Covariance eigenvalue clipping: Definition & Examples

Covariance eigenvalue clipping is a spectral denoising method that replaces a group of sample eigenvalues judged to be noise with one common value, while retaining the other eigenvalues and the sample eigenvectors. In the trace-preserving version, that common value is the average of the noise group. This can prevent a portfolio optimiser from treating noisy small eigenvalues as exceptionally cheap risk directions.

The name also appears in numerical matrix repair, where it can mean imposing an eigenvalue floor. Here it means noise-bulk averaging; the two operations solve different problems.

Flatten the noise group, keep the directions

Write a sample covariance or correlation matrix using its eigenvalue decomposition:

S=U\operatorname{diag}(\lambda_1,\ldots,\lambda_p)U^\top.

Let $\mathcal N$ be a nonempty set of indices classified as noise. Set

\bar\lambda_{\mathcal N}=\frac{1}{|\mathcal N|}\sum_{i\in\mathcal N}\lambda_i, \qquad d_i=\begin{cases} \bar\lambda_{\mathcal N},& i\in\mathcal N,\\ \lambda_i,& i\notin\mathcal N, \end{cases} \qquad \widetilde S=U\operatorname{diag}(d_1,\ldots,d_p)U^\top.

Because the replacement group's sum is unchanged,

\operatorname{tr}(\widetilde S)=\operatorname{tr}(S)

. Individual diagonal entries need not survive. The construction retains an eigenbasis

U

, although after several eigenvalues become equal, the basis within that eigenspace is no longer unique. Bun, Bouchaud and Potters describe this rule as recipe 3, equation (5).

Three assets expose the diagonal problem

Consider the following stipulated correlation matrix, with dimensionless entries:

C=\begin{pmatrix} 1&0.8&0\\ 0.8&1&0\\ 0&0&1 \end{pmatrix}.

Its eigenvalues are $0.2$ , $1$ and $1.8$ . The corresponding unit-length directions are the A–B spread $(1,-1,0)/\sqrt2$ , asset C alone $(0,0,1)$ , and the A+B common direction $(1,1,0)/\sqrt2$ .

For the example only, designate the two lower eigenvalues as noise. No observations, sample-size ratio or fitted random-matrix threshold establish that classification. Their average is

(0.2+1)/2=0.6

. Keeping

1.8

gives

\widetilde C=\begin{pmatrix} 1.2&0.6&0\\ 0.6&1.2&0\\ 0&0&0.6 \end{pmatrix}.

The trace remains

3

, and the matrix is positive definite, but its diagonal is

(1.2,1.2,0.6)

. It is therefore not yet a correlation matrix. Reading its

0.6

A–B entry as a correlation would mix up covariance with correlation.

To restore unit diagonal, let $D=\operatorname{diag}(\widetilde C_{11},\widetilde C_{22},\widetilde C_{33})$ and calculate

C_{\mathrm{clean}}=D^{-1/2}\widetilde C D^{-1/2} =\begin{pmatrix} 1&0.5&0\\ 0.5&1&0\\ 0&0&1 \end{pmatrix}.

In particular,

0.6/\sqrt{1.2\times1.2}=0.5

. This diagonal rescaling preserves positive semidefiniteness when all diagonal entries are positive. It is not an orthogonal similarity transformation: the final eigenvalues are

0.5

1

and

1.5

, not

0.6

0.6

and

1.8

. The original eigenvectors happen to remain eigenvectors in this example; that need not hold after general diagonal rescaling.

Stipulated three-asset example. Sample eigenvalues 0.2, 1 and 1.8 become 0.6, 0.6 and 1.8 after averaging two supplied noise modes. Restoring unit diagonal gives eigenvalues 0.5, 1 and 1.5. The A–B entry moves from 0.8 to 0.6 to 0.5; every stage has trace 3.

Original stipulated example. Bulk averaging preserves trace but changes the diagonal; restoring unit diagonal changes the eigenvalues again. Noise membership is supplied, not inferred from data.

For a unit-length A–B spread exposure, the estimated directional variance goes from $0.2$ to $0.6$ before normalisation and then to $0.5$ . These are three different risk matrices. If cleaning correlations for assets with supplied volatility estimates, convert the final correlation matrix back to covariance using those volatilities; otherwise asset risk levels have been changed as well.

Where the noise labels come from

A common heuristic uses the Marchenko–Pastur law. Under a standard independent, unit-variance noise reference, the limiting upper spectral edge is

(1+\sqrt{p/n})^2

, with

p

variables and

n

observations under the reference's normalisation. One may retain eigenvalues above a selected edge and average the rest.

That edge is model dependent and asymptotic, not an exact finite-sample signal test. Serial dependence, estimated volatility scaling and a nonidentity population bulk can undermine a literal threshold. Genuine structure can remain inside the bulk, while retaining outliers unchanged leaves their estimation error intact. Bun, Bouchaud and Potters, §7.2.2, discuss these limitations. The three-asset calculation demonstrates the transformation only; it provides no evidence that clipping improves forecasting or portfolio performance.

Similar-looking operations with different targets

An eigenvalue floor uses $d_i=\max(\lambda_i,\varepsilon)$ . A floor of $0.6$ in the example produces $(0.6,1,1.8)$ and trace $3.4$ : it does not average the lower pair. For a real symmetric matrix, flooring at zero is the nearest positive semidefinite matrix in Frobenius norm, without a unit-diagonal constraint. Higham gives the relevant theorem. Our starting matrix is already positive definite, so that repair would leave it unchanged.
Nonlinear covariance shrinkage estimates eigenvalue-specific corrections from the spectrum rather than assigning every bulk direction one constant. Bulk clipping is itself a piecewise nonlinear spectral rule, but it is not the fitted nonlinear shrinkage estimator. Both can retain the sample eigenvectors; neither retention alone establishes optimality. The loss-dependent oracle framework appears in §§6.1–6.2.

For the portfolio consequence, Mean–Variance Optimisation derives inverse-covariance weights and explains their sensitivity to unstable inputs. Its public page is a preview; the full lesson requires membership.

Sources

Bun, J., Bouchaud, J.-P. and Potters, M. (2016), Cleaning correlation matrices, Risk, April 2016: “Five cleaning recipes,” equation (5) for clipping and equations (7)–(8) for the rotationally invariant alternative.
Bun, J., Bouchaud, J.-P. and Potters, M. (2016 preprint; published 2017), Cleaning large correlation matrices: tools from random matrix theory, §§6.1–6.2 and 7.2.2.
Higham, N. J. (2021), What Is the Nearest Positive Semidefinite Matrix?, the Frobenius-norm theorem and its eigenvalue-floor construction.

Flatten the noise group, keep the directions

Three assets expose the diagonal problem

Where the noise labels come from

Similar-looking operations with different targets

Sources

Related terms

Eigenvalue decomposition

Marchenko–Pastur law

Nonlinear covariance shrinkage

Sample covariance