Nonlinear Eigenvalue Shrinkage Under Frobenius Loss

Why Change Eigenvalues but Keep Eigenvectors?

In high dimensions, sample covariance eigenvalues can be substantially distorted even when the sample eigenvectors provide a useful coordinate system. A rotation-equivariant estimator respects that coordinate system: it transforms in the same way as the sample covariance when the variables are rotated. Under squared Frobenius loss, the key question is then how to choose an estimated variance along each sample eigenvector.

The answer starts with an oracle calculation. If the population covariance were known, we could find the best estimator among matrices diagonal in the sample eigenbasis. This oracle is not generally available as an estimator, because it uses the unknown population covariance. It does, however, identify the target eigenvalues that data-driven nonlinear shrinkage methods seek to estimate.

For a matrix function to be rotation-equivariant, it must satisfy \(T(QSQ^T)=QT(S)Q^T\) for every orthogonal \(Q\). When \(S\) has distinct eigenvalues, this condition forces \(T(S)\) to be diagonal in the eigenbasis of \(S\). To see why, consider changing the sign of one eigenvector: this leaves \(S\) unchanged, so equivariance requires the output to be unchanged too. Any off-diagonal entry involving that eigenvector would change sign and must therefore be zero. If an eigenvalue is repeated, rotations within its eigenspace leave \(S\) unchanged; equivariance requires the output to be a scalar multiple of the identity within that eigenspace. Thus spectral estimators retain the sample eigenspaces, while choosing new eigenvalues within them.

The Frobenius-Loss Oracle

Write the population covariance in sample coordinates as \(A=U^T\Sigma U\). Orthogonal invariance of the Frobenius norm gives \(\|\hat{\Sigma}-\Sigma\|_F^2=\|\operatorname{diag}(d_1,\ldots,d_p)-A\|_F^2\). The diagonal and off-diagonal entries contribute separately. In particular, the choice of \(d_i\) affects only the squared difference between \(d_i\) and \(A_{ii}\).

Proof. Put \(A=U^T\Sigma U\). The Frobenius norm is unchanged by multiplication on either side by an orthogonal matrix, so the loss equals \(\|\operatorname{diag}(d_1,\ldots,d_p)-A\|_F^2\). Expanding it entry by entry gives \[ \sum_{i=1}^p(d_i-A_{ii})^2+\sum_{i\ne j}A_{ij}^2. \] The second sum does not depend on any \(d_i\). Each term in the first sum is uniquely minimized at \(d_i=A_{ii}=u_i^T\Sigma u_i\). If \(\Sigma\) is positive semidefinite, then \(u_i^T\Sigma u_i\geq 0\), as required. \(\square\)

The oracle eigenvalue is the population variance in the direction of a sample eigenvector; it is not, in general, the population eigenvalue matched by rank. The sample eigenvalue \(\lambda_i\) does not appear in the pointwise minimization. It matters in practice because an estimator must infer the oracle target from the observed spectrum and its relation to the sample eigenvectors.

Identity Population Covariance: The Limiting Correction

Now suppose the observations are independent \(N(0,I_p)\) vectors and \(S=XX^T/n\), where \(X\) is a \(p\times n\) matrix of independent standard normal entries. Let \(p/n\to c\), with \(0<c<1\). The population covariance \(I_p\) is positive definite. Since the Gaussian data matrix has full row rank with probability one when \(p<n\), there are no rank-induced zero sample eigenvalues in this setting. This qualification matters: for a singular population covariance, zero sample eigenvalues can occur even when \(p/n<1\). In a general limiting model, keeping the population spectrum bounded away from zero is an additional condition if one wants the limiting sample spectrum bounded away from zero.

For this identity model, every unit vector \(u_i\) satisfies \(u_i^TI_pu_i=1\). Thus the oracle eigenvalue is exactly one for every sample eigenvector, at every finite dimension. The sample eigenvalues themselves do not concentrate at one: their limiting empirical distribution is the Marchenko–Pastur distribution. The following calculation shows explicitly how the nonlinear spectral correction recovers the oracle value throughout the limiting bulk.

Proof. The boundary value of the Stieltjes transform for the Marchenko–Pastur density, with the stated sign convention, is \[ m(x+i0)=\frac{1-c-x+i\sqrt{(b-x)(x-a)}}{2cx},\qquad a<x<b. \] Substitution gives \[ 1-c-cx\,m(x+i0) =\frac{1-c+x-i\sqrt{(b-x)(x-a)}}{2}. \] Therefore \[ |1-c-cx\,m(x+i0)|^2 =\frac{(1-c+x)^2+(b-x)(x-a)}{4}. \] To simplify the numerator, use \(a+b=2(1+c)\) and \(ab=(1-c)^2\). Expanding the product yields \((b-x)(x-a)=-x^2+(a+b)x-ab\). Adding \((1-c+x)^2=x^2+2(1-c)x+(1-c)^2\) cancels the \(x^2\) and constant terms. The coefficient of \(x\) is \(2(1-c)+(a+b)=2(1-c)+2(1+c)=4\). Hence the sum is \(4x\), so the squared modulus is \(x\) and \(\delta(x)=x/x=1\). This agrees with the exact finite-sample oracle \(u_i^TI_pu_i=1\). \(\square\)

Interpretation, Limits, and a Common Pitfall

The identity calculation is a transparent benchmark for nonlinear shrinkage. The oracle uses the sample eigenvectors but assigns the same population variance to every direction. In more general covariance models, the oracle values \(u_i^T\Sigma u_i\) need not be equal, and their limiting relationship to sample eigenvalues depends on the population spectrum and the eigenvector–eigenvalue relationship. The identity result should not be read as a universal rule to replace all sample eigenvalues by one: it is specific to identity population covariance and Frobenius loss.

A common mistake is to assume that the eigenvalue with rank \(i\) in the sample must estimate the population eigenvalue with the same rank. The Frobenius oracle calculation shows why this need not be true: it targets the population variance in the sample eigenvector direction, not a rank-matched population eigenvalue. Another mistake is to claim that \(c<1\) alone rules out zero sample eigenvalues for every population covariance. It rules out rank-induced zeros when the population covariance is positive definite and the data matrix has full row rank; a singular population covariance can still produce zeros. The limiting correction derived here uses a nonsingular identity population and interior points of its limiting spectral support.