Organizations: Macao Polytechnic University, Macao · College of Textile Science and Engineering, Wuyi University, Jiangmen, China · Guangdong Provincial Key Laboratory of Novel Polyamide 6 Functional Fiber Materials Research and Application, Jiangmen, China
Resampling is the standard remedy for weight degeneracy in particle filters, but in high dimensions it distorts the spectrum of the ensemble covariance in a way that is computable from the importance weights. This paper studies the inverse problem: recovering the pre-resampling spectrum from the observed weights and the empirical spectrum alone, and using it without biasing the filter. We first show that the forward spectral law extends to random importance weights once one conditions on the weights together with the realized offspring counts, rather than on the weights alone. We then construct a regularized free-deconvolution estimator, prove its consistency with an explicit error budget at a fixed resolution scale, and certify model mismatch through a computable forward residual. The repaired spectrum enters the filter only as the preconditioner of a Metropolis-adjusted rejuvenation move, so that spectral error affects mixing efficiency but not the invariant target. For the full filtering cycle, we derive a geometrically summable error recursion under stable scalar dynamics and an explicit law for the cost of abstaining from repair. Pre-registered simulations show that the inversion reduces the spectral error by an order of magnitude, that the corrected preconditioner recovers about 90% of the oracle's sampling efficiency in an anisotropic regime, and that the cost of abstention depends on when abstentions occur rather than on how many. The repair recovers eigenvalue laws, not eigenvectors, and is certified only within an explicitly stated exogenous-weight regime.
Figures & tables
Fig. 1: The spectrally corrected filtering cycle and its error recursion. (a) One cycle: propagation (Section 8.2 ), selection by resampling, which acts as a free multiplicative distortion with the law Tk+1 determined by the exogenous weights (Theorem 3 ), and regularized spectral inversion driven by the plug-in law Tk+1 (Theorems 4 – 6 ). The supervisor accepts the repair or abstains and passes on the uncorrected spectrum (Proposition 3 , Algorithm 1, step 9). Solid arrows carry spectral laws, dashed arrows carry certificates and the control signal, and the dotted arrow marks the ideal law μk+1 , which is not computed. With the switch held on abstain, successive selections compound; on accept, each cycle is reset. (b) The first-order comparison system majorizing the error ηk=W1(Hk+,Hk⋆,+) : the recursion ηk+1≤aˉ2ηk+Bk+1θ has pole aˉ2<1 (Theorem 8 (a)), where Bk+1θ is the per-step budget capped at the resolution scale (Section 6.4 ). Panel (b) depicts an upper bound, not the actual error dynamics.
Fig. 2: Static spectral distortion and repair (EXP-1). (a) Empirical spectral distribution after resampling, pooled over 50 repetitions at N=2000 , α=1 , against the predicted law H0⊠Tc(ν) for multinomial and residual resampling; markers on the vertical axis compare the empirical zero-atom mass (o) with the closed form (x). (b) Second spectral moment against the predicted count second moment, for the six weight-profile and scheme groups at N=2000 ; the line is the certificate of Corollary 2 . (c) Median W1 inversion error against N at α=1 for the oracle, conditional plug-in, and realized-count inverters (solid: multinomial; dashed: residual; bars: interquartile range); the dotted line is the pre-registered limit of the uncorrected spectrum.
Fig. 3: Diagnostic certificates and identifiability (EXP-2). (a) Mismatch statistic εmodel across the nine settings at N=2000 , α=1 ; the statistic rises sharply only for the wrong mechanism and for strongly correlated particles. (b) Median inversion error against the zero-atom mass of the distortion law, H=δ1 , N=1000 , colored by log10κinv ; conditioning degrades as the zero atom grows (Remark 3 ).
Fig. 4: Identifiability and supervisor activity on the (c,α) grid. (a) Map level: median estimated condition number κinv ; conditioning degrades toward large c and heavy-tailed weights (small α ), where the zero atom is largest. (b) Pipeline level: activation fraction of the comparative clause of the supervisor; the framed cell is the deployment setting of EXP-3, (c,α)=(0.75,3) . (c) Distribution of the supervisor’s decision ratio BN/δfb over the 660 steps of EXP-3 (median 0.146 , maximum 0.824 ); no step crosses the activation threshold 1 (dashed).
Fig. 5: Full-cycle behavior (EXP-3). (a) Median error ηk=W1(Hk+,Hk⋆,+) over 20 repetitions with interquartile bands, for no correction, conditional plug-in, oracle, and plug-in with supervisor; dotted lines mark the weight shocks. (b) Forced abstention with the same number of abstentions, AK=6 , clustered at the end of the horizon or spread uniformly. (c) Error against the geometric abstention load Λk , with the bound of Proposition 3 (red).
Preconditioner
κ
ESS/step
Acc.
ESS/s
max∣z∣
I
42.4
3.26×10−4
0.237
1355
3.23
Src+εI
19.6
5.45×10−4
0.226
2242
1.60
lift (H)+εI
5.6
8.72×10−4
0.237
2046
1.88
oracle Σ+εI
2.4
9.72×10−4
0.240
1794
2.41
Table 1: Metropolis preconditioning in Regime A ( d=225 , c=0.15 , κ(Σ)=50 , robs=200 ). Condition number of the preconditioned posterior covariance, slowest-direction ESS per step, acceptance rate, ESS per second at equal wall clock, and the maximum ∣z∣ -score of the target-displacement test.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Fig. 6: Count-law checks. (a) Empirical probability mass function of the counts against the mixed Poisson law, N=2000 , α=1 . (b) Heavy-tailed weights, α=0.3 , on logarithmic axes. (c) Empirical zero-atom share of the centered resampled matrix against the exact count-rank identity 1−(#{nj>0}−1)/d (circles) and the limiting value of Corollary 2 (crosses); all points lie on the diagonal.
Fig. 7: Complete comparison for Proposition 1. (a) The active-column fraction ps is ordered by the convex order and not by r2 ; equal r2 can give different ps . (b) Two laws with r2=2 and equal second moment produce different limiting spectra and zero atoms. (c) The scheme-specific excess over the weighted ensemble: ca12 for multinomial resampling and ca12EπU for residual resampling.
Fig. 8: Inversion sensitivity. (a) The fixed regularization schedule γ0/N against the scale-free schedule in the strong-anisotropy regime; dashed lines mark the true values. (b) Relative difference in W1 between the two schedules, per experiment and step. (c) Activation counts of the comparative clause under three versions of the discretization-floor estimate, at three bandwidths; version 3 is the one reported.
Fig. 9: Fallback error against N at c=0.75 . (a) Median δfb with interquartile bands and bootstrap slope intervals, which contain zero for both weight profiles. (b) Worst grid-resolution ratio over repetitions; values above 1 indicate that the resolution cap is reached. (c) Dependence on the bandwidth.
Fig. 10: Ledoit–Wolf nonlinear shrinkage as an external baseline. (a) Median W1 error of each inverter, 900 paired draws. (b) Number of paired draws in which Ledoit–Wolf is closer to H0 than each inverter. (c) W1 error against the mass that Ledoit–Wolf assigns to the common null value.
Fig. 11: Burn-in diagnostics for the Metropolis chains of Table 1. (a) The Robbins–Monro step-size controller settles before the end of burn-in (dashed). (b) Tail drift ratio of each controller at the end of burn-in, against the pre-stated threshold 0.2 . (c) Burn-in length in units of the integrated autocorrelation time, per chain, against the pre-stated floor of five.
Fig. 12: Activation zone of the comparative clause. (a) Median decision ratio BN/δfb on the (c,α) grid at the deployment bandwidth; large markers indicate activation. (b) Activated fraction against c at three bandwidths.
Fig. 13: Passive monitor experiment. (a) ESS, zero-atom mass, and effective dimension relative to their steady-state means; dotted lines mark the weight shocks. (b) Zero-atom mass against ESS; steady and shock steps form two separated clusters on one curve. (c) Shock-to-steady ratios against the pre-registered threshold 1.5 for a silent ESS; the ESS falls by a factor of 3.23 , so the prediction fails.
Resampling is the standard remedy against weight degeneracy in particle filters, yet its effect on the spectrum of the particle second-moment matrix is poorly understood. We study a single resampling step for a d-dimensional ensemble of N particles with d/N -> c. For any unbiased scheme, the resampled empirical second-moment matrix dominates its pre-resampling counterpart in conditional convex order, while the conditional mean is preserved; the excess second spectral moment equals the Frobenius energy of the resampling perturbation. For multinomial resampling, the empirical count law converges to the mixed-Poisson law determined by the limiting weight law. For residual resampling, the count law is determined by the joint limit of floor-fraction pairs: we classify all residual limits of weight sequences converging to delta_1 in law as a one-parameter family between the deterministic identity and the full multinomial count law. Each count law feeds a common covariance-side spectral map, yielding exact consequences: a closed-form zero-eigenvalue mass, a rank-and-spectral-gap transition with a universal lower edge benchmark, computable regular edges, unbounded right support, and scheme-specific second-moment costs. The effective sample size determines one bulk statistic but cannot determine the limiting spectrum. Pre-registered numerical experiments confirm the theory at every level.
Deming Song, Chi Kin Lam
the Faculty of Applied Sciences, Macao Polytechnic University, Macao, China
The ensemble Gaussian mixture filter (EnGMF) is a powerful, convergent particle filter capable of medium-to-high dimensional non-linear filtering. The EnGMF relies on a resampling step that can generate physically unrealistic posterior samples, that would subsequently produce physically meaningless forecasts. This work introduces the discriminator-informed resampling procedure, that augments the posterior resampling step with a discriminator that accepts or rejects candidate particles based on their physical plausibility. In this work these discriminators are learned through a normalizing flow approach. Numerical experiments on both the Ikeda map and the Lorenz '63 system show that discriminator informed resampling procedure consistently reduces error relative to the standard EnGMF in low-ensemble regimes.
Particle filters are a standard tool for nonlinear state estimation, but their resampling step is discrete, preventing gradient-based learning in variational sequential Monte Carlo. We introduce Differentiable Systematic Resampling (DSR), a temperature-controlled relaxation of systematic resampling, that preserves the CDF-ordered, banded structure of systematic resampling while enabling full gradient flow. DSR converges to exact systematic resampling as the temperature vanishes, and we prove a pointwise exponential convergence rate for the induced bias. Compared to optimal-transport-based differentiable resampling, DSR avoids iterative solvers and has substantially lower computational overhead. Experiments on stochastic dynamical systems and real-world handwriting data show that DSR achieves comparable or superior filtering and dynamics learning performance.
Fredrik Cumlin, Saikat Chatterjee
Department of Information Science, KTH Royal Institute of Technology, Stockholm, Sweden