{
  "id": "marchenko-pastur-law",
  "title": "Marchenko–Pastur Law",
  "lang": "en",
  "source_lang": "en",
  "summary": "The Marchenko–Pastur law describes the limiting eigenvalue distribution of large sample covariance matrices when dimension and sample size grow proportionally.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "Marchenko–Pastur distribution",
    "Marčenko–Pastur law",
    "MP law"
  ],
  "generated_at": "2026-11-06",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/marchenko-pastur-law/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/marchenko-pastur-law/",
      "json": "https://aiwiki.page/zh/marchenko-pastur-law.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The Marchenko–Pastur law is a fundamental result in random matrix theory describing the limiting distribution of the eigenvalues of large sample covariance matrices. When the number of measured variables and the number of observations increase at comparable rates, the eigenvalues generally remain spread over a nontrivial interval, even when the underlying variables are independent and have equal variance. The law specifies this spread and, when there are more variables than observations, the proportion of eigenvalues equal to zero. It was established by Vladimir Marchenko and Leonid Pastur in 1967."
    },
    {
      "heading": "Setting and statement",
      "text": "Let \\(X\\) be a real \\(p\\times n\\) matrix, whose columns represent \\(n\\) observations of \\(p\\) variables. Assume that its entries are independent and identically distributed, with mean zero and variance \\(\\sigma^2>0\\). Define\n\n\\[S_n=\\frac{1}{n}XX^{\\mathsf T},\\]\n\nwhere \\(X^{\\mathsf T}\\) denotes the transpose. For complex entries, the transpose is replaced by the conjugate transpose, and the variance means \\(\\mathbb E|X_{ij}|^2\\).\n\nThe matrix \\(S_n\\) is an uncentered sample covariance matrix. Its eigenvalues \\(\\lambda_1,\\ldots,\\lambda_p\\) are nonnegative. Their empirical spectral distribution is\n\n\\[\\mu_{S_n}=\\frac{1}{p}\\sum_{i=1}^{p}\\delta_{\\lambda_i},\\]\n\nwhere \\(\\delta_t\\) is a unit point mass at \\(t\\). This distribution gives equal weight to every eigenvalue, including any zeros.\n\nSuppose that\n\n\\[p,n\\longrightarrow\\infty, \\qquad \\frac{p}{n}\\longrightarrow c\\in(0,\\infty).\\]\n\nIn the standard independent, identically distributed model, finite variance is sufficient for \\(\\mu_{S_n}\\) to converge almost surely, in the sense of weak convergence, to a deterministic probability distribution \\(\\mu_{c,\\sigma^2}\\).\n\nThis means that, for every bounded continuous function \\(f\\),\n\n\\[\\frac{1}{p}\\sum_{i=1}^{p}f(\\lambda_i) \\longrightarrow \\int f(x)\\,\\mu_{c,\\sigma^2}(dx)\\]\n\nalmost surely. The theorem concerns the distribution of the eigenvalues collectively; it does not, by itself, establish convergence of the largest or smallest eigenvalue."
    },
    {
      "heading": "Density and point mass at zero",
      "text": "Define the spectral edges\n\n\\[a=\\sigma^2(1-\\sqrt c)^2, \\qquad b=\\sigma^2(1+\\sqrt c)^2.\\]\n\nThe Marchenko–Pastur distribution is\n\n\\[\\mu_{c,\\sigma^2}(dx) = \\left(1-\\frac{1}{c}\\right)_+\\delta_0(dx) + \\frac{\\sqrt{(b-x)(x-a)}}{2\\pi c\\sigma^2x} \\mathbf 1_{(a,b)}(x)\\,dx,\\]\n\nwhere \\(u_+=\\max(u,0)\\), and \\(\\mathbf 1_{(a,b)}\\) is the indicator function of the interval \\((a,b)\\).\n\nThe continuous part has a density supported between \\(a\\) and \\(b\\). Its total mass is \\(1\\) when \\(c\\leq1\\), and \\(1/c\\) when \\(c>1\\).\n\nFewer variables than observations: \\(0<c<1\\)\n\nThere is no point mass at zero, and the lower edge is strictly positive:\n\n\\[a>0.\\]\n\nNevertheless, the limiting eigenvalues are not all equal to the population variance \\(\\sigma^2\\). Sampling variability produces a persistent spectral spread because the dimension increases with the sample size.\n\nEqual asymptotic dimensions: \\(c=1\\)\n\nThe support is \\([0,4\\sigma^2]\\), and the density becomes\n\n\\[f(x)= \\frac{1}{2\\pi\\sigma^2} \\sqrt{\\frac{4\\sigma^2-x}{x}}, \\qquad 0<x<4\\sigma^2.\\]\n\nIt diverges as \\(x^{-1/2}\\) near zero, but the divergence is integrable. There is no point mass at zero.\n\nMore variables than observations: \\(c>1\\)\n\nThe law has a point mass of size \\(1-1/c\\) at zero. This follows from rank:\n\n\\[\\operatorname{rank}(XX^{\\mathsf T})\\leq n.\\]\n\nThus, when \\(p>n\\), at least \\(p-n\\) eigenvalues are exactly zero, and their proportion approaches \\(1-1/c\\). The continuous component remains separated from zero because \\(a>0\\)."
    },
    {
      "heading": "Normalization conventions",
      "text": "Different references use different aspect ratios and matrix normalizations. The density must be interpreted together with these choices.\n\nFor \\(X\\in\\mathbb R^{p\\times n}\\), this article uses\n\n\\[c=\\frac pn, \\qquad S_n=\\frac1nXX^{\\mathsf T},\\]\n\nand assigns mass \\(1/p\\) to each eigenvalue.\n\nThe companion Gram matrix\n\n\\[G_n=\\frac1nX^{\\mathsf T}X\\]\n\nhas the same nonzero eigenvalues as \\(S_n\\), but it is \\(n\\times n\\). Consequently, its empirical distribution has different weights and a different number of zero eigenvalues. At finite dimensions,\n\n\\[\\mu_{G_n} = \\frac pn\\,\\mu_{S_n} + \\left(1-\\frac pn\\right)\\delta_0.\\]\n\nWhen \\(p>n\\), the negative coefficient in this algebraic identity cancels part of the zero mass already present in \\(\\mu_{S_n}\\); the resulting measure is still a probability measure.\n\nSome presentations instead normalize by \\(p\\) or define the aspect ratio as \\(n/p\\). These choices rescale the edges or reverse the parameter convention. Apparent discrepancies between formulas often arise from these differences."
    },
    {
      "heading": "Interpretation in high-dimensional statistics",
      "text": "If the columns of \\(X\\) have population covariance \\(\\sigma^2I_p\\), where \\(I_p\\) is the identity matrix, every population eigenvalue equals \\(\\sigma^2\\). In fixed dimension, the sample covariance matrix approaches the population covariance as \\(n\\) increases.\n\nThe proportional-growth regime is different. With \\(p/n\\to c>0\\), noise alone produces sample eigenvalues throughout the Marchenko–Pastur bulk. A broad spectrum therefore does not necessarily indicate unequal population variances or genuine correlation structure.\n\nFor example, if \\(\\sigma^2=1\\) and \\(p/n\\to1/4\\), then\n\n\\[a=\\frac14, \\qquad b=\\frac94.\\]\n\nThus, under an isotropic independent-noise model, the bulk extends from \\(0.25\\) to \\(2.25\\), despite every population eigenvalue being \\(1\\).\n\nAs \\(c\\) tends to zero, both edges approach \\(\\sigma^2\\), and the limiting distribution approaches \\(\\delta_{\\sigma^2}\\). This connects the high-dimensional regime to the classical regime in which the number of observations greatly exceeds the number of variables.\n\nCentering observations\n\nIn applications, covariance matrices are usually formed after subtracting the sample mean. If \\(\\bar x\\) is the mean observation vector, the centered matrix with normalization \\(1/n\\) is\n\n\\[S_n^{\\mathrm{cent}} = \\frac1n\\sum_{j=1}^{n}(x_j-\\bar x)(x_j-\\bar x)^{\\mathsf T} = S_n-\\bar x\\bar x^{\\mathsf T}.\\]\n\nCentering changes the matrix by rank at most one. It therefore does not change the limiting empirical spectral distribution. Replacing \\(1/n\\) by \\(1/(n-1)\\) also leaves the limiting law unchanged, although both choices affect finite-sample quantities."
    },
    {
      "heading": "Moments and transform characterization",
      "text": "The mean and variance of the Marchenko–Pastur distribution, viewed as a distribution over eigenvalues, are\n\n\\[\\int x\\,\\mu_{c,\\sigma^2}(dx)=\\sigma^2,\\]\n\nand\n\n\\[\\int(x-\\sigma^2)^2\\,\\mu_{c,\\sigma^2}(dx) = c\\sigma^4.\\]\n\nThe second formula measures spectral dispersion; it is distinct from the variance \\(\\sigma^2\\) of the data entries.\n\nFor an integer \\(k\\geq1\\), its \\(k\\)-th moment is\n\n\\[\\int x^k\\,\\mu_{c,\\sigma^2}(dx) = \\sigma^{2k} \\sum_{j=1}^{k} \\frac1k \\binom{k}{j}\\binom{k}{j-1}c^{j-1}.\\]\n\nIn particular,\n\n\\[m_1=\\sigma^2,\\qquad m_2=\\sigma^4(1+c),\\qquad m_3=\\sigma^6(1+3c+c^2).\\]\n\nThese are moments of the limiting measure. Convergence of high-order empirical moments requires stronger assumptions than weak convergence alone.\n\nAnother characterization uses the Stieltjes transform\n\n\\[m(z)=\\int\\frac{1}{x-z}\\,\\mu_{c,\\sigma^2}(dx), \\qquad \\operatorname{Im}z>0.\\]\n\nIt satisfies\n\n\\[c\\sigma^2z\\,m(z)^2 + \\bigl[z+\\sigma^2(c-1)\\bigr]m(z) +1=0.\\]\n\nThe relevant solution has positive imaginary part in the upper half-plane and satisfies \\(m(z)\\sim-1/z\\) at infinity. The density can be recovered from its boundary values:\n\n\\[f(x)=\\frac1\\pi \\lim_{\\eta\\downarrow0}\\operatorname{Im}m(x+i\\eta)\\]\n\nat interior points of the continuous spectrum."
    },
    {
      "heading": "Why the law appears",
      "text": "Two principal approaches explain the limiting distribution.\n\nThe moment method\n\nThe normalized trace\n\n\\[\\frac1p\\operatorname{tr}(S_n^k) = \\frac1p\\sum_{i=1}^{p}\\lambda_i^k\\]\n\nis the \\(k\\)-th empirical spectral moment. Expanding the trace gives sums of products of entries of \\(X\\).\n\nIndependence and zero means eliminate many terms. Among those remaining, certain index patterns dominate as \\(p\\) and \\(n\\) increase proportionally. Their counting yields the moment formula above. In treatments that begin with sufficiently strong moment assumptions, concentration arguments establish convergence of the empirical moments; truncation arguments help extend the distributional result to weaker assumptions.\n\nResolvent methods\n\nThe resolvent is the matrix\n\n\\[R(z)=(S_n-zI_p)^{-1}.\\]\n\nIts normalized trace is the empirical Stieltjes transform:\n\n\\[m_n(z)=\\frac1p\\operatorname{tr}R(z).\\]\n\nResolvent identities and concentration estimates lead to a limiting self-consistency equation. In the isotropic independent-entry case, this becomes the quadratic equation for \\(m(z)\\). Resolvent methods also extend naturally to nonidentity population covariance matrices."
    },
    {
      "heading": "Applications and extensions",
      "text": "Principal component analysis\n\nIn principal component analysis, sample covariance eigenvalues quantify the variance associated with principal components. The Marchenko–Pastur law provides a benchmark for the bulk spectrum generated by high-dimensional isotropic noise.\n\nThe upper edge \\(b\\) is a limiting bulk boundary, not an exact finite-sample significance threshold. In particular, a sample eigenvalue slightly above \\(b\\) does not by itself establish a non-noise component. Appropriate inference requires a specified model and finite-sample or fluctuation theory.\n\nSpiked covariance models\n\nA spiked covariance model introduces a small number of population eigenvalues different from the background value. Such finite-rank changes generally leave the limiting bulk distribution unchanged, but sufficiently strong spikes can produce isolated sample eigenvalues.\n\nFor unit background variance, under standard spiked-model assumptions, a population eigenvalue \\(\\ell>1+\\sqrt c\\) produces an upper outlier whose limit is\n\n\\[\\ell\\left(1+\\frac{c}{\\ell-1}\\right).\\]\n\nA spike below this threshold does not separate from the upper bulk edge. This illustrates why a genuine population component need not appear as a visibly isolated sample eigenvalue.\n\nRegularization and numerical conditioning\n\nWhen \\(c\\) approaches \\(1\\), the lower spectral edge approaches zero. Under conditions ensuring convergence of the extreme eigenvalues, this corresponds to poor conditioning of sample covariance matrices. For \\(c>1\\), these matrices are necessarily singular.\n\nThe law helps analyze spectral regularization, including the effect of replacing \\(S_n\\) by \\(S_n+\\alpha I_p\\). It also appears in asymptotic calculations for ridge regression, estimation risk, and linear inverse problems.\n\nGaussian and non-Gaussian data\n\nFor Gaussian entries, the unnormalized matrix \\(XX^{\\mathsf T}\\) has a Wishart distribution. The Marchenko–Pastur law gives its proportional-dimensional spectral limit after normalization.\n\nGaussianity is not essential. The same limiting law holds for a wide class of independent-entry distributions. This robustness is an example of universality, although bulk convergence, edge convergence, and fluctuation laws require different assumptions."
    },
    {
      "heading": "Scope and limitations",
      "text": "The standard law assumes an isotropic population and suitable independence. If the population covariance has a nontrivial limiting spectrum, the sample spectrum generally follows a deformed Marchenko–Pastur law rather than the density given above.\n\nDependence between observations or coordinates, unequal entry variances, and sparse sampling can also change the limit. If the entries have infinite variance, the usual normalization and limiting distribution need not apply.\n\nBulk convergence does not control individual extremes. A vanishing fraction of very large eigenvalues can leave the empirical spectral limit unchanged. Under stronger assumptions, such as a finite fourth moment in the standard independent-entry model, extreme-eigenvalue results establish convergence of the largest eigenvalue to \\(b\\), and, for \\(c<1\\), of the smallest eigenvalue to \\(a\\).\n\nFinally, the law is asymptotic. Finite matrices exhibit deviations from the limiting density and fluctuations near the edges. Those questions require additional results beyond the Marchenko–Pastur theorem itself."
    }
  ],
  "links": [
    "random-matrix-theory",
    "eigenvalues-and-eigenvectors",
    "covariance-matrix",
    "matrix",
    "expected-value",
    "variance",
    "conjugate-transpose",
    "almost-surely",
    "probability-distribution",
    "indicator-function",
    "probability-density-function",
    "matrix-rank",
    "gram-matrix",
    "identity-matrix",
    "sample-mean",
    "matrix-trace",
    "principal-component-analysis",
    "regularization",
    "ridge-regression"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}