Physics. Dept. 2009 The statistics of eigenvalue distributions in large random matrices Øyvind Ryan November 2009 Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Random matrices I Have numerous applications: wireless communications, nance, nuclear physics. I In many applications, large matrices are assumed. I It turns out that asymptotic results on the eigenvalue distributions (i.e. results which hold in the limit when the matrices grow large) can be found more easily than in the nite matrix regime. I These asymptotic results can also be put into a matrix-based probability framework. I The asymptotic results can also be applied successfully as approximations to the nite matrix cases. Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Denition With the empirical eigenvalue distribution of an L × L hermitian random matrix T we mean the (random) function FT (λ) = #{i |λi ≤ λ} , L (1) where λi are the (random) eigenvalues of T. Questions addressed in random matrix theory: 1. When Tn is a random matrix ensemble, does FTn converge to a non-random function as n → ∞, and how can one express this function? 2. How does the statistics of the largest and smallest eigenvalue of Tn behave as n → ∞? 3. Given independent random matrix ensembles Sn , Tn , when do FSn Tn and FSn +Tn converge to a non-random function as n → ∞, and how can this be expressed in terms of FSn and FTn ? Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions I Remarkably, for many ensembles of random matrices, one has that FTn becomes "less random" when dimensions increase, meaning that more or less the same eigenvalue distribution is observed for dierent realizations when n is large. I Also, for many ensembles of random matrices An , Bn , when An and Bn are independent, combinations such as An Bn and An + Bn exhibit eigenvalue distributions which, for large dimensions and for all realizations, can be completely determined from that of An and Bn . I This last statement is very much in contrast to the fact that the eigenvalues of A + B or AB in general can't be determined from that of A and B. The statement is thus something which applies only for certain large random matrices. In random matrix theory, there are two methods which are widely used to infer on the empirical eigenvalue distribution: the Stieltjes transform approach, and the moment approach. Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions The Stieltjes transform approach Denition The Stieltjes transform of FT is the function Z 1 dFT (x ) = tr((T − z I)−1 ) mFT (z ) = x −z I The distribution function FT can be recovered from the Stieltjes transform mFT , using the so-called Stieltjes inversion formula. I Many limit eigenvalue distributions of random matrices can be characterized by the fact that their Stieltjes transforms are solutions to certain equations. Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions The moment approach Denition The k'th moment of FT is the number Z Tk = x n dFT (x ) = tr(Tk ). P 1/2k One can show that, if T2k = ∞, then FT is uniquely determined by the sequence of moments {Tk }∞ k =0 . Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Gaussian matrices I It turns out that when Xn is n × n and Gaussian, the eigenvalue distribution of Dn Xn and Dn + Xn depends only on that of Dn for large n. I Similarly, the eigenvalue distribution of Dn can be determined from that of Dn Xn or Dn + Xn . This is important in applications, since Xn often represents some sort of noise, so that the result can be used for denoising [1]. Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 I Statistics of eigenvalue distributions The way we compute the eigenvalue distribution of Dn Xn and Dn + Xn from that of Dn and Xn can be put into a framework for a new type of probability theory, where I I I I matrices take the role as random variables, the trace takes the role as expectation, eigenvalue distributions take the role as the distributions of the random variables, the Gaussian law has a central role as a central limit in a new type of central limit theorem. This framework is called free probability [2]. I Useful way to think about free probability: Two independent random matrices, where the eigenvectors for one of them point in each direction with equal probability, then the two matrices are "free" (to be dened). Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions The full circle law Let Xn = √1n Yn where Yn is n × n and has i.i.d. complex standard Gaussian entries. When n grows large, the eigenvalue distribution grows towards the central limit in free probability, and is called the full circle law. Here for n = 500. 1 0.8 0.6 0.4 0.2 0 −0.2 −0.4 −0.6 −0.8 −1 −1 −0.5 0 0.5 1 plot(eig( (1/sqrt(1000)) * (randn(500,500) + j*randn(500,500)) ),'kx') Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions The semicircle law 35 30 25 20 15 10 5 0 −3 −2 −1 0 1 2 3 A = (1/sqrt(2000)) * (randn(1000,1000) + j*randn(1000,1000)); A = (sqrt(2)/2)*(A+A'); hist(eig(A),40) Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions The Marchenko Pastur law What happens with the eigenvalues of N1 Xn XH n when Xn is an n × N random matrix with standard complex Gaussian entries? I One can show that · µµ ¶p ¶¸ X X ak 1 H Xn Xn = c l (π̂)−1 + E tr . N N 2k π̂∈NC2p I k Convergence is "almost sure", i.e. we have very accurate eigenvalue prediction when the matrices are large. When Nn → c, the eigenvalue distribution converges to the Marchenko Pastur law with parameter c, denoted µ Nn , with density p (x − a)+ (b − x )+ 1 + µc f (x ) = (1 − ) δ(x ) + , (2) c 2π cx √ √ where (z )+ = max(0, z ), a = (1 − c )2 and b = (1 + c )2 . Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Four dierent Marchenko Pastur laws µ Nn are drawn. 1.6 c=0.5 c=0.2 c=0.1 c=0.05 1.4 1.2 Density 1 0.8 0.6 0.4 0.2 0 0.5 1 Øyvind Ryan 1.5 2 2.5 3 The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions The non-asymptotic case I Free probability is a tool for computing the asymptotic eigenvalue distributions for large Gaussian matrices. I One also has important results in the nite matrix regime when one replaces limit eigenvalue distributions with mean eigenvalue distributions. I For instance, one can nd the mean eigenvalue distribution of D + X/DX when D is deterministic and X is Gaussian. I The author is implementing a machinery for computing such mean eigenvalue distributions [3]. Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Vandermonde matrices Consider N × L Vandermonde matrices V 1 ··· −j ω1 e ··· 1 V= √ .. .. . N . e −j (N −1)ω1 · · · of the form 1 e −j ωL .. . (3) e −j (N −1)ωL I The entries in V lie on the unit circle I The ωi are called phase distributions, and are assumed i.i.d. on [0, 2π) I We assume N , L → ∞ at the same rate with limN →∞ NL = c. The normalizing factor √1 is included to ensure limiting N asymptotic behaviour. Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Various results exist on the distribution of the determinant of Vandermonde matrices, but there are many open problems (below, VH V is used since V is rectangular in general): I How can we nd the singular value distribution of V? I How can we infer on the moments of D from the mixed moments DVH V? I How can we express the eigenvalue distribution of combinations of independent Vandermonde matrices? I Asymptotic results? If X is an N × N standard, complex, Gaussian matrix, then the asymptotic capacity is ¡ ¡1 ¢¢ limN →∞ N1 log XXH = ´ N ³ 2 det I + ρ ¡√ ¢2 ¡√ ¢2 2 log2 1 + ρ − 14 4ρ + 1 − 1 − log4ρ2 e 4ρ + 1 − 1 . We are not aware of similar asymptotic expressions for the determinant/capacity of Vandermonde matrices. Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Main result [4] 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[tr(D1 (N )VH VD2 (N )VH V · · · × Dn (N )VH V)]. (4) N →∞ It turns out that for a large family of Vandermonde matrices, the limits (4) exist, and the limit depend only on the spectra of V and Di (N ). Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Eigenvalue distributions of Vandermonde matrices 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 H 10 0 0 1 2 3 4 5 6 7 8 9 10 H 1 N XX , (a) V V, with V a 1600 × 800 Vander- (b) with X an 800 × 1600 commonde matrix with uniform phase dis- plex, standard, Gaussian matrix. tribution. Figure: Histogram of mean eigenvalue distributions. Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Main results I Vandermonde matrices display almost sure convergence. I Vandermonde matrices do not have compactly supported limiting eigenvalue distributions (contrary to Gaussian matrices). I Characterization of when a "mixed moment" ³ ´ lim tr D1 (N )ViH1 Vi2 · · · Dn (N )ViH2n−1 Vi2n N →∞ (5) is computable from only knowledge of the spectra of V1 , V2 , ... I Characterization of all "convolution operations", such as nding the spectrum of V1H V1 + V2H V2 from that of V1 and V2 when they are independent and large. Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Open questions I How can one establish an analytical machinery for Vandermonde matrices, such as a transform with similar properties as the R-transform in free probability? I Is there a general criterion for saying exactly when the spectrum of A + B/AB can be determined from that of A and B? Gaussian matrices are bi-unitarily invariant, which is a sucient criterion. But it may not be necessary. Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Software implementation Software implementation [3] which computes I Any asymptotic moment of Vandermonde matrices I Any asymptotic moment of Gaussian matrices I Any moment of nite Gaussian matrices I Any combination of independent matrices of the above types The software implementation is in Matlab and can generate formulas for all moments in latex as well as compute the moments numerically. The implementation I Iterates through partitions, I Computes many dierent equivalence relations on partitions, I performs Fourier-Motzkin elimination Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions I This talk is available at http://folk.uio.no/oyvindry/talks.shtml I My publications are listed at http://folk.uio.no/oyvindry/publications.shtml THANK YOU! Øyvind Ryan The statistics of eigenvalue distributions in large random ma Physics. Dept. 2009 Statistics of eigenvalue distributions Ø. Ryan and M. Debbah, Channel capacity estimation using free probability theory, IEEE Trans. Signal Process., vol. 56, no. 11, pp. 56545667, November 2008. F. Hiai and D. Petz, The Semicircle Law, Free Random Variables and Entropy. American Mathematical Society, 2000. Ø. Ryan, Documentation for the Random Matrix Library, 2009, http://i.uio.no/~oyvindry/rmt/doc.pdf. Ø. Ryan and M. Debbah, Asymptotic behaviour of random vandermonde matrices with entries on the unit circle, IEEE Trans. on Information Theory, vol. 55, no. 7, pp. 31153148, 2009. Øyvind Ryan The statistics of eigenvalue distributions in large random ma