the Marchenko-Pastur law
/ Marchenko-Pastur = mar-CHEN-ko PAS-toor /
The Marchenko-Pastur law is the semicircle law's sibling for sample covariance matrices: it is the universal limiting density of the eigenvalues of a high-dimensional sample covariance matrix when the true covariance is the identity. It answers the practical question that haunts every high-dimensional statistician: if I estimate covariance from data that is pure isotropic noise, what does the eigenvalue spectrum of my estimate look like, and where does noise end and signal begin?
Let S = (1/n) X X^* be a p-by-p sample covariance matrix built from a p-by-n matrix X of iid entries with mean 0 and variance 1, and let both p and n grow with aspect ratio y = p/n converging to a constant in (0, infinity). Then the empirical spectral distribution of S converges to the Marchenko-Pastur law. For y <= 1 it is the absolutely continuous density rho(x) = (1/(2 pi y x)) sqrt((b - x)(x - a)) on the interval [a, b], where the edges are a = (1 - sqrt(y))^2 and b = (1 + sqrt(y))^2. For y > 1 (more variables than samples, so S is singular) the matrix has p - n exact zero eigenvalues, contributing an atom of mass 1 - 1/y at 0, plus the same continuous density on [a, b]. The edges a and b are the spectral support: when y = 1 they are 0 and 4, when y is small they pinch toward 1 (recovering the classical S -> identity). The Stieltjes transform of the MP law satisfies a quadratic self-consistent equation, the covariance analogue of the semicircle's.
The Marchenko-Pastur law is the workhorse of high-dimensional statistics. It is the null model: if your data really were unstructured noise, your sample covariance eigenvalues would fill [a, b] and nothing more, so any eigenvalue lying above the upper edge b is statistical evidence of genuine signal (a real direction of variance) rather than sampling artifact — this is the logic behind the spiked-model and the BBP transition, and behind cleaning empirical covariance matrices in finance. A caveat: the clean MP law assumes the population covariance is the identity (isotropic noise) and entries with finite variance. A general population covariance Sigma produces a deformed law — the free multiplicative convolution of MP with the spectrum of Sigma — and heavy-tailed entries break the law entirely. And the upper edge b is the bulk edge; the largest eigenvalue's fluctuation around b is again Tracy-Widom, on scale n^(-2/3).
A portfolio analyst has p = 500 stock returns over n = 1000 days, so y = 0.5. The Marchenko-Pastur upper edge is b = (1 + sqrt(0.5))^2 approximately 2.91. Any eigenvalue of the sample correlation matrix below 2.91 is statistically indistinguishable from noise and should be regularised; the one or two eigenvalues far above 2.91 (a market-wide 'mode') are the real, tradeable signal. The MP edge is the noise cutoff.
Eigenvalues above the MP upper edge b = (1+sqrt(y))^2 are signal; everything inside is noise.
The MP law assumes isotropic population covariance (identity). For a non-identity Sigma the limiting spectrum is the free multiplicative convolution of MP with Sigma's spectrum, not MP itself; using the plain MP edges as a noise cutoff is only valid for genuinely white noise.