Boulder 2008 Applications and fundamental results on random Vandermonde matrices Øyvind Ryan July 2008 Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Free probability I I I I I Sheds new light on results in random matrix theory: If the eigenvalue distributions of independent n × n-random matrices An , Bn exhibit some form of convergence, and one of them exhibits unitary invariance, then the eigenvalue distribution of An + Bn can be computed in the large n-limit. The limit eigenvalue distribution of An + Bn is the additive free convolution of the limiting eigenvalue distributions of An and Bn . Free convolution can be expressed nicely in terms of the limit moments of the matrices. Free convolution can also be expressed nicely in terms of equations involving the Stieltjes transform of the involved measures. The Poisson distribution has its analogue in the free Poisson distribution (the Marchenko Pastur law) in free probability theory. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Does other types of random matrices (i.e. non-unitarily invariant) t into a similar framework? We have investigated this for Vandermonde matrices [1, 2], which are widely used. They have the form 1 ··· 1 x1 · · · xL V= .. .. .. . . . N −1 N −1 x1 · · · xL It is straightforward to show that square Vandermonde matrices have determinant Y (xl − xk ). det(V) = 1≤k <l ≤N In particular, V is nonsingular if the xk are dierent. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Various results exist on the distribution of the determinant of Vandermonde matrices (Gaussian entries (Metha), entries with B-distribution (Selberg)), but there are many open problems (below, VH V is used since V is rectangular in general): I How can³we nd the ¡ H ¢k ´ moments of the Vandermonde matrices (i.e. trL V V ) (not the determinant itself)? I I I Deconvolution problem: How to estimate the moments of D from mixed moments DVH V? Mixed moments of independent Vandermonde matrices? Asymptotic results? If X is an N × N standard, complex, Gaussian matrix, then ¡ ¡1 ¢¢ H limN →∞ N1 log det I + ρ XX = ´ 2 N ³ ¢2 ¢2 ¡ ¡√ 1 √ 2 log2 1 + ρ − 4 4ρ + 1 − 1 − log4ρ2 e 4ρ + 1 − 1 . (which is the expression for the capacity). We are not aware of similar asymptotic expressions for the determinant/capacity of Vandermonde matrices. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Random Vandermonde matrices We will consider Vandermonde matrices V of dimension N × L of the form 1 ··· 1 −j ω · · · e −j ωL 1 e 1 V= √ (1) .. . . . . . . N . e −j (N −1)ω1 · · · e −j (N −1)ωL (i.e. we assume that the xi lie on the unit circle). The ωi are called phase distributions. We will limit the study of Vandermonde matrices to cases where I I The phase distributions are i.i.d. The asymptotic case N , L → ∞ with limN →∞ NL = c. The normalizing factor √1 is included to ensure limiting N asymptotic behaviour. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Where can such Vandermonde matrices appear? Consider a multi-path channel of the form: h(τ ) = L X αi g (τ − τi ) i =1 αi are i.d. Gaussian random variables with power Pi , I τi are uniformly distributed delays over [0, T ], I g is the low pass transmit lter. I L is the number of paths In the frequency domain, the channel is given by: I c (f ) = L X αi G (f )e −j2πf τi i =1 We suppose the transmit lter to be ideal (G (f ) = 1). Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Sampling the continuous frequency signal at fi = i W N (N is the number of frequency samples) where W is the bandwidth, our model becomes α1 n1 1 . . r = VP 2 (2) .. + .. , αL nN where V is a random Vandermonde matrix of the type (1), and I P is the L × L diagonal power matrix, I ni is independent, additive, white, zero mean Gaussian noise of variance √σ . N Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Main result Denition Dene Kρ,ω,N = RN 1 n+1−|ρ| × (0,2π)|ρ| Qn jN (ω −ω ) b(k −1) b(k ) 1−e k =1 1−e j (ωb(k −1) −ωb(k ) ) , (3) d ω1 · · · d ω|ρ| , where ωρ1 , ..., ωρ|ρ| are i.i.d. (indexed by the blocks of ρ), all with the same distribution as ω , and where b(k ) is the block of ρ which contains k (where notation is cyclic, i.e. b(−1) = b(n)). If the limit Kρ,ω = lim Kρ,ω,N N →∞ exists, then Kρ,ω is called a Vandermonde mixed moment expansion coecient. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Main result 2 Assume that I {Dr (N )}1≤r ≤n are diagonal L × L matrices which have a joint limit distribution as N → ∞, I L N → c. We would like to express the limits Mn = lim E [trL (D1 (N )VH VD2 (N )VH V · · · × Dn (N )VH V)]. (4) N →∞ It turns out that this is feasible when all Vandermonde mixed moment expansion coecients Kρ,ω exist. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde For convenience, dene £ ¡¡ ¢n ¢¤ mn = (cM )n = c limN →∞ E trL D(N )VH V , dn = (cD )n = c limN →∞ trL (Dn (N )) , (5) Theorem Assume D1 (N ) = D2 (N ) = · · · = Dn (N ). When ω = u, m1 = d1 m2 = d2 + d12 m3 = d3 + 3d2 d1 + d13 m4 = d4 + 4d3 d1 + 8/3d22 + 6d2 d12 + d14 m5 = d5 + 5d4 d1 + 25/3d3 d2 + 10d3 d12 + 40/3d22 d1 + 10d2 d13 + d15 m6 = d6 + 6d5 d1 + 12d4 d2 + 15d4 d12 + 151/20d32 + 50d3 d2 d1 +20d3 d13 + 11d23 + 40d22 d12 + 15d2 d14 + d16 m7 = d7 + 7d6 d1 + 49/3d5 d2 + 21d5 d12 + 497/20d4 d3 + 84d4 d2 d1 +35d4 d13 + 1057/20d32 d1 + 693/10d3 d22 + 175d3 d2 d12 +35d3 d14 + 77d23 d1 + 280/3d22 d13 + 21d2 d15 + d17 . Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Comparison The Gaussian equivalent of this is m1 m2 m3 m4 m5 m6 d1 d2 + d12 d3 + 3d2 d1 + d13 d4 + 4d3 d1 + 3d22 + 6d2 d12 + d14 d5 + 5d4 d1 + 5d3 d2 + 10d3 d12 + 10d22 d1 + 10d2 d13 + d15 d6 + 6d5 d1 + 6d4 d2 + 15d4 d12 + 3d32 + 30d3 d2 d1 +20d3 d13 + 5d23 + 10d22 d12 + 15d2 d14 + d16 m7 = d7 + 7d6 d1 + 7d5 d2 + 21d5 d12 + 7d4 d3 + 42d4 d2 d1 +35d4 d13 + 21d32 d1 + 21d3 d22 + 105d3 d2 d12 +35d3 d14 + 35d23 d1 + 70d22 d13 + 21d2 d15 + d17 , (6) 1 H H when we replace V V with N XX , with X an L × N complex, standard, Gaussian matrix. = = = = = = Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Sketch of proof We can write h ³ ´i E trL D1 (N )VH VD2 (N )VH V · · · Dn (N )VH V as L−1 P i1 ,...,in j1 ,...,jn (7) E ( D1 (N )(j1 , j1 )VH (j1 , i2 )V(i2 , j2 ) D2 (N )(j2 , j2 )VH (j2 , i3 )V(i3 , j3 ) .. . (8) Dn (N )(jn , jn )VH (jn , i1 )V(i1 , j1 )) The (j1 , ..., jn ) give rise to a partition ρ of {1, ..., n}, where each block ρj consists of equal values, i.e. ρj = {k |jk = j }. This ρ will actually represent the ρ used in the denition of Kρ,ω,n . The rest of the proof goes by carefully computing this limit quantity using much combinatorics. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Comparisons Denote by µµ ¶ ¶∗n λ λ ν(λ, α) = lim 1− δ − 0 + δα n→∞ n n the (classical) Poisson distribution of rate λ and jump size α. Denote also by ¶ ¶¢n µµ λ λ δ − 0 + δα µ(λ, α) = lim 1− n→∞ n n the free Poisson distribution of rate λ and jump size α (the Marchenko Pastur law). Denote also µc = µ( c1 , c ), νc = ν(c , 1). Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Comparisons 2 Corollary Assume that V has uniformly distributed phases. Then the limit moment h ³³ ´n ´i Mn = lim E trL VH V N →∞ satsies the inequality φ(a1n ) ≤ Mn ≤ 1 E (a2n ), c where a1 ∼ µc , a2 ∼ νc . In particular, equality occurs for m = 1, 2, 3 and c = 1. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Comparisons 3 1 1 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 1 2 3 4 5 6 7 8 9 10 0 0 1 2 3 4 5 6 7 8 9 10 H 1 N XX , H with X an 800 × 1600 com(a) V V, with V a 1600 × 800 Van- (b) dermonde matrix with uniformly dis- plex, standard, Gaussian matrix. tributed phases. Figure: Histogram of mean eigenvalue distributions. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Other results I Mixed moments of (more than one) independent Vandermonde matrices. I Generalized ¡ ¢Vandermonde matrices: These have the form V = e j αk βl 1≤k ≤N ,1≤l ≤L . It is known that V is nonsingular i all αk are dierent, and all βl are dierent. [1, 2] also contain results on the asymptotics of generalized Vandermonde matrices. I Exact moments of lower order Vandermonde matrices. Reveals slower convergence (O ( n1 , contrary to O ( n12 ) for Gaussian counterpart). I Computation of the asymptotic moments when the phase distribution is not uniform. Phase distributions with continous density, and phase distributions with singularities. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Application 1: Estimation of the number of paths Return to the multi-path channel model (2). For simplicity, set W = T = 1, so that the phase distribution of the Vandermonde matrix is uniform. We take K observations of (2) and form the observation matrix Y = [r1 · · · rK ] (1) (K ) α1 1 . = VP 2 .. ··· .. . α1 .. . αL ··· αL (1) (K ) (1) (K ) n1 .. + . ··· .. . n1 .. . nN ··· nN (1) , (K ) (9) It is now possible to combine the deconvolution result for Vandermonde matrices with known deconvolution results for Gaussian matrices to estimate L from a number of observations (assuming P is known). All values of L are tried, and the one which "best matches" the observed values is chosen: Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Estimation of the number of paths 2 Proposition Assume that V has uniformly distributed phases, and let mPi be the ^i ) the moments of the sample moments of P, and mR̂i = trN (R covariance matrix ^ = 1 YYH . R K N , c2 = NL , and c3 = KL . Then Dene also c1 = K £ ¤ E mR̂ = c2 mP1 + σ 2 µ ¶ h i 1 2 E mR̂ = c2 1 − mP2 + c2 (c2 + c3 )(mP1 )2 N +2σ 2 (c2 + c3 )mP1 + σ 4 (1 + c1 ) h i E mR̂3 = ··· Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Estimation of the number of paths 3 70 70 Estimate of L Actual value of L 50 50 40 40 Estimate of L Actual value of L L 60 L 60 30 30 20 20 10 10 0 0 10 20 30 40 50 60 Number of observations 70 80 90 (a) K = 1 100 0 0 10 20 30 40 50 60 Number of observations 70 80 90 100 (b) K = 10 Figure: √ Estimate for the number of paths. Actual value of L is 36. Also, σ = 0.1, N = 100. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Application 2: Wireless capacity Analysis For a general matrix W, the mean capacity is dened as ¡ ¡ ¢¢ CN = N1 E log2 det IN + σ12 WWH ¡ ¡ ¡ ¢¢¢ R ¡ P 1 H = N1 N = log2 1 + k =1 E log2 1 + σ 2 λk WW ¢ t µ(dt ) (10) H where µ is the mean empirical eigenvalue distribution of WW . P k +1 t k , Substituting the Taylor series log2 (1 + t ) = ln12 ∞ k =1 (−1) k we obtain P (−1)k +1 mk (µ)ρk (11) , CN = ln12 ∞ k =1 k where ρ is SNR, and where mk (µ) = 1 σ2 Z t k d µ(t ) for k ∈ Z+ However, if W is a Vandermonde matrix, many more moments are required for precise estimation of capacity than we can provide with the formulas for the rst 7 moments. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Wireless capacity Analysis 2 3 Asymptotic capacity sample capacity 2.5 Capacity 2 1.5 1 0.5 0 0 1 2 3 4 5 ρ 6 7 8 9 10 ¡ ¢ Figure: Several realizations of the capacity N1 log2 det I + ρ N1 XXH when X is standard, complex, Gaussian. Matrices of size 36 × 36 were used. The known expression for the asymptotic capacity is also shown. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde 3 3 2.5 2.5 2 2 Capacity Capacity Wireless capacity Analysis 3 1.5 1.5 1 1 0.5 0.5 0 0 1 2 3 4 5 ρ 6 7 8 (a) ¡Realizations H¢ 1 when N log2 det I + ρVV has uniform phase distribution. 9 10 0 0 1 2 3 4 5 ρ 6 7 8 9 10 of (b) of ¡Realizations ¢ ω N1 log2 det I + ρVVH when ω has a certain non-uniform phase distribution. Figure: Several realizations of the capacity for Vandermonde matrices for two dierent phase distributions. Matrices of size 36 × 36 were used. Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Open questions I I I I I Do Vandermonde matrices display almost sure convergence? How can one establish an analytical machinery for Vandermonde matrices, such as a transform with similar properties as the R-transform in free probability? Have Vandermonde matrices compactly supported limiting eigenvalue distributions? Central limit theorem Initial calculations suggest that the limit moments of a sum VωH1 ,c Vω1 ,c + VωH2 ,c Vω2 ,c (where Vω1 ,c , Vω2 ,c are independent, ω1 , ω2 are phase distributions, c is the limit aspect ratio) equals the limit moments of VωH3 ,d Vω3 ,d (for certain ω3 , d). What are the properties of such "additive Vandermonde convolution". Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde I This talk is available at http://heim.i.uio.no/∼oyvindry/talks.shtml. I My publications are listed at http://heim.i.uio.no/∼oyvindry/publications.shtml THANK YOU! Øyvind Ryan Applications and fundamental results on random Vandermon Boulder 2008 Applications and fundamental results on random Vandermonde Ø. Ryan and M. Debbah, Random Vandermonde matrices-part I: Fundamental results, Submitted to IEEE Trans. on Information Theory, 2008. , Random Vandermonde matrices-part II: Applications, Submitted to IEEE Trans. on Information Theory, 2008. Øyvind Ryan Applications and fundamental results on random Vandermon