Many large datasets are panels: many series recorded over the same long stretch of time. Examples are hundreds of brain regions in resting-state fMRI, or hundreds of stocks. Two things are often true of such data at once. First, each series has long memory: correlations between distant time points fade as a slow power law rather than exponentially. Second, the quantity we actually analyze is a nonlinear function of some hidden, better-behaved signal. Absolute or squared returns are functions of a hidden volatility; signal envelopes and thresholded activations are functions of a hidden activity level. Random matrix theory is the standard tool for interpreting the eigenvalues of a large sample covariance matrix, and it is how practitioners separate real structure from noise when the number of series $p$ is comparable to the number of time points $n$. Its standard formulas assume independent time points. Long memory breaks that assumption. Nonlinearity breaks it in a further way that, as far as this file shows, no one has worked out.
This note proposes to derive and test the eigenvalue distribution of $\frac1n YY^\top$ when $Y_{i,t}=G(X_{i,t})$. Here $X$ is a Gaussian panel with long memory in time and correlation across series, and $G$ is a nonlinearity that need not be invertible. The core idea is to expand $G$ in Hermite polynomials, which is the Wiener chaos expansion: an orthogonal "Fourier series" for functions of Gaussian variables. The expansion splits the panel into uncorrelated layers, one per polynomial order $k$. Layer $k$ has its own memory, which is long whenever $k\alpha<1$, and its own pattern of cross-series correlation. The eigenvalue spectrum is then governed by a coupled equation with one term per layer. The predictions are concrete. The first-order spectrum should be computable from a handful of Hermite coefficients and one memory exponent. The power-law tail of the spectrum should reveal the memory of the lowest long-memory layer. The few largest eigenvalues, which come from the lowest frequencies, are where the "keep only the leading term" simplification should fail.
Core Claim
Standardize $X$ so that each entry has unit variance, with $\mathrm{Cov}(X_{i,s},X_{j,t})=\Sigma_{ij}\,\rho(t-s)$ and $\rho(h)\sim |h|^{-\alpha}$, $\alpha\in(0,1)$. Write $G=\sum_{k\ge q}\frac{c_k}{k!}H_k$, where $H_k$ is the $k$-th Hermite polynomial and $q$ is the Hermite rank (the order of the first nonzero coefficient). Hermite polynomials of different orders are uncorrelated under Gaussian inputs, which gives an exact identity:
$$\mathrm{Cov}(Y_{i,s},Y_{j,t})=\sum_{k\ge q}\frac{c_k^2}{k!}\,\Sigma_{ij}^{\,k}\,\rho(t-s)^{k}.$$
The covariance of the whole panel is therefore a sum of Kronecker products $\sum_k A_k\otimes B_k$. In each term, $A_k=\frac{c_k^2}{k!}\Sigma^{\circ k}$ is the cross-series correlation matrix raised entrywise to the power $k$, and $B_k$ is the Toeplitz matrix built from $\rho(h)^k$. A Toeplitz matrix is one whose entries depend only on the time gap, which is what a stationary series produces.
For the linear case $G(x)=x$, the covariance is a single separable product $\Sigma\otimes T$. A nonlinear $G$ applied to cross-correlated long-memory data destroys that separability. Layer $k$ carries long memory only if $k\alpha<1$. Its cross-series structure is the $k$-th entrywise power of $\Sigma$, which flattens factor structure as $k$ grows.
We claim three things. The first is a deterministic equivalent for the first-order spectrum. A deterministic equivalent is a non-random function that the empirical Stieltjes transform $m_n(z)=\frac1p\mathrm{tr}(\frac1nYY^\top-zI)^{-1}$ converges to; the Stieltjes transform is the standard generating function from which the eigenvalue density is read off. We expect the limit to have the standard form for a covariance built from a sum of separable pieces:
$$\bar Q(z)=\Big(-z\big(I_p+\textstyle\sum_k \tilde e_k(z)A_k\big)\Big)^{-1},\quad \tilde e_k=\tfrac1n\mathrm{tr}\,B_k\Big(-z\big(I_n+\textstyle\sum_j e_j(z)B_j\big)\Big)^{-1},\quad e_k=\tfrac1n\mathrm{tr}\,A_k\bar Q(z).$$
The limit is $m(z)=\frac1p\mathrm{tr}\bar Q(z)$. This form is our conjecture to be proved, and the fact that it matches the true nonlinear matrix is a "Gaussian equivalence" claim, not something given. The technical obstacle is specific to long memory. The largest eigenvalues of $B_k$ grow like $n^{1-k\alpha}$ when $k\alpha<1$, because they are partial sums of $\rho^k$. Standard resolvent arguments assume bounded operator norms, so they do not apply directly.
The second claim concerns the spectral tail. The temporal symbol of $B_k$ behaves like $|\lambda|^{k\alpha-1}$ near frequency zero. Its eigenvalue distribution therefore has a power-law tail with exponent $1/(1-k\alpha)$. The heaviest such tail belongs to the smallest $k$ with $c_k\neq0$. We predict that this tail carries through to the bulk of the sample spectrum. If so, the spectral tail of a single panel estimates $q\alpha$, which per-series estimators cannot separate from $d_Y$.
The third claim is that the reduction principle fails at the edge. The handful of top eigenvalues come from the lowest temporal frequencies. These are the same coarse-scale coefficients for which High order chaotic limits of wavelet scalograms under long-range dependence proved that quadratic functionals escape the leading Hermite term. That paper finds a phase transition between Gaussian, Hermite-of-order-$(q-1)$ and Rosenblatt limits, depending on sample size relative to scale. We predict that the locations of these eigenvalues still match the Gaussian surrogate (a Gaussian matrix with the same covariance). Their fluctuations should not match it, because they depend on fourth-order structure of the low-chaos layers that a Gaussian surrogate lacks.
This sharpens the tension raised by Risk Convergence of Centered Kernel Ridge Regression With Large Dimensional Data. That paper linearizes inner-product kernels by Taylor expansion under independent data and finds that once centered, the choice of kernel stops mattering. Its reported minimum risk of 0.3627 on Communities and Crime is the same across linear, polynomial, sigmoid and exponential kernels. Our prediction is the opposite under long memory. The nonlinearity stays relevant because its quadratic and cubic layers keep their own long memory whenever $k\alpha<1$. On the fMRI data described below, $\alpha\approx0.30$, so layers 1 through 3 are all long-memory and layer 4 is not.
Relation to Prior Work
Prior work falls into four groups by the role it plays here.
The direct predecessors are the univariate chaos papers of Clausel, Roueff, Taqqu and Tudor. Large scale behavior of wavelet coefficients of non-linear subordinated processes with long memory establishes how wavelet coefficients of $G(X_t)$ behave at large scales. Estimation du paramètre de longue mémoire de séries temporelles non linéaires shows that the nonlinearity changes the limits of memory estimators. Large scale reduction principle and application to hypothesis testing studies when partial sums reduce to the leading Hermite term. The scalogram paper shows that this reduction fails for quadratic statistics. All of them treat a single series and no matrix spectra. We reuse their chaos calculus and their scalogram phase transition as the predicted mechanism for edge eigenvalues. The new questions are what the cross-series entrywise powers $\Sigma^{\circ k}$ do, and what happens at proportional $p/n$.
The supply of resolvent methods comes from Hessian Eigenspectra of More Realistic Nonlinear Models and the kernel ridge paper. Both derive Stieltjes-transform deterministic equivalents for nonlinear models, and both assume independent samples. Consistent Estimation of the High-Dimensional Efficient Frontier applies the same tools to finance. It corrects the minimum-variance portfolio variance by $(1-c)^{-1}$ and tests on short-memory CCC-GARCH data and S&P 500 returns. We change the data assumption, not the toolbox.
Several papers motivate the problem. Estimation of large covariance and precision matrices from temporally dependent observations and Large covariance/correlation matrix estimation for long memory temporal data document long memory in a real panel: HCP fMRI with $n=1190$, $p=907$, and autocorrelation decay $\alpha\approx0.30$. They also show that the sample mean converges only at rate $n^{-\alpha/2}$. Their methods are entrywise thresholding and CLIME, which rely on sparsity and do not study spectra. On the finance side, Estimating Long Memory in Volatility shows that volatility proxies are noisy nonlinear functions of a latent long-memory Gaussian process. That is exactly our setting with $G(x)=|x|$-type maps.
Two results constrain the design. Djah et al. show that diagrammatic and orthogonal-polynomial decompositions diverge outside Gaussian measure, so we restrict to Gaussian subordination; long-memory stochastic volatility still qualifies, since its extra noise is independent. Samorodnitsky et al. show that bounded nonlinearities cannot remove long memory. This is consistent with our layer-by-layer accounting.
Research Program
The theory has four parts.
The first part proves the deterministic equivalent. We would split each $B_k$ into a bounded part plus a slowly growing number of low-frequency directions. The bounded part would be handled by standard resolvent identities, extended so that each row is a nonlinear Gaussian functional rather than a Gaussian vector. Concentration of the needed quadratic forms should follow from Gaussian-functional variance bounds; the scalogram paper's chaos calculus supplies the fourth-cumulant diagram sums. The low-frequency directions would be treated by perturbation as outliers.
The second part derives the tail exponent. It also derives the effect of centering. Subtracting the sample mean removes approximately the zero-frequency direction, which is the largest temporal direction. Under long memory, this rank-one change should visibly remove an outlier rather than vanish, as it does in the independent case. This is the open question flagged by Shu and Nan's slow sample-mean rate.
The third part characterizes the fluctuations of the top eigenvalues. For independent rows, we expect a Gaussian central limit across series, but with a variance set by the kurtosis of the Hermite-process limits rather than the Gaussian value. For factor-driven rows, we expect possibly non-Gaussian limits.
The fourth part turns the theory into estimators. The first is a Stieltjes-fit estimator. It fits $\alpha$ and the first few $c_k^2$ by matching the empirical $m_n(z)$ to $m(z)$ on a grid of $z$ in the upper half-plane, where both quantities are smooth and stable. This recovers the memory of the latent signal even when $G$ cannot be inverted. Two things identify the layers: their different temporal symbols, and their different cross-series structures $\Sigma^{\circ k}$. Neither is available from a single series. The second estimator is an eigenvalue shrinker in the style of Optimal shrinkage of eigenvalues in the spiked covariance model. It is recomputed under the new bulk law, because the Marchenko–Pastur edges it assumes no longer exist when the bulk has a power-law tail.
Baselines
The naive reference point is the Marchenko–Pastur law with independent columns: predicted bulk edges $[(1-\sqrt c)^2,(1+\sqrt c)^2]$, plus the $(1-c)^{-1}$ correction from the efficient-frontier paper. It is what practitioners actually use. It will be run on identical matrices, with no tuning. The proposal beats it if its predicted spectrum is closer to the observed one, measured by Wasserstein distance between predicted and observed eigenvalue distributions and by the error in $m(z)$. The comparison must favor the proposal at every tested memory level $d\ge0.25$. The efficient-frontier paper reports only qualitative frontier shifts, so there is no reported result on Lacuna for spectral-fit error or a numerical GMV correction.
The strongest competitor is a linear long-memory model fitted directly to $Y$. In this approach $Y$ is treated as $\Sigma_Y\otimes T_Y$. $T_Y$ is estimated from the pooled periodogram with a local Whittle exponent as in Hurvich et al., and the separable deterministic equivalent is plugged in. It ignores the chaos structure but captures the observed memory. When $\Sigma=I$, Gaussian equivalence implies that this model is exactly right for the bulk. It will receive the same data and the same grid-search budget over its parameters as the Stieltjes-fit. Beating it means a lower spectral-fit error when $\Sigma$ has factor structure and $G$ is non-monotone, and better calibrated predictions for the top eigenvalues.
The Gaussianization baseline applies a rank-based normal-score transform to each series and then fits the linear long-memory model. The sparse baseline is generalized thresholding and CLIME as tuned by Shu and Nan's gap-block cross-validation. Their HCP analysis reports 1.47% marginal and 2.71% direct connections, but these are descriptive rather than accuracy figures. For memory estimation, the per-series baseline is wavelet or local Whittle estimates of $d_Y$, averaged across series. The same estimates are then mapped to $d_X$ using the true $q$ in an oracle version and an estimated $q$ in a realistic version. The Hurvich et al. table values are not given in the file, so we report no number for this baseline.
The central ablation truncates the chaos sum at $K=1,2,3$ and at full order. $K=1$ is the reduction-principle model that keeps only the leading Hermite term. A second ablation replaces the true nonlinear $Y$ with its Gaussian surrogate, which has the same covariance. If the bulk matches but the fluctuations of the top eigenvalues do not, that isolates the edge claim.
Why not just…?
Why not just fit a separable long-memory model to the transformed data? This is the real threat. With independent series, we expect it to match our bulk prediction exactly, and we concede that case in advance. Its weak point is the cross-series pattern. For non-monotone $G$ such as $|x|$ or $H_2$, the true covariance mixes $\Sigma$, $\Sigma^{\circ 2}$ and $\Sigma^{\circ 3}$, each paired with a different memory. A single Kronecker product can fit neither the spectral tail nor the factor eigenvalues. The ablation over factor strength and $G$ measures the size of this misfit. If the misfit falls within Monte Carlo error, we drop the bulk claim and keep only the edge and estimation claims.
Why not just Gaussianize each series with a rank transform? For monotone $G$ this recovers $X$ up to marginal estimation error and should match or beat our method, so we treat monotone $G$ as a control rather than a target. Squaring, absolute value and two-sided thresholding cannot be inverted from the marginals. These are the maps that matter for volatility proxies, and our main claims are restricted to them.
Why not just pool per-series wavelet or local Whittle estimates? Pooling reduces variance, but every series reports $d_Y$, and converting to $d_X$ requires the Hermite rank $q$. The file notes that $q$ is fragile in small samples, and the scalogram paper shows that wavelet statistics can converge to non-leading chaos limits. The panel adds information a single series lacks, namely the entrywise powers $\Sigma^{\circ k}$. The comparison against the oracle-$q$ and estimated-$q$ pooled baselines tests whether that information helps.
Why not just threshold? Thresholding needs sparsity. Factor-driven stock volatility and spiked models are dense, which is why the efficient-frontier paper and the spiked-shrinkage paper use spectral corrections. We include thresholding to show where sparsity holds (fMRI) and where it fails (stocks).
Proposed Experiments
The synthetic study generates $X$ with fractionally integrated temporal dependence, $d\in\{0.1,0.25,0.35,0.45\}$, and $n$ up to 8192. Cross-series correlation $\Sigma$ is either the identity or a model with one to five factors. The nonlinearities are $G\in\{x,\ H_2,\ |x|,\ \mathbf 1\{x>u\},\ H_1+H_q\}$; the last is the Clausel benchmark family. The aspect ratio is $c\in\{0.25,0.5,1,2\}$. We measure four things: the Stieltjes-transform error on a complex grid, the Wasserstein distance between predicted and observed eigenvalue distributions, the fitted tail exponent compared with $1/(1-q\alpha)$, and the top-eigenvalue fluctuations compared with the Gaussian surrogate.
Two checks are decisive. The first is whether the tail exponent tracks $q\alpha$ rather than $\alpha$ for $|x|$, which has rank 2. The second is whether the variance of the top eigenvalues departs from the Gaussian surrogate in the regime where $q\alpha<1$, and agrees with it when $q\alpha>1$.
The first real-data test is semi-synthetic, using HCP resting-state fMRI as described by Shu and Nan ($n=1190$, $p=907$, $c\approx0.76$, $\alpha\approx0.30$). Real BOLD series serve as the latent $X$, and controlled maps $G$ are applied: amplitude, square and threshold. Ground truth for $G$ is known while the memory and cross-series structure are real. The second real test uses the absolute intraday returns of the 200 S&P 500 stocks from 2017 to 2022 in the efficient-frontier paper, a natural non-invertible map of latent volatility. On held-out time blocks separated by a gap, as in Shu and Nan's validation scheme, we score covariance shrinkage by out-of-sample Frobenius and Stein loss. We also check that the fitted latent memory is consistent with the univariate $H\approx0.6$ reported for the 1990s index in Estimating the Fractal Dimension of the S&P 500 Index Using Wavelet Analysis, acknowledging that the period and asset set differ.
The project fails cleanly in either of two cases. One is that the separable fit to $Y$ matches the chaos model under factor structure across all nonlinearities. The other is that the top-eigenvalue fluctuations match the Gaussian surrogate. Either outcome would tell us which of the three claims survives.
References
- High order chaotic limits of wavelet scalograms under long--range dependence
- Risk Convergence of Centered Kernel Ridge Regression with Large Dimensional Data
- Large scale behavior of wavelet coefficients of non-linear subordinated processes with long memory
- Estimation du paramètre de longue mémoire de séries temporelles non linéaires
- Large scale reduction principle and application to hypothesis testing
- Hessian Eigenspectra of More Realistic Nonlinear Models
- Consistent estimation of the high-dimensional efficient frontier
- Estimation of large covariance and precision matrices from temporally dependent observations
- Large covariance/correlation matrix estimation for long memory temporal data
- Estimating Long Memory in Volatility
- Feynman graph representation of the perturbation series for general functional measures
- Function-indexed empirical processes based on an infinite source Poisson transmission stream
- Optimal shrinkage of eigenvalues in the spiked covariance model
- ESTIMATING THE FRACTAL DIMENSION OF THE S&P 500 INDEX USING WAVELET ANALYSIS