Auto-Differentiable Filtering Methods

• Auto-differentiable filtering is a method that transforms traditional recursive Bayesian filters into differentiable computation graphs, enabling gradient propagation through prediction, update, and regularization steps.
• It leverages smooth approximations, gradient surgery, and analytic differentiation to replace non-differentiable operations, thereby improving numerical stability and adaptive inference.
• This approach integrates learned generative updates and multimodal sensor fusion, enhancing applications in robotics, cosmography, point-cloud denoising, and more.

Searching arXiv for recent and foundational papers on auto-differentiable filtering and related differentiable filtering methods. Auto-differentiable filtering denotes a class of methods in which a filtering algorithm, or a filter-like inference operator, is represented as a differentiable computation graph so that gradients can propagate through prediction, update, regularization, and, in some cases, adaptive discretization itself. In the cited literature, the term spans differentiable Bayesian filters for latent-state estimation, auto-differentiable ensemble and particle methods, square-root and adaptive Kalman variants, differentiable digital filters, and domain-specific constructions such as ray-guided source reconstruction in strong-lensing analysis (Kloss et al., 2020, Adrian et al., 21 Mar 2026, Enzi et al., 29 Jun 2026).

1. Conceptual basis

The canonical starting point is the recursive Bayesian filter,

$$\mathrm{bel}(\mathbf{x}_t)=\eta\, p(\mathbf{z}_t \mid \mathbf{x}_t)\int p(\mathbf{x}_t \mid \mathbf{x}_{t-1}, \mathbf{u}_{t-1})\, \mathrm{bel}(\mathbf{x}_{t-1})\, d\mathbf{x}_{t-1},$$

with the familiar prediction and update decomposition. Auto-differentiable filtering converts this recursion into a neural-network layer or recurrent computation graph, so that sensor models, process models, and noise models can be trained end-to-end by gradient descent rather than fixed analytically or tuned manually (Kloss et al., 2020).

Within this formulation, differentiability is not an aesthetic property but an algorithmic one. A differentiable filter exposes gradients of the loss with respect to internal filtering quantities such as process parameters, observation encoders, covariance models, proposal mechanisms, or adaptive-grid transforms. In multimodal settings, this preserves an explicit belief representation and uncertainty propagation while allowing raw observations such as images, tactile streams, and proprioceptive signals to enter through learned modules (Lee et al., 2020).

The same logic extends beyond classical hidden-state estimation. Several papers replace hard geometric or discretization decisions by smooth surrogates or smooth coordinate transforms: collision checking becomes a differentiable proxy score, point-cloud denoising is constrained by differentiable rendering, and adaptive source-plane meshing is replaced by a smooth warping of a uniform grid (Zhi et al., 2021, Zhou et al., 2024, Enzi et al., 29 Jun 2026). This suggests that auto-differentiable filtering is best understood as a computational principle: preserve the filtering or regularization semantics while eliminating gradient-breaking operations wherever possible.

2. Mechanisms used to obtain differentiability

One recurring mechanism is the substitution of discontinuous structure by smooth maps. In strong-lensing source reconstruction, the ray-guided transformed uniform grid constructs adaptivity through a coordinate transformation

$\vec y=\vec g(\vec x),$

with components built from empirical cumulative distribution functions of back-traced rays and then smoothed by polynomial fitting and cubic splines; the source remains defined on a uniform grid, and the adaptive behavior is transferred into a differentiable mapping (Enzi et al., 29 Jun 2026). In DiffCo, hard collision queries are replaced by a smooth kernel-based score over configuration space, so collision avoidance enters optimization as a differentiable proxy constraint rather than as a non-differentiable geometric oracle (Zhi et al., 2021). In 3DMambaIPF, the filtering objective is augmented by a differentiable rendering loss so that gradients from rendered views act directly on 3D point positions (Zhou et al., 2024).

A second mechanism is gradient surgery rather than forward-pass modification. For particle filtering, one line of work keeps the forward algorithm unchanged and only alters the backward graph through stop-gradient corrections on resampling weights. This yields automatic-differentiation estimators of classical particle-filter score quantities without changing the forward Monte Carlo estimator (Ścibior et al., 2021). Related implementations package alternative resampling relaxations—non-differentiable, soft, optimal-transport, stop-gradient, and kernel-mixture—behind a unified PyTorch interface so that the differentiability choice becomes a modular algorithmic component (Brady et al., 29 Oct 2025).

A third mechanism is analytic differentiation of numerically stable primitives. For square-root Kalman filters, the problematic object is the triangularization step, normally computed by QR. The robust approach differentiates the Gramian identity

$LL^\top = MM^\top$

rather than the QR factors themselves, thereby avoiding non-uniqueness and singular Jacobians in rank-deficient settings (Corenflos, 13 Mar 2026). For direct-form digital filters, the filter is rewritten in state-space form and the backward pass is expressed as another recursion of the same type, giving closed-form gradients with respect to inputs, coefficients, and initial conditions (Yu et al., 18 Nov 2025). In analytic cosmography, numerical quadrature is replaced by Carlson symmetric elliptic integrals whose derivatives remain within the same special-function family, making the entire layer GPU-friendly and auto-differentiable (Karchev, 2022).

A fourth mechanism is to replace fixed parametric updates by learned generative or attention-based updates. DnD Filter conditions a diffusion model on the predicted state and current observation features, so the Bayesian update is executed as a reverse denoising process rather than a Kalman gain computation (Wan et al., 3 Mar 2025). DiffPF applies the same principle to particle filtering by replacing weighted-resampling updates with equally weighted sampling from a conditional diffusion posterior (Wan et al., 21 Jul 2025). In multimodal soft-robot estimation, a-MDF replaces the conventional gain with an Attention Gain module; the paper explicitly frames the conventional gain mechanism as remaining non-differentiable and limiting adaptability, and uses attention to make it learnable and context-sensitive (Liu et al., 2023).

3. Recursive Bayesian filters, data assimilation, and Kalman-family methods

The most direct instantiations are differentiable variants of classical recursive estimators. A comparative study of dEKF, dUKF, dMCUKF, and differentiable particle-filter variants shows that the process model, sensor model, and noise models can all be learned inside the recursive structure, with negative log-likelihood favored when uncertainty calibration matters, heteroscedastic noise often beneficial, and sequence length 10 recommended as a practical training default (Kloss et al., 2020). The same literature emphasizes that dEKF is usually the simplest and most numerically stable entry point, whereas dUKF and dMCUKF are preferable under stronger nonlinearities (Kloss et al., 2020).

Multimodal differentiable filters preserve the same predict-update skeleton while learning how heterogeneous sensors should interact. In contact-rich manipulation, differentiable EKF and differentiable PF architectures are combined with feature fusion, unimodal weighted fusion, and crossmodal weighted fusion; the result is accuracy comparable to unstructured LSTM baselines while retaining modality-specific uncertainty semantics and interpretability (Lee et al., 2020). For soft robots, a-MDF pushes this further into latent-space ensemble filtering, where a transformer-style dynamics model propagates latent ensembles and attention replaces the analytic gain in the update step (Liu et al., 2023).

Ensemble methods are particularly important in high-dimensional data assimilation. AD-EnKF treats the full ensemble Kalman recursion as a computation graph and maximizes an EnKF likelihood surrogate,

$$\widehat{L}(\theta)=\sum_{t=1}^T \log \mathcal{N}\big(y_t; H\widehat{m}_t,\; H\widehat{C}_t H^\top + R\big),$$

with gradients propagated through the forecast and analysis steps via reverse-mode autodiff and reparameterized process noise (Chen et al., 2021). In the linear-Gaussian setting, the paper proves $O(N^{-1/2})$ convergence for both the likelihood estimate and its gradient, establishing a consistency result for the autodiff-trained EnKF surrogate (Chen et al., 2021).

A more general data-assimilation framework extends this logic to joint learning of state, dynamics, and filter parameters from partial, noisy observations. Its approximate observation-likelihood objective uses

$S_t = H_t C_t^{\theta,\phi} H_t^\top + R_t,$

and supports AD-3DVar-$C$, AD-EnKF, and AD-Ens3DVar as instances of the same framework (Adrian et al., 21 Mar 2026). The practical conclusion is not that one method dominates universally, but that ensemble-based variants tend to be more accurate while 3DVar-based variants are cheaper and sometimes preferable when the observation structure or computational budget is restrictive (Adrian et al., 21 Mar 2026).

Projection filters provide a different route to differentiability. Instead of propagating moments or particles, the filter density is projected onto a finite-dimensional exponential-family manifold

$$p_\theta(x)=\exp\!\left(c(x)^\top\theta-\psi(\theta)\right),$$

and automatic differentiation is used to recover the expected natural statistics and Fisher metric through

$$\eta(\theta)=\nabla_\theta \psi(\theta), \qquad g(\theta)=\nabla_\theta^2 \psi(\theta),$$

with sparse-grid quadrature supplying the log-partition function in multidimensional settings (Emzir et al., 2021). This yields finite-dimensional stochastic dynamics for the manifold parameters while avoiding the explicit curse of dimensionality of tensor-product quadrature (Emzir et al., 2021).

Finally, numerical robustness inside Kalman-family methods has itself become a differentiation problem. The square-root Kalman literature shows that all learning-relevant outputs depend on the triangularized factor only through its Gramian; differentiating $LL^\top=MM^\top$ yields exact JVPs and VJPs even when the QR factor is non-unique or $\vec y=\vec g(\vec x),$0 is rank-deficient (Corenflos, 13 Mar 2026). This moves auto-differentiable filtering from an implementation convenience toward a statement about stable differential calculus for the linear-algebra core of filtering.

4. Particle filters, diffusion updates, and the resampling problem

Particle filtering remains the most acute test case for differentiability because the resampling step is discrete. In a standard particle filter, the ancestor indices are sampled from a categorical distribution determined by the normalized weights, so infinitesimal parameter changes can induce discontinuous lineage changes. The stop-gradient correction method addresses this by leaving the forward pass unchanged while modifying the backward graph; the corrected weights recover classical Fisher-identity and Poyiadjis-style score estimators under automatic differentiation, and repeated autodiff also yields the corresponding second-order estimator (Ścibior et al., 2021).

The practical ecosystem around this idea is now broad. PyDPF implements non-differentiable resampling, soft resampling, optimal-transport resampling, stop-gradient resampling, and kernel-mixture resampling in a unified PyTorch framework (Brady et al., 29 Oct 2025). The package does not resolve the trade-offs; rather, it makes them explicit. Non-differentiable resampling is fast but biased in the backward pass, soft resampling introduces a controllable bias-variance compromise, optimal transport is smooth but $\vec y=\vec g(\vec x),$1 and potentially unstable, stop-gradient resampling has stronger theoretical support for certain objectives but can be high variance, and kernel-mixture resampling regularizes the particle cloud at the price of additional design choices (Brady et al., 29 Oct 2025).

Recent work replaces importance weighting altogether with learned posterior samplers. DnD Filter uses a DDPM-style reverse diffusion process conditioned on the predicted state and observation features, thereby removing the requirement for a Gaussian parametric update and allowing nonlinear posterior updates in state estimation (Wan et al., 3 Mar 2025). Its stagewise training strategy is reported to give a 20× speedup over straightforward iterative training, and on visual odometry the method achieves a 25% improvement in estimation accuracy relative to state-of-the-art differentiable filters (Wan et al., 3 Mar 2025).

DiffPF pushes the same logic into particle filtering itself. Instead of weighting and resampling particles, it learns a conditional diffusion model that samples equally weighted particles from an approximation to the filtering posterior conditioned on predicted particles and the current observation (Wan et al., 21 Jul 2025). The reported gains are substantial: an 82.8% improvement in estimation accuracy on a highly multimodal global localization benchmark and a 26% improvement on KITTI visual odometry compared to state-of-the-art differentiable filters, while still operating in real time at 59 Hz with 5 diffusion steps on disk tracking and around 46.5 Hz on KITTI (Wan et al., 21 Jul 2025). Conceptually, these methods recast the update step as conditional generative modeling, which is a distinct departure from both Kalman-style Gaussian corrections and importance-weighted SMC.

5. Filtering-like operators beyond classical state estimation

Auto-differentiable filtering also appears in problems where “filtering” is better understood as adaptive regularized inference over structured domains. In strong gravitational lens modelling, the ray-guided transformed uniform grid defines the source as a Gaussian process on a regular grid,

$\vec y=\vec g(\vec x),$2

and then warps the coordinates through smoothed marginal eCDFs of back-traced rays (Enzi et al., 29 Jun 2026). The result is adaptive source resolution without Delaunay tessellations or Voronoi binning, preservation of FFT-based Gaussian-process regularization, and full compatibility with JAX/NumPyro autodiff (Enzi et al., 29 Jun 2026). Empirically, the method reaches comparable fit quality with about a factor of two fewer pixels per dimension, increases ELBOs at fixed pixel count, and only mildly changes the ELBO difference between lens models with and without substructure (Enzi et al., 29 Jun 2026).

In robotics, DiffCo turns collision checking into a smooth filter-like constraint over configuration space. It learns a sparse multi-label kernel perceptron and evaluates a polyharmonic radial-basis collision score whose gradients are backpropagated through optimization (Zhi et al., 2021). The score is explicitly described as a pseudo-distance or margin score rather than a true geometric distance, but it is analytically differentiable and often 10–100x faster to compute than geometric collision checking in trajectory-optimization settings (Zhi et al., 2021).

Point-cloud denoising provides another variant. 3DMambaIPF combines iterative point-cloud filtering with a differentiable rendering loss, using a Mamba-based encoder-decoder to process large point clouds patchwise while image-space view consistency constrains boundary geometry (Zhou et al., 2024). The method reports state-of-the-art performance on datasets with typically up to 50K points and demonstrates scalability to models with about 500K points; its ablations further indicate that adding view loss improves performance consistently, with 32 rendered views performing best in the reported study (Zhou et al., 2024). Here, auto-differentiable filtering means that the denoising criterion is not only a 3D reconstruction loss but also a differentiable rendering-based surface-consistency loss.

Classical signal processing has also been absorbed into this program. A general treatment of direct-form digital filters derives closed-form reverse-mode rules, including gradients with respect to initial conditions, from the state-space recurrence underlying type-II transposed direct forms (Yu et al., 18 Nov 2025). Custom C++/CUDA implementations in PyTorch are reported to achieve at least 1000× speedup over naive Python implementations, and for low-order filters the exact time-domain method with analytical gradients is faster than frequency-domain approximations (Yu et al., 18 Nov 2025).

A related, filtering-adjacent development occurs in cosmological inference. Analytic $\vec y=\vec g(\vec x),$3CDM cosmography replaces quadrature-based distance calculations by Carlson symmetric elliptic integrals implemented in PyTorch, so gradients propagate through redshift-distance relations inside variational-inference pipelines (Karchev, 2022). The reported application scales to up to $\vec y=\vec g(\vec x),$4 mock type Ia supernovae, illustrating how auto-differentiable scientific primitives can serve the same role for Bayesian inference that differentiable filters serve for state estimation (Karchev, 2022).

6. Limitations, misconceptions, and research directions

A common misconception is that “differentiable” implies exact or unbiased gradients. The particle-filter literature makes the opposite point. Non-differentiable resampling can be fast but biased in the backward pass; soft resampling, optimal transport, and kernel-mixture schemes exchange bias, variance, and computational cost in different ways; and even the stop-gradient correction, although theoretically aligned with classical score estimators, does not eliminate particle degeneracy or the difficulty of proposal learning (Brady et al., 29 Oct 2025, Ścibior et al., 2021).

A second misconception is that a smooth proxy is interchangeable with the underlying physical quantity. DiffCo’s score is not a true geometric distance, only a pseudo-distance or margin in kernel space (Zhi et al., 2021). In RTU-grid lens modelling, the field remains Gaussian after transformation but with covariance $\vec y=\vec g(\vec x),$5, so adaptivity changes the effective correlation structure rather than merely redistributing pixels (Enzi et al., 29 Jun 2026). In diffusion-based filters, the update is flexible precisely because it is learned; if the process prior is poor or the conditional diffusion model is mis-specified, posterior sampling quality degrades accordingly (Wan et al., 21 Jul 2025).

There are also explicit modeling restrictions. The general data-assimilation framework that jointly learns states, dynamics, and filter parameters relies on linear observation operators $\vec y=\vec g(\vec x),$6 and on a Gaussian approximation to the forecast distribution in its likelihood construction (Adrian et al., 21 Mar 2026). Analytic cosmography covers cases in which $\vec y=\vec g(\vec x),$7 is a polynomial of degree at most 4, not arbitrary time-varying dark-energy models (Karchev, 2022). The square-root Kalman differentiation rule is exact for objectives that depend on the factor only through its Gramian, which is precisely why the Gramian differential replaces direct QR differentiation (Corenflos, 13 Mar 2026).

The broader trajectory of the field is nevertheless clear. Online, label-free adaptation under noise-statistics drift is now feasible through optimal-transport objectives in adaptive Kalman filtering (He et al., 9 Aug 2025). Adaptive discretization no longer has to break autodiff, as shown by smooth ray-guided coordinate warps in strong lensing (Enzi et al., 29 Jun 2026). Learned posterior samplers can now replace hard weighting-and-resampling pipelines in multimodal filtering problems (Wan et al., 21 Jul 2025). This suggests that future work will continue to merge three previously conflicting goals: expressive non-Gaussian inference, numerical stability of the underlying filter primitives, and end-to-end differentiability suitable for large-scale scientific and robotic deployment.