{"id":"b55058b5-5d7d-4dab-86f7-d2e6c8cfb21b","arxiv_id":"2608.07754","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Frequency-locked resonate-and-fire neurons are recast as a complex state-space model with phase outputs, enabling FFT-based parallel training and a new bridge to hyperdimensional computing, subject to a derivation gap.","lead":"A new state-space formulation for resonate-and-fire spiking neurons, called the phase state-space model, encodes information in the phases of frequency-locked oscillations and claims parallel, surrogate-free training. The paper demonstrates the idea on a copying task and sequential FashionMNIST, combining a short-time Fourier transform, attention, and hyperdimensional computing readouts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 14's convolution index is one step off from the recurrence in Eq. 13, so the FFT/convolution view is not shown to compute the same U(t) as the discrete view; the central three-view equivalence claim is unsupported as written.","rationale":"The reader's verdict of REJECT is supported, but the specific weakest_assumption they chose (the Dirac-spike/phase-encoding input assumption) is not the most load-bearing concern. The most direct and unambiguous flaw is the off-by-one inconsistency between the discrete recurrence (Eq. 13) and the convolutional view (Eq. 14). This is a purely internal mathematical error: unrolling Eq. 13 for two steps immediately shows that Eq. 14 as written includes the current input term that the recurrence does not. Since the paper's headline contribution is the three equivalent execution views, and the FFT/parallel training view depends on the correctness of Eq. 14, this invalidates the central claim as presented. It is not a disagreement with external consensus but a failure of internal consistency. The reader's chosen weakest assumption is a valid limitation of the model (real inputs, finite-width spikes, or multiple spikes per period would break the phase-spike encoding), but it is more about the model's scope than about the correctness of the equivalence within the stated model. I therefore partially agree with the reader: the verdict REJECT is appropriate, and the off-by-one is a concrete, checkable flaw. A straightforward re-derivation or code comparison will settle whether the flaw is a typo or a substantive error; if it is a typo, the paper could be revised and resubmitted, but as submitted the central claim is not established.","tokens_in":8512,"tokens_out":15616,"duration_ms":142297,"concrete_test":"Run the provided Julia package (PhasorNetworks.jl) and independently implement the discrete recurrence (Eq. 13) and the convolution (Eq. 14) for L=4 timesteps with fixed random phase inputs and kernel K[n]=e^{knT}; compare the resulting U[n] sequences element-by-element. Also symbolically unroll Eq. 13 for n=1,2,3 and compare with the convolution sum in Eq. 14. If the package's outputs match Eq. 14 but not Eq. 13, the equivalence is genuinely false. If they match Eq. 13, then Eq. 14 is a typo and the claim could be repaired by correcting the convolution index to sum_{j=0}^{n-1} and re-running the numerical precision comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the continuous, discrete, and convolutional/FFT views produce identical R&F potentials. The derivation of the discrete view (Eq. 13) is: U[n+1] = e^{kT} U[n] + B U'(theta_n). Unrolling from U[0]=0 gives U[1] = B U'(theta_0) and U[2] = e^{kT} B U'(theta_0) + B U'(theta_1), so U[n] = sum_{j=0}^{n-1} e^{k(n-j)T} B U'(theta_j). However, Eq. 14 writes U[n] = sum_{j=0}^{n} K[n-j] H[j] with K[n]=e^{knT} and H[j]=B U'(theta_j). For n=1, Eq. 14 gives U[1] = e^{kT} B U'(theta_0) + B U'(theta_1), which disagrees with Eq. 13 by the extra term B U'(theta_1). The correct causal convolution is sum_{j=0}^{n-1} K[n-j]H[j], or equivalently K[0] must be zero. Because the FFT/parallel training view is built directly on Eq. 14, the claimed equivalence between the parallel and recurrent modes is not established by the text. This is an internal mathematical inconsistency, not a matter of disagreement with prior work. The paper's reference to Fig. 1(d-f) cannot resolve the issue here because the figures and numerical results are not included, and the empirical section reports no quantitative results. The underlying idea may be repairable by correcting the index, but as submitted the three-view equivalence claim is not proven.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Phase State Space Model (phi-SSM), a reformulation of resonate-and-fire (R&F) neurons as a complex-valued linear state-space model with a diagonal transition matrix. It derives three execution modes: a continuous differential view, a discrete recurrence sampled at the resonant period, and a convolutional/FFT view. The paper further interprets the discrete update in terms of hyperdimensional (HD) computing operations (bundle, bind, similarity) and presents a sequence-processing network combining an STFT adapter, an attention-like module, and a codebook readout. The central claimed contribution is that the three views produce identical neural potentials, enabling parallel, surrogate-free training of R&F networks, and the paper reports qualitative ablation results on a copy task and sequential FashionMNIST.","tokens_in":8922,"tokens_out":8970,"duration_ms":85748,"significance":"If the three-view equivalence were correct, the paper would provide a useful bridge between spiking R&F networks and parallel SSM training, avoiding surrogate gradients while retaining biological features of R&F neurons. The derivation is transparent and builds on a standard linear-systems core, and the open-source release is a practical strength. However, the central equivalence as written contains a causal-index error in the convolution, and the HD interpretation discards magnitude information without proof. The empirical section reports no quantitative results, so the practical claims are not yet substantiated. The underlying idea is promising and likely repairable, but the present manuscript does not establish its main claim.","major_comments":[{"comment":"Equation (14) is not the unrolled form of Eq. (13). With U[0]=0, Eq. (13) gives U[1]=B U'(theta_0) and U[2]=e^{kT} B U'(theta_0) + B U'(theta_1), i.e. U[n] = sum_{j=0}^{n-1} e^{k(n-j)T} B U'(theta_j). Equation (14) instead sums j=0..n, which adds the non-causal term B U'(theta_n) to U[n]. Because the FFT/parallel training view is built directly on Eq. (14), the claimed exact equality of the discrete and convolutional views is not established as written. Please correct the convolution limits (or set K[0]=0) and re-derive the kernel accordingly.","section":"2.4, Eq. (14)"},{"comment":"Equation (20) presents the HD bundle/bind form as an equality, but the underlying complex recurrence in demodulated coordinates is V[n+1] = e^{lambda T} |U[n]| e^{i pi Y[n]} + |B| e^{lambda delta t} e^{i pi (1-theta)}, where the two terms generally have different magnitudes. The phase of this weighted sum is not equal to bundle(Y[n], 1-theta) as defined in Eq. (17), which operates on unit phasors. The phrase 'subject to magnitude decay' does not remove the magnitude weighting; the equality holds only if magnitudes are equal or the bundle operation is redefined to include magnitudes. Please either prove the equality under an explicit normalization condition or reframe Eq. (20) as an analogy rather than an equivalence.","section":"2.5, Eq. (20)"},{"comment":"The paper asserts numerical agreement 'within numerical precision' and improved performance from the STFT adapter and attention module, but the submitted text contains no figures and no quantitative results: no accuracies, losses, error bars, or task details are reported. Consequently the demonstration and the three-view equivalence claim are unverifiable. Please include the actual figure content or tabulated results, with task definitions and hyperparameters.","section":"3.4 and Fig. 1(d-f)"},{"comment":"The exact equivalence among the three views is derived for an input that is a sequence of Dirac-delta spikes, at most one per neuron per resonant period, with spike timing mapped from phase by Eq. (11). This restriction is not stated as a limitation. For general real-valued inputs the continuous-time dynamics are not represented by the discrete/convolutional kernels unless the ZOH adapter of Sec. 3.1 is used. Please state this input assumption explicitly and discuss its scope, since the abstract claims compatibility with both real and spiking inputs.","section":"2.4, input assumption"}],"minor_comments":[{"comment":"The boundary convention is unclear: with theta=1, Eq. (11) gives delta t=0 and t_spike=T, which coincides with the next sample point; please clarify how spikes exactly at sample boundaries are handled.","section":"2.4, Eq. (11)"},{"comment":"The attention scores A_ij are not normalized across j, since the softmax is deliberately avoided; calling these 'attention scores' without noting that the sum over j need not be 1 may mislead readers.","section":"3.2, Eq. (24)"},{"comment":"Several references are incomplete: [5] lacks a year and venue, and [13] and [14] would benefit from full bibliographic details (venue, publication status).","section":"References"},{"comment":"The caption says the convolutional/FFT view 'requires the storage of all timesteps previously solved iteratively,' but in parallel training all timesteps are computed simultaneously; the wording should distinguish inference-time storage from training-time computation.","section":"Fig. 1(c) caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is single-author and draws substantially on the author's prior work [13,14] for the HD-computing connection; this is not improper, but the novelty relative to those papers should be clarified. The empirical section is currently too thin for a full archival paper, and the index error in Eq. (14) must be fixed before the central equivalence claim can be accepted. If the authors provide corrected derivations and quantitative results, the revised version could be suitable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Wilkie,\n\nQuick take: the paper has a real idea—a phase-domain SSM for resonate-and-fire neurons with a discrete update at the resonant period and an explicit hyperdimensional-computing interpretation—and it deserves a revision, not a dismissal. But the central equivalence claim, the thing that makes the parallel training view credible, is not actually proven as written.\n\nThe genuinely new parts are the phase-of-spike encoding, the fixed-frequency relative-phase readout, and the HD bundling/binding interpretation. Those are not in the P-SpikeSSM or Huber et al. papers cited. The mathematical core reduces to standard diagonal-SSM discretization, but the packaging is original and worth taking seriously.\n\nNow the soft spots. Equation 14 is off by one. Unrolling Eq. 13 from U[0]=0 gives U[n] = sum_{j=0}^{n-1} e^{k(n-1-j)T} H[j], but Eq. 14 has K[n-j] with K[n]=e^{knT}, which adds an extra e^{kT} factor and includes a j=n term B U'(theta_n). So the FFT/convolution view as written does not compute the same U(t) as the recurrence. This is likely fixable by redefining K or shifting indices, but the claimed exact equality of the three views is unsupported. The 'identical within numerical precision' figures don't help here because they are not included in the arXiv text, and there are no numbers in the body.\n\nEquation 20's HD interpretation also drops magnitudes: bundling uses only phases, while the actual recurrence carries decay and input magnitudes. It's a useful intuition, not an exact equivalence, and the paper should say so.\n\nThe Dirac-delta input assumption—one spike per neuron per resonant period, exact timing encoded in phase—is stated nowhere as a limitation and is not validated against continuous-time simulation with realistic spikes. The empirical section is very thin: two tasks, 20 epochs, no numbers, no error bars, no hyperparameter details, just a reference to a figure that isn't present.\n\nWho is this for? People working on spiking SSMs, R&F networks, and HD computing. The architecture ideas—STFT adapter, attention via HD similarity—are interesting but need proper evaluation.\n\nRecommendation: send it to peer review, but require a fixed convolution index, a clean statement of what the three-view equivalence actually is, and real quantitative results with code. The off-by-one is a small fix, but it is load-bearing because the whole pitch is the parallel view.","headline":"A genuinely interesting phase-domain formulation of R&F SSMs, but the central three-view equivalence is unproven as written due to an off-by-one convolution index, and the empirical section is too thin to verify anything.","tokens_in":9422,"tokens_out":4089,"would_cite":false,"duration_ms":39617,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes the Phase State Space Model (φ-SSM), which recasts resonate-and-fire spiking networks so continuous, discrete, and convolutional execution views compute identical potentials, enabling parallel, surrogate-free training.","keywords":["spiking neural networks","state-space models","resonate-and-fire","phase encoding","relative phase","hyperdimensional computing","parallel training","surrogate-free learning"],"falsifier":"Run the continuous-time R&F equations (Eqn. 6) with a high-accuracy ODE solver using finite-width current pulses or two spikes within one resonant period, then compare that solution to the discrete recurrence and the FFT convolution; any disagreement beyond numerical tolerance would show the three-view equivalence holds only for ideal single Dirac spikes per period, not for the general inputs the paper claims to support.","tokens_in":8278,"feed_emoji":"⚡","tokens_out":8211,"duration_ms":71051,"temperature":0.7,"pith_summary":"This paper aims to give resonate-and-fire (R&F) spiking networks—neurons whose complex-valued potentials oscillate and emit spikes—the same parallel-training advantage that state-space models bring to recurrent networks. The author's central move is to lock layers of R&F neurons to a shared resonant frequency, making the relative phases between neurons constant over time and allowing spikes to be read as phase values rather than as discontinuous events. On this basis the paper defines the Phase State Space Model ($\\varphi$-SSM), claiming that the same network can be executed as a continuous differential equation, a discrete recurrence, or a convolutional/FFT kernel, with outputs identical within numerical precision. If true, spiking networks can be trained in parallel without surrogate gradient kernels while keeping recurrent, biologically plausible inference. The paper demonstrates the idea by building a network that combines a short-time Fourier transform adapter, attention, and a codebook readout, all communicating via relative phases.","feed_headline":"Phase state-space model makes spiking networks train in parallel","feed_subtitle":"Resonate-and-fire networks run recurrently yet train through a fast convolution with no surrogate spikes.","key_machinery":"The load-bearing object is the relative-phase representation of a frequency-locked R&F layer: $\\theta_{\\mathrm{rel}}(u,\\omega,t)=\\angle(u e^{-i\\omega t})$, which removes the global rotation of the resonant frequency and leaves a time-invariant phase code. The mechanism that connects the three execution views is the phase-to-time mapping $\\delta t(\\theta,k)=(2\\pi/\\Im(k))(1/2-\\theta/2)$, which converts a spiking input's decoded phase into the exact offset used in $U'(T,\\theta,k)=e^{k\\delta t}$. Substituting this into the linear R&F update gives $U[n+1]=e^{kT}U[n]+B\\,U'(T,\\theta,k)$, and unrolling that recurrence yields the convolution $U[n]=(K*H)[n]$ with kernel $K[n]=e^{k n T}$, which the FFT evaluates in parallel. The same discrete-time kernel also supplies the continuous and recurrent views, since all three are algebraic rearrangements of the same exponential transition.","core_discovery":"The central claim is that the phase of an R&F neuron's complex potential is a sufficient output to define a spiking state-space model, and that this phase-based model has three exactly equivalent execution views. By fixing the resonant frequency $\\omega$ of a layer, the relative phase $\\theta_{\\mathrm{rel}}(u,\\omega,t)=\\angle(u e^{-i\\omega t})$ is invariant through time, so a spike emitted at phase $\\theta$ can be converted into a contribution at the next sample time $T=2\\pi/\\omega$ via $U'(T,\\theta,k)=e^{k\\delta t(\\theta,k)}$, with $\\delta t(\\theta,k)=(2\\pi/\\Im(k))(1/2-\\theta/2)$. This yields the discrete recurrence $U[n+1]=e^{kT}U[n]+B\\,U'(T,\\theta,k)$, which unrolls into a causal convolution $U[n]=(K*H)[n]$ with kernel $K[n]=e^{k n T}$, computable in parallel by FFT. The paper reports that the continuous, discrete, and FFT implementations give outputs of $U(t)$ that agree within numerical precision, and uses the convolutional view to train networks on a copying task and sequential FashionMNIST. In the author's framing, this is an R&F-native alternative to surrogate-gradient training that retains the SSM's triple view and connects naturally to hyperdimensional computing.","pith_inferences":["Editorial inference: if the three-view equivalence is exact for Dirac-delta inputs, the same construction should transfer to other linear oscillatory neuron models whose envelopes follow a complex exponential, giving a general recipe for parallel training of resonant recurrent networks beyond resonate-and-fire neurons.","Editorial inference: the phase-based spike code suggests a testable hardware prediction—an analog physical oscillator implementing the continuous view should match the FFT training view as long as the resonator remains linear and frequency-locked.","Editorial inference: because the FFT view computes all timesteps in $O(L \\log L)$ work per layer, the $\\varphi$-SSM could make very long spiking sequences trainable end-to-end, a regime where backpropagation-through-time is impractical.","Editorial inference: treating spikes as phase symbols turns the spiking network into a differentiable vector-symbolic architecture, so techniques from memory capacity and error correction in HD computing may apply directly to the trained representations."],"forward_implications":["The same network can be executed as a continuous differential system, a discrete recurrence, or an FFT convolution, with outputs matching to numerical precision, so one network serves both real-time hardware and parallel GPU training.","Gradients flow through relative phases at every discrete time step for every neuron, including silent ones, without surrogate spiking kernels.","Spikes in a $\\varphi$-SSM layer can be interpreted as sparse phase communication, because relative phases of frequency-locked R&F neurons are invariant through time.","The $\\varphi$-SSM recurrence is equivalent to hyperdimensional computing operations (bundling, binding, similarity), linking recurrent SSM training tools to HD computing techniques.","The inclusion of an STFT adapter, attention, and codebook readout improves performance on the copying task and sequential FashionMNIST relative to ablations."],"supporting_citations":[{"why":"Supplies the state-space model theory of continuous, discrete, and convolutional views that the φ-SSM extends to R&F networks.","marker":"[5]"},{"why":"Defines the resonate-and-fire neuron model and its complex-valued differential update, the substrate of the paper.","marker":"[7]"},{"why":"Prior adaptation of SSM methods to R&F networks that the paper's phase-based formulation builds on and differentiates from.","marker":"[6]"},{"why":"Provides the frequency-locking relative-phase invariance and R&F implementations of HD operations that underlie the φ-SSM derivation.","marker":"[13]"},{"why":"Defines the FHRR bundling, binding, and similarity operations used to connect the φ-SSM to HD computing.","marker":"[9]"},{"why":"Describes surrogate gradient learning, the standard approach the paper's surrogate-free gradient method is designed to replace.","marker":"[12]"},{"why":"Shows an STFT implemented in neuromorphic hardware, motivating the STFT adapter layer using multi-compartment R&F neurons.","marker":"[15]"},{"why":"Supplies the residual and attentional vector-symbolic architecture used for the attention module built from φ-SSM projections.","marker":"[14]"}],"fun_headline_variants":["Phase encoding unlocks parallel training for spiking networks","Resonate-and-fire networks get surrogate-free parallel training","Spiking nets train via FFT with phase state-space model","No surrogate needed: phase SSM parallelizes spiking training","Phase view converts spiking networks to parallel convolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The discrete and convolutional kernels assume each input is a Dirac-delta spike carrying exactly one decoded phase per neuron per resonant period; if inputs are real-valued currents, finite-width pulses, or contain multiple spikes per period, the kernelized views are not proven to match the continuous R&F dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Phase encoding unlocks parallel training for spiking networks","Resonate-and-fire networks get surrogate-free parallel training","Spiking nets train via FFT with phase state-space model","No surrogate needed: phase SSM parallelizes spiking training","Phase view converts spiking networks to parallel convolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1278,"prompt_tokens":915,"completion_tokens":363,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":285}},"tokens_in":531,"tokens_out":363,"duration_ms":3278,"temperature":1.0,"reasoning_tokens":285,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:20:54.743028+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the continuous-time R&F equations (Eqn. 6) with a high-accuracy ODE solver using finite-width current pulses or two spikes within one resonant period, then compare that solution to the discrete recurrence and the FFT convolution; any disagreement beyond numerical tolerance would show the three-view equivalence holds only for ideal single Dirac spikes per period, not for the general inputs the paper claims to support.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the state-space model theory of continuous, discrete, and convolutional views that the φ-SSM extends to R&F networks."},{"cited_title":"Hyperdimensional Computing Provides a Programming Paradigm for Oscillatory Systems","cited_arxiv_id":"2312.11783","evidence_quote":"Provides the frequency-locking relative-phase invariance and R&F implementations of HD operations that underlie the φ-SSM derivation."},{"cited_title":"A Survey on Hyperdimensional Computing aka Vector Symbolic Architectures, Part II: Applications, Cognitive Models, and Challenges","cited_arxiv_id":"2112.15424","evidence_quote":"Defines the FHRR bundling, binding, and similarity operations used to connect the φ-SSM to HD computing."},{"cited_title":"Residual and Attentional Architectures for Vector-Symbols","cited_arxiv_id":"2207.08953","evidence_quote":"Supplies the residual and attentional vector-symbolic architecture used for the attention module built from φ-SSM projections."}],"review_version":1}