential
scaling dependence between mean switching time and applied current in thermally
activated scenarios. The exact analytical scaling was reduced to quadratures and
an analytical approximation was suggested. The resulting analytical scaling de-
pendence to the standard form log(τ) ∝ ξ(1− I/IC)β and the current dependence
of the exponent β was studied (5.45). The exponent β was found to depend non-
linearly on the applied current intensity, similar to [147]. In the uniaxial macrsopin
limit D → 0, the constant β = 2 result was recovered. For D 6= 0, β was found to
depend on both the applied current intensity and the precise value of D.
For OOP dynamics, applied currents greater than IOOP = (2/π)√D + 1/ sinω,
were found to give rise to stable fixed points in the macrospin’s energy dynamics.
159
This is consistent with the description of a stable limit cycle, interpreted as an
OOP precessional state. We predict that stable OOP precessions are possible only
in one of the two out-of-plane directions, selected by the direction of the applied
current. Furthermore, by comparing our results to those obtained via CEOA
methods to study the threshold currents for magnetic switching, we predict the
occurence of hysteretic transitions between IP and OOP stable states for effective
tilts larger than a critical tilt ωC = arctan(1/√D), which has been observed in very
recent experiments [130]. For tilts ω < ωC , we predict that magnetic switching
will take place since the threshold current for onset of stable OOP precessionary
states is expected to be larger than that required for destabilization of the initial
IP basin. Overall, this leads to a very simple condition that spin-valves or MTJs
must satisfy to behave like a spin-torque oscillator (STO): ηref/ηpol <√D. Our
theory agrees with numerical results and represents a starting point for testing
how well the macrospin approximation captures the magnetization dynamics in
real world devices.
Upon exploring the thermal contribution to oscillator linewidth broadening,
we observe the existence of a critical temperature TC separating a regime where
phase noise dominates decoherence and one where decoherence is the result of both
phase and amplitude noise. The former cannot be accounted for by our CEOA
theory and is a result of the full complexity of the LLG dynamics. This is in
agreement with self-oscillator theory, where a transition temperature is predicted
to exist between a phase noise dominated regime at large temperatures and one
limited by thermal deflections about the equilibrium magnetic trajectory at low
temperatures [18, 161].
Our methodology is similar to that proposed by Slavin, Tiberkevich and Kim [12,
160
146, 148]. However, instead of approaching the multiscaling analysis by study-
ing the complex oscillatory amplitude of the macrospin’s dynamics using a self-
oscillator equation, we focused on the macrospin’s diffusion over its energy land-
scape. The loss of generality in doing so is compensated by new insights into the
macrospin’s dynamical characteristics capable of describing parameter spaces of
the macrospin model previously considered analytically unobtainable.
Last but not least, in the final chapter we have shown how to bypass the
limits of the constant energy orbit averaged theory and numerically reconstruct
the thermal escape properties of the macrospin dynamics. In doing so, we have
framed the sLLGS dynamics within the more general context of noisy dynamical
systems driven by nonconservative force fields. The most probable escape paths
MPEPs of such systems are in fact known to often exhibit sudden transitions due
to parametric instabilities. We have looked into whether the macrospin model may
exhibit such behavior and have found that the small size of the Landau damping
constant guarantees that such parametric instabilities will not occur.
Overall, the sLLGS dynamics studied throughout this thesis represent a first
approximation of the micromagnetic dynamics exhibited by thin (< 10 nm) mag-
netic layers . The choice of driving the system via DC currents was paramount
towards the numerical stability of the stochastic integrators employed. Upon in-
troducing other dynamical degrees of freedom (i.e. AC currents, non-constant MS,
etc.) the dynamics will likely become chaotic adding an entire layer of difficulty
to their analytical treatment. This should, however, not discourage the curious re-
searcher. The exploration of chaotic regimes to the macrospin dynamics might in
fact allow for more efficient techniques for destabilizing the IP basins allowing for
either faster switching transients or, conversely, more energy efficient destabiliza-
161
tion. It remains to be understood how far the macrospin model can truly be taken
as a proxy for the much richer micromagnetic dynamics. The true dynamics of a
magnetic element might in fact transition between regimes where the macrospin
model is applicable only on short timescales. Ultimately, the development of effi-
cient integration methods capable of evolving non-coherent magnetic states subject
to the thermal noise will be crucial towards answering such problematics.
162
Appendix A
Constant Energy Orbit Averaging
of sLLGS Equation
The procedure outlined in the main text considers the total time differntial of
the macrospin’s reference landscape energy:
ε = 2 (Dmzmz −mxmx) , (A.1)
where mx and mz are to be substituted in by employing the sLLGS dynamical
equations (1.28) which we rewrite for convenience.
163
mx = Dmzmy
+ α[(Inx +mx)(1−m2
x) + Inzmx(√
1− n2mz − nmy) +Dmxm2z
]my = −(D + 1)mxmz
+ α[mymz(Inz
√1− n2 +Dmz)−mxmy(Inx +mx) + Inzn(1−m2
y)]
mz = mxmy
− α[(Inz√
1− n2 +Dmz)(1−m2z) +mxmz(Inx +mx) + Inznmzmy
].
(A.2)
Plugging the sLLGS equation into the time differential for the macropin’s en-
ergy then results in:
ε = −2α
[Inz√
1− n2(Dmz −Dm3
z +m2xmz
)+ Inx
(Dmxm
2z +mx −m3
x
)+ Inznmy
(Dm2
z −m2x
)+(D2m2
z −D2m4z +m2
x −m4x + 2Dm2
xm2z
) ], (A.3)
which, collecting all terms Dm2z −m2
x = ε, simplifies to:
ε = −2α[I cos(ω)(1 + ε)mx + I sin(ω) cos(ψ)(D − ε)mz
+ I cos(ω) sin(ψ)εmy +D2m2z +m2
x − ε2] (A.4)
We remind the reader that our choice of notation was such that nx = cos θ,
164
nz = sin θ and n = sinψ represent the various polar and azimuthal tilts of the net
effective spin-polarization axis’ orientation with respect to the macropin’s easy-
and hard-axes.
Deriving now the energy evolution equation corresponding to IP (5.22) and
OOP (5.49) dynamics is a straightforward manner. In particular, the symmetry
of constant energy trajectories (see Figure 1.1) immediately makes sure that the
third term appearing in (A.4) is zero.
We now proceed to show how to compute the averages appearing in (5.22) and
(5.49) explicitely.
A.1 IP dynamics: 〈mx〉 and 〈m2x〉
The next step is to show how the necessary averages can be computed explicitly.
As an example, consider 〈mx〉; its average is given by:
〈mx〉 =1
T (ε)
∫ T
0
dtm0x(t), (A.5)
where the integration is over time and one uses the expression for m0x in terms
of its Jacobi elliptic function. We can express the same integral in terms of the
geometrical parametrization (5.21). In fact
〈mx〉 =1
T (ε)
∫ T
0
dtm0x(t) =
1
T (ε)
∮dw
dm0z/dw
mz0
m0x(w). (A.6)
165
Putting together the expression for mz0 from (5.2) and, again, employing the
geometrical parametrizations from (5.21), we obtain
〈mx〉 = − 4
T (ε)√D + 1
∫ acosh(1/γ−)
0
dwcosh(s)√
1− γ2− cosh2(s)
= − 2π
T (ε)√D + 1
= −π2
√D
D + 1− γ2−
1
K(1− γ2−). (A.7)
Repeating the procedure for the 〈m2x〉 term in (5.22),
〈m2x〉(ε) =
1
T(ε)
∫ T
0
dtm2x(t) =
4
T(ε)
∫ acosh(1/γ−)
0
ds|∂sm0z
m0x
|(m0x)
2
=4
T (ε)γ2−
√D
D + 1(D + 1− γ2
−)
∫ acosh(1/γ−)
0
dscosh(s)√
1− γ2− cosh2(s)
=4
T (ε)√D + 1
√D
D + 1− γ2−η1
=D
D + 1− γ2−
E(1− γ2−)
K(1− γ2−), (A.8)
where we have used:
T (ε) = 4
√D + 1− γ2
−
D(D + 1)η0(γ−) = 4
√D + 1− γ2
−
D(D + 1)K(1− γ2
−) (A.9)
A.2 OOP dynamics: 〈mz〉 and 〈m2z〉
To compute the constant energy orbit averages in (5.49), we write the integrals
using the geometric parametrizations (5.51):
166
〈mz〉m0 =±
T(ε)
∫ T
0
dtmz(t) =±4
T(ε)
∫ acosh(1/γ+)
0
ds|∂sm0z
m0z
|m0z
=±4
T(γ+)
γ+√D(D + 1)
∫ acosh(1/γ+)
0
dscosh(s)√
1− γ2+ cosh2(s)
=±π
2√D(1− γ2
+) + 1
1
K[1− γ2+].
(A.10)
Proceeding analogously for 〈m2z〉:
〈m2z〉m0 =
1
T(ε)
∫ T
0
dtm2z(t) =
4
T(ε)
∫ acosh(1/γ+)
0
ds|∂sm0z
m0z
|(m0z)
2
=4
T(γ+)
γ2+√
D(D + 1)√
1 +D(1− γ2+)
∫ acosh(1/γ+)
0
dscosh2(s)√
1− γ2+ cosh2(s)
=1
1 +D(1− γ2+)
E[1− γ2+]
K[1− γ2+],
(A.11)
where, as stated in the main text, E[x] is the complete elliptic integral of the second
kind.
In both derivations we have taken advantage of eqns. (15) and (24) to write
the period as a function of γ+. Written explicitly, the period reads:
T(ε) =4√
D(1 + ε)K[
D − εD(1 + ε)
] = 4
√1 +D(1− γ2
+)
D(D + 1)K[1− γ2
+]. (A.12)
167
A.3 Proof of Special Elliptic Integral Identities
In this section we’d like to prove a non-trivial elliptic integral identity used in
simplifying the expressions for (5.28) and (5.52).
e(x) ≡ xE(1− 1
x2) = E(1− x2) (A.13)
k(x) ≡ 1
xK(1− 1
x2) = K(1− x2) (A.14)
can be immediately proven by considering the defining differential equation satis-
fied by K(x), namely [179]:
∂xK(x) =E(x)− (1− x)K(x)
2x(1− x). (A.15)
Computing then explicitely the derivative of k(x) using definition (A1) and
rearranging, one finds that:
∂xk(x) =e(x)− (1− x)k(x)
2x(1− x), (A.16)
in other words they satisfy the same differential relation. The identical procedure
can be performed on e(x) thus proving the assertions made.
168
Appendix B
Orbit averaging of a Stratonovich
Equation
There are several advantages in adopting a Stratonovich convention when writ-
ing the dynamical equations. First, it is the most natural way of modeling a
physical process where the Gaussian noise represents the short correlation time
limit of a colored noise process: by the Wong-Zakai theorem [92], such a limit
of multiplicative noise converges to Statonovich calculus. Second, a Stratonovich
interpretation follows the conventional rules of calculus in dealing with functions
of a stochastic variable. Third, many conventional numerical schemes used to sim-
ulate Langevin equations (such as the Heun scheme adopted for this work) evolve
towards the Stratonovich solution.
The Stratonovich formulation of a stochastic differential equation (SDE), how-
ever, fails to accurately represent the correlation between multiplicative terms and
the specific noise realization [88]. To average the multiplicative noise terms over
constant energy orbits, we take advantage of the fact that sums of Gaussian ran-
169
dom variables∑
i µixi (where xi are standard 0 mean and variance 1 Gaussian vari-
ables) behave like a single Gaussian variable x with variance given by the square
sum of the individual variances µ2 =∑
i µ2i . Since the multiplicative noise terms
B(m) W appearing in our sLLGS equations are state-dependent, the Gaussian
variable summation cannot be employed due to the temporal correlation between
the state-dependent variances B2(m) and the specific noise realization W.
This problem can be avoided by converting the sLLGS equations into their Ito
representation. The multplicative noise terms of (5.49) become (DmzBxj−mxBzj)·
Wj (with summation over repeated indices). The state-dependent variances are
now uncorrelated with respect to the noise realization, and so a summation of
Gaussian random variables can now be employed. Averaging over constant energy
orbits then leads, after a bit of algebra, to the noise term appearing in (5.49).
Altering the multiplicative noise convention can generally alter the qualitative
nature of the solution to the stochastic differential equation. To maintain con-
sistency between Ito and Stratonovich models, the drift term must be modified
to ensure that Boltzmann equilibrium is obtained at long times in the absence of
non-conservative forces (in our case, the applied current). The fundamental rea-
son is that the SDE is simply a model of the underlying dynamics subject to two
constraints: the chosen form of the thermal noise and the steady-state equilbrium
Boltzmann distribution [164, 184]. In the absence of applied currents, (5.49) can
be written more concisely as:
〈∂tε〉 = [−αf(ε) + h(ε)] +
√2α
ξf(ε) · W (B.1)
170
with
f(ε) = 2[D(D + 1)〈m2
z〉+ ε(1 + ε)], (B.2)
where h(ε) represents the extra modification necessary in the drift term to retain
all physically relevant Boltzmann relaxation properties. Deriving the Ito Fokker-
Planck equation relative to such a dynamic then gives:
∂tρ = ∂ε
[(αf(ε)− h(ε) +
α
ξ∂εf(ε))ρ+
α
ξf(ε)∂ερ
]. (B.3)
Upon imposing h(ε) ≡ αξ∂εf(ε), the steady-state solution reduces to the simple
form ρeq(ε) ∝ exp[−ξ ε] as expected.
Employing the previously derived expression for 〈m2z〉 from Appendix A, h(ε)
is found to be (in terms of the auxiliary variable γ):
h(ε) =α
ξ
D(1− γ2+) + 1
1− γ2+
×[1−
(D(1− γ2
+) + 2
D(1− γ2+) + 1
)E[1− γ2
+]
K[1− γ2+]
+1
γ2+(2− γ2
+)
(E[1− γ2
+]
K[1− γ2+]
)2]
+α
ξ
D(1 + γ2+) + 1
D(1− γ2+) + 1
, (B.4)
which can be shown to lead to a negligible correction of the drift dynamics (≈
0.1α/ξ ≈ 10−5 since typical parameter values are α ∼ 0.01 and ξ ∼ 100).
171
Bibliography
[1] M.N. Baibich, J.M. Broto, A. Fert, F. Nguyen Van Dau, F. Petroff, P. Eitenne,
G. Creuzet, A. Friederich, and J. Chazelas, Phys. Rev. Lett. 61, 2472 (1988).
[2] G. Binasch, P. Grunberg, F. Saurenbach, and W. Zinn, Phys. Rev. B 39, 4828
(1988).
[3] J. Slonczewski, J. Magn. Magn. Mater. 159, L1-L7 (1996).
[4] L. Berger, Phys. Rev. B 54, 9353 (1996).
[5] Y. Tserkovnyak, A. Brataas, and G.E.W. Bauer, Phys. Rev. Lett. 88, 117601
(2002).
[6] W.H. Rippard, M.R. Pufall, and S.E. Russek, Phys. Rev. B 74, 224409 (2006).
[7] A.V. Nazarov, H M. Olson, H.Cho, K.Nikolaev, Z.Gao, S.Stokes, and B.B.
Pant, Appl. Phys. Lett. 88, 162504 (2006).
[8] Q. Mistral, J.-V. Kim, T. Devolder, P. Crozat, C. Chappert, J.A. Katine,
M.J. Carey, and K. Ito, Appl. Phys. Lett. 88, 192507 (2006).
[9] V.Tiberkevich, A. Slavin, and J.-V. Kim, Appl. Phys. Lett. 91, 192506 (2007).
172
[10] C. Boone, J.A. Katine, J.R. Childress, J. Zhu, X. Cheng, and I.N. Krivorotov,
Phys. Rev. B 79, 140404 (2009).
[11] S. Mizukami, Y. Ando, and T. Miyazaki, Jpn. J. Appl. Phys. 40, 580 (2001).
[12] Joo-Von Kim, V. Tiberkevich and A.N. Slavin, Phys. Rev. Lett. 100, 017207
(2008).
[13] J.C. Slonczewski and J.Z. Sun, J. Magn. Magn. Mater. 310, 169 (2007).
[14] A.L. Chudnovskiy, J. Swiebodzinski, and A. Kamenev, Phys. Rev. Lett. 101,
066601 (2008).
[15] Y. Tserkovnyak, A. Brataas, and G.E. Bauer, J. Magn. Magn. Mater. 320,
1282 (2008).
[16] D.C. Ralph and M.D. Stiles, J. Magn. Magn. Materials 320, 1190 (2008).
[17] S. Hernandez and R.H. Victora, Appl. Phys. Lett. 97, 062506 (2010).
[18] J. C. Sankey, I. N. Krivorotov, S. I. Kiselev, P. M. Braganca, N. C. Emley, R.
A. Buhrman, and D. C. Ralph, Phys. Rev. B 72, 224427 (2005).
[19] T.J. Silva and W.H. Rippard, J. Magn. Magn. Mater. 320, 1260 (2008).
[20] K.V. Thadani, G. Finocchio, Z.-P. Li, O. Ozatay, J.C. Sankey, I.N. Krivorotov,
Y.-T. Cui, R.A. Buhrman, and D.C. Ralph, Phys. Rev. B 78, 024409 (2008).
[21] V.S. Pribiag, G. Finocchio, B. Williams, D.C. Ralph, and R.A. Buhrman,
Phys. Rev. B 80, 180411 (2009).
[22] P. M. Braganca, B.A. Gurney, B.A. Wilson, J.A. Katine, S. Maat, and J.R.
Childress, Nanotechnology 21, 235202 (2010).
173
[23] S. Bonetti, V. Tiberkevich, G. Consolo, G. Finocchio, P. Muduli, F. Mancoff,
A. Slavin, and J. Akerman, Phys. Rev. Lett. 105, 217204 (2010).
[24] L. Liu, C.-F. Pai, D.C. Ralph, and R.A. Buhrman, Phys. Rev. Lett. 109,
186602 (2012).
[25] L. Fu, Z. X. Cao, S. Hemour, K. Wu, D. Houssameddine, W. Lu, S. Pistorius,
Y. S. Gui, and C.-M. Hu, Appl. Phys. Lett. 101, 232406 (2012).
[26] G. Tatara and H. Kohno, Phys. Rev. Lett. 92, 086601 (2004).
[27] Z. Li and S. Zhang, Phys. Rev. B 69, 134416 (2004).
[28] S.S.P. Parkin, M. Hayashi, and L. Thomas, Science 320, 190 (2008).
[29] G. Tatara, H. Kohno, and J. Shibata, Phys. Rep. 468, 213 (2008).
[30] G.S.D. Beach, M. Tsoi, and J. L. Erskine, J. Magn. Magn. Mater. 320, 1272
(2008).
[31] K.M.D. Hals, A.K. Nguyen, and A. Brataas, Phys. Rev. Lett. 102, 256601
(2009).
[32] C. Burrowes, A.P. Mihai, D. Ravelosona, J.-V. Kim, C. Chappert, L. Vila, A.
Marty, Y. Samson, F. Garcia-Sanchez, L.D. Buda-Prejbeanu, I. Tudosa, E.E.
Fullerton, and J.-P, Attan. Nat. Phys. 6, 17 (2010).
[33] J. A. Katine, F. J. Albert, and R. A. Buhrman, Phys. Rev. Lett. 84, 3149–3152
(2000).
[34] J. Grollier, V. Cros, A. Hamzic, J.M. George, H. Jaffres, A. Fert, G. Faini, J.
Ben Youssef, and H. Legall, Appl. Phys. Lett. 78, 3663 (2001).
174
[35] J. Z. Sun, Phys. Rev. B 62, 1 (2000).
[36] S. Mangin, D. Ravelosona, J.A. Katine, M.J. Carey, B.D. Terris, and E.E.
Fullerton, Nat. Mater. 5, 210 (2006).
[37] B. Ozyilmaz, A.D. Kent, D. Monsma, J.Z. Sun, M.J. Rooks, and R.H. Koch,
Phys. Rev. Lett. 91, 067203 (2003).
[38] S. Urazhdin, N.O. Birge, W.P. Pratt, and J. Bass, Phys. Rev. Lett. 91, 146803
(2003).
[39] E.B. Myers, F.J. Albert, J.C. Sankey, E. Bonet, R.A. Buhrman, and D.C.
Ralph, Phys. Rev. Lett. 89, 196801 (2002).
[40] T. Seki, S. Mitani, K. Yakushiji, and K. Takanashi, Appl. Phys. Lett. 88,
172504 (2006).
[41] I. N. Krivorotov, N. C. Emley, A. G. F. Garcia, J. C. Sankey, S. I. Kiselev,
D. C. Ralph and R. A. Buhrman, Phys. Rev. Lett. 93, 166603 (2004).
[42] D. Houssameddine, U. Ebels, B. Delaet, B. Rodmacq, I. Firastrau, F. Ponthe-
nier, M. Brunet, C. Thirion, J.-P. Michel, L. Prejbeanu-Buda, M.-C. Cyrille,
O. Redon and B. Dieny, Nature Materials 2007 6, 447-453 (2007).
[43] J.C. Slonczewski, US patent 5,695,864 (1997).
[44] K.L. Wang, J.G. Alzate, and P.K. Amiri, J. Phys. D: Appl. Phys. 46, 074003
(2013).
[45] Survey, U. S. Coast And Geodetic. Principal Facts of the Earth’s Magnetism
and Methods of Determining the True Meridian and the Magnetic Declination.
London: Forgotten Books. (Original work published 1909).
175
[46] Memoires sur l’Electricite et le Magnetisme, Histoire de l’Academie Royale
des Sciences, (1785-1789).
[47] J.C. Maxwell, On Physical Lines of Force, Philosophical Magazine (1861);
A Dynamical Theory of the Electromagnetic Field, Phil. Trans. of the Royal
Society of London, (1865).
[48] L.D. Landau and E.M. Lifshitz, Phys. Z. Sowietunion 8, 153 (1935).
[49] W.F. Brown, Jr.,J. App. Phys. 49 3, 1937–1942 (1978).
[50] G. Bertotti, Hysteresis in Magnetism - For Physicists, Material Scientists and
Engineers. Electromagnetism, Academic Press (1998).
[51] A.H. Morrish, The Physical Principles of Magnetism. New York, Wiley (1975).
[52] D. Spisak and J. Hafner, Phys. Rev. B 55, 8304 (1997).
[53] D. Spisak and J. Hafner, J. Mag. Mag. Mat., 168 3, (1997).
[54] M.H. Levitt, Spin Dynamics: Basics of Nuclear Magnetic Resonance. Wiley,
2nd Edition (2008).
[55] E.U. Condon and G.H. Shortley, The Theory of Atomic Spectra. Cambridge
University Press (1935).
[56] V. Kambersky, Can. J. Phys. 48, 2906 (1970).
[57] V. Kambersky, Czechoslovak Journal of Physics, Section B 26, 1366 (1976).
[58] A. Misra and R. H. Victora, Phys. Rev. B 73, 172414 (2006).
176
[59] A. Y. Dobin and R. H. Victora, Phys. Rev. Lett. 92, 257204 (2004); J. Appl.
Phys. 95, 7139 (2004).
[60] M. C. Hickey and J. S. Moodera, Phys. Rev. Lett. 102, 137601 (2009).
[61] J. Foros, A. Brataas, Y. Tserkovnyak, and G. E. W. Bauer, Phys. Rev. Lett.
95, 016601 (2005); Phys. Rev. B 78, 140402 (2008).
[62] T.L. Gilbert, Phys. Rev. 100, 1234 (1955).
[63] B. Ozyilmaz, A. D. Kent, J. Z. Sun, M. J. Rooks, and R. H. Koch, Phys. Rev.
Lett. 93, 176604 (2004).
[64] E.B. Myers, D.C. Ralph, J.A. Katine, R.N. Louie, and R.A. Buhrman, Science
285, 867 (1999).
[65] T. Dunn and A. Kamenev, J. Appl. Phys. 115, 233906 (2014).
[66] M.D. Stiles and A. Zangwill, Phys. Rev. B 66, 014407 (2002).
[67] S. Zhang, P.M. Levy, and A. Fert, Phys. Rev. Lett. 88, 236601 (2002).
[68] H. Kubota, A. Fukushima, K. Yakushiji, T. Nagahama, S. Yuasa, K. Ando,
H. Maehara, Y. Nagamine, K. Tsunekawa, D. D. Djayaprawira, N. Watanabe,
and Y. Suzuki, Nat. Phys. 4, 37 (2008).
[69] I. Theodonis, N. Kioussis, A. Kalitsov, M. Chshiev, and W. H. Butler, Phys.
Rev. Lett. 97,237205 (2006).
[70] M.A. Zimmler, B. Ozyilmaz, W. Chen, A.D. Kent, J.Z. Sun, M.J. Rooks, and
R.H. Koch, Phys. Rev. B 70, 184438 (2004).
177
[71] J.C. Sankey, Y.-T. Cui, J.Z. Sun, J.C. Slonczewski, R.A. Buhrman, and D.C.
Ralph, Nat. Phys. 4, 67 (2008).
[72] R. H. Koch, J. A. Katine, and J. Z. Sun Phys. Rev. Lett 92, 088302 (2004).
[73] S. Petit, C. Baraduc, C. Thirion, U. Ebels, Y. Liu, M. Li, P. Wang and B.
Dieny, Phys. Rev. Lett. 98, 077203 (2007).
[74] A. M. Deac, A. Fukushima, H. Kubota, H. Maehara, Y. Suzuki, S. Yuasa, Y.
Nagamine, K. Tsunekawa, D. D. Djayaprawira, and N. Watanabe. Nat. Phys.
4, 803 (2008).
[75] W. F. Brown, Phys. Rev. 135, 5 (1963).
[76] A. Kamenev Field Theory of Non-Equilibrium Systems, Cambridge University
Press, (2011).
[77] Aditi Mitra, So Takei, Yong Baek Kim, and A. J. Millis, Phys. Rev. Lett. 97,
236808 (2006).
[78] U. Ebels, D. Houssameddine, I. Firastrau, D. Gusakova, C. Thirion, B. Dieny,
and L. D. Buda-Prejbeanu, Phys. Rev. B 78, 024436 (2008).
[79] I. Firastrau, D. Gusakova, D. Houssameddine, U. Ebels, M.-C. Cyrille, B.
Delaet, B. Dieny, O. Redon, J.-Ch. Toussaint, and L. D. Buda-Prejbeanu,
Phys. Rev. B 78, 024437 (2008)
[80] C. Papusoi, B. Delaet, B. Rodmacq, D. Houssameddine, J.-P. Michel, U.
Ebels, R. C. Sousa, L. Buda-Prejbeanu and B. Dieny, App. Phys. Lett. 95
072506 (2009)
178
[81] H. Liu, D. Bedau, D. Backes, J. A. Katine, J. Langer, and A. D. Kent, App.
Phys. Lett. 97, 242510 (2010).
[82] H. Liu, D. Bedau, D. Backes, J. Katine, and A. D. Kent, App. Phys. Lett.
101, 032403 (2012).
[83] L. Ye, D. B. Gopman, L. Rehm, D. Backes, G. Wolf, T. Ohki, A. F. Kirichenko,
I. V. Vernik, O. A. Mukhanov, and A. D. Kent, J. App. Phys. 115, 17C725
(2014).
[84] P.-B. He,Z.-D. Li, A.-L. Pan, Q. Wan, Q.-L. Zhang, R.-X. Wang, Y.-G. Wang,
W.-M. Liu and B.-S. Zou, Phys. Rev. B 78, 054420 (2008).
[85] P.-B. He,Z.-D. Li, A.-L. Pan, Q.-L. Zhang, Q. Wan, R.-X. Wang, Y.-G. Wang,
W.-M. Liu and B.-S. Zou, J. Appl. Phys. 105, 043908 (2009).
[86] Z. Hou, Z. Zhang, J. Zhang, and Y. Li, APL 99, 222509 (2011).
[87] C.-M. Lee, J.-S. Yang, and T.-H. Wu, IEEE Trans. Mag. 47, 649 (2011).
[88] I. Karatsas and S. Shreve, Brownian Motion and Stochastic Calculus, 2nd
ed.(Springer-Verlag, New York, 1997).
[89] J. L. Garcia-Palacios and F. J. Lazaro, Phys. Rev. B 68, 22 (1998).
[90] R. Zwanzig, Nonequilibrium Statistical Mechanics (Oxford Univerity Press,
Oxford, UK, 2001).
[91] B. Øksendal, Stochastic Differential Equations Springer, New York, (2000).
[92] E. Wong and M. Zakai, Ann. Math. Statist. 36 (1965), p. 1560.
179
[93] A. Kolmogorov, ”On Analytical Methods in the Theory of Probability”, 448-
451, (1931).
[94] J.C. Crocker and D.G. Grier, Phys. Rev. Lett. 73, 352 (1994).
[95] J.C. Crocker and D.G. Grier, J. Colloid Interface Sci. 179, 298 (1996).
[96] H. K. Janssen, Z. Phys. B 23, 377 (1976).
[97] B. Sun, J. Lin, E. Darby,A.Y. Grosberg and D.G. Grier, Phys. Rev. E 80,
010401R (2009).
[98] B. Sun, D.G. Grier and A.Y. Grosberg, Phys. Rev. E 82, 021123 (2010).
[99] R.S. Maier and D.L. Stein, J. Stat. Phys. 83, 291 (1996).
[100] R.S. Maier and D.L. Stein, Phys. Rev. E 48, 931 (1993).
[101] M.I. Dykman, D.G. Luchinsky, P.V.E. McClintock, and V.N. Smelyanskiy,
Phys. Rev. Lett. 77, 5229 (1996).
[102] M.I. Dykman, E. Mori, J. Ross and P.M. Hunt, J. CHem. Phys. 100, 5735
(1994).
[103] M.I. Dykman, M. Millonas and V.N. Smelyanskiy, Phys. Lett. A 195, 53
(1994).
[104] S. Chandrashekhar, Rev. Mod. Phys. 15, 1 (1943).
[105] H. Nyquist, Phys. Rev. 32, 110–113 (1928).
[106] H.B. Callen, T.A. Welton, Phys. Rev. 83, 34–40 (1951).
180
[107] H. Poincare, ”Sur les courbes definies par une equation differentielle”, Oeu-
vres 1, Paris (1892)
[108] I. Bendixson, Acta Mathematica 24, Issue 1, pp 1-88 (1901).
[109] M. Suzuki, Adv. Chem. Phys. 46, 195 (1981).
[110] B. Caroli, C. Caroli, and B. Roulet, J. Stat. Phys. 26, 83, (1981).
[111] R. Blackmore and B. Shizgal, Phys. Rev. A 31, 1855 (1985).
[112] R.C. Desai and R. Zwanzig, J. Stat. Phys. 19, 1 (1978).
[113] K. Komentani and H. Shimizu, J. Stat. Phys. 13, 473 (1975).
[114] J.J. Brey, J.M. Casado, and M. Morillo, Physica A 128, 497 (1984).
[115] H. Haken, Physics Letters A 53, 1 (1975); Rev. Mod. Phys. 47, 175 (1975).
[116] H. Grabert and M.S. Green, Phys Rev. A 19, 1747 (1979).
[117] H. Dekker, Phys. Rev. A 19, 2102 (1979).
[118] M.F. Wehner and W.G. Wolfer, Phys. Rev. A 27, 2663 (1983).
[119] D.L. Ermak and H. Buckholtz, J. Comput. Phys. 35, 169 (1980).
[120] N.G. Kampen, J Stat. Phys. 17, 71 (1977).
[121] H. Tomita, A. Ito, and H. Kidachi, Prog. Theor. Phys. 56, 786 (1976).
[122] P.D. Lax, Commun. Pure Appl. Math. 6, 231 (1953).
[123] G.E. Forsyth and W.R. Wasow, Finite Difference Methods for Partial Dif-
ferential Equation, Wiley, New York (1967).
181
[124] V. Palleschi, F. Sarri, G. Marcozzi, and M.R. Torquati, Phys. Lett. A 146,
378 (1990).
[125] http://en.wikipedia.org/wiki/ASCI_Red
[126] NVIDIA CUDA webpage, http://www.nvidia.com/object/cuda_home.
html
[127] P.E. Kloeden and E. Platen, Numerical Solution of Stochastic Differential
Equations. No. 23 in Application of Mathematics, springer, Berlin (1995).
[128] I. Gikhman, A. Skorokhod, Stochastic Differential Equations, Springer,
Berlin (1972).
[129] D. Bedau, H. Liu, J. Z. Sun, J. A. Katine, E. E. Fullerton, S. Mangin and
A. D. Kent, Appl. Phys. Lett. 107, 262502 (2010).
[130] L. Ye, G. Wolf, D. Pinna, G. D. Chaves, A. D. Kent, arXiv:1408.4494 (2014).
[131] N. Corporation, NVIDIA CUDA programming guide v6, available from
NVIDIA CUDA webpage, http://www.nvidia.com/object/cuda_home.
html, (2014).
[132] M. Januszewski, M. Kostur, Comp. Phys. Comm. 181, 183 (2010).
[133] H. Nguyen, GPU Gems 3 (Addison-Wesley Professional, 2007).
[134] P. L’Ecuyer, Operations Research 44, 5 (1996), pp. 816-822.
[135] W.H. Press, S.A. Teukolsky, W.T. Vetterling and B.P. Flannery, Numerical
recipes in C: The art of scientific computing (Cambridge University Press,
1992).
182
[136] T. Taniguchi and H. Imamura, Phys. Rev. B 83, 054432 (2011).
[137] T. Taniguchi and H. Imamura, Phys. Rev. B 85, 18440 (2012).
[138] W. H. Butler, T. Mewes, C. K. A. Mewes, P. B. Visscher, W. H. Rippard, S.
E. Russek, Ranko Heindl arXiv:1202.2621
[139] D. M. Apalkov and P. B. Visscher, Phys. Rev. B 72, 180405R (2005).
[140] W. T. Coffey, Phys. Rev. B 68, 3249 (1998).
[141] W. T. Coffey, Y.P. Kalmykov, J.T. Waldron The Langevin Equation with Ap-
plications in Physics, Chemistry, and Electrical Engineering (World Scientific
Pub Co Inc, 1996).
[142] S. I. Denisov and A. N. Yunda, Physica B 245, 282 (1998).
[143] C. N. Scully, P. J. Cregg, D. S. F. Crothers, Phys. Rev. B 45, 474 (1992).
[144] D. A. Garanin, Phys. Rev. E 54, 3250 (1996).
[145] G. Bertotti, C. Serpico, I.D. Mayergoyz, A. Magni, M. d’Aquino, and R.
Bonin, Phys. Rev. Lett. 94 127206 (2005).
[146] A. Slavin, V. Tiberkevich, IEEE Trans. Mag. 44, 7 (2008).
[147] T. Taniguchi, Y. Utsumi, M. Marthaler, D. S. Golubev and H. Imamura,
Phys. Rev. B 87, 054406 (2013).
[148] A. Slavin, V. Tiberkevich, IEEE Trans. Mag. 45, 4 (2009).
[149] S. Strogatz, Nonlinear Dynamics and Chaos (Addison-Wesley, Reading, MA,
1994).
183
[150] S. Urazhdin, V. Tiberkevich, and A. Slavin Phys. Rev. Lett. 105, 237204
(2010).
[151] M. Quinsat, D. Gusakova, J. F. Sierra, J. P. Michel, D. Houssameddine,
B. Delaet, M.-C. Cyrille, U. Ebels, B. Dieny, L. D. Buda-Prejbeanu, J. A.
Katine, D. Mauri, A. Zeltser, M. Prigent, J.-C. Nallatamby and R. Sommet,
App. Phys. Lett. 97, 182507 (2010).
[152] J. F. Sierra, M. Quinsat, F. Garcia-Sanchez, U. Ebels, I. Joumard, A. S.
Jenkins, B. Dieny, M.-C. Cyrille, A. Zeltser and J. A. Katine, App. Phys.
Lett. 101, 062407 (2012).
[153] F. M. de Aguiar, A. Azevedo, and S. M. Rezende, Phys. Rev. B, 75 132404
(2007).
[154] K. Newhall and E. Vanden-Eijnden, J. Appl. Phys. 113, 184105 (2013).
[155] D. Pinna, D. L. Stein, and A. D. Kent, IEEE Trans. Mag. 49, 7 (2013).
[156] D. Pinna, A. D. Kent, and D. L. Stein, Phys. Rev. B 88, 104405 (2013).
[157] D. Pinna, D. L. Stein, and A. D. Kent, Phys. Rev. B 90, 174405 (2014).
[158] M. Abramowitz, I. Stegun, Handbook of Mathematical Functions (Dover Pub-
lications, 1964).
[159] D. Pinna, Aditi Mitra, D. L. Stein and A. D. Kent, Appl. Phys. Lett. 101,
262401 (2012).
[160] T. Taniguchi, App. Phys. Exp. 7, 053004 (2014).
[161] J.-V. Kim, Solid State Physics, vol. 63, pp.217-294, (Academic Press, 2012).
184
[162] G. Bertotti, I. Mayergoyz, and C. Serpico, Nonlinear Magnetization Dynam-
ics in Nanosystems (Elsevier, Oxford, UK, 2009).
[163] A. W. C. Lau and T. C. Lubensky, Phys. Rev. E 76, 011123 (2007).
[164] G. Pesce, A. McDaniel, S. Hottovy, J. Wehr and G. Volpe, Nature Commu-
nications Nov 12 (2013)
[165] A. Brataas, A. D. Kent and H. Ohno, Nature Materials 11, 372 (2012).
[166] D. Pinna, D. L. Stein, and A. D. Kent, J. Appl. Phys. 114, 033901 (2013).
[167] H.A. Kramers, Physica VII, 4 (1940)
[168] D.E. Nikonov, G.I. Bourianoff, G. Rowlands, and I.N. Krivorotov, J. Appl.
Phys. 107, 113910 (2010).
[169] O.J. Lee, D.C. Ralph, and R.A. Buhrman, App. Phys. Lett. 99, 102507
(2011).
[170] G.E. Rowlands, T. Rahman, J. A. Katine, J. Langer, A. Lyle, H. Zhao, J.G.
Alzate, A.A. Kovalev, Y. Tserkovnyak, Z.M. Zeng, H.W. Jiang, K. Galatsis,
Y.M. Huai, P. Khalili Amiri, K.L. Wang, I.N. Krivorotov, and J.-P. Wang,
App. Phys. Lett. 98. 102509 (2011).
[171] Junbo Park, D.C. Ralph, and R.A. Buhrman, App. Phys. Lett. 103, 252406
(2013).
[172] M. I. Freidlin and A. D. Wentzell Random Perturbations of Dynamical Sys-
tems (Springer, New York, 1991).
[173] E. Weinan and E. Vanden-Eijnden, Phys. Rev. B 66, 052301 (2002).
185
[174] E. Weinan, W. Ren, E. Vanden-Eijnden, J. Chem. Phys 126, 164104 (2007).
[175] P. Ao, Communications in Theoretical Physics 49, 1073 (2008).
[176] C. Kwon, P. Ao, and D.J. Thouless, Proc. Nat’l Acad. Sci. 102, 13029 (2005).
[177] R.S. Maier and D.L. Stein, Phys. Rev. Lett. 69, 3691 (1992).
[178] R.S. Maier and D.L. Stein, Phys. Rev. Lett. 71, 1783 (1993).
[179] E. T. Whittaker, G. N. Watson, A Course in Modern Analysis 4th Ed. (Cam-
bridge University Press, Cambridge, UK 1927).
[180] R. Kubo, N. Hashitsume, Supp. Prog. Theor. Phys. 46, pp. 210-220 (1970).
[181] W. Rumelin, SIAM Journal on Numerical Analysis 19, 3 (1982), pp. 604-613
[182] R. V. Kohn, M. G. Reznikoff, and E. Vanden-Eijnden, J. Nonlinear Sci. 15,
pp. 223–253 (2005).
[183] Y. Saito and T. Mitui, Ann. Inst. Statist. Math. 45, 3 (1993).
[184] A.W.C. Lau and T.C Lubensky, Phys. Rev. E 76, 011123 (2007).
186