The BBP Eigenvalue Phase Transition

From a Population Spike to a Sample Outlier

In high-dimensional principal component analysis, a population covariance matrix can contain a direction of elevated variance without its corresponding sample eigenvalue necessarily standing apart from the noise spectrum. The BBP phase transition describes the threshold separating these behaviors. Its location depends on both the strength of the population spike and the limiting ratio of dimension to sample size.

We study a Gaussian rank-one model. Let \(p,n\to\infty\) with \(p/n\to\gamma\), where \(0<\gamma<\infty\). Let \(v\in\mathbb R^p\) be a deterministic unit vector, and let the population covariance have one eigenvalue \(\alpha>1\) and all remaining eigenvalues equal to one. The sample covariance is formed from \(n\) independent observations with this covariance and known mean zero.

Write the data matrix as \(Y=\Sigma^{1/2}Z\), where \(Z\) is \(p\times n\) with independent standard normal entries. The sample covariance is \(S=YY^\mathsf{T}/n\). Without a spike, the empirical eigenvalue distribution converges to the Marchenko–Pastur law. Its upper edge is

The question is whether the largest eigenvalue of \(S\) converges to this edge or separates above it.

A Secular Equation for Eigenvalues Above the Bulk

Set \(W=ZZ^\mathsf{T}/n\) and \(A=\Sigma^{1/2}\), so that \(S=AWA\). For any \(\lambda\) outside the spectrum of \(W\), a determinant identity reduces the eigenvalue problem to a scalar equation.

To verify the identity, note that \(A^{-2}=I_p+(\alpha^{-1}-1)vv^\mathsf{T}\). Multiplication on both sides by \(A^{-1}\) gives \(\det(\lambda I_p-S)=\det(A)^2\det(\lambda A^{-2}-W)\). Applying the matrix determinant lemma to the rank-one change from \(\lambda I_p-W\) yields

When \(\lambda\) is outside the spectrum of \(W\), the first two determinant factors are nonzero, so the bracket must vanish. Rearranging gives the stated equation. This finite-sample reduction is useful because it isolates the effect of the spike in one quadratic form.

For a fixed \(\lambda>b_\gamma\), the Gaussian matrix \(W\) is orthogonally invariant. Conditional on its eigenvalues, its eigenvectors are uniformly oriented. Consequently the quadratic form in the equation converges to the normalized trace of the resolvent: \(v^\mathsf{T}(\lambda I_p-W)^{-1}v\to m_\gamma(\lambda)\), where \(m_\gamma(\lambda)=\int(\lambda-x)^{-1}\,dF_\gamma(x)\) and \(F_\gamma\) is the Marchenko–Pastur limit. In this region the resolvent is bounded; conditional quadratic-form fluctuations vanish as \(p\to\infty\), and the Marchenko–Pastur limit gives convergence of the normalized trace.

The Threshold and the Outlier Location

The Marchenko–Pastur transform on the real line above the upper edge satisfies \(\lambda=1/m_\gamma(\lambda)+1/(1-\gamma m_\gamma(\lambda))\). The physical solution is the one that tends to zero as \(\lambda\to\infty\). Equivalently,

This function decreases from its edge value \(1+1/\sqrt{\gamma}\) to \(1\) as \(\lambda\) increases from \(b_\gamma\) to infinity. The limiting secular equation requires it to equal \(\alpha/(\alpha-1)=1+1/(\alpha-1)\). Thus there is a solution strictly above the edge exactly when \(1+1/(\alpha-1)<1+1/\sqrt{\gamma}\).

For a supercritical spike, solve the limiting equation explicitly. Put \(y=\lambda m_\gamma(\lambda)\). The transform identity gives \(y=1+m_\gamma(\lambda)/(1-\gamma m_\gamma(\lambda))\), so \(m_\gamma(\lambda)=(y-1)/(1+\gamma(y-1))\). At an outlier, \(y=\alpha/(\alpha-1)\). Substitution then gives

The threshold follows from the edge value of the transform: the required level \(y=\alpha/(\alpha-1)\) lies strictly between \(1\) and \(1+1/\sqrt{\gamma}\) precisely when \(\alpha>1+\sqrt{\gamma}\). At equality, the solution meets the bulk edge, not a point strictly beyond it.

These limiting equations locate a separated eigenvalue but do not, by themselves, prove that the largest eigenvalue stays at the edge when the spike is subcritical. For \(\lambda\) above the largest eigenvalue of \(W\), the finite-sample function \(\lambda v^\mathsf{T}(\lambda I_p-W)^{-1}v\) decreases with \(\lambda\): in an eigenbasis of \(W\), each positive-eigenvalue term is a nonnegative weight times \(\lambda/(\lambda-\lambda_j(W))\), which decreases. The secular equation therefore has no root a fixed distance above the edge when its limiting level is unattainable there. Meanwhile, \(S-W\) has rank at most two, so the spike does not change the limiting empirical eigenvalue distribution; eigenvalues from that bulk remain arbitrarily close to its upper edge. These facts give edge convergence in the subcritical and critical cases.

Worked Examples

Why the Phase Transition Matters

A population direction can be stronger than the background variance and still fail to produce a visually distinct sample eigenvalue. This distinction matters when using the spectrum to detect latent structure: an eigenvalue at the bulk edge is not, by itself, evidence that a population spike has generated an outlier. The threshold depends on \(\gamma\), so increasing dimension relative to sample size raises the spike strength needed for separation.

The result concerns eigenvalue separation. It does not describe the alignment of the corresponding sample eigenvector with the population direction; that is a separate question. In applications, eigenvalue detection and directional estimation should not be treated as interchangeable conclusions.

Common Mistake

• Using the outlier formula below threshold. Substituting any \(\alpha>1\) into \(\alpha(1+\gamma/(\alpha-1))\) does not guarantee a sample outlier. The formula describes the physical solution of the limiting secular equation only when \(\alpha>1+\sqrt{\gamma}\). Below that threshold, the largest eigenvalue converges to \(b_\gamma\).
• Using the reciprocal aspect ratio. Here \(\gamma=p/n\), not \(n/p\). Reversing the ratio changes both the bulk edge and the critical spike.
• Calling the critical case an outlier. At \(\alpha=1+\sqrt{\gamma}\), the limiting location equals the edge. Separation requires a strict inequality.