Adaptive Stopping Criterion for Stochastic Estimation

• Adaptive Stopping Criterion is a data-driven rule for terminating iterative computations based on real-time error monitoring and bias-variance balancing.
• It employs methods such as the Lepskiĭ principle and balancing techniques to select optimal parameters without prior knowledge of jump-activity indices.
• Empirical studies show that this approach achieves near-minimax rates and robust performance in high-frequency stochastic process estimation.

An adaptive stopping criterion is a data-driven or statistically principled rule for terminating an iterative or sequential computation, estimation, or experimental procedure, with the goal of achieving near-optimal trade-offs between risk, computational efficiency, and accuracy. Such criteria are central in modern statistical estimation, iterative numerical algorithms, model selection, active learning, and high-frequency stochastic process inference. Adaptive stopping rules respond to observable features of the process itself—such as empirical errors, residuals, or estimator instability—rather than relying on fixed, problem-specific thresholds. This yields estimators or policies that are robust to unknown regularity, noise, or activity indices, and can often achieve (nearly) minimax rates of convergence without prior knowledge of key problem parameters.

1. Fundamental Setup and Motivating Example

A canonical setting is the high-frequency observation of a multivariate Lévy process over a fixed interval, for the purpose of estimating the continuous covariation component $C$ in the Lévy–Khintchine decomposition. Let $X=(X_t)_{t\ge0}$ denote a $d$-dimensional Lévy process, observed discretely at times $t_j=j/n$, $j=1,\dots,n$, on $[0,1]$. The process has triplet $(b,C,F)$, where $b$ is drift, $C$ the covariance matrix of the Gaussian part, and $F$ a Lévy measure. The characteristic function of increments is

$$\phi_n(u)=\mathbb{E}[e^{i\langle u, X_{j/n} - X_{(j-1)/n}\rangle}] = \exp\{\tfrac{1}{n}\Psi(u)\}.$$

A family of spectral estimators $\widehat C^{12}_n(U)$ for an entry of $C$ is indexed by a frequency parameter $U>0$. Each estimator exhibits a bias–variance trade-off: higher $U$ decreases deterministic bias but increases stochastic error, with the minimax-optimal rate depending on the unknown jump activity $r\in[0,2)$. The challenge is to adaptively select the data-dependent $U$ yielding the best risk without knowing $r$ or the optimal index in advance (Papagiannouli, 2020).

2. Lepskiĭ-type Adaptive Stopping Rule

The Lepskiĭ principle provides a formal methodology for adaptive parameter selection in nested families of estimators. Given a discretized grid of frequency parameters $U_0 < U_1 < \cdots < U_K$, and corresponding estimators $\widehat C^{12}_{n,j}$, the Lepskiĭ rule selects the maximal $j^*$ such that all subsequent estimators are statistically indistinguishable from $\widehat C^{12}_{n,j^*}$ up to an explicit stochastic error bound. Formally, define a pseudo-metric $d(x, y) = |x-y|$, and a deterministic upper bound $s_n(U_j)$ on the stochastic error, monotone in $j$. The rule is:

$$j^* = \inf\left\{j \le K-1: \exists\, k \le j\,\,\textrm{s.t.}\,\, d(\widehat C^{12}_{n,j+1},\,\widehat C^{12}_{n,k}) \leq C\,s_n(U_{j+1})\right\} \wedge K.$$

Alternatively, with a minimax rate $w_n(j)$ and a constant $A$:

$$j^* = \inf\left\{j \le K-1: \forall\, k = j+1,\ldots,K,\,\, d(\widehat C^{12}_{n,j},\,\widehat C^{12}_{n,k}) \le 2A\,w_n(k)\right\} \wedge K.$$

This rule can be implemented algorithmically via nested pairwise tests. The index $j^*$ selected adaptively achieves a risk within a constant multiple of the oracle index (if that were known), and thus yields nearly optimal performance (Papagiannouli, 2020).

3. Stochastic Error Control and Oracle Start

A nontrivial challenge is the non-monotonicity and possible explosion of stochastic error bounds for small or large values of the family parameter—here, the frequency $U$. The process includes the following steps:

• Uniform deviation inequalities for the empirical characteristic function $\widehat\phi_n(u)$, proved using Talagrand’s inequality and entropy bounds, provide high-probability control over the stochastic error.
• A truncated inverse is defined for $\widehat\phi_n(u)$ to avoid division by small values, ensuring numerical stability and valid stochastic bounds.
• To avoid instability at low frequencies, the procedure identifies an “oracle start” value:

$$U_{\mathrm{start}}^{\mathrm{oracle}} = \inf\left\{ U > 0 : |\phi_n(U)| \leq 1/2 \right\},$$

with $s_n(U)$ monotonic for $U \ge U_{\mathrm{start}}^{\mathrm{oracle}}$. In practice, the empirical version $\widehat U_{\mathrm{start}}^{\mathrm{oracle}}$ is used (Papagiannouli, 2020).

This regime ensures that the adaptive selection principle operates in the range where the stochastic error bound is monotone and the Lepskiĭ rule is valid.

4. Balancing Principle and Theoretical Guarantees

The balancing principle equates deterministic bias and stochastic error terms to locate the frequency (or general parameter) where the mean-squared error is minimized:

• Bias, for $r \in (1,2]$, is $d(U) = M 2^{r/2} U^{2 - r}$.
• Data-driven stochastic bound is $$\tilde s_n(U) = \frac{2C \gamma(n)}{U^2} w(U)^{-1} / |\tilde\phi_n(U)|$$, with $\gamma(n) = (n \log n)^{1/2}$, $w(U)$ a weight.
• The balancing point $U_{\mathrm{bal}}$ satisfies $\tilde s_n(U) = d(U)$, yielding the rate-optimal trade-off.

The adaptive estimator at frequency $U_{\widehat j}$ (selected via the stopping rule) satisfies, with high probability,

$$|\widehat C^{12}_{n, \widehat j} - C^{12}| \leq 5\,\tilde s_n(U_{\mathrm{bal}}),$$

thus attaining the minimax rate up to constants (Papagiannouli, 2020).

5. Empirical Performance and Practical Aspects

Empirical studies with bivariate Lévy processes (with covariance $$C = \begin{pmatrix}2 & 1 \ 1 & 1\end{pmatrix}$$) and both finite- ($r_i < 1$) and infinite-variation ($r_i > 1$) jumps confirm key properties:

• For finite-variation jumps, the adaptive estimator is stable across a substantial frequency range and matches the true parameter value well.
• For infinite-variation jumps, instability at low frequencies necessitates the start-up cutoff, after which the adaptive criterion successfully avoids regions of poor estimator behavior.
• No a priori knowledge of the jump-activity index or characteristic function decay is required; only quasi-monotonicity and data-driven error bounds.
• The approach generalizes to high-frequency financial or insurance data where such stochastic processes are commonly modeled (Papagiannouli, 2020).

6. Broader Context and Related Methodologies

The Lepskiĭ-type adaptive stopping rule is part of a larger landscape of regularization and early stopping strategies in nonparametric estimation and inverse problems. Its core features—balancing bias and stochastic variability, and adaptivity to unknown smoothness or activity indices—are shared by analogous procedures in nonparametric regression, functional estimation, and spectral regularization. The methodology is robust to various sources of ill-posedness and scalable to high-dimensional settings, provided the computational tractability of empirical characteristic function evaluations and the implementation of recursive testing are ensured.

A key strength of this approach is its simultaneous minimax adaptivity (over a scale of smoothness or jump-activity indices) and finite-sample validity under mild sufficient conditions. Its use of data-driven stochastic bounds and error control avoids the need for tuning via held-out data or prior parameter calibration. The criteria and proofs rely on advanced probabilistic tools (e.g., concentration inequalities, empirical process theory), underscoring the centrality of high-confidence error quantification in modern adaptive stopping rules for stochastic processes (Papagiannouli, 2020).

---

References: All technical results, definitions, and empirical findings are from (Papagiannouli, 2020).