Statistical decision theory for the adaptive best-basis covariance estimator of locally stationary processes

Mallat, Papanicolaou and Zhang (Annals of Statistics, 1998) proposed a data-adaptive estimator for the covariance operator of a nonstationary 'locally stationary' process observed in few realizations: select a best orthonormal local-cosine basis from a dyadic binary-tree dictionary by maximizing the estimated band-limited covariance energy, keep only the diagonal (an approximate Karhunen–Loève) coefficients when the number of realizations is small, and smooth these over the time–frequency plane. The paper established approximation-theoretic results (operator-norm compressibility of locally stationary covariances in adapted local-cosine bases) and exact second-moment identities for sample covariance coefficients, but no consistency statement, convergence rates, or risk bounds for the estimator itself; no conditions under which the estimated best basis converges to the true best basis; and no data-driven, theoretically grounded rule for the truncation bandwidth $K$ or the time–frequency smoothing variance. The problem is to supply this missing statistical theory: consistency and operator-norm convergence rates over the general locally stationary class of the source paper, convergence conditions for the data-selected basis, and oracle-type guarantees (matching the performance of the best fixed tuning chosen with knowledge of the true covariance) for data-driven tuning of $K$ and the smoothing bandwidth. Later work solved closely related questions in reformulated settings (penalized macrotile model selection, cross-validated bandwidths in the Dahlhaus local-stationarity framework, high-dimensional convergence rates for non-adaptive time-varying covariance estimates), but the precise guarantees for this estimator and class remain open.

Background

A stationary time series has a covariance operator that is diagonalized by the Fourier basis, which is why classical spectral estimation and the periodogram work: the covariance is compressed to a diagonal of spectral values. For a nonstationary process whose second-order structure drifts with time, no fixed basis diagonalizes the covariance operator, and estimating the full matrix from few realizations suffers severe variance accumulation. Classical attempts to describe time-varying second-order structure go back to Priestley's evolutionary spectrum (1965); the modern statistical treatment of local stationarity was set by Dahlhaus (1997), who models processes that are locally close to stationary and estimates time-varying parameters with local windows.

Mallat, Papanicolaou and Zhang (1998) gave an operator-theoretic, basis-adaptive alternative. They define a zero-mean process to be locally stationary if there exists an orthonormal local-cosine basis — cosine atoms multiplied by smooth window functions of length $l_p$ centered at $a_p$, with the window lengths varying slowly across time — in which the covariance operator $T$ has matrix entries that decay rapidly away from the diagonal in both time and frequency; the decay is controlled by constants $Q_n$ indexed by a decay order $n$. Here the operator norm measures the largest singular value, while the Hilbert–Schmidt norm is the square root of the sum of squared matrix entries. Their key deterministic results state that such a covariance is well approximated in operator norm by a sparse band truncation $B_K$ keeping only coefficients within a time–frequency distance $K$ of the diagonal, with error decaying faster than any power of $K$, and that covariance operators with smooth time-varying symbols (pseudodifferential operators, i.e. time-varying filters of white noise) satisfy the definition. On the statistical side they derive exact second-moment identities for sample covariance coefficients in any orthonormal basis, which quantify the variance penalty for estimating many entries from few realizations. These ingredients motivate their algorithm: search a large dictionary of local-cosine bases built from admissible dyadic binary trees (depth limited by a minimal window ramp of size $\eta$) by maximizing the estimated band-limited Hilbert–Schmidt energy, a criterion computable by bottom-up dynamic programming, a recursive optimization over the tree, in $O(M\log_2 M)$ operations for a signal of length $M$; with $N$ realizations and $N$ small, keep only the diagonal coefficients (an approximate Karhunen–Loève estimator — the Karhunen–Loève basis being the eigenbasis of the covariance operator, so a diagonal in a well-adapted basis mimics it — guaranteed positive semidefinite); and reduce the remaining variance by smoothing the estimated diagonal energies over the time–frequency plane with factorized Gaussian kernels whose scales follow the tile geometry. Numerical experiments showed order-of-magnitude error reductions relative to naive sample covariance for small $N$.

What the 1998 paper did not supply is the statistical decision theory for this pipeline. Consistency of best-basis covariance estimation was available only for special models in the earlier best-basis literature (the Donoho–Mallat–von Sachs 1996 line, cited by the paper); the general locally stationary class, the convergence of the data-selected basis, the choice of $K$ and of the smoothing kernel variance, the effect of post-hoc window modifications, and the behavior for processes whose diagonal energies are not smooth in the time–frequency plane were all left open. Neighboring communities later built parallel theories: the wavelet-process school (Nason, von Sachs and co-workers) estimates time-varying wavelet spectra with pointwise adaptive bandwidths; the Dahlhaus school proves oracle-type bandwidth selection (a data-driven bandwidth performing nearly as well as the best fixed one, chosen with knowledge of the true process) by cross-validation for local M-estimators, i.e. maximum-likelihood-type estimators computed in moving time windows; high-dimensional statistics proves operator-norm-consistent covariance estimation for i.i.d. data. None of these analyzes the Mallat–Papanicolaou–Zhang estimator in its own generality.

Problem Statement

For the adaptive covariance-estimation pipeline of Mallat, Papanicolaou and Zhang (1998) — selection of a best local-cosine basis from a dyadic binary-tree dictionary by maximizing the estimated band-limited Hilbert–Schmidt covariance energy, followed by estimation of the (possibly diagonal, $K$-truncated) covariance coefficients in that basis and Gaussian smoothing of the diagonal energies over the time–frequency plane — establish the missing statistical decision theory over the general class of locally stationary processes defined by the paper (existence of an adapted local-cosine basis with slowly varying window lengths and off-diagonal decay of the covariance matrix entries):

(1) Consistency and rates: prove or disprove that, as the number of realizations $N$ and/or the signal length $M$ grow, the estimator converges to the true covariance operator in operator norm, and derive finite-sample or asymptotic operator-norm risk bounds for the full two-stage (basis selection plus coefficient estimation and smoothing) procedure, including the regimes with very small $N$ (down to $N=1$) for which the method was designed.

(2) Basis convergence: determine whether, and if so under what precise conditions, the estimated best basis converges to the true (optimal) basis — equivalently, when data-driven basis selection is stable — in the general locally stationary class rather than a special model, or prove that such convergence can fail.

(3) Data-driven tuning: determine whether a data-driven, theoretically grounded rule exists for choosing the truncation parameter $K$ and the time–frequency smoothing kernel variance with a guarantee (for example an oracle-type inequality) that the resulting operator-norm risk attains the bias–variance optimum attainable over these tuning parameters, and if so provide it, or prove that no such rule exists; and characterize when smoothing is inappropriate (highly irregular diagonal energies) rather than bias-inducing.

Each of (1)–(3) is a required component of the objective, and each is independently resolvable: a rigorous affirmative result (a proof or explicit construction) or a rigorous negative result (a disproof or counterexample) resolves that component on its own, so partial results stand as partial resolution. The objective as a whole is resolved only when all three components are addressed. All three components serve the single scientific objective of placing this adaptive estimator on the same decision-theoretic footing as classical covariance and spectral estimation; secondary acknowledged gaps that may be addressed within the same theory are the quantification of the effect of post-selection window modifications on the approximation constants and on estimation risk, and the computational-statistical tradeoffs for very large $M$ with small $N$.

References

1. S. Mallat, G. Papanicolaou, Z. Zhang, 'Adaptive covariance estimation of locally stationary processes', The Annals of Statistics 26(1), 1998. DOI: 10.1214/aos/1030563977, URL: https://doi.org/10.1214/aos/1030563977
2. D. L. Donoho, S. Mallat, R. von Sachs, 'Estimating covariances of locally stationary processes: consistency of best basis methods', Proceedings of Third International Symposium on Time-Frequency and Time-Scale Analysis (TFTS-96), 1996. DOI: 10.1109/tfsa.1996.547482, URL: https://doi.org/10.1109/tfsa.1996.547482
3. D. L. Donoho, S. Mallat, R. von Sachs, Y. Samuelides, 'Locally stationary covariance and signal estimation with macrotiles', IEEE Transactions on Signal Processing, 2003. DOI: 10.1109/TSP.2002.808116, URL: https://doi.org/10.1109/TSP.2002.808116
4. M. B. Priestley, 'Evolutionary spectra and non-stationary processes', Journal of the Royal Statistical Society Series B, 1965. DOI: 10.1111/j.2517-6161.1965.tb01488.x, URL: https://doi.org/10.1111/j.2517-6161.1965.tb01488.x
5. R. Dahlhaus, 'Fitting time series models to nonstationary processes', The Annals of Statistics 25(1), 1997. DOI: 10.1214/aos/1034276620, URL: https://doi.org/10.1214/aos/1034276620
6. S. Van Bellegem, R. von Sachs, 'On adaptive estimation for locally stationary wavelet processes and its applications', International Journal of Wavelets, Multiresolution and Information Processing, 2004. DOI: 10.1142/S0219691304000603, URL: https://doi.org/10.1142/S0219691304000603
7. S. Van Bellegem, R. von Sachs, 'Locally adaptive estimation of evolutionary wavelet spectra', The Annals of Statistics, 2008. DOI: 10.1214/07-AOS524, URL: https://doi.org/10.1214/07-AOS524
8. A. Cardinali, G. P. Nason, 'Locally stationary wavelet packet processes: basis selection and model fitting', Journal of Time Series Analysis, 2017. DOI: 10.1111/jtsa.12230, URL: https://doi.org/10.1111/jtsa.12230
9. S. Richter, R. Dahlhaus, 'Cross validation for locally stationary processes', The Annals of Statistics, 2019. DOI: 10.1214/18-AOS1743, URL: https://doi.org/10.1214/18-AOS1743
10. R. Dahlhaus, S. Richter, 'Adaptation for nonparametric estimators of locally stationary processes', Econometric Theory, 2022. DOI: 10.1017/S0266466622000500, URL: https://doi.org/10.1017/S0266466622000500
11. D.-N. Zhang, W. B. Wu, 'Convergence of covariance and spectral density estimates for high-dimensional locally stationary processes', The Annals of Statistics, 2021. DOI: 10.1214/20-AOS1954, URL: https://doi.org/10.1214/20-AOS1954
12. N. El Karoui, 'Operator norm consistent estimation of large-dimensional sparse covariance matrices', The Annals of Statistics, 2008. DOI: 10.1214/07-AOS559, URL: https://doi.org/10.1214/07-AOS559
13. G. Motta, C. M. Hafner, R. von Sachs, 'Locally stationary factor models: identification and nonparametric estimation', Econometric Theory, 2011. DOI: 10.1017/s0266466611000053, URL: https://doi.org/10.1017/s0266466611000053
14. F. M. Garcia, I. M. G. Lourtie, 'Estimation of locally stationary covariance matrices from data', Proceedings of ICASSP 2003, 2003. DOI: 10.1109/icassp.2003.1201787, URL: https://doi.org/10.1109/icassp.2003.1201787
15. M. Niedźwiecki, M. Ciołek, Y. Kajikawa, 'On adaptive covariance and spectrum estimation of locally stationary multivariate processes', Automatica, 2017. DOI: 10.1016/j.automatica.2017.04.033, URL: https://doi.org/10.1016/j.automatica.2017.04.033