{"id":"a99eb365-1ca0-4473-b116-4dfca7cd44b1","arxiv_id":"2411.15638","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"StateMixNN learns particle-filter transition and proposal densities as Gaussian mixtures parameterized by neural networks, trained only on the observation likelihood, and reports improved state recovery on Lorenz 96 and Kuramoto systems.","lead":"StateMixNN is a new method that learns both the way a hidden system evolves and the best guessing distribution for a particle filter, using two neural networks that output Gaussian mixture densities and training only on observed data. It matters because it offers a path to filtering in state-space models where the transition dynamics are unknown, a common situation in tracking, finance, ecology, and weather.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training objective in Eq. (11) is not the log-likelihood: sum-of-log-weights is not log-sum-exp, so the claim of likelihood-based identification of the transition and proposal is unsupported.","rationale":"The reader's CONDITIONAL verdict remains appropriate, but the most load-bearing flaw is more specific than the identifiability concern. The identifiability of f versus π would be substantially mitigated if the training loss were the true marginal log-likelihood: asymptotically, the likelihood of the observations depends on f and g but not on the proposal π, so π mainly affects variance. The paper's Eq. (11) instead makes π directly enter a sum-of-log-weights objective, so f can be traded against π in a way that is not a finite-sample artifact. This explains both the acknowledged identifiability issue and the absence of any validation of the learned transition against the true kernel. The empirical RIMSE results are plausible and consistently reported, and the method is clearly specified apart from this issue, so I would not reject the paper outright. I would require (a) correcting Eq. (10)/(11) or explicitly stating that the loss is a surrogate, (b) releasing code so the implemented loss can be checked, and (c) reporting a comparison between the learned transition and the true transition in the two test systems. With those changes the central state-recovery claim could stand; without them, the paper's central theoretical justification is unsupported.","tokens_in":21858,"tokens_out":13012,"duration_ms":135054,"concrete_test":"Re-run the Lorenz96 experiment (K=100, T=100, S∈{6,10}, same hyperparameters) with the loss replaced by the standard filtered log-likelihood estimator ℓ = Σ_t [logsumexp(log w_t) − log K], and compare against the results in Figs. 4–5. Additionally, evaluate the learned transition density f̂ on a grid of (x_{t-1}, x_t) pairs and compare it with the true Gaussian-noise transition of Eq. (14) via KL divergence or a two-sample test. If RIMSE remains below 1 and f̂ approaches the true kernel, the central claim survives. If RIMSE degrades and/or f̂ is far from the true kernel, the reported gains come from the sum-of-log-weights heuristic plus a compensating proposal, not from likelihood-based learning of the state dynamics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2, Eq. (11), states that after resampling at every step the training objective is ℓ(θ|y1:T) ∝ Σ_t Σ_k log w_t^(k). The standard log-marginal-likelihood estimator of the SIR particle filter is Σ_t log(Σ_k w_t^(k)) − T log K. For K>1 these are not proportional: the log of a sum is not the sum of logs. The paper's supporting statement, 'the log-likelihood is maximised when all weights are equal,' is a property of the sum-of-log-weights surrogate, not of the marginal likelihood of the observation series. Consequently, the learned θ(f) and θ(π) are not maximum-likelihood estimates. With the stated loss, the objective can be increased by making the ratio f/π large at sampled particles, so the learned transition density can adapt to the proposal and to observed y values rather than to the true dynamics. This makes the Section 4.3 identifiability concern concrete: the coupling between f and π is not an asymptotic artifact but is built into the loss. The central claim that StateMixNN learns state dynamics 'given only the observation model' therefore rests on an incorrect likelihood equation. If the implementation actually uses the log-sum-exp loss, this is an exposition error; as written, the training objective is a heuristic that rewards weight concentration around observations, not likelihood maximization.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes StateMixNN, a differentiable particle filter (DPF) in which the transition distribution and the proposal distribution of a state-space model are both approximated by multivariate Gaussian mixtures whose means and diagonal covariances are outputs of neural networks. The networks are trained by maximizing an estimated parameter log-likelihood using the stop-gradient DPF of Sciberior and Wood, requiring only observations and a known observation model. The method is evaluated on stochastic Lorenz 96 and Kuramoto oscillator systems, with reported relative mean squared error (RIMSE) against a bootstrap particle filter, across varying numbers of particles, series lengths, state noise levels, and state dimensions. The central empirical claim is that StateMixNN improves hidden-state recovery relative to the bootstrap filter and the improved auxiliary particle filter, especially in highly nonlinear regimes.","tokens_in":22125,"tokens_out":4269,"duration_ms":40103,"significance":"If the method works as claimed, it would be a practically valuable contribution: it offers a way to perform filtering when the transition kernel is unknown, a setting where standard particle filters cannot be applied directly. The empirical study is thorough, with 200 independent runs and 95% intervals for two nonlinear systems and four experimental axes. The paper also provides a useful discussion of mixture parametrizations, conditioning updates, and likelihood degeneracy. However, the significance hinges on the training objective being a genuine likelihood estimator and on the learned transition being the actual dynamics; both points are currently not established.","major_comments":[{"comment":"The training objective in Eq. (11) is not the particle-filter log-likelihood. After resampling at every step, the standard unbiased estimator of the log marginal likelihood is sum_t log( (1/K) sum_k w_t^(k) ), i.e., a sum over time of log-sum-exp terms, whereas Eq. (11) is sum_t sum_k log w_t^(k). For K > 1 these are not proportional: the log of a sum is not the sum of logs. The statement in Section 3.2 that 'the log-likelihood is maximised when all weights are equal' is a property of the sum-of-log-weights surrogate, not of the marginal likelihood of the observation series. Consequently, the learned theta^(f) and theta^(pi) are not maximum-likelihood estimates, and the claim in the abstract and Section 1 that the method is 'trained targeting the log-likelihood' is unsupported. If the implementation actually computes log-sum-exp, this is an exposition error; as written, the loss is a heuristic that rewards weight concentration around sampled particles and can be increased by making the ratio f/pi large at sampled locations, allowing the learned transition to adapt to the proposal and observations rather than to the true dynamics. This issue is load-bearing because the central claim of learning state dynamics 'given only the observation model' rests on the likelihood interpretation of the training objective.","section":"Section 3.2, Eq. (11)"},{"comment":"The learned transition kernel is never validated against the true transition kernel. Section 4.3 acknowledges that changes in the log-likelihood can be attributed either to the transition or to the proposal, and that the two distributions interact in the weight computation; the paper assumes the alternating scheme resolves this, but alternating conditional updates only attribute gradient changes to the updated network, they do not identify the pair (f, pi). Because the importance weights depend on f and pi only through the ratio f/pi, many pairs give the same likelihood. The reported RIMSE improvements in Figures 4-11 could come from a proposal that compensates for an inaccurate transition rather than from learning the actual dynamics. A direct validation of the learned f against the true kernel (e.g., a KL divergence or predictive comparison on held-out transitions) is necessary to support the claim that StateMixNN learns state dynamics; without it, the method's advantage over baselines that require the true transition is not attributable to correct dynamics learning.","section":"Section 4.3 and Section 5"}],"minor_comments":[{"comment":"The abstract claims the method 'significantly improves recovery of the hidden state in comparison with the state-of-the-art', but the conclusion (Section 6) states that StateMixNN 'performs similarly to PropMixNN', a method that requires the true transition. The abstract should be qualified to indicate that the significant improvements are relative to bootstrap and improved auxiliary particle filters, not to PropMixNN.","section":"Abstract and Section 6"},{"comment":"In Eq. (12), the proposal density in the last term is written as pi(xt|...) with xt rather than x_t^(k); this is a typographical inconsistency that may confuse readers.","section":"Section 3.2, Eq. (12)"},{"comment":"The term 'relative improvement in MSE' is used for RIMSE = MSE_method / MSE_baseline, which is a ratio rather than an improvement; consider renaming it 'relative MSE' or defining the improvement as 1 - RIMSE to avoid ambiguity.","section":"Section 5.1"},{"comment":"The telescoping observation batching in Algorithm 4 is a heuristic, and the claim that it prevents behavior where only the start and end of the series are well represented is not empirically demonstrated; a sensitivity analysis or an ablation for the batch schedule would strengthen the presentation.","section":"Section 4.4"}],"recommendation":"major_revision","confidential_remarks":"The Eq. (11) issue is the key technical point. If the authors can show that the implementation actually uses a log-sum-exp estimator, then the paper may be salvageable by correcting the equations and the surrounding text. If the loss is indeed sum-of-log-weights, the method should be reframed as a heuristic contrastive objective, and the claims about likelihood-based identification and learning the true transition should be substantially weakened. The missing validation of the learned transition is the other main obstacle. The empirical work is otherwise solid and well presented, and the paper fits the scope of Signal Processing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"StateMixNN is a sensible engineering extension of the authors' PropMixNN: it learns both the transition kernel and the proposal as Gaussian mixtures whose parameters come from neural networks, and it trains everything end-to-end inside a stop-gradient differentiable particle filter. The experiments are systematic and the results are consistent: across Lorenz 96 and Kuramoto, with varying particles, series length, noise, and dimension, StateMixNN beats BPF and IAPF on state MSE. That is real evidence the method works as a heuristic.\n\nThe problem is the training loss. Equation (11) says ℓ(θ|y1:T) ∝ Σ_t Σ_k log w_t^(k). That is not the log-likelihood. The standard SIR marginal-likelihood estimator is Σ_t log Σ_k w_t^(k) (up to constants), and the log of a sum is not the sum of logs. The paper's objective is a weight-concentration surrogate, not a likelihood. So the learned transition and proposal are not maximum-likelihood estimates. This matters because the paper's central claim—learning the hidden dynamics 'given only the observation model'—rests on likelihood-based identification. With the stated loss, the transition density can adapt to the proposal and to the observations, rather than to the true dynamics. Section 4.3 acknowledges an identifiability problem, but the alternating scheme only stabilizes optimization; it does not resolve the coupling.\n\nOther soft spots: the learned transition is never compared with the true kernel; there is no baseline that also learns the transition; the abstract says 'significantly improves' over state-of-the-art, but Section 6 says StateMixNN 'performs similarly' to PropMixNN, which requires the true transition. The Gumbel-Softmax temperature is unspecified, and no code is released. These are addressable.\n\nCredit where due: the method is clearly specified, the experiments are thorough, and jointly learning f and π is a sensible step beyond PropMixNN. A clean derivation of the correct loss, an ablation separating the contributions of the two networks, and a comparison against a learned-transition baseline would make the paper much stronger. I would not cite it in its current form, but it deserves peer review because the empirical method is plausible and the conceptual error is correctable. If you use it in a reading group, it is a good case study in how easily a surrogate loss gets mislabeled as a likelihood.","headline":"A promising empirical method whose central likelihood claim is mathematically wrong; fix Eq. (11) and validate the learned transition before trusting the 'learns dynamics' story.","tokens_in":22666,"tokens_out":4650,"would_cite":false,"duration_ms":40386,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M05","62F15","65C05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"StateMixNN recovers hidden states without knowing the dynamics","keywords":["state-space models","particle filters","differentiable particle filters","Gaussian mixtures","proposal distribution learning","transition kernel learning","neural networks","sequential Monte Carlo"],"falsifier":"In the Lorenz 96 or Kuramoto synthetic setting, train StateMixNN and then compare the learned transition mixture $\\hat f$ with the known true transition density over a grid of $(x_{t-1}, x_t)$ pairs. If $\\hat f$ is close to the true $f$ and filtering MSE stays low, the paper's attribution is supported; if $\\hat f$ is far from the true $f$ while MSE remains low, the improvement is a proposal artifact.","tokens_in":21620,"feed_emoji":"🧠","tokens_out":9701,"duration_ms":85017,"temperature":0.7,"pith_summary":"StateMixNN sets out to solve a practical version of the filtering problem: recover a hidden state when the transition dynamics of the system are unknown, given only the observation model and an observation series. It replaces the unknown transition kernel and the proposal distribution of a particle filter by two equally weighted multivariate Gaussian mixtures whose means and diagonal covariances are outputs of dense neural networks. Training maximizes the estimated log-likelihood through a stop-gradient differentiable particle filter, so no hidden-state labels are needed. In experiments on the chaotic Lorenz 96 and Kuramoto oscillator systems, StateMixNN reports relative mean-square error below 1 against the bootstrap filter across tested particle counts, series lengths, noise levels, and dimensions, and it outperforms the improved auxiliary particle filter without requiring the transition kernel.","feed_headline":"StateMixNN recovers hidden states without knowing the dynamics","feed_subtitle":"Trained on observation likelihood alone, it outperforms standard filters on chaotic systems.","key_machinery":"The load-bearing object is the particle weight equation $w_t = g(y_t|x_t) f(x_t|x_{t-1}) / \\pi(x_t|x_{t-1}, y_t)$, together with the two networks that produce $f$ and $\\pi$. Each network is a multilayer perceptron whose final layer emits, for each of $S$ mixture components, a $d_x$-dimensional mean and a $d_x$-dimensional covariance scale vector; the covariance is diagonal, so $C^{(s)} = \\mathrm{diag}(c^{(s)})^2$, and the mixture weights are fixed equal. The transition network takes only the previous particle, preserving Markovianity; the proposal network takes the previous particle and the current observation. Training uses the stop-gradient differentiable particle filter so gradients of the log-likelihood flow through resampling, with the Gumbel-Softmax reparameterization for the mixture component draw and the reparameterization trick for each Gaussian. Parameters are learned in an alternating scheme, proposal conditional on transition and then transition conditional on proposal, with telescoping observation batches to avoid likelihood degeneracy.","core_discovery":"The paper's central claim is that by estimating both the transition distribution $f(x_t|x_{t-1})$ and the proposal distribution $\\pi(x_t|x_{t-1}, y_t)$ from the observation series alone, one can estimate the hidden state given only the observation model $g$. This is new because standard particle filters and learned proposals such as PropMixNN assume $f$ is known; StateMixNN drops that requirement by training a transition network and a proposal network in alternation. The two distributions are equal-weight mixtures of diagonal multivariate Gaussians, with component parameters generated by multilayer perceptrons, and training targets the observation log-likelihood via a stop-gradient differentiable particle filter. The reported result is that StateMixNN obtains relative mean-square error values below 1 compared with the bootstrap filter for the Lorenz 96 and Kuramoto systems across the tested ranges, with larger gains in highly nonlinear regimes and with more mixture components.","pith_inferences":["Editorial inference: the reported state-recovery gain does not by itself prove that the learned transition kernel is the true one; the paper's own identifiability discussion implies a check that compares the learned kernel with known dynamics in a synthetic setting.","Editorial inference: a natural stress test beyond the paper is to freeze the trained transition network and run filtering with a bootstrap proposal; if performance collapses, the learned proposal is doing the lifting rather than the learned dynamics.","Editorial inference: the architecture is deliberately simple, so the same likelihood-based training could be tried with heavier-tailed mixture components, such as Student-t mixands, to test whether Gaussian mixtures limit performance on heavy-tailed chaotic transitions."],"forward_implications":["Filtering becomes possible for systems whose dynamics are unknown, as long as the observation model is specified, because the transition kernel no longer needs to be supplied.","Trained particle filters can beat both the bootstrap filter and an observation-aware auxiliary filter on chaotic, multimodal systems, with the margin growing as state noise increases.","The learned proposal effectively carries the observation information that the transition alone lacks, which should help exactly when the bootstrap filter degenerates: diffuse transition, informative observation.","The training target needs only the observation series, so the method applies to real-world datasets where hidden states are never recorded."],"supporting_citations":[{"why":"Provides the bootstrap particle filter used as the RIMSE baseline and as the initial transition estimate in training.","marker":"[8]"},{"why":"Supplies the stop-gradient differentiable particle filter through which gradients of the log-likelihood are computed.","marker":"[13]"},{"why":"Defines the improved auxiliary particle filter (IAPF) that StateMixNN is compared against as an observation-aware proposal baseline.","marker":"[20]"},{"why":"Introduces PropMixNN, the predecessor method that learns only the proposal and is used as a baseline requiring known transition dynamics.","marker":"[23]"},{"why":"Supplies the likelihood-based parameter estimation setup and the log-weight numerical stabilization used in the training loss.","marker":"[11]"}],"fun_headline_variants":["StateMixNN learns hidden dynamics from data alone","Neural particle filter picks up transition and proposal","No known equations? Neural filter reveals hidden states","Chaotic systems: StateMixNN outperforms without fixed dynamics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's load-bearing premise is that maximizing the observation log-likelihood can separate the transition kernel from the proposal, so the learned transition is the true dynamics rather than one half of a compensating pair $(f, \\pi)$ that only makes the filter's weights look good.","fun_headline_variants_meta":{"raw":{"variants":["StateMixNN learns hidden dynamics from data alone","Neural particle filter picks up transition and proposal","No known equations? Neural filter reveals hidden states","Chaotic systems: StateMixNN outperforms without fixed dynamics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000331,"raw_usage":{"total_tokens":1796,"prompt_tokens":852,"completion_tokens":944,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":882}},"tokens_in":468,"tokens_out":944,"duration_ms":8837,"temperature":1.0,"reasoning_tokens":882,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:05:05.789124+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the Lorenz 96 or Kuramoto synthetic setting, train StateMixNN and then compare the learned transition mixture $\\hat f$ with the known true transition density over a grid of $(x_{t-1}, x_t)$ pairs. If $\\hat f$ is close to the true $f$ and filtering MSE stays low, the paper's attribution is supported; if $\\hat f$ is far from the true $f$ while MSE remains low, the improvement is a proposal artifact.","supporting_citations":[{"cited_title":"Novel approach to nonlinear and non- Gaussian Bayesian state estimation","cited_arxiv_id":null,"evidence_quote":"Provides the bootstrap particle filter used as the RIMSE baseline and as the initial transition estimate in training."},{"cited_title":"Bugallo, and P etar M","cited_arxiv_id":null,"evidence_quote":"Defines the improved auxiliary particle filter (IAPF) that StateMixNN is compared against as an observation-aware proposal baseline."},{"cited_title":"E nd- to-end learning of gaussian mixture proposals using di fferentiable particle filters and neural networks","cited_arxiv_id":null,"evidence_quote":"Introduces PropMixNN, the predecessor method that learns only the proposal and is used as a baseline requiring known transition dynamics."},{"cited_title":"Cambridge University Press, 1st edition, 2013","cited_arxiv_id":null,"evidence_quote":"Supplies the likelihood-based parameter estimation setup and the log-weight numerical stabilization used in the training loss."}],"review_version":1}