CMA 2007
Applications of free probability and random matrix theory Øyvind Ryan
December 2007
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Some important concepts from classical probability Random variables are functions (i.e. they commute w.r.t. multiplication) with a given p.d.f. (denoted f ) Expectation (denoted E ) is integration Independence Additive convolution (∗) and the logarithm of the Fourier transform Multiplicative convolution Central limit law, with special role of the Gaussian law ∗n Poisson distribution Pc : The limit of 1 − nc δ(0) + cn δ(1) as n → ∞. Divisibility: For a given a, find i.i.d. b1 , ..., bn such that fa = fb1 +···+bn . Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Can we find a more general theory, where the random variables are matrices (or more generally, operators), with their eigenvalue distribution (or spectrum) taking the role as the p.d.f.? What are the analogues to the above mentioned concepts for this theory? What are the applications of such a theory?
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Free probability Free probability was developed as a probability theory for random variables which do not commute, like matrices The random variables are elements in a unital ∗-algebra (denoted A), typically B(H), or Mn (C). Expectation (denoted φ) is a normalized linear functional on A. The pair (A, φ) is called a noncommutative probability space. For matrices, φ will be the normalized trace trn , defined by n
trn (a) =
1X aii . n i =1
For random matrices, we set φ(a) = τn (a) = E (trn (a)) is defined by. Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
What is the "central limit" for large matrices? We will attempt to make a connection with classical probability through large random matrices. We would like to define random matrices as "independent" if all entries in one are independent from all entries in the other. Assume that X1 , ..., Xm are n × n i.i.d. complex matrices, and τn (Xi ) = 0, τn (Xi2 ) = 1. What is the limit when m → ∞ in X1 + · · · + Xm √ ? m
If Xi = √1n Yi where Yi has i.i.d. complex standard Gaussian entries, then X1 + · · · + Xm √ ∼ X, m where X = √1n Y and Y has i.i.d. complex standard Gaussian entries. Therefore, matrices with complex standard Gaussian entries are central limit candidates. Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
The full circle law What happens when n is large? The eigenvalues converge to what 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
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
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
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
The Mar˘chenko Pastur law What happens with the eigenvalues of N1 XX H when X is an n × N random matrix with standard complex Gaussian entries? The eigenvalue distribution converges to the Mar˘chenko Pastur law with parameter Nn , denoted µ Nn . Let f µc be the p.d.f. of µc . Then p
(x − a)+ (b − x)+ , (1) 2πcx √ √ where (z)+ = max(0, z), a = (1 − c)+ and a = (1 + c)+ . f
µc
1 (x) = (1 − )+ δ(x) + c
The matrices N1 XX H occur most frequently as sample covariance matrices: N is the number of observations, and n is the number of parameters in the system.
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Four different Mar˘chenko 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
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Derivation of the limiting distribution for
√1 XXH N
When x is standard complex Gaussian, we have that E |x|2p = p!.
A more general statement concerns a random matrix √1N XXH , where X is an n × N random matrix with independent standard complex Gaussian entries. It is known [HT] that p 1 X k(ˆπ ) l(ˆπ ) 1 H √ XX N n , = p τn N n N π∈Sp
where π ˆ is a permutation in S2p constructed in a certain way from π, and k(ˆ π ), l (ˆ π ) are functions taking values in {0, 1, 2, ...}. Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
One can show that this equals p X X ak 1 H √ XX . = 1+ τn N 2k N π ˆ ∈NC k 2p
The convergence is "almost sure", which means that we have very accurate eigenvalue prediction when the matrices are large.
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Motivation for free probability One can show that for the Gaussian random matrices we considered, the limits φ Ai1 B j1 · · · Ail B jl = lim trn Ain1 Bnj1 · · · Ainl Bnjl n→∞
exist. If we linearly extend the linear functional φ to all polynomials in A and B, the following can be shown:
Theorem If Pi , Qi are polynomials in A and B respectively, with 1 ≤ i ≤ l , and φ(Pi (A)) = 0, φ(Qi (B)) = 0 for all i, then φ (P1 (A)Q1 (B) · · · Pl (A)Ql (B)) = 0. This motivates the definition of freeness, which is the analogy to independence. Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Definition of freeness
Definition A family of unital ∗-subalgebras (Ai )i ∈I is called a free family if aj ∈ Aij ⇒ φ(a1 · · · an ) = 0. (2) i1 6= i2 , i2 6= i3 , · · · , in−1 6= in φ(a1 ) = φ(a2 ) = · · · = φ(an ) = 0
A family of random variables ai is called a free family if the algebras they generate form a free family.
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
The free central limit theorem
Theorem If a1 , ..., an are free and self-adjoint, φ(ai ) = 0, φ(ai2 = 1, supi |φ(aik )| < ∞ for all k,
√ then the sequence (a1 + · · · + an )/ n converges in distribution to the semicircle law. In free probability, the semicircle law thus √ has the role of the 1 Gaussian law. it’s density is density 2π 4 − x 2
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Similarities between classical and free probability 1
Additive convolution ⊞: The p.d.f. of the sum of free random variables. The role of the logarithm of the Fourier transform is now taken by the R-transform, which satisfies Rµa ⊞µb (z) = Rµa (z) + Rµb (z).
2
The S-transform: Transform on probability distributions which satisfies Sµa ⊠µb (z) = Sµa (z)Sµb (z)
3
Poisson distributions have their analogy in the free Poisson distributions: These are given by the Mar˘cenko Pastur laws µc with parameter c, which also can be written as the limit of ⊞n as n → ∞ 1 − cn δ(0) + nc δ(1)
4
Infinite divisibility: There exists an analogy to the Lévy-Hinčin formula for infinite divisibility in classical probability.
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Main usage of free probability in my papers Let A and B be random matrices. How can we make a good prediction of the eigenvalue distribution of A when one has the eigenvalue distribution of A + B and B? Simplest case is when one assumes that B is Gaussian (Noise). What about the eigenvectors? Assume that we have the eigenvalue distribution of 1 H N (R + X )(R + X ) , where R and X are n × N random matrices, with X Gaussian. If the columns of R are realizations of some random vector r , what is the covariance matrix E (ri rj∗ )? Have use for multiplicative free convolution with the Mar˘chenko Pastur law. This has an efficient implementation.
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Channel capacity estimation
The following is a much used observation model in MIMO systems: ˆ i = √1 (H + σXi ) H n
(3)
where n is the number of receiving and transmitting antennas, ˆ i is the n × n measured MIMO matrix, H H is the n × n MIMO channel and
Xi is the n × n noise matrix with i.i.d zero mean unit variance Gaussian entries.
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Channel capacity estimation With free probability we can estimate the eigenvalues of n1 HHH ˆ i . This helps us estimate the channel based on few observations H capacity: The capacity of a channel with channel matrix H and signal to noise ratio ρ = σ12 is given by C
= =
1 1 H log det I + 2 HH n nσ n 1X 1 log(1 + 2 λl ) n σ
(4) (5)
l=1
where λl are the eigenvalues of n1 HHH .
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Observation model
Form the compound observation matrix h i ˆ 1, H ˆ 2 , ..., H ˆL , ˆ 1...L = √1 H H L from the observations ˆ i = √1 (H + σXi ) , H n
(6)
Using free probability, one can with high accuracy estimate the ˆ 1...LH ˆH . eigenvalues of n1 HHH from the eigenvalues of H 1...L
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Free capacity estimation for channel matrices of various rank 4.5
4
Capacity
3.5 True capacity, rank 3 Cf, rank 3 True capacity, rank 5 Cf, rank 5
3
True capacity, rank 6 Cf, rank 6 2.5
2
1.5 0
10
20
30
40 50 60 Number of observations
70
80
90
100
Figure: The free probability based estimator for various number of observations. σ 2 = 0.1 and n = 10. The rank of H was 3, 5 and 6. Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
Application areas
digital communications, nuclear physics, mathematical finance Situations in these fields, can often be modelled with random matrices. When the matrices get large, free probability theory is an invaluable tool for describing the asymptotic behaviour of many systems. Other types of matrices which are of interest are random unitary matrices and random Vandermonde matrices.
Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
List of papers Free Deconvolution for Signal Processing Applications. Submitted to IEEE Trans. Inform. Theory. arxiv.org/cs.IT/0701025. Multiplicative free Convolution and Information-Plus-Noise Type Matrices. Submitted to Ann. Appl. Probab. arxiv.org/math.PR/0702342. Channel Capacity Estimation using Free Probability Theory. Submitted to IEEE. Trans. Signal Process. arxiv.org/abs/0707.3095. Random Vandermonde Matrices-Part I: Fundamental results. Work in progress. Random Vandermonde Matrices-Part II: Applications to wireless applications. Work in progress. Applications of free probability in finance. Estimation of the covariance matrix itself (not only it’s eigenvalue distribution). 2008. Øyvind Ryan
Applications of free probability and random matrix theory
CMA 2007
Applications of free probability and random matrix theory
References
[HT]: "Random Matrices and K-theory for Exact C ∗ -algebras". U. Haagerup and S. Thorbjørnsen. citeseer.ist.psu.edu/114210.html. 1998. This talk is available at http://heim.ifi.uio.no/∼oyvindry/talks.shtml. My publications are listed at http://heim.ifi.uio.no/∼oyvindry/publications.shtml
Øyvind Ryan
Applications of free probability and random matrix theory