Moderate-Deviation Sharpening

• Moderate-deviation sharpening is the systematic refinement of leading-order asymptotic corrections in regimes where deviations shrink slower than CLT but remain smaller than LDP scales.
• It leverages techniques such as martingale decompositions, randomized concentration inequalities, and change-of-measure strategies to precisely quantify finite-sample impacts on estimator distributions and operational tradeoffs.
• The approach has significant applications in statistical estimation, autoregressive models, quantum resource theories, and stochastic PDEs, providing actionable insights for both theoretical analysis and practical implementations.

Moderate-deviation sharpening refers to the systematic, quantitative refinement of the leading-order (“first nontrivial”) corrections to probability asymptotics, estimator distributions, or operational tradeoffs in the moderate-deviation regime. This regime, which interpolates between the central limit theorem (CLT) scale and the large deviation principle (LDP) scale, features deviations shrinking to zero but on a scale large enough that classical variance-dominated approximations require systematic corrections. Recent developments in moderate-deviation sharpening have enabled a precise characterization of finite-sample effects, pivotal estimator tails, resource-conversion trade-offs, and entropy expansions in both classical and quantum settings. The theory leverages advanced techniques including martingale decompositions, randomized concentration inequalities, skeleton methods for SPDEs, and majorisation theory.

1. Foundations of the Moderate-Deviation Regime

The moderate-deviation regime is defined by deviation magnitudes that decay to zero slower than the CLT scale but remain sublinear compared to typical LDP scaling. For i.i.d. sums, a moderate sequence $\{t_n\}$ satisfies $t_n \to 0$ and $n t_n^2 \to \infty$. This regime enables expansion of probabilities and estimator distributions in powers of $t_n$, yielding not only leading-order rates but also capturing the first systematic corrections (the “sharpening”). In resource theories, this occurs for error levels $\epsilon_n = \exp(- n t_n^2)$ and conversion rates approaching their optimum at $r_n = r_\infty + O(t_n)$ (Chubb et al., 2018), and, in statistical models, for normalized test statistics exceeding moderate thresholds $x_n \le o(n^\gamma)$ (Shao et al., 2014).

2. Moderate-Deviation Sharpening in Self-Normalized Processes

Shao and Zhou’s theory (Shao et al., 2014) establishes sharp moderate-deviation inequalities for self-normalized sums $T_n = S_n / V_n$ and generalizations including Studentized $U$-statistics. The main result provides explicit relative-error expansions: $$\frac{P(S_n / V_n \ge x)}{1 - \Phi(x)} = 1 + O\left((1+x)^3 \frac{\sum_{i} E|X_i|^3}{(\sum_{i} E X_i^2)^{3/2}}\right)$$ uniformly for $$0 \le x \le (\sum_i E X_i^2)^{1/2} / (\sum_i E|X_i|^3)^{1/3}$$.

For general self-normalized statistics of the form $T_n = (W_n + D_{1n}) / \{V_n^2(1+D_{2n})\}^{1/2}$, two-sided exponential inequalities are established: $$P(T_n \ge x) \le (1-\Phi(x)) \exp\{C_3 L_{n,x}\}(1+C_4 R_{n,x}) + P(xD_{1n} > V_n/4) + P(x^2 D_{2n} > 1/4)$$ where $L_{n,x}$ and $R_{n,x}$ quantify moment and remainder contributions. With $D_{1n}=D_{2n}=0$, this reduces to a uniform $o(1)$ approximation in the relative-error sense.

For Studentized $U$-statistics, under moment assumptions, the tail expansion reads

$$\frac{P(T_n \ge x)}{1 - \Phi(x)} = 1 + O\left((\frac{\sigma_p}{\sigma})^p (1+x)^p n^{1-p/2} + (\sqrt{a_m} + \frac{\sigma_h}{\sigma})(1+x)^3 n^{-1/2}\right)$$

valid for $x$ up to $$c \min\{(\sigma/\sigma_p) n^{1/2-1/p}, n^{1/6}/a_m^{1/6}\}$$, with all relevant symbols as in (Shao et al., 2014).

These sharp expansions are facilitated by randomized concentration inequalities (via Stein’s method), moment truncations, change-of-measure arguments, and martingale decompositions, removing previous dependence on exponentially integrable moments and substantially weakening required conditions.

3. Moderate-Deviation Sharpening in Statistical Estimation and Autoregressive Models

For bifurcating autoregressive processes (BAR($p$)), moderate-deviation sharpening is realized in precise deviation inequalities and an exact MDP for least-squares estimators under minimal tail assumptions (Djellout et al., 2012). Writing

$$Z_n := \frac{\sqrt{\tau_n}}{b_{\tau_n}} (\hat{\theta}_n - \theta)$$

with $b_n \to \infty$, $b_n / \sqrt{n} \to 0$, the normalized estimators satisfy a full MDP with speed $b_{\tau_n}^2$ and quadratic “good” rate function,

$$I_\theta(x) = \frac12 x^T (\Gamma^{-1} \otimes L) x$$

where $\Gamma$ and $L$ encode model and noise covariance structure.

Key technical steps include:

• Martingale representation and bracket control, with super-exponential convergence of normalized bracket processes.
• Sharp exponential inequalities for truncated martingales, verified via Puhalskii’s theorem.
• Exponential approximation arguments establishing full MDPs for estimation errors, with no additional order terms.

This leads to deviation results holding uniformly for all $n$ and error levels, with tail bounds that interpolate seamlessly between the CLT and LDP regimes, and rate functions precisely matching the limiting Gaussian law structure.

4. Moderate-Deviation Sharpening in Branching Processes and Random Walks

In supercritical branching random walks, moderate-deviation sharpening has uncovered double-exponential asymptotics for the lower tails of maximum displacement distributions (Chen et al., 2018). For the maximum $M_n$ at generation $n$, and sequences $\ell_n \to \infty$, $\ell_n = O(n)$, the asymptotic probability

$$P\big(M_n \leq m_n - \ell_n\big) = \exp\left\{-\exp\big[\beta \ell_n (1 + o(1))\big]\right\}$$

where $m_n = x^* n - \frac{3}{2\theta^*}\log n$ is the centered maximum and $\beta$ is an explicit constant, captures the subleading double-exponential decay in the moderate regime. The analysis extends to situations with heavy-tailed or bounded step size distributions and leverages change-of-measure, truncation strategies, and population subadditivity techniques to make polynomial prefactors negligible. The resulting rate functions are exact up to negligible corrections, sharpening previous moderate-deviation results whose valid $\ell_n$ range was much smaller.

5. Moderate-Deviation Sharpening in Quantum Information and Resource Theory

In quantum information, moderate-deviation sharpening governs finite blocklength corrections to rates in state transformations, channel coding, and simulation tasks (Ramakrishnan et al., 2021). The expansions take the form

$$\frac{1}{n} D_h^{\epsilon_n}(\rho^{\otimes n}\|\sigma^{\otimes n}) = D(\rho\|\sigma) \pm \sqrt{2V} a_n + O(a_n)$$

for moderate $\epsilon_n = \exp(- n a_n^2)$, with $a_n \to 0$, $n a_n^2 \to \infty$. In resource interconversion governed by (thermo-)majorisation, the optimal rate satisfies

$r_n = r_\infty - \kappa t_n + o(t_n)$

for error $\epsilon_n = \exp(-n t_n^2)$, with $\kappa$ determined by entropy and variance ratios of the involved states (Chubb et al., 2018). Notably, resonance conditions ($D_e=1$ or $D_{\mathrm{th}}=1$) induce effective reversibility—$\kappa = 0$—allowing the asymptotic optimal rate to be achieved with negligible error even at finite $n$.

These results rely on tight relations between hypothesis-testing relative entropy and smoothed max-information, and the translation of classical moderate-deviation bounds to quantum settings via spectrum and majorisation combinatorics.

6. Sharpening in Moderate-Deviation Principles for Stochastic PDEs

For stochastic fractional conservation laws, moderate-deviation sharpening is achieved by identifying the exact (quadratic-variational) rate function and speed $A(\varepsilon)^2$ for deviations of scaled solution differences

$$z^\varepsilon(t,x) = \frac{u^\varepsilon(t,x) - \bar{u}(t,x)}{A(\varepsilon)}$$

where $\sqrt{\varepsilon} A(\varepsilon) \to 0$, $A(\varepsilon) \to \infty$ (Behera et al., 2023). The moderate-deviation principle (MDP) constructed via weak convergence methods (Budhiraja-Dupuis) rigorously interpolates between the CLT ($A(\varepsilon)=1$) and LDP ($A(\varepsilon) = \varepsilon^{-1/2}$) scales. The established MDP is sharp in the sense of exactly matching speeds and rate functionals, but the cited work does not pursue higher-order Edgeworth corrections or explicit error terms.

7. Methodological Innovations and Implications

Moderate-deviation sharpening has hinged on several technical advances:

• Randomized concentration inequalities (via Stein’s method and coupling) (Shao et al., 2014).
• Martingale decompositions and uniform bracket control for dependent or heteroskedastic processes (Djellout et al., 2012).
• Controlled change-of-measure and skeleton mapping in stochastic analysis (Behera et al., 2023).
• Majorisation-based spectral analysis and entropic tail estimates in resource theories (Chubb et al., 2018, Ramakrishnan et al., 2021).

These techniques have made possible refined, non-asymptotic characterizations of estimator error, test critical-value coverage, bootstrap accuracy, finite-blocklength channel rates, and operational transformations in both classical and quantum systems.

In summary, moderate-deviation sharpening systematically exposes and quantifies the leading corrections to asymptotic limit theorems across probability, statistics, dynamical systems, and quantum information, under minimal integrability and structural assumptions. Such results have direct significance for practical finite-sample inference, resource tradeoff optimization, and foundational probability theory.