Marchenko-Pastur Law

• Marchenko-Pastur Law is a fundamental result in random matrix theory that characterizes the asymptotic spectral distribution of sample covariance matrices.
• It provides explicit formulas and universal applicability for matrices with independent or weakly dependent entries, underpinning modern high-dimensional statistics.
• The law connects matrix models with orthogonal polynomials and free probability, yielding critical insights for signal processing and statistical physics applications.

The Marchenko–Pastur law is a fundamental result in random matrix theory describing the limiting spectral distribution (ESD) of sample covariance matrices formed from large collections of independent or weakly dependent high-dimensional random vectors. It is a key theoretical tool in mathematical statistics, high-dimensional signal processing, random matrix theory, and free probability. The law provides explicit formulas for the asymptotic empirical distribution of eigenvalues, supports universality in various structural models, and has deep connections with combinatorics, operator algebras, and statistical physics.

1. Definition and Explicit Formula

Let $X$ be a $p \times n$ random matrix with i.i.d. entries (mean zero, variance $\sigma^2$), and consider the (possibly rescaled) sample covariance matrix: $S = \frac{1}{n} X X^\top.$ Assume $p/n \to c \in (0, \infty)$ as $n, p \to \infty$. The empirical spectral distribution $\mu_{S}$ of $S$ converges almost surely to the nonrandom Marchenko–Pastur (MP) law with parameter $c$: $$d\mu_c(x) = \max\{1 - 1/c, 0\}\, \delta_0(\mathrm dx) + \frac{\sqrt{(b-x)\,(x-a)}}{2\pi c x}\, \mathbf{1}_{[a,b]}(x)\,\mathrm dx,$$ where $p \times n$0 and $p \times n$1 (Lu et al., 2014, Yaskov, 2021). For $p \times n$2, there is an atom at zero of mass $p \times n$3.

The MP law can also be described via its Stieltjes transform, which solves the quadratic equation: $p \times n$4 This analytic characterization is essential for rigorous proofs and extends naturally to generalized and free probability contexts (Szpojankowski, 2015).

2. Universal Domain of Validity and Generalizations

Independence, Dependence, and Universality

The convergence to the MP law does not fundamentally depend on Gaussianity; when the entries of $p \times n$5 are independent and suitably standardized (mean zero, variance one, finite fourth moment), the MP law holds universally for a substantial class of spectral statistics (Lu et al., 2014, Bloemendal et al., 2013). The law admits extensions to a range of settings with weak dependence:

• Block-independent and tensor models: If each vector is partitioned into blocks with intra-block dependence but small maximum block size, or formed as symmetric tensors of independent variables with certain growth control ($p \times n$6), the MP law still holds for the sample covariances (Yaskov, 2021, Bryson et al., 2019).
• Weak correlations: For Gaussian data with general covariance matrices whose entrywise correlations decay to zero as $p \times n$7 for any $p \times n$8, all spectral moments converge to those of the MP law; this is a sharp threshold (Fleermann et al., 2022).
• Dependent entries with mild conditions: For generic dependent ensembles, e.g. random matrices with weak intra-row and short-range inter-row dependence, convergence (at least in expectation) is preserved when decorrelation and moment conditions are met (O'Rourke, 2012).

A table summarizing selected generalizations:

Model Type	Key Condition	MP Law Validity
i.i.d. entries	Mean 0, variance 1, $p \times n$9	Yes, for $\sigma^2$0
Block-dependent	Max block size $\sigma^2$1	Yes, (Bryson et al., 2019)
Homogeneous random tensors	$\sigma^2$2 (optimal scaling)	Yes, (Yaskov, 2021, Bryson et al., 2019)
Uniformly decaying correlation	$\sigma^2$3 with $\sigma^2$4	Yes, (Fleermann et al., 2022)

For sample covariance matrices from time series or structured ensembles, see sections below.

3. Combinatorial Structure and Connections

Moment Formula and Laguerre Polynomials

The $\sigma^2$5th moment of the MP law is

$\sigma^2$6

This formula is directly matched by the leading term in the sum of powers of the roots of Laguerre polynomials $\sigma^2$7 under the scaling $\sigma^2$8, solidifying the intrinsic connection between matrix models, orthogonal polynomials, and free convolution (Kornyik et al., 2016).

Moreover, the expectation of the characteristic polynomial $\sigma^2$9 matches the monic Laguerre polynomial, with the roots of the mean characteristic polynomial equidistributed according to the MP law.

Free Probability and Regression Characterization

Within the framework of free probability, the MP law corresponds to the free Poisson distribution and belongs to the free Meixner class. It is uniquely characterized by a constant conditional expectation of certain noncommutative linear forms (compressions and inverse twists) built from free, positive, self-adjoint elements $S = \frac{1}{n} X X^\top.$0 in a $S = \frac{1}{n} X X^\top.$1-probability space: $S = \frac{1}{n} X X^\top.$2 with $S = \frac{1}{n} X X^\top.$3 and $S = \frac{1}{n} X X^\top.$4 free. This regression collapse ties the MP law to a noncommutative analog of the Lukacs–Wesołowski characterization of the gamma law (Szpojankowski, 2015).

4. Extensions to Structured and Ensemble-specific Settings

Graph Ensembles

For sparse random bipartite biregular graphs, the symmetrized spectrum of the normalized adjacency matrix converges to a symmetrized MP law, and the spectrum of the associated biadjacency block matches the classical MP density under appropriate scaling and mild degree growth conditions (Dumitriu et al., 2013, Yang, 2017).

Random Band Matrices

In band matrix ensembles with bandwidth increasing appropriately with $S = \frac{1}{n} X X^\top.$5, the limiting ESD converges to the MP law. The mechanism involves reducing the general integral equation for the Stieltjes transform to the quadratic MP equation for the $S = \frac{1}{n} X X^\top.$6 case, formalizing the universality of the MP law beyond classical full-matrix or sparse models (Jana et al., 2016).

Time Series and High-dimensional Periodograms

For high-dimensional discrete stationary time series:

• When the innovation process has independently sampled coordinates and the system has simultaneous diagonalizability, both covariance and lagged autocovariance matrices have a limiting ESD described by generalizations of the MP law, given by a functional equation for the Stieltjes transform involving population spectral measures (Liu et al., 2013).
• For Daniell-smoothed periodograms $S = \frac{1}{n} X X^\top.$7 of high-dimensional (possibly non-diagonalizable) time series, when $S = \frac{1}{n} X X^\top.$8, $S = \frac{1}{n} X X^\top.$9, and $p/n \to c \in (0, \infty)$0, the ESD of $p/n \to c \in (0, \infty)$1 at any frequency converges to the MP law of parameter $p/n \to c \in (0, \infty)$2 (Deitmar, 2024).

A notable implication is that in high-dimensional regimes, periodogram-based spectral density estimators are generically inconsistent due to Marchenko–Pastur-style spreading.

5. Local Spectral Laws, Rigidity, and Delocalization

Local Marchenko–Pastur Laws

Deep refinements establish that, with optimal accuracy, the empirical eigenvalue density of $p/n \to c \in (0, \infty)$3 converges locally to the MP density at all scales down to $p/n \to c \in (0, \infty)$4 up to the hard edge ($p/n \to c \in (0, \infty)$5) (Cacciapuoti et al., 2012, Kafetzopoulos et al., 2022). Isotropic variants of the local law provide sharp bounds for all quadratic forms,

$p/n \to c \in (0, \infty)$6

for deterministic unit vectors $p/n \to c \in (0, \infty)$7, $p/n \to c \in (0, \infty)$8 in the bulk, and $p/n \to c \in (0, \infty)$9 the MP Stieltjes transform (Bloemendal et al., 2013).

Eigenvalue Rigidity and Delocalization

The local laws enable eigenvalue rigidity (closeness of individual eigenvalues to their classical locations at $n, p \to \infty$0), and delocalization of eigenvectors (entries are at most $n, p \to \infty$1 in the bulk) (Kafetzopoulos et al., 2022, Yang, 2017, Cacciapuoti et al., 2012).

6. Robustness, Universality, and Nonlinear Spectral Statistics

The MP law extends to robust covariance estimators. For instance, Tyler's M-estimator, a scale-invariant robust scatter estimator, shares the same limiting ESD (after rescaling) as the sample covariance in the high-dimensional limit under Gaussian or elliptical populations (Zhang et al., 2014).

In nonlinear settings, such as the empirical Kendall's tau matrix for i.i.d. vector-valued random data, the limiting ESD is an affine transformation of the MP law, making it manifestly universal for even certain $n, p \to \infty$2-statistic-based random matrices (Bandeira et al., 2016).

Similarly, in $n, p \to \infty$3-deformed ensembles (e.g., the $n, p \to \infty$4-Laguerre unitary ensemble in the $n, p \to \infty$5 scaling regime), the limiting law exhibits a $n, p \to \infty$6-deformation of the MP density, with a phase transition—recovering the classical MP law as $n, p \to \infty$7 (Byun et al., 14 Jan 2026).

7. Methodologies and Proof Techniques

Key methodologies include:

• Resolvent (Stieltjes transform) analysis: Fundamental for establishing both global and local laws, reduction to quadratic fixed-point equations, and deriving analytic properties of the limiting spectral distribution (Lu et al., 2014, Deitmar, 2024).
• Moment methods and combinatorics: Enumeration of noncrossing pairings, catalan and Narayana numbers, and explicit matching with Laguerre polynomial root statistics (Kornyik et al., 2016).
• Concentration inequalities: Control of quadratic forms and trace functionals via Hanson–Wright and related bounds, necessary for general dependence models (Yaskov, 2021, Bryson et al., 2019).
• Feynman diagrams and diagrammatic expansions: Widespread in physics, provide an alternative derivation of the law and highlight the universality under minimal moment assumptions (Lu et al., 2014).
• Large deviation principles and equilibrium problems: Used for advanced settings, such as $n, p \to \infty$8-deformations, yielding a full LDP and explicit variational characterizations (Byun et al., 14 Jan 2026).

• (Szpojankowski, 2015): A constant regression characterization of a Marchenko–Pastur law
• (Yaskov, 2021): Marchenko–Pastur law for a random tensor model
• (Zhang et al., 2014): Marchenko–Pastur Law for Tyler's M-estimator
• (Lu et al., 2014): Universal Asymptotic Eigenvalue Distribution of Large $n, p \to \infty$9 Random Matrices — A Direct Diagrammatic Proof
• (Bryson et al., 2019): Marchenko–Pastur law with relaxed independence conditions
• (Deitmar, 2024): Marchenko–Pastur laws for Daniell smoothed periodograms
• (Kafetzopoulos et al., 2022): Local Marchenko–Pastur law at the hard edge of the Sample Covariance ensemble
• (Fleermann et al., 2022): Large Sample Covariance Matrices of Gaussian Observations with Uniform Correlation Decay
• (Cacciapuoti et al., 2012): Local Marchenko–Pastur Law at the Hard Edge of Sample Covariance Matrices
• (Dumitriu et al., 2013): The Marčenko–Pastur law for sparse random bipartite biregular graphs
• (Bandeira et al., 2016): Marčenko–Pastur Law for Kendall's Tau
• (Yang, 2017): Local Marchenko–Pastur Law for Random Bipartite Graphs
• (Liu et al., 2013): On the Marčenko–Pastur law for linear time series
• (Agliari et al., 2018): A novel derivation of the Marchenko–Pastur law through analog bipartite spin-glasses
• (O'Rourke, 2012): A note on the Marchenko–Pastur law for a class of random matrices with dependent entries
• (Kornyik et al., 2016): On the moments of roots of Laguerre-polynomials and the Marchenko–Pastur law
• (Bloemendal et al., 2013): Isotropic Local Laws for Sample Covariance and Generalized Wigner Matrices
• (Jana et al., 2016): Distribution of singular values of random band matrices; Marchenko–Pastur law and more
• (Byun et al., 14 Jan 2026): $\mu_{S}$0-deformation of the Marchenko–Pastur law

This spectrum of results establishes the Marchenko–Pastur law as a universal invariant for the bulk eigenvalue distribution of a wide class of large random matrices, unifying random matrix theory, free probability, and high-dimensional statistics.