Small Codes for Magic State Distillation

Mark Howard∗

Department of Physics and Astronomy, University of Sheffield, Sheffield, UK andInstitute for Quantum Computing and Department of Applied Mathematics,

University of Waterloo, Waterloo, Ontario, Canada, N2L 3G1

Hillary DawkinsInstitute for Quantum Computing and Department of Physics and Astronomy,

University of Waterloo, Waterloo, Ontario, Canada, N2L 3G1

Magic state distillation is a critical component in leading proposals for fault-tolerant quantumcomputation. Relatively little is known, however, about how to construct a magic state distillationroutine or, more specifically, which stabilizer codes are suitable for the task. While transversalityof a non-Clifford gate within a code often leads to efficient distillation routines, it appears to notbe a necessary condition. Here we have examined a number of small stabilizer codes and highlighta handful of which displaying interesting, albeit inefficient, distillation behaviour. Many of thesedistill noisy states right up to the boundary of the known undististillable region, while some distilltoward non-stabilizer states that have not previously been considered.

I. INTRODUCTION

Most efforts towards building a large-scale quantumcomputer use error-correcting codes to protect the quan-tum information. However, no matter what code is cho-sen, the set of gates that are transversal (i.e. manifestlyfault-tolerant) will be non-universal [1], meaning theseoperations are insufficient for useful quantum compu-tation. Therefore, some additional resource will be re-quired to supplement the limited set of operations. Themost promising technique for circumventing this issue isto supplement the non-universal gate set with a supplyof special resource states, known as magic states. Havingaccess to pure magic states enables implementation of anadditional unitary operation, which ultimately providesus with a universal gate set. The magic state distillationprotocol, as introduced by Knill [2] and Bravyi and Ki-taev [3], provides a method of producing these resourcestates through an iterative procedure in which less puremagic states are consumed to produce a higher puritymagic state using only stabilizer operations.

Here we use small stabilizer codes, which means thatthe number of qubits, the number of measurements andthe number of nonlocal operations involved in each roundof distillation are all small. While these are attractive fea-tures, it seems that using a small code limits the amountof purification that can occur per round of distillation.For practical purposes, one typically quantifies how manyraw magic states are required to produce a single magicstate of very high purity (e.g., infidelity of no more than10−10). Applying this metric one typically finds thatsmall codes fare much worse than larger codes like the 15-qubit Reed-Muller code [3]. If instead we rank codes bytheir range of applicability, i.e., the threshold noise rateabove which purification no longer occurs, then smaller

∗ m.howard@sheffield.ac.uk

codes seem to fare at least as well as larger codes. Al-though this may not be a pressing issue experimentally,since it seems reasonable to expect raw state prepara-tion infidelity on the order of 10−2–10−4, it is of founda-tional interest. Bravyi and Kitaev highlighted the factthat tight magic state distillation routines imply a sharptransition, as noise is decreased, from a circuit that isclassically efficiently simulable to one that enables uni-versal quantum computation. The question of tight dis-tillability for qutrit states was addressed in [4, 5].

Efficient codes for magic state distillation typically ex-hibit quadratic (p 7→ O(p2)) or cubic (p 7→ O(p3)) sup-pression of the error parameter p. The codes listed hereall exhibit linear error suppression and consequently arenot competitive with existing routines in terms of effi-ciency. Nevertheless we feel that exploring the landscapeof codes that achieve distillation is still worthwhile. Re-ichardt [9] has previously summarized a handful of knowndistillation routines exhibiting the best thresholds.

FIG. 1. The stabilizer octahedron inscribed within the Blochsphere: Six Pauli eigenstates form the vertices of an octahe-dron. States within the octahedron are provably undistillableso the best one can hope for is to distill states up to the bound-ary. The two pure non-stabilizer states singled out here, |H〉and |T 〉, were shown to be distillable by Bravyi and Kitaev[3].

arX

iv:1

512.

0476

5v2

[qu

ant-

ph]

4 M

ar 2

016

2

A number of the codes presented below achieve tightdistillation right up to the boundary of the stabilizer oc-tahedron – the convex hull of Pauli eigenstates depictedin Fig. 1 – whose interior contains states that are prov-ably undistillable. We also present codes that converge,upon iteration, to states that are not those that are typi-cally considered i.e., the H-type and T -type magic states.None of our codes require twirling (which diagonalizes thestate in the {|H〉, |H⊥〉} or {|T 〉, |T⊥〉} basis) betweenrounds.

II. BACKGROUND

Here we briefly summarize notation and terminologyassociated with both magic state distillation and withthe type of quantum codes that we have used.

A. Magic state distillation, thresholds and yields

Magic state distillation routines are described in termsof stabilizer error correcting codes, which in turn are de-scribed by a set of generators {Gi} consisting of Paulioperators. The protocol (see e.g. [6]) consists of itera-tively applying the following steps:

1. Prepare n copies of the input state ρ⊗nin , where n isthe size of the code

2. Perform Pauli measurements corresponding to eachof the n − k generators Gi, and postselect on thedesired outcome

3. Optionally perform a Clifford transformation basedon the measurement outcome.

When successful, the output state(s) will be purified inthe direction of the target magic state. Typically k = 1and consequently an [[n, k = 1,distance]] stabilizer codeis used (see [10–12] for k ≥ 2 however).

For simplicity, one can assume that input states sufferfrom depolarizing noise and take the form

ρM = (1− p)|M〉〈M |+ p12

2, (1)

where |M〉 is the target magic state. The maximumvalue of p for which distillation is possible even in prin-ciple, which we call poct, is when ρM crosses the bound-ary into the stabilizer octahedron. With reference toFig 1, |H〉-type states are of the form |H〉 = (|0〉 +

eiπ/4|1〉)/√

2 or any image of this under a Clifford gate,while |T 〉-type states have Bloch vectors of the form

(±1,±1,±1)/√

3. The simple geometry allows us to find

poct = (1 − 1/√

2) ≈ 0.2929 for |H〉-type states and

poct = (1− 1/√

3) ≈ 0.4226 for T -type states.

1 2 3

(a) Graph with labelled vertices

|+〉

|+〉

|+〉

|Γ〉 =

∑a,b,c

(−1)ab+bc|a,b,c〉2√2

(b) The graph state preparation circuit for the above graph

FIG. 2. Creating the graph state associated with a givengraph amounts to performing controlled-Z rotations betweenqubits whose vertices are connected by an edge in the graph.

The yield [8] of a magic state distillation routine isdefined as

Y (p, pout) =∏

k=1...N

p(k)s

n(2)

where N is the total number of iterations needed to ob-tain pout starting at initial error rate p, and p

(k)s is the

probability of success on the kth iteration. This quan-tity relates to the efficiency/resource overhead of a magicstate distillation routine.

B. Codeword stabilized quantum codes

The Pauli measurements associated with a stabilizercode can be written in binary symplectic notation so thatG = (Xx1 ⊗ Xx2 ⊗ . . . ⊗ Xxn)(Zz1 ⊗ Zz2 ⊗ . . . ⊗ Zzn)becomes (x1, x2, . . . , xn|z1, z2, . . . , zn). An n-qubit stabi-lizer code with n generators has k = 0 and consequentlyspecifies a 1-dimensional subspace of Hilbert space i.e, astabilizer state. By applying local Cliffords (the unitariesthat map Pauli operators to Pauli operators under conju-gation), a generic stabilizer state (X|Z) can be broughtto the form (1n|Γ) where Γ is the adjacency matrix of thegraph. When stabilizer states take this particular formthey are called graph states, sometimes denoted |Γ〉.

The majority of magic state distillation routines pre-sented here will be described by codeword stabilized(CWS) codes (see [13]). In this representation, the codeis given by a graph state written as an adjacency matrix,Γ ∈ Zn×n2 , as well as a classical codeword, w ∈ Zn2 . Wemay choose the logical basis state |0L〉 of our code to bethe state

|0L〉 = |Γ〉 =∑

x∈Zn2

ixT Γx|x〉. (3)

3

Subsequently the logical |1L〉 operator is given by

|1L〉 = Zw|0L〉 =∑

x∈Zn2

ixT Γx+2wT x|x〉, (4)

where w is the codeword and Zw denotes the product ofZ operators

Zw = Zw1 ⊗ ...⊗ Zwn . (5)

The unitary encoding/decoding associated with thesecodes is quite straightforward as illustrated in Fig. 2b –it involves controlled-Z rotations applied to qubits whosevertices are connected by edges in the graph. A con-venient way of visualizing CWS codes is to display thegraph and highlight the subset of vertices 1 ≤ i ≤ n suchthat wi = 1 see Fig. 3,4 and 6.

III. RESULTS

All stabilizer codes are local Clifford equivalent to someCWS stabilizer code. Consequently CWS stabilizer codesform a subset of all stabilizer codes and so checking allcombinations of graphs, Γ, and codewords, w, may stillmiss stabilizer codes that are useful for distillation. Nev-ertheless, for n ≤ 4 qubits we iterated over all graphson n vertices and for n ∈ {5, 6} qubits we iterated overnon-isomporphic graphs on n vertices. We were primar-ily focused on recording those CWS codes that achievetight distillation but we also noted a number of non-tightcodes that distilled to target states other than |H〉.

With the exception of one code, the codes that wepresent will be depicted graphically. This concise repre-sentation is possible because of the CWS formalism thatwe described in II B. All the codes that we present con-verge, upon repeated iteration, to a pure state on thesurface of the Bloch sphere. We give this point in co-ordinates (x, y, z) where x = Tr(|M〉〈M |X) etc. Codesthat we describe below as being tight obey the followingproperty: all non-stabilizer states in the same quadrantas |M〉 are distillable. For instance, if |M〉 has Blochvector (x, 0, z) with x, z > 0 then all states ρ satisfyingTr(Xρ) + Tr(Zρ) > 1 converge to |M〉 under repeatediteration. (We confirmed this numerically by taking arandom sample of 1000 points from the relevant region).

A. Codes achieving tight distillation

A 3-qubit code with generators

G1 Z I ZG2 X Z XZL X X YXL I X Z

(6)

(a) This code converges to(x, y, z) = (sin θ, 0, cos θ) where

θ = arctan√

(√

5− 1)/2

(b) This code converges to(x, y, z) = (0.66796, 0., 0.7442)

(c) This code converges to(x, y, z) = (0.81281, 0, 0.58252)

(d) This code converges to(x, y, z) = (0.64969, 0, 0.7602)

(e) This code converges to (x, y, z) = (1, 0, 1)/√

2

FIG. 3. Codeword stabilized quantum codes with tight distil-lation thresholds. The classical codeword w associated witheach graph is the binary vector with a “1” in lighter pinkpositions and “0” in darker blue positions.

corresponding to the logical basis states

|0L〉 =1

2

10−i0010i

, |1L〉 =1

2

−i0100−i0−1

(7)

distills an equatorial state in the y-z plane with Blochcoordinates (0,−.83929,−.54369) up to a tight errorthreshold of poct = .276921.

Other codes achieving tight distillation are given inpictorial form in Fig. 3 and Fig. 4. It is a straightforwardexercise to recover the generators and logical operatorsif necessary. The 5-qubit code in Fig. 3(e) distills to the

4

(a) This code converges to(x, y, z) =

(0.84893, 0., 0.52851)

(b) This code converges to(x, y, z) = (0.63544, 0., 0.77215)

(c) This code converges to(x, y, z) = (0.81281, 0., 0.58252)

(d) This code converges to(x, y, z) = (0.84534, 0., 0.53423)

(e) This code converges to(x, y, z) = (0.58252, 0., 0.81281)

FIG. 4. Codeword stabilized quantum codes with tight distil-lation thresholds. The classical codeword w associated witheach graph is the binary vector with a “1” in lighter pinkpositions and “0” in darker blue positions.

|H〉-type magic state but is more efficient than the (alsotight) 7-qubit Steane code. The efficiency of these codesis compared in Fig. 5.

Y (p, 10−4)

10−23

10−37

10−51

10−65

10−79

0.05 0.10 0.15 0.20 0.25

p

FIG. 5. Curves showing the yield (efficiency) of varioustight distillation routines as a function of input noise rate p.From bottom to top we have (i) the 7-qubit Steane code asapplied by Reichardt [14] (ii) the 5-qubit code in Fig. 3(e),(iii) 3-qubit code from Fig. 3(a) and (iv) the 3-qubit code inEq. (6).

B. Codes not achieving tight distillation

Here we highlight a small number of codes that, eventhough they do not achieve tight distillation, we still findto be noteworthy.

The |T 〉-type states (depicted in Fig. 1) were shownto be distillable by Bravyi and Kitaev [3] using the per-fect [[5, 1, 3]] code. We are not aware of any additionalroutines for |T 〉 states that have subsequently been devel-oped. This is in marked contrast to |H〉-type distillationfor which a number of codes have been found. For thisreason we note the existence of a 4-qubit code in Fig. 6(b)that also converges to |T 〉-type states. A visual compari-son with the [[5, 1, 3]] code in Fig. 6(a) indicates that theyare closely related. Unfortunately, the threshold for thenew code is worse than that of the [[5, 1, 3]] code. Find-ing tight distillation routines for the |T 〉 direction wasalready known to be more complex than the |H〉 casebecause of a no-go theorem in [15].

In Fig. 6(c) we depict a code that distills |π/3〉 =

(|0〉 + eiπ/3|1〉)/√

2. This is particulary interesting asthe associated gate U = diag(1, eiπ/3) is not transversalfor any stabilizer code [17] which prevents it from beingdistillable by the most commonly used distillation tech-

5

niques.

It is of interest to find distillation routines that distill|V 〉-type magic states, which look like (x, y, z) = (3

5 , 0,45 )

in the Bloch sphere picture. Supplementing Cliffordswith such states leads to a set of gates – the V -basis – thatis highly efficient for gate synthesis [7]. We did not findsuch a code but in Fig. 6(d) we depict a code that con-verges to a nearby state (x, y, z) = (0.60965, 0., 0.79267).

(a) This is the well-known 5-qubitcode in CWS form, which Bravyiand Kitaev showed distilled |T 〉

(b) This code also distills to |T 〉 albeit lessefficiently and with a worse threshold than theabove code. Note that a trivial Z correction

must be applied in between rounds

(c) This code converges to a stateClifford-equivalent to

|π/3〉 = (|0〉+ eiπ/3|1〉)/√

2. The limiting state

has Bloch vector (x, y, z) = ( 12, 0,√3

2)

(d) This code converges to(x, y, z) = (0.60965, 0., 0.79267)

FIG. 6. Codeword stabilized quantum codes that do notachieve tight thresholds

IV. CONCLUSIONS

We have presented a collection of qubit magic statedistillation schemes using small stabilizer codes, most ofwhich achieve tight distillation up to the edge of the sta-bilizer octahedron. Two codes that are not tight werestill noted because they converge to |T 〉 and |π/3〉 re-spectively. Our distillation routines converge to equato-rial states, which means that they can be used to imple-ment rotations about a Pauli axis using standard half-teleportation techniques [3]. We have noted that ineffi-ciency of our routines relates to linear error suppressionand leave as open question whether there is some way ofboosting their efficiency by combining them with othercodes or techniques. Another possible avenue is to usethese codes to convert non-stabilizer states of one type toanother, as was done in e.g., [16]. For example, if we haveaccess to almost pure |H〉 states (after using the Reed-Muller code, say) then input these to the |π/3〉 routine,the output is an almost pure state somewhere between|H〉 and |π/3〉. Further analysis of the relative merits ofmethods such as this is left for future work.

6

V. ACKNOWLEDGMENTS

The authors acknowledge financial support from theGovernment of Canada through NSERC through the

CGS-M program and the discovery grant program, aswell as the U.S. Army Research Office through GrantNo. W911NF-14-1-0103, and FQXI. MH was supportedby EPSRC (grant EP/M024261/1).

[1] B. Eastin and E. Knill, “Restrictions on TransversalEncoded Quantum Gate Sets ” Phys. Rev. Lett. 102,110502 (2009).

[2] E. Knill, “Quantum computing with realistically noisydevices” Nature 434, 39 (2005).

[3] S. Bravyi and A. Kitaev, “Universal quantum com-putation with ideal Clifford gates and noisy ancillas”Phys. Rev. A 71, 022316 (2005).

[4] H. Dawkins, M. Howard, “Qutrit Magic State DistillationTight in Some Directions” Phys. Rev. Lett 115, 030501(2015).

[5] H. Anwar, E. T. Campbell, and D. E. Browne, “Qutritmagic state distillation ” New J. Phys. 14, 063006 (2012).

[6] E. T. Campbell and D. E. Browne, “On the Structure ofProtocols for Magic State Distillation” Lecture Notes inComputer Science 5906, 20 (2009).

[7] A. Bocharov, Y. Gurevich and K. M. Svore, “Efficientdecomposition of single-qubit gates into V basis circuits”Phys. Rev. A 88, 012313 (2013).

[8] E. T. Campbell, H. Anwar, and D. E. Browne, “Magic-State Distillation in All Prime Dimensions Using Quan-tum Reed-Muller Codes” Phys. Rev. X 2, 041021 (2012).

[9] Ben W. Reichardt “Quantum universality by state dis-tillation” Quantum Inf. Comput. 9:1030-1052, (2009)

[10] A. M. Meier, B. Eastin, E. Knill, “Magic-state distilla-tion with the four-qubit code” Quantum Information &Communication 13, 195 (2013).

[11] S. Bravyi, and J. Haah, “Magic-state distillation withlow overhead” Phys. Rev. A 86, 052329 (2012).

[12] C. Jones, “Multilevel distillation of magic states for quan-tum computing” Phys. Rev. A 87, 042305 (2013).

[13] A. Cross, G. Smith, J. A. Smolin, B. Zeng, “CodewordStabilized Quantum Codes” Information Theory, IEEETransactions on 55, 433-438 (2009).

[14] B. W. Reichardt, “Quantum Universality from MagicStates Distillation Applied to CSS Codes” Quantum In-formation Processing 4, 251 (2005).

[15] E. T. Campbell and D. E. Browne, “Bound Statesfor Magic State Distillation in Fault-Tolerant QuantumComputation” Phys. Rev. Lett. 104, 030503 (2010).

[16] G. Duclos-Cianci, and K. Svore, “Distillation of non-stabilizer states for universal quantum computation”Phys. Rev. A 88, 042325 (2013).

[17] Jonas T. Anderson, Tomas Jochym-O’Connor, “Clas-sification of transversal gates in qubit stabilizer codes”arXiv:1409.8320 (2014).