Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
1
Advances in Random Matrix Theory: Let there be tools
Alan Edelman
MIT: Dept of Mathematics, Computer Science AI Laboratories
World Congress, Bernoulli Society Barcelona, Spain
Wednesday July 28, 20042/1/2005
Brian Sutton, Plamen Koev, Ioana Dumitriu,
Raj Rao and others
2
Message Ingredient: Take Any important mathematics Then Randomize! This will have many applications! We can’t keep this in the hands of specialists
anymore: Need tools!
2/1/2005
3
Tools
So many applications … Random matrix theory: catalyst for 21st century
special functions, analytical techniques, statistical techniques
In addition to mathematics and papers Need tools for the novice! Need tools for the engineers! Need tools for the specialists!
2/1/2005
4
Themes of this talk
Tools for general beta What is beta? Think of it as a measure of (inverse)
volatility in “classical” random matrices. Tools for complicated derived random matrices Tools for numerical computation and simulation
2/1/2005
Precise statements require n→∞ etc.
5
Let S = random symmetric Gaussian MATLAB: A=randn(n); S=(A+A’)/2;
S known as the Gaussian Orthogonal Ensemble n=20; s=30000; d=.05; e=[]; %gather up eigenvalues im=1; %imaginary(1) or real(0) for i=1:s,
v=eig(a)'; e=[e v]; end
axis('square'); axis([-1.5 1.5 -1 2]); hold on; t=-1:.01:1; plot(t,sqrt(1-t.^2),'r');
Wigner’s Semi-Circle The classical & most famous rand eig theorem
Normalized eigenvalue histogram is a semi-circle %matrix size, samples, sample dist
a=randn(n)+im*sqrt(-1)*randn(n);a=(a+a')/(2*sqrt(2*n*(im+1)));
hold off; [m x]=hist(e,-1.5:d:1.5); bar(x,m*pi/(2*d*n*s));
χ3χG
Gχ2
χ
2/1/2005
χ6
G
χ5
G χ6
χ4
G χ5
G χ4
2
χ3
χ1 G 1
Same eigenvalue distribution as GOE: O(n) storage !! O(n2) compute
6
sym matrix to tridiagonal form
G4β
χ
7
χ6β
G
χ5β
G χ6β
χ4β
G χ5β
χ3β
χ
χ2β
G 3β
χβ
G χ2β
G χβ
General beta beta:
2/1/2005
1: reals 2: complexes 4: quaternions
Bidiagonal Version corresponds To Wishart matrices of Statistics
8
Tools Motivation: A condition number problem
The Polynomial Method 9 trick
2/1/2005
Jack & Hypergeometric of Matrix Argument MOPS: Ioana Dumitriu’s talk
The tridiagonal numerical 10
9
Tools Motivation: A condition number problem
The Polynomial Method 9 trick
2/1/2005
Jack & Hypergeometric of Matrix Argument MOPS: Ioana Dumitriu’s talk
The tridiagonal numerical 10
10
Numerical Analysis: Condition Numbers
κ(A) = “condition number of A” If A=UΣ κ(A) = σmax/σmin . Alternatively, κ(A) = √λ max (A’A)/√λ min (A’A) One number that measures digits lost in finite
precision and general matrix “badness” Small=good Large=bad
The condition of a random matrix??? 2/1/2005
V’ is the svd, then
11
Von Neumann & co. Solve Ax=b via x= (A’A) -1A’ b
M ≈A-1 Matrix Residual: ||AM-I||2 ||AM-I||2< 200κ2 n ε
How should we estimate κ?
Assume, as a model, that the elements of A areindependent standard normals!
≈
2/1/2005
12
Von Neumann & co. estimates (1947-1951)
“For a ‘random matrix’ of order n the expectation value has been shown to be about n”
Goldstine, von Neumann
√10n” Bargmann, Montgomery, vN
Goldstine, von Neumann
X ∞
P(κ
2 / 2 / 2 3
4 2 x x e x x y − − + =
Distribution of κ/n
13
→∞
Experiment with n=2002/1/2005
Random cond numbers, n
14
Finite n
n=10 n=25
n=50 n=100 2/1/2005
15
Condition Number Distributions P(κ/n > x) ≈ 2/x P(κ/n2 > x) ≈ 4/x2
Real n x n, nÆ∞
Complex n x n, nÆ∞
Generalizations:
•β: 1=real, 2=complex
•finite matrices
•rectangular: m x n
2/1/2005
2/1/2005 16
Real n x n, nÆ∞
Complex n x n, nÆ∞
Condition Number Distributions P(κ/n > x) ≈ 2/x P(κ/n2 > x) ≈ 4/x2
Square, nÆ∞: P(κ/nβ > x) ≈ (2ββ-1/Γ(β))/xβ (All Betas!!) General Formula: P(κ>x) ~Cµ/x β(n-m+1),
where µ= βWm-1,n+1 (β) and C is a known geometrical constant.
1F1((β/2)(n+1), ((β/2)(n+m-1); -(x/2)Im-1)
(n-m+1)/2th moment of the largest eigenvalue of
Density for the largest eig of W is known in terms of from which µ is available
Tracy-Widom law applies probably all beta for large m,n. Johnstone shows at least beta=1,2.
17
Tools Motivation: A condition number problem
The Polynomial Method 9 trick
2/1/2005
Jack & Hypergeometric of Matrix Argument MOPS: Ioana Dumitriu’s talk
The tridiagonal numerical 10
18
Multivariate Orthogonal Polynomials &
The important special functions of the 21st century Begin with w(x) on I
∫ pκ(x)pλ(x) ∆(x)β ∏i w(xi)dxi = δκλ Jack Polynomials orthogonal for w=1
on the unit circle. Analogs of xm 2/1/2005
Hypergeometrics of Matrix Argument
Ioana Dumitriu’s talk
192/1/2005
Multivariate Hypergeometric Functions
202/1/2005
Multivariate Hypergeometric Functions
21
Wishart (Laguerre) Matrices! 2/1/2005
2/1/2005 22
Plamen’s clever idea
23
Tools Motivation: A condition number problem
The Polynomial Method 9 trick
2/1/2005
Jack & Hypergeometric of Matrix Argument MOPS: Ioana Dumitriu’s talk
The tridiagonal numerical 10
2/1/2005 24
Mops (Dumitriu etc.) Symbolic
A=randn(n); S=(A+A’)/2; trace(S^4)
det(S^3)
25
Symbolic MOPS applications
2/1/2005
β=3; hist(eig(S))
26
Symbolic MOPS applications
2/1/2005
A=randn(m,n); hist(min(svd(A).^2))
272/1/2005
Smallest eigenvalue statistics
28
Tools Motivation: A condition number problem
9 trick
2/1/2005
Jack & Hypergeometric of Matrix Argument MOPS: Ioana Dumitriu’s talk The Polynomial Method -- Raj! The tridiagonal numerical 10
29
Courtesy of the Polynomial Method 2/1/2005
RM Tool – Raj!
2/1/2005 30
∑ ∫ ∞
=
∞ −
= − Γ
= 1 0
1 1 1 ) (
1 ) ζ( k
s u
s
k du
e u
s s
6 π
4 1
3 1
2 1 1 ) ζ 2 (
2
2 2 2 = + + + + = …
31
The Riemann Zeta Function
On the real line with x>1, for example
May be analytically extended to the complex plane, with singularity only at x=1.
2/1/2005
2/1/2005 32
-3 -2 -1 0 ½ 1 2 3
∑ ∫ ∞
=
∞ −
= − Γ
= 1 0
1 1 1 ) (
1 ) ζ( k
x u
x
k du
e u
x x
All nontrivial roots of ζ(x) satisfy Re(x)=1/2. (Trivial roots at negative even integers.)
The Riemann Hypothesis
2/1/2005 33
-3 -2 -1 0 ½ 1 2 3
The Riemann Hypothesis
∑∫∞
=
∞ −
=−Γ
=10
1 11)(
1)ζ(k
xu
x
kdu
eu
xx
All nontrivial roots of ζ(x) satisfy Re(x)=1/2. (Trivial roots at negative even integers.)
ζ(x)| along Re(x)=1/2 Zeros =.5+i* 14.134725142 21.022039639 25.010857580 30.424876126 32.935061588 37.586178159 40.918719012 43.327073281 48.005150881 49.773832478 52.970321478 56.446247697 59.347044003
|
34
Computation of Zeros 10^k+1
through 10^k+10,000 for k=12,21,22.
See http://www.research.att.com/~amo/zeta_tables/
A=randn(n)+i*randn(n); S=(A+A’)/2;
2/1/2005
Odlyzko’s fantastic computation of
Spacings behave like the eigenvalues of
352/1/2005
Nearest Neighbor Spacings & Pairwise Correlation Functions
2/1/2005 36
Painlevé Equations
37
Spacings
Normalize so that average spacing = 1.
jth) Conjecture: These functions are the same for random
matrices and Riemann zeta
2/1/2005
Take a large collection of consecutive zeros/eigenvalues.
Spacing Function = Histogram of consecutive differences (the (k+1)st – the kth) Pairwise Correlation Function = Histogram of all possible differences (the kth – the
38
Tools Motivation: A condition number problem
The Polynomial Method 9 trick
2/1/2005
Jack & Hypergeometric of Matrix Argument MOPS: Ioana Dumitriu’s talk
The tridiagonal numerical 10
39
Everyone’s Favorite Tridiagonal
-21 1
1-21 1-2
… ……
… …
1 n2
d2 dx2
2/1/2005
40
Everyone’s Favorite Tridiagonal
-21 1
1-21 1-2
… ……
… …
1 n2
d2 dx2
1 (βn)1/2+
G
G G
dW β1/2+
2/1/2005
41
Stochastic Operator Limit
,
N(0,2)χ χN(0,2)χ
χN(0,2)χ χN(0,2)
nβ2 1 ~H
β
β2β
2)β(n1)β(n
1)β(n
β n
⎟⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜⎜
⎝
⎛
−−
−
β 2 x
dx d
2
2
+−
,G β
2HH nn β n +≈
∞
… …
…
2/1/2005
, dW
422/1/2005
Largest Eigenvalue Plots
43
MATLAB
beta=1; n=1e9; opts.disp=0;opts.issym=1;alpha=10; k=round(alpha*n^(1/3)); % cutoff parametersd=sqrt(chi2rnd( beta*(n:-1:(n-k-1))))';H=spdiags( d,1,k,k)+spdiags(randn(k,1),0,k,k);H=(H+H')/sqrt(4*n*beta);eigs(H,1,1,opts)
2/1/2005
44
Tricks to get O(n9) speedup • •
•
general purpose eigensolvers.) •
convergence of the largest eigenvalue) •
structures and numerical choices.) • 1/3
determined numerically by the top k × k segment of n. (This is aninteresting mathematical statement related to the decay of the Airyfunction.)
2/1/2005
Sparse matrix storage (Only O(n) storage is used) Tridiagonal Ensemble Formulas (Any beta is available due to the tridiagonal ensemble) The Lanczos Algorithm for Eigenvalue Computation ( This allows the computation of the extreme eigenvalue faster than typical
The shift-and-invert accelerator to Lanczos and Arnoldi (Since we know the eigenvalues are near 1, we can accelerate the
The ARPACK software package as made available seamlessly in MATLAB (The Arnoldi package contains state of the art data
The observation that if k = 10n , then the largest eigenvalue is
45
Level Densities
2/1/2005
46
Open Problems
The distribution for general beta Seems to be governed by a convection-diffusion equation
2/1/2005
47
Random matrix tools!
2/1/2005
/ColorImageDict > /JPEG2000ColorACSImageDict > /JPEG2000ColorImageDict > /AntiAliasGrayImages false /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 150 /GrayImageDepth -1 /GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages true /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict > /GrayImageDict > /JPEG2000GrayACSImageDict > /JPEG2000GrayImageDict > /AntiAliasMonoImages false /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 300 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict > /AllowPSXObjects true /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXOutputIntentProfile (None) /PDFXOutputCondition () /PDFXRegistryName (http://www.color.org) /PDFXTrapped /False
/Description >>> setdistillerparams> setpagedevice