Nonlinear History Sampling

• Nonlinear History Sampling is a methodology where sample selection depends nonlinearly on the evolving empirical history, diverging from traditional linear or memoryless schemes.
• It leverages adaptive feedback mechanisms across domains such as MCMC, event-driven signal processing, and stochastic control to enhance sampling efficiency.
• Rigorous theoretical guarantees, including unbiasedness and variance reduction, underpin its success in complex simulations, rare-event analysis, and modern machine learning applications.

Nonlinear history sampling refers to a class of methodologies in which the sampling process explicitly depends on the evolving trajectory or empirical history, inducing nonlinearities in the sampling law that are often distinct from traditional memoryless or linear schemes. These frameworks span discrete, continuous, stochastic, and physical domains, and arise in areas such as Markov Chain Monte Carlo (MCMC), event-driven signal processing, rare-event simulation, system identification, and uncertainty propagation in dynamical systems. This survey synthesizes rigorous developments in nonlinear history sampling, focusing on the mathematical formulations, core mechanisms, algorithmic strategies, theoretical guarantees, and applications—anchored throughout by contemporary arXiv literature and foundational results.

1. Mathematical Foundations of Nonlinear History Sampling

Nonlinear history sampling fundamentally diverges from classical, linear sampling paradigms by making the sample-selection rule an explicit (and typically nonlinear) functional of the past observation process or empirical measures.

Event-Dependent Sampling in Signals

Gluskin defined event-dependent sampling as a rule whereby the next sampling time $t_{k}$ is a nonlinear function $\Phi$ of the full past trajectory: $$t_k = \Phi\left(f(\tau) : 0 \leq \tau \leq t_{k-1}\right),\quad k=1,2,\ldots$$ Classic examples include level-crossing and zero-crossing samplers, which generate event times whenever the signal crosses a predefined level, making the sampling set adapt nonlinearly to the input path (Gluskin, 2010). Such rules induce strong nonlinearities at the level of the sampling grid itself.

History-Driven Targets in Graph MCMC

In discrete state-space Monte Carlo, nonlinear history sampling is exemplified by the introduction of a history-driven target (HDT) distribution $\boldsymbol{\pi}[\mathbf{x}]$, where $\mathbf{x}$ is the empirical measure of past visits: $\pi_i[x] \propto \mu_i\, (x_i/\mu_i)^{-\alpha}$ Any reversible or non-reversible MCMC kernel can then be “target-replaced” to use $\pi[x]$ at each step, creating an adaptive, self-repellent mechanism that biases against over-visited states (Hu et al., 23 May 2025). This realizes a nonlinear feedback between empirical history and stochastic transition structure.

Importance Sampling and Variational Control

In continuous-time stochastic processes (diffusions), the Gibbs variational principle equates sampling from a posterior path law (conditioned on an observed nonlinear potential $H$) with solving a stochastic optimal control problem: $$dX_t = b(X_t, t) dt + \sigma(X_t, t) dW_t + a(X_t, t) u^*(X_t, t) dt$$ where the control $u^*$ is functionally dependent on the past and future of $\Phi$0, encoding a fully nonlinear, history-aware adaptation of the dynamics to the desired target measure (Raginsky, 2024).

These examples highlight the ubiquity of nonlinear, history-dependent transformations in advanced sampling theory.

2. Rigorous Characterizations and Algorithmic Principles

The design and analysis of nonlinear history sampling schemes are anchored in certain structural principles:

• Self-repellency / Adaptivity: Mechanisms boost the sampling weight or activity of historically under-represented components and suppress over-sampled ones, manifesting as explicit nonlinearity in the state/empirical frequency dependence (as in HDT MCMC (Hu et al., 23 May 2025)).
• Locality and Scalability: Efficient constructions require only local or compressed history information, particularly in large state or graph spaces (see LRU-cache techniques in HDT (Hu et al., 23 May 2025)).
• Fixed-point and Unbiasedness: The empirical process is designed to stabilize around the original target distribution (or law), and rigorous ODE/SA theory proves global convergence (see Lyapunov and ODE arguments for HDT (Hu et al., 23 May 2025)).
• Variance Reduction: Central limit theorems quantify the efficacy of nonlinear history adaptation in variance reduction, often showing that variance scales inversely with the strength of the history-coupling parameter (e.g., $\Phi$1 reduction in HDT (Hu et al., 23 May 2025)).
• Minimization / Control Principle: In continuous-path settings, the nonlinear sampling law emerges from free-energy minimization or stochastic optimal control (Gibbs variational, HJB) (Raginsky, 2024).

Designing such schemes involves either (i) direct, theoretically-motivated nonlinear feedback rules, or (ii) end-to-end learning of optimal history-sensitive samplers in complex settings (e.g., graph neural networks (Feldman et al., 9 Apr 2025)).

3. Main Methodologies

A selection of rigorous nonlinear history sampling techniques is summarized below:

Framework	Core Nonlinearity	Domain
History-Driven Target (HDT) (Hu et al., 23 May 2025)	Nonlinear target $\Phi$2 in empirical measure	Graph MCMC
Event-Dependent Sampling (Gluskin, 2010)	Sampling times $\Phi$3 depend on $\Phi$4's history	Signal processing
Nonlinear Control Sampling (Raginsky, 2024)	Path-wise feedback drift computed via HJB/FPK PDE	Diffusions
FLASH Adaptive Sampling (Feldman et al., 9 Apr 2025)	Learnable, data-driven selection of history	TGNNs (dynamic graphs)
Momentum-based History-Guided Sampling (HiGS) (Sadat et al., 26 Sep 2025)	EMA of past model predictions steers sampling	Diffusion models
Burst SINDy/HAVOK (Champion et al., 2018)	Block/burst sampling strategies with scale-specific adaptation	System identification

History-Driven Target MCMC (HDT)

The unique HDT form emerges from scale-invariance, local dependence, fixed-point, and self-repellency requirements, yielding

$\Phi$5

Algorithmically, at each MCMC step, the static target $\Phi$6 in the kernel or acceptance ratio is replaced by $\Phi$7 computed from the (possibly approximate) empirical counts, and the empirical measure is updated recursively. An LRU caching scheme provides scalable memory efficiency (Hu et al., 23 May 2025).

Event-Dependent Sampling for Signal Integration

Sampling times determined by dynamic features of the function, such as level or zero crossers, generate nonlinear sampling sets that converge to classical Lebesgue integrals as level spacing vanishes. Such schemes are realized analogically using comparators and integrators and enable clock-free, event-driven measurements (Gluskin, 2010).

Adaptive Neural and Graph Sampling

In TGNNs, FLASH learns a parametric scoring function over neighbor histories, transforming static sampling into a learnable, fully history-dependent process. Loss functions incentivize the sampler to favor history fragments that improve prediction, and empirical results establish superiority over static heuristics for dynamic link prediction (Feldman et al., 9 Apr 2025).

Stochastic-Optimal Nonlinear Control for Path-Sampling

Mitter-Newtown–type variational formulations recast posterior path-sampling in diffusions as minimization of relative entropy plus path cost, yielding an optimal feedback drift $\Phi$8 computed from the gradient of a log-partition function (solving the corresponding HJB or Feynman-Kac PDE) (Raginsky, 2024).

4. Theoretical Guarantees and Performance

Major nonlinear history sampling constructions admit strong theoretical quantification of their properties:

• Convergence and Unbiasedness: The stochastic-approximation perspective (ODE averaging, Lyapunov arguments) proves that empirically driven nonlinear samplers (such as HDT) have the unique fixed point at the intended target law and achieve almost-sure convergence (Hu et al., 23 May 2025).
• Variance Reduction (CLT): The limiting covariance of the empirical measure in HDT is reduced by a known analytic factor, and as $\Phi$9, the variance can be made arbitrarily small. This demonstrates practical near-zero variance performance with minimal computational overhead (Hu et al., 23 May 2025).
• Expressivity over Static Heuristics: Adaptive history-based sampling (e.g., FLASH) is strictly more expressive than uniform or truncation-based heuristics, able to capture temporal dependencies and complex structural patterns in dynamic graphs that are provably impossible for fixed rule-based samplers (Feldman et al., 9 Apr 2025).
• Optimality in Control-based Path Sampling: Gibbs variational path-sampling realizes minimal free energy, equivalence to stochastic optimal control, and unification of classical importance sampling, time-reversal, and Schrödinger bridge frameworks within a rigorous information-theoretic paradigm (Raginsky, 2024).

5. Applications Across Scientific and Engineering Domains

Nonlinear history sampling is applied in diverse advanced contexts:

• Graph and Network Science: Efficient, variance-reduced sampling of nodes in massive discrete spaces for network statistics, distributed optimization, and inference, leveraging scalable nonlinear feedback mechanisms (HDT, LRU cache) (Hu et al., 23 May 2025).
• Event-Driven Signal Processing: Hardware-level Lebesgue integration, frequency estimation, and quantization in absence of external clocks or uniform sampling, using event-driven nonlinear samplers (Gluskin, 2010).
• Rare-Event Simulation in SPDEs: Hybrid Monte Carlo with trajectory-dependent bias functionals for efficiently sampling rare or extreme histories in stochastic field theories, using nonlinear importance sampling and reweighting for high-order statistics (Margazoglou et al., 2018).
• System Identification and Model Discovery: Nonlinear burst/block sampling strategies in SINDy/HAVOK compress multiscale data acquisition while preserving the identifiability of fast and slow dynamical modes (Champion et al., 2018).
• Machine Learning-Driven Adaptive History Sampling: End-to-end, data-adaptive, history-based sampling modules (FLASH) embedded in temporal graph neural networks show significant gains in sparse, heterogeneous, and highly nonstationary environments (Feldman et al., 9 Apr 2025).
• Control of Nonlinear Dynamical Systems: History- and memory-based neural operators enable real-time predictor-based feedback in nonlinear systems with input delays and non-uniform sampling, with sharp computational and tracking advantages (Bhan et al., 31 Mar 2026).
• Uncertainty Quantification in Nonlinear Systems: Variational LSTMs with augmented (history-rich) inputs and Monte Carlo dropout propagate both aleatoric and epistemic uncertainty in nonlinear response histories with high fidelity and efficiency (Sapkota et al., 2 Apr 2026).

6. Limitations, Open Problems, and Research Directions

Despite substantial advances, several foundational challenges remain:

• Extension from Memoryless to True Memory-Based Nonlinearities: Most current rigorous theory covers either memoryless or short-history nonlinearities, with deep open problems in establishing sampling/reconstruction theorems for genuinely nonlocal, history-dependent sampling operators (e.g., Volterra-series type) (0812.3066).
• Practical Caching and Scalability: For very large state spaces, precise maintenance of full empirical history is intractable; approximate strategies (LRU caches, compressed sketches) introduce estimation bias/variance trade-offs whose theoretical characterization is ongoing (Hu et al., 23 May 2025).
• Non-invertible and Noisy Nonlinearities: Extension to non-invertible samplers (clipping, dead zones), and robust recovery under sampling noise and hardware imperfections, remain active domains for theoretical and practical work (0812.3066).
• Complexity and Implementation: Nonlinear projection, feedback law computation (e.g., from PDEs in the control-theoretic approach), large-scale variational neural operators, and efficient iterative schemes for history-based inversion pose ongoing engineering and algorithmic challenges (Raginsky, 2024, Bhan et al., 31 Mar 2026).
• Theoretical Unification: Deeper connections between nonlinear MCMC feedbacks, event-driven sampling in signals, and continuous-time control-theoretic history sampling await further integration—especially in domains such as machine learning and high-dimensional statistical physics.

7. Summary Table of Major Nonlinear History Sampling Paradigms

Reference	Core Domain	Nonlinearity Mechanism	Theoretical Guarantee
(Hu et al., 23 May 2025)	Discrete MCMC	Self-repellent empirical feedback	Unbiasedness; CLT w/ variance gain
(Gluskin, 2010)	Signal processing	Sampling times depend on function	Lebesgue convergence, clock-free
(Raginsky, 2024)	SDE path sampling	Drift adapts via HJB/FPK control	Variational optimality, exact law
(Feldman et al., 9 Apr 2025)	TGNNs	Learned scoring of history	Strict improvement over heuristics
(Sadat et al., 26 Sep 2025)	Diffusion models	EMA/corrected prediction	Discretization/variance improvement
(Sapkota et al., 2 Apr 2026)	Structural Q.	Augmented/sequential stochastic input	Full uncertainty propagation

The nonlinear history sampling paradigm thus unifies a diverse array of advanced sampling techniques, each leveraging path- or history-dependent feedback to outperform classical approaches in adaptivity, efficiency, or statistical accuracy across a spectrum of scientific and engineering domains.