localization from incomplete noisy distance...
TRANSCRIPT
-
Localization from Incomplete Noisy DistanceMeasurements
Adel Javanmard and Andrea Montanari
Stanford University
July 27, 2011
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 1 / 41
-
A chemistry question
Which physical conformations are produced by given chemical bonds?
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 2 / 41
-
. . .more questions. . .
Manifold Learning Sensor Net. Localization
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 3 / 41
-
General `geometric inference' problem
Given partial/noisy information about distances.
Reconstruct the points positions.
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 4 / 41
-
Connection with matrix completion
What happens to matrix completion if the probabilty of revealingentry i ; j depends on the value of that entry?
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 5 / 41
-
Connection with matrix completion
What happens to matrix completion if the probabilty of revealingentry i ; j depends on the value of that entry?
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 5 / 41
-
This talk
0.5
0
‐0.5
‐0.5 0 0.5
R.G.G. G(n ; r)
x1; � � � ; xn 2 [�0:5; 0:5]d
r � �(logn=n)1=d
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 6 / 41
-
This talk
0.5
0
‐0.5
‐0.5 0 0.5
R.G.G. G(n ; r)
x1; � � � ; xn 2 [�0:5; 0:5]dr � �(log n=n)1=d
adversarial noise
j~d2ij � d2ij j � �
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 7 / 41
-
Approaches
Triangulation
Multidimensional scaling
Divide and conquer (Singer 2008)
SDP relaxations (Biswas, Ye 2004)
Manifold learning (ISOMAP, LLE, HLLE, . . . )
Little quantitative theory, especially in presence of noise
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 8 / 41
-
Approaches
Triangulation
Multidimensional scaling
Divide and conquer (Singer 2008)
SDP relaxations (Biswas, Ye 2004)
Manifold learning (ISOMAP, LLE, HLLE, . . . )
Little quantitative theory, especially in presence of noise
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 8 / 41
-
Approaches
Triangulation
Multidimensional scaling
Divide and conquer (Singer 2008)
SDP relaxations (Biswas, Ye 2004)
Manifold learning (ISOMAP, LLE, HLLE, . . . )
Little quantitative theory, especially in presence of noise
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 8 / 41
-
Approaches
Triangulation
Multidimensional scaling
Divide and conquer (Singer 2008)
SDP relaxations (Biswas, Ye 2004)
Manifold learning (ISOMAP, LLE, HLLE, . . . )
Little quantitative theory, especially in presence of noise
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 8 / 41
-
Approaches
Triangulation
Multidimensional scaling
Divide and conquer (Singer 2008)
SDP relaxations (Biswas, Ye 2004)
Manifold learning (ISOMAP, LLE, HLLE, . . . )
Little quantitative theory, especially in presence of noise
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 8 / 41
-
Approaches
Triangulation
Multidimensional scaling
Divide and conquer (Singer 2008)
SDP relaxations (Biswas, Ye 2004)
Manifold learning (ISOMAP, LLE, HLLE, . . . )
Little quantitative theory, especially in presence of noise
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 8 / 41
-
Approaches
Triangulation
Multidimensional scaling
Divide and conquer (Singer 2008)
SDP relaxations (Biswas, Ye 2004)
Manifold learning (ISOMAP, LLE, HLLE, . . . )
Little quantitative theory, especially in presence of noise
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 8 / 41
-
Outline
1 SDP relaxation and robust reconstruction
2 Lower bound
3 Rigidity theory and upper bound
4 Discussion
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 9 / 41
-
SDP relaxation and robust reconstruction
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 10 / 41
-
Preliminary notes
Positions can be reconstructed up to rigid motions
NP-hard [Saxe 1979]
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 11 / 41
-
Optimization formulation
minimizenXi=1
kxik22; xi 2 Rd
subject to���kxi � xj k22 � ed2ij ��� � �
Nonconvex
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 12 / 41
-
Optimization formulation
minimizenXi=1
kxik22; xi 2 Rd
subject to���kxi � xj k22 � ed2ij ��� � �
Nonconvex
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 12 / 41
-
Optimization formulation
minimizenXi=1
Qii
subject to���Qii � 2Qij +Qjj � ed2ij ��� � �Qij = hxi ; xj i
Nonconvex
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 13 / 41
-
Optimization formulation (better notation)
minimize Tr(Q)
subj:to���hMij ;Qi � ed2ij ��� � �Qij = hxi ; xj i
Mij = eij eTij ;
eij = (0; : : : ; 0; +1|{z}i
; 0; : : : ; 0; �1|{z}j
; 0; : : : ; 0)
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 14 / 41
-
Semide�nite programing relaxation
minimize Tr(Q)
subj:to���hMij ;Qi � ed2ij ��� � �((((
(((Qij = hxi ; xj i Q � 0
Mij = eij eTij ;
eij = (0; : : : ; 0; +1|{z}i
; 0; : : : ; 0; �1|{z}j
; 0; : : : ; 0)
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 15 / 41
-
Semide�nite programing relaxation
SDP-based Localization
Input : Distance measurements edij , (i ; j ) 2 GOutput : Low-dimensional coordinates x1; : : : ; xn 2 Rd
1: Solve the following SDP problem:
minimize Tr(Q),
s.t.��hMij ;Qi � ed2ij �� � �, (i ; j ) 2 G ,Q � 0.
2: Eigendecomposition Q = U�UT ;
3: Top d e-vectors X = Ud�1=2d ;
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 16 / 41
-
Robustness?
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 17 / 41
-
Robustness?
Theorem (Javanmard, Montanari '11)
Assume r � 10pd(logn=n)1=d . Then, w.h.p.,
d(X ; X̂ ) � C1(nrd)5 �r4
;
Further, there exists a set of `adversarial' measurements such that
d(X ; X̂ ) � C2 �r4
:
d(X ; X̂ ) � 1n
nXi=1
kxi � bxikJavanmard, Montanari (Stanford) Localization Problem July 27, 2011 18 / 41
-
Lower bound
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 19 / 41
-
Proof: Lower bound
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 20 / 41
-
Proof: Lower bound
T : [�0:5; 0:5]d ! Rd+1
T (t1; t2; � � � ; td) = (R sin t1R
; t2; � � � ; td ;R(1� cos t1R)); R =
r2
p�
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 21 / 41
-
Rigidity theory and upper bound
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 22 / 41
-
Uniqueness , Global rigidity
Global rigidity
Assume noiseless measurements.Is the reconstruction unique? (up to rigid motions)
Depends both on G and on (x1; : : : ; xn)Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 23 / 41
-
Generic lobal rigidity: Characterization
Theorem (Connelly 1995; Gortler, Healy, Thurston, 2007)
(G ; fxig) is generically globally rigid in Rd,(G ; fxig) admits a stress matrix , with rank() = n � d � 1.
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 24 / 41
-
Stress matrix: : : imagine putting springs on the edges : : :
!ij
Equilibrium x1; : : : ; xn :
[force on i ] =Xj2@i
!ij (xj � xi ) = 0
ij = !ij , ii = �P
j2@i !ijJavanmard, Montanari (Stanford) Localization Problem July 27, 2011 25 / 41
-
Stress matrix: : : imagine putting springs on the edges : : :
!ij
Equilibrium x1; : : : ; xn :
[force on i ] =Xj2@i
!ij (xj � xi ) = 0
ij = !ij , ii = �P
j2@i !ijJavanmard, Montanari (Stanford) Localization Problem July 27, 2011 25 / 41
-
In�nitesimal rigidity
Consider a continuos motion preserving distances instantaneously
(xi � xj )T ( _xi � _xj ) = 0; 8(i ; j ) 2 E
Trivial motions
_xi = Axi + b; A = �AT 2 Rd�d
�
rotation translationJ
J]
De�nition
(G ; fxig) is in�nitesimally rigid if rotations and translations are theonly in�nitesimal motions.
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 26 / 41
-
In�nitesimal rigidity
Consider a continuos motion preserving distances instantaneously
(xi � xj )T ( _xi � _xj ) = 0; 8(i ; j ) 2 E
Trivial motions
_xi = Axi + b; A = �AT 2 Rd�d
�
rotation translationJ
J]
De�nition
(G ; fxig) is in�nitesimally rigid if rotations and translations are theonly in�nitesimal motions.
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 26 / 41
-
Rigidity matrix
(xi � xj )T ( _xi � _xj ) = 0; 8(i ; j ) 2 E
Rigidity matrix: RG;X �
264 _x1..._xn
375 = 0
dim(null(RG;X )) � d(d � 1)2| {z }A
+ d|{z}b
=
d + 1
2
!
(G ; fxig) is in�nitesimally rigid if rank(RG;X ) = nd ��d+1
2
�.
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 27 / 41
-
Rigidity matrix
(xi � xj )T ( _xi � _xj ) = 0; 8(i ; j ) 2 E
Rigidity matrix: RG;X �
264 _x1..._xn
375 = 0
dim(null(RG;X )) � d(d � 1)2| {z }A
+ d|{z}b
=
d + 1
2
!
(G ; fxig) is in�nitesimally rigid if rank(RG;X ) = nd ��d+1
2
�.
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 27 / 41
-
Rigidity matrix
(xi � xj )T ( _xi � _xj ) = 0; 8(i ; j ) 2 E
Rigidity matrix: RG;X �
264 _x1..._xn
375 = 0
dim(null(RG;X )) � d(d � 1)2| {z }A
+ d|{z}b
=
d + 1
2
!
(G ; fxig) is in�nitesimally rigid if rank(RG;X ) = nd ��d+1
2
�.
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 27 / 41
-
(Idea of the) proof of the upper bound
Noise is analogous to stretching/compressing the edges
Global/in�nitesimal rigidity: Rank of Stress/Rigidity matrix.
Needed: Quantitative rigidity theory(Rank ) bounds on singular values)
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 28 / 41
-
A mechanical analogy
Graph I Graph II
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 29 / 41
RGG1_aristotelian.aviMedia File (video/avi)
RGG2_aristotelian.aviMedia File (video/avi)
-
A mechanical analogy
U (X ) =X
(i ;j )2E
1
2(kxi � xj k2 � d2ij )2 �
Xi
f Ti xi
_X = �rXU
xi + �xi equilibrium positions in the presence of force
(
Id +RG;XRTG;X ) � �x � f
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 30 / 41
-
A mechanical analogy
U (X ) =X
(i ;j )2E
1
2(kxi � xj k2 � d2ij )2 �
Xi
f Ti xi
_X = �rXU
xi + �xi equilibrium positions in the presence of force
(
Id +RG;XRTG;X ) � �x � f
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 30 / 41
-
Steps of the proof
Solution of SDP ! QGram matrix ! Q0 (Q0 = XXT , X = [x1jx2j � � � jxn ]T )
Q = Q0 + S
= Q0 +XYT +YXT + S?
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 31 / 41
-
Steps of the proof
Solution of SDP ! QGram matrix ! Q0 (Q0 = XXT , X = [x1jx2j � � � jxn ]T )
Q = Q0 + S
= Q0 +XYT +YXT + S?
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 31 / 41
-
Steps of the proof
Solution of SDP ! QGram matrix ! Q0 (Q0 = XXT , X = [x1jx2j � � � jxn ]T )
Q = Q0 + S
= Q0 + XYT +YXT| {z }
Controlled by rigidity matrix
+ S?|{z}by stress matrix
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 32 / 41
-
Steps of the proof
Lemma
For a stress matrix � 0,
kS?k� � �max()�min(jhu ;x i?)
jE j�:
Lemma
�min(jhu ;x i?) � C1(nrd)�2r4; �max() � C2(nrd)2:
Lemma
kXY T +YXTk1 � C (nrd)5n2
r4:
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 33 / 41
-
Of independent interest
Lemma
jhu ;x i? � C1(nrd)�3r2 Ljhu ;x i?
(Manifold learning folklore: � L2)
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 34 / 41
-
Discussion
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 35 / 41
-
Discussion : Manifold learning
Data set Proximity graph
~dij = kxi � xj kRN ; dij = dM(xi ; xj )
� / r4
r20(r0 � curvature radius)
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 36 / 41
-
Discussion : Manifold learning
Data set Proximity graph
~dij = kxi � xj kRN ; dij = dM(xi ; xj )
� / r4
r20(r0 � curvature radius)
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 36 / 41
-
Discussion : Manifold learning
Data set Proximity graph
~dij = kxi � xj kRN ; dij = dM(xi ; xj )
� / r4
r20(r0 � curvature radius)
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 36 / 41
-
Manifold learning: Reconstruction error
d(X ; X̂ ) � C (nrd)5
r20
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 37 / 41
-
Maximum Variance Unfolding
[cf. Weinberger and Saul, 2006]
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 38 / 41
-
Optimization formulation
maximizenX
1�i ;j�n
kxi � xj k2
subj:to���kxi � xj k2 � d2ij ��� � � 8(i ; j ) 2 EnXi=1
xi = 0
Nonconvex
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 39 / 41
-
Semide�nite programing relaxation
maximize Tr(Q)
subj:to���hMij ;Qi � d2ij ��� � �uTQu = 0
Q � 0
Mij = eij eTij ;
eij = (0; : : : ; 0; +1|{z}i
; 0; : : : ; 0; �1|{z}j
; 0; : : : ; 0)
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 40 / 41
-
Conclusion
Numerous examples of geometric inference.
Many open problems.
Thanks!
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 41 / 41
-
Conclusion
Numerous examples of geometric inference.
Many open problems.
Thanks!
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 41 / 41
-
Conclusion
Numerous examples of geometric inference.
Many open problems.
Thanks!
Javanmard, Montanari (Stanford) Localization Problem July 27, 2011 41 / 41
SDP relaxation and robust reconstructionLower boundRigidity theory and upper boundDiscussion