Marchenko–Pastur law: Definition & Examples

The Marchenko–Pastur law describes the limiting distribution of eigenvalues of a large sample covariance matrix under an independent, equal-variance noise model. Its practical message is striking: even when the population covariance is the identity matrix, sample eigenvalues spread across a wide band. A portfolio optimiser can mistake this sampling variation for reliably low-risk or high-risk directions.

The noise model behind the band

Let $X$ have $n$ observation rows and $p$ variable columns, with independent, identically distributed entries of mean zero and variance $\sigma^2$ . Use the known-zero-mean covariance estimator

S=\frac{1}{n}X^\top X.

As $p,n\to\infty$ with $p/n\to c\in(0,\infty)$ , the empirical eigenvalue distribution—giving weight $1/p$ to each of the $p$ eigenvalues—converges to the Marchenko–Pastur distribution. Its continuous part has density

f_c(\lambda)=\frac{\sqrt{(\lambda_+-\lambda)(\lambda-\lambda_-)}}{2\pi c\sigma^2\lambda}, \qquad \lambda_-<\lambda<\lambda_+,

where

\lambda_\pm=\sigma^2(1\pm\sqrt c)^2.

The density is zero outside this interval. If

c>1

, the distribution also has probability mass

1-1/c

concentrated exactly at zero. The continuous part then integrates to

1/c

, rather than one. Lalley's covariance-matrix theorem develops the iid model; Adamczak's Definition 2.2 states the distribution including its zero mass.

Here

c

means variables divided by observations. Some sources use the reciprocal ratio. Check the convention before transferring an edge formula.

One hundred assets, four hundred observations

Consider a hypothetical universe of $p=100$ independent, standardised return series observed on $n=400$ independent dates. The population variance is one in every direction. With $c=100/400=0.25$ ,

\lambda_-=(1-0.5)^2=0.25, \qquad \lambda_+=(1+0.5)^2=2.25.

Noise alone therefore produces a limiting band whose upper edge is nine times its lower edge. This is a band for sample eigenvalues, not a range of different population variances, and the density within it is not uniform.

A seeded simulation of 100 independent unit-variance variables over 400 observations compared with the Marchenko–Pastur density between 0.25 and 2.25. Support bands widen as p/n increases; at p/n=2, half the limiting probability is a separate mass at zero.

Figure: Original Gaussian simulation with seed 20261030, known zero mean and divisor n. The lower panel compares limiting support bands and reports zero mass separately; it does not plot that mass as a density height.

The endpoints are asymptotic edges, not finite-sample confidence limits. Even correctly generated finite noise matrices can have eigenvalues beyond them. An eigenvalue just above 2.25 does not by itself establish an economic factor. Laloux and colleagues explicitly discuss this finite-sample blurring in their study of financial correlation matrices.

When there are more variables than observations

With

p=200

and

n=100

c=2

. The positive continuous support is approximately

[0.1716,5.8284]

, but half the limiting probability is at zero. The positive lower edge does not imply an invertible matrix.

This also follows from matrix rank:

X^\top X

has rank at most

n

. With iid Gaussian entries it has rank

n

almost surely, leaving

p-n=100

zero eigenvalues in this known-mean example. The nonzero eigenvalues cannot compensate for the missing directions when an optimiser tries to invert

S

Normalisation matters. Ordinary sample covariance usually estimates means and divides by

n-1

. For Gaussian observations, the corresponding reference has

n-1

independent degrees of freedom: use

p/(n-1)

with that divisor. Centring the 200-variable, 100-observation example reduces the rank to at most 99, leaving at least 101 zeros. Changing the divisor rescales eigenvalues; it cannot restore rank.

A benchmark for cleaning, with a model attached

Comparing an observed spectrum with this noise benchmark helps diagnose estimation error before using covariance eigenvectors as portfolio risk directions. It motivates covariance eigenvalue clipping, but the law alone does not choose a replacement rule or establish its performance.

For actual returns, serial dependence, changing volatility, unequal population variances and genuine cross-asset relationships can alter the reference spectrum. A sample correlation matrix also estimates its volatility scaling from data; it is not literally the known-variance matrix used above. Departures from the band can reflect model mismatch as well as structure, while weak structure can remain inside the bulk. Any cleaning method still needs evaluation on later observations.

For the underlying risk interpretation, Eigenvalues and Covariance Matrices explains why an eigenvalue is the variance of a unit-length eigenportfolio. The preview is public; the full lesson requires membership.

Sources

Steven Lalley, Bulk Spectrum: Wigner and Marchenko–Pastur Theorems, §2, Theorem 5 and its iid extension: covariance normalisation and the limiting density.
Radosław Adamczak (2011), On the Marchenko–Pastur and circular laws for some classes of random matrices with dependent entries, §2.2, Definition 2.2: full distribution including the zero atom; the introduction explains why dependence needs separate assumptions.
Laurent Laloux, Pierre Cizeau, Jean-Philippe Bouchaud and Marc Potters (1999), Noise Dressing of Financial Correlation Matrices, pp. 1467–1468: portfolio motivation, the independent-return benchmark and finite-sample edges.

The noise model behind the band

One hundred assets, four hundred observations

When there are more variables than observations

A benchmark for cleaning, with a model attached

Sources

Related terms

Covariance eigenvalue clipping

Eigenvalue

Matrix rank

Sample covariance