Sparse Approximations Nick Harvey University of British Columbia TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: AAA
Approximating Dense Objects by Sparse Objects Floor joists Wood Joists Engineered Joists
Approximating Dense Objects by Sparse Objects Bridges Masonry Arch Truss Arch
Approximating Dense Objects by Sparse Objects Bones Human Femur Robin Bone
Mathematically Can an object with many pieces be approximately represented by fewer pieces? Independent random sampling usually does well Theme of this talk: When can we beat random sampling? 6 -1 4 5 7 6 -1 5 -3 2 8 1 Dense Matrix Sparse Matrix Dense Graph Sparse Graph
Talk Outline Vignette #1: Discrepancy theory Vignette #2: Singular values and eigenvalues Vignette #3: Graphs Theorem on “Spectrally Thin Trees”
Discrepancy Given vectors v1,…,vn2Rd with kvikp bounded. Want y2{-1,1}n with ki yivikq small. Eg1: If kvik1·1 then E ki yi vik1 · Eg2: If kvik1·1 then 9y s.t. ki yi vik1 · Spencer ‘85: Partial Coloring + Entropy Method Gluskin ‘89: Sidak’s Lemma Giannopoulos ‘97: Partial Coloring + Sidak Bansal ‘10: Brownian Motion + Semidefinite Program Bansal-Spencer ‘11: Brownian Motion + Potential function Lovett-Meka ‘12: Brownian Motion Non-algorithmic Algorithmic
Discrepancy Given vectors v1,…,vn2Rd with kvikp bounded. Want y2{-1,1}n with ki yivikq small. Eg1: If kvik1·1 then E ki yi vik1 · Eg2: If kvik1·1 then 9y s.t. ki yi vik1 · Eg3: If kvik1·¯, kvik1·±, and ki vik1·1, then 9y with ki yi vik1 · Harvey ’13: Using Lovasz Local Lemma. Question: Can log(±/¯2) factor be improved?
Talk Outline Vignette #1: Discrepancy theory Vignette #2: Singular values and eigenvalues Vignette #3: Graphs Theorem on “Spectrally Thin Trees”
Partitioning sums of rank-1 matrices Let v1,…,vn2Rd satisfy i viviT=I and kvik2·±. Want y2{-1,1}n with ki yiviviTk2 small. Random sampling: E ki yiviviTk2 · . Rudelson ’96: Proofs using majorizing measures, then nc-Khintchine Marcus-Spielman-Srivastava ’13: 9y2{-1,1}n with ki yiviviTk2 · . 2
Partitioning sums of matrices Given dxd symmetric matrices M1,…,Mn2Rd with i Mi=I and kMik2·±. Want y2{-1,1}n with ki yiMik2 small. Random sampling: E ki yiMik2 · Also follows from nc-Khintchine. Ahlswede-Winter ’02: Using matrix moment generating function. Tropp ‘12: Using matrix cumulant generating function.
Partitioning sums of matrices Given dxd symmetric matrices M1,…,Mn2Rd with i Mi=I and kMik2·±. Want y2{-1,1}n with ki yiMik2 small. Random sampling: E ki yiMik2 · Question: 9y2{-1,1}n with ki yiMik2 · ? Conjecture: Suppose i Mi=I and kMikSch-1·±. 9y2{-1,1}n with ki yiMik2 · ? MSS ’13: Rank-one case is true Harvey ’13: Diagonal case is true (ignoring log(¢) factor) False!
Partitioning sums of matrices Given dxd symmetric matrices M1,…,Mn2Rd with i Mi=I and kMik2·±. Want y2{-1,1}n with ki yiMik2 small. Random sampling: E ki yiMik2 · Question: Suppose only that kMik2·1. 9y2{-1,1}n with ki yiMik2 · ? Spencer/Gluskin: Diagonal case is true
Column-subset selection Given vectors v1,…,vn2Rd with kvik2=1. Let st.rank=n/ki viviTk2. Let . 9y2{0,1}n s.t. i yi=k and (1-²)2 · ¸k( i yiviviT ). Spielman-Srivastava ’09: Potential function argument Youssef ’12: Let . 9y2{0,1}n s.t. i yi=k, (1-²)2 · ¸k( i yiviviT ) and ¸1( i yiviviT ) · (1+²)2.
Column-subset selection up to the stable rank Given vectors v1,…,vn2Rd with kvik2=1. Let st.rank=n/ki viviTk2. Let . For y2{0,1}n s.t. i yi=k, can we control ¸k( i yiviviT ) and ¸1( i yiviviT ) ? ¸k can be very small, say O(1/d). Rudelson’s theorem: can get ¸1 · O(log d) and ¸k>0. Harvey-Olver ’13: ¸1 · O(log d / log log d) and ¸k>0. MSS ‘13: If i viviT =I, can get ¸1 · O(1) and ¸k>0.
Talk Outline Vignette #1: Discrepancy theory Vignette #2: Singular values and eigenvalues Vignette #3: Graphs Theorem on “Spectrally Thin Trees”
weighted degree of node c Graph Laplacian 5 10 Graph with weights u: a c d 2 1 b a b c d Laplacian Matrix: a 7 -2 -5 3 -1 16 -10 10 negative of u(ac) b Lu = D-A = c weighted degree of node c d Effective Resistance from s to t: voltage difference when each edge e is a (1/ue)-ohm resistor and a 1-amp current source placed between s and t = (es-et)T Luy (es-et) Effective Conductance: cst = 1 / (effective resistance from s to t)
Spectral approximation of graphs Edge weights u Edge weights w 5 -1 4 6 7 6 -1 -5 5 -3 2 8 -8 1 10 Lu = Lw = ®-spectral sparsifier: Lu ¹ Lw ¹ ®¢Lu
Ramanujan Graphs Suppose Lu is complete graph on n vertices (ue=1 8e). Lubotzky-Phillips-Sarnak ’86: For infinitely many d and n, 9w2{0,1}E such that e we=dn/2 (actually Lw is d-regular) and MSS ‘13: Holds for all d¸3, and all n=c¢2k. Friedman ‘04: If Lw is a random d-regular graph, then 8²>0 with high probability.
Arbitrary graphs Spielman-Srivastava ’08: For any graph Lu with n=|V|, 9w2RE such that |support(w)| = O(n log(n)/²2) and Proof: Follows from Rudelson’s theorem MSS ’13: For any graph Lu with n=|V|, 9w2RE such that we 2 £(²2) ¢ N ¢ (effective conductance of e) |support(w)| = O(n/²2) and
Spectrally-thin trees Question: Let G be an unweighted graph with n vertices. Let C = mine (effective conductance of edge e). Want a subtree T of G with . Equivalent to Goddyn’s Conjecture ‘85: There is a subtree T with Relates to conjectures of Tutte (‘54) on nowhere-zero flows, and to approximations of the traveling salesman problem.
Spectrally-thin trees Question: Let G be an unweighted graph with n vertices. Let C = mine (effective conductance of edge e). Want a subtree T of G with . Rudelson’s theorem: Easily gives ® = O(log n). Harvey-Olver ‘13: ® = O(log n / log log n). Moreover, there is an efficient algorithm to find such a tree. MSS ’13: ® = O(1), but not algorithmic.
Talk Outline Vignette #1: Discrepancy theory Vignette #2: Singular values and eigenvalues Vignette #3: Graphs Theorem on “Spectrally Thin Trees”
Spectrally Thin Trees Proof overview: Given an (unweighted) graph G with eff. conductances ¸ C. Can find an unweighted tree T with Proof overview: Show independent sampling gives spectral thinness, but not a tree. ► Sample every edge e independently with prob. xe=1/ce Show dependent sampling gives a tree, and spectral thinness still works.
Matrix Concentration Theorem: [Tropp ‘12] Let Y1,…,Ym be independent, PSD matrices of size nxn. Let Y=i Yi and Z=E [ Y ]. Suppose Yi ¹ R¢Z a.s. Then
Independent sampling Let Suppose Mi ¹ Z. Define sampling probabilities xe = 1/ce. It is known that e xe = n–1. Claim: Independent sampling gives T µ E with E [|T|]=n–1 and Theorem [Tropp ‘12]: Let M1,…,Mm be nxn PSD matrices. Let D(x) be a product distribution on {0,1}m with marginals x. Let Suppose Mi ¹ Z. Then Define Me = ce¢Le. Then Z = LG and Me ¹ Z holds. Setting ®=6 log n / log log n, we get whp. But T is not a tree! Yi = Xi Mi Y = \sum_i Yi Z = E[Y] = E[\sum_i Yi] = \sum_i E[Xi Mi] = \sum_i xi Mi. Assume that Yi <= R*Z as, which is equivalent to Mi <= R*Z a.s. It turns out that R = 1. Laplacian of the single edge e Properties of conductances used
Spectrally Thin Trees Proof overview: Given an (unweighted) graph G with eff. conductances ¸ C. Can find an unweighted tree T with Proof overview: Show independent sampling gives spectral thinness, but not a tree. ► Sample every edge e independently with prob. xe=1/ce Show dependent sampling gives a tree, and spectral thinness still works. ► Run pipage rounding to get tree T with Pr[ e2T ] = xe = 1/ce
Pipage rounding [Ageev-Svirideno ‘04, Srinivasan ‘01, Calinescu et al. ‘07, Chekuri et al. ‘09] Let P be any matroid polytope. E.g., convex hull of characteristic vectors of spanning trees. Given fractional x Find coordinates a and b s.t. line z x + z ( ea – eb ) stays in current face Find two points where line leaves P Randomly choose one of those points s.t. expectation is x Repeat until x = ÂT is integral x is a martingale: expectation of final ÂT is original fractional x. ÂT1 ÂT6 ÂT2 x ÂT3 ÂT5 ÂT4
Pipage rounding and concavity Say f : Rm ! R is concave under swaps if z ! f( x + z(ea-eb) ) is concave 8x2P, 8a, b2[m]. Let X0 be initial point and ÂT be final point visited by pipage rounding. Claim: If f concave under swaps then E[f(ÂT)] · f(X0). [Jensen] Let E µ {0,1}m be an event. Let g : [0,1]m ! R be a pessimistic estimator for E, i.e., Claim: Suppose g is concave under swaps. Then Pr[ ÂT 2 E ] · g(X0).
Chernoff Bound Chernoff Bound: Fix any w, x 2 [0,1]m and let ¹ = wTx. Define . Then, Claim: gt,µ is concave under swaps. [Elementary calculus] Let X0 be initial point and ÂT be final point visited by pipage rounding. Let ¹ = wTX0. Then Bound achieved by independent sampling also achieved by pipage rounding
Matrix Pessimistic Estimators Theorem [Tropp ‘12]: Let M1,…,Mm be nxn PSD matrices. Let D(x) be a product distribution on {0,1}m with marginals x. Let Suppose Mi ¹ Z. Let Then and . Pessimistic estimator Main Theorem: gt,µ is concave under swaps. Bound achieved by independent sampling also achieved by pipage rounding
Spectrally Thin Trees Proof overview: Given an (unweighted) graph G with eff. conductances ¸ C. Can find an unweighted tree T with Proof overview: Show independent sampling gives spectral thinness, but not a tree. ► Sample every edge e independently with prob. xe=1/ce Show dependent sampling gives a tree, and spectral thinness still works. ► Run pipage rounding to get tree T with Pr[ e2T ] = xe = 1/ce
Matrix Analysis Matrix concentration inequalities are usually proven via sophisticated inequalities in matrix analysis Rudelson: non-commutative Khinchine inequality Ahlswede-Winter: Golden-Thompson inequality if A, B symmetric, then tr(eA+B) · tr(eA eB). Tropp: Lieb’s concavity inequality [1973] if A, B Hermitian and C is PD, then z ! tr exp( A + log(C+zB) ) is concave. Key technical result: new variant of Lieb’s theorem if A Hermitian, B1, B2 are PSD, and C1, C2 are PD, then z ! tr exp( A + log(C1+zB1) + log(C2–zB2) ) is concave.
Questions Can Spencer/Gluskin theorem be extended to matrices? Can MSS’13 be made algorithmic? Can MSS’13 be extended to large-rank matrices? O(1)-spectrally thin trees exist. Can one be found algorithmically? Are O(1)-spectrally thin trees helpful for Goddyn’s conjecture?