Finite-Dimensional Recursions for Small-Noise Expansions in Nonlinear Filtering

Scientific Impact Assessment

Core Contribution. This paper provides the computational machinery for a small-system-noise asymptotic expansion of the nonlinear filter, complementing the author's earlier existence/justification result (Kurisaki 2026b). The central achievement is showing that the expansion coefficients — which are a priori conditional expectations of polynomial functionals and iterated observation integrals — can be computed by a *finite-dimensional* recursive system of SDEs driven by the observation process, extending the Kalman–Bucy/extended-Kalman filter. The key novelty is a "finite closure mechanism": terminal-time differentiation of smoothing fields successively replaces them with lower-degree endpoint covariance factors, and this degree strictly decreases, guaranteeing termination. A crucial technical byproduct is the rigorous treatment of *non-adapted* stochastic integrals (arising because Fubini + Wick produce integrands depending on observations beyond the integration time) as L²-limits of discrete sums, plus a generalized Itô formula for differentiating them with respect to terminal time. The paper also extends the expansion to the conditional characteristic function (finite Gaussian–polynomial/Edgeworth form) and demonstrates that dimension grows polynomially, not exponentially, in the system dimension.

Methodological Rigor. For a probability-theory paper the rigor is high. The convergence of discretized sums (Proposition 4.1 via Girsanov, BDG, Hölder), the induction-on-degree proof of the generalized Itô formula (Theorem 4.12), and the closure/polynomial-growth results (Theorems 4.17, Proposition 4.18) are carefully constructed. Lemma 4.11's explicit separated-variable representation of the smoothing covariance is the load-bearing structural fact and is proved cleanly. The 1D worked example in Section 3, with the full third-order derivation in the supplement, gives concrete confidence that the abstract machinery actually closes. The main gap is empirical: a single 1D Aït–Sahalia-type experiment, at third order only, with a truncation caveat (Remark 6.1) because the model violates the standing boundedness assumptions.

Potential Impact. Nonlinear filtering is foundational to signal processing, control, data assimilation (weather/geophysics), and quantitative finance. The paper positions itself squarely against the curse of dimensionality afflicting particle filters and chaos/PDE methods, where cost grows exponentially. If the polynomial-growth claim scales in practice, this is a genuinely attractive alternative in the high-signal-to-noise / small-diffusion regime. However, adoption is bounded by (a) the small-noise restriction, (b) the fact that multidimensional performance is asserted (via the D_r(d₁) bounds in the Discussion) but never numerically demonstrated, and (c) the heavy technical entry cost. The characteristic-function extension broadens applicability beyond conditional means to full posterior reconstruction, which is a meaningful conceptual step.

Timeliness & Relevance. High-dimensional filtering remains a live bottleneck, so a method promising polynomial scaling addresses a real need. That said, small-noise expansions are a comparatively niche corner of the filtering literature, and this paper is the second in a self-referential sequence rather than a response to a broad community demand.

Strengths. (1) The finite-closure argument is elegant and non-obvious; the degree functional turning an apparently exploding tree into a terminating recursion is the paper's intellectual core. (2) The rigorous construction of non-adapted integrals and the generalized Itô formula are contributions of independent technical interest. (3) Clear pedagogical scaffolding — the 1D case precedes the general theorem. (4) Polynomial-vs-exponential dimensional scaling is a concrete, quantified advantage.

Limitations. (1) Thin empirical validation — one low-dimensional model, no high-dimensional demonstration to substantiate the headline scaling advantage where it would matter most. (2) Restricted to the small-noise regime; accuracy degrades for strongly nonlinear/non-Gaussian large-noise problems where particle filters dominate. (3) Very heavy dependence on the author's own two prior papers (2026a, 2026b) — this is a link in a chain rather than a self-contained contribution. (4) No code released. (5) The improvement shown (37% RMSE reduction vs. first order) still leaves the estimate roughly on par with the particle-filter reference, so the practical payoff of higher-order terms is suggestive rather than compelling. (6) Practical implementation burden (deriving the tree of coefficient equations) grows quickly with order, as the supplement's sprawling third-order system illustrates.

Other observations. The work is squarely theoretical/methodological, of specialist difficulty requiring years of stochastic-analysis expertise. It does not challenge or replicate prior claims. Its most likely mode of influence is as a reusable technical building block (the non-adapted integral apparatus and closure argument) for a small community of filtering theorists, plus the author's own planned follow-ups on minimal recursions and cumulant reorganization.

Overall this is a competent, rigorous, and genuinely novel piece of applied-probability theory with a clear structural insight, but its impact is tempered by niche scope, small-noise restriction, minimal empirical demonstration, and its position within a self-referential research program.