{"id":"21e33ec7-d360-4cd7-9fd8-24f0552d4987","arxiv_id":"2412.12104","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Reservoir computing can exploit time-variant, non-echo-state dynamics when a nonlinear readout transforms reservoir states into time-invariant outputs.","lead":"This paper proposes generalized reservoir computing, a framework in which the reservoir itself does not need to respond identically to identical inputs, as long as a trained nonlinear readout turns its time-varying states into stable outputs. The authors show that chaotic and oscillatory systems, including Lorenz 96 and a spin-torque oscillator, can then perform tasks such as emulating the Kuramoto-Sivashinsky equation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that any TV reservoir can be made TI by a nonlinear readout ('in general', TI Transformation section) is unproven and false for phase-modulated oscillators, where input and time are entangled by a 2π phase ambiguity.","rationale":"The reader's CONDITIONAL verdict is right: the paper has a plausible central idea and nontrivial demonstrations, but the abstract makes a universal claim outrunning the evidence. The single most load-bearing point is the existence of a TI readout for arbitrary TV reservoirs. The paper provides a proof only for two separable cases and uses MLP fits elsewhere; no theorem or counterexample analysis is provided. I identified a concrete counterexample class: phase-modulated oscillators. In that class, the state is a periodic function of the cumulative input, so the state history is invariant under shifting any input by 2π; a TI readout cannot recover the original input value. This directly contradicts the unqualified 'in general' and shows that a necessary condition (input observability, roughly that the input does not act through a wrapping coordinate without a known initial phase) is missing. The paper's own Supplementary S2 gives a negative result for linear readouts but no positive characterization for nonlinear ones; Supplementary S6 describes detrending/envelope extraction as heuristic post-processing, further indicating the lack of a general constructive method. This does not invalidate the specific experiments: the NARMA10, memory-capacity, and embedding results may well be correct for the chosen parameters and input ranges. It does mean the central claim should be restated as conditional on an input-observability condition, and the analytical/numerical evidence should be framed as existence proofs for example classes rather than a general framework. Therefore the reader's CONDITIONAL verdict is unchanged; the paper should condition acceptance on either a theorem characterizing when the TI transformation exists or a restriction of the claims.","tokens_in":26672,"tokens_out":26642,"duration_ms":256405,"concrete_test":"Analytical check: for the model x_t=(cos θ_t, sin θ_t), θ_{t+1}=θ_t+ω+u_t, write down the two trajectories generated by input histories u and u' with u'_0=u_0+2π and u'_i=u_i for i>0. Verify that (x_t, x_{t-1}, ...) are identical for all t≥1, whereas the target u_0 (or u'_{t-1} for memory task) differs. This proves no readout f can be simultaneously TI and correct, settling the generality question without simulation. If the authors intend the claim for bounded inputs of width <2π, they must state and justify that restriction; the current text does not.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim ('removes the limitation', abstract; 'in general', TI Transformation section) rests on the existence of a readout f(x_t, x_{t-1}, ...) that maps a TV reservoir state trajectory to a time-invariant function of input history. No existence theorem is stated; the analytical examples (Fig. 2a,b) cover only time-dependence separable from the input, e.g., an oscillator whose phase rotates while the input modulates the radius. Consider a minimal TV system: a phase-modulated oscillator, x_t = (cos θ_t, sin θ_t), θ_{t+1} = θ_t + ω + u_t, with u_t a real input. The state history reveals only θ_t modulo 2π. If two input histories differ by u'_0 = u_0 + 2π and u'_i = u_i for i>0, then θ'_t = θ_t + 2π for all t≥1, so the entire observable state trajectory is identical, yet the memory target u_0 differs by 2π. Therefore no function of the state history can produce a correct TI output for both inputs. This is not a contrivance: phase/frequency modulation is a common way inputs act on oscillators, including spin-torque oscillators; the paper neither states an input-range restriction narrower than 2π nor an observability condition that excludes this case. The quadratic readout f(x)=X²+Y² in Fig. 2a works only because the input modulates the radius (a non-wrapping coordinate), not the phase. Hence the unqualified 'in general' is unsupported and likely false.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes generalized reservoir computing (GRC), in which the echo state property (ESP) is required only of the final output yhat_t, not of the reservoir state x_t, and a nonlinear readout with memory is used as a 'TI transformation' to remove explicit time dependence from a time-variant (TV) reservoir. Two analytically solved examples are given (a periodically oscillating reservoir whose radius encodes the input, and a Lissajous-knot reservoir whose Z-coordinate contains a small input term), followed by numerical demonstrations using an ESN with positive conditional Lyapunov exponent on NARMA10, memory-capacity experiments on the Lorenz and Rössler models and on a real spin-torque oscillator, and attractor-embedding tasks in which a chaotic Lorenz96 reservoir emulates Rössler, Lissajous, and Kuramoto-Sivashinsky targets. The central claim is that any time-variant dynamical system can serve as a computational resource, with conventional RC as a special case.","tokens_in":27085,"tokens_out":9465,"duration_ms":90420,"significance":"The clean analytic examples, the self-contained TIPC derivations in Supplementary S1, and the careful initial-condition checks (node-wise ESP index, global ESP index) are strengths; the real-device spin-torque-oscillator experiment is valuable, and the supplementary material honestly reports where the embedded attractors only approximately match their targets. If the general TI-transformation claim were established, the framework would substantially broaden physical reservoir computing. However, the general existence statement is unproven and, as written, false, so the paper is better understood as a demonstration for a useful class of TV reservoirs than as a general theory of computation with arbitrary time-variant dynamics.","major_comments":[{"comment":"The passage 'The TV state can be transformed into TI outputs by nonlinearity with memory f(xt, xt-1, ...) in general' is the load-bearing premise of the paper, but it is not proven and, as stated, is false. Consider a phase-modulated oscillator with state xt = (cos theta_t, sin theta_t) and theta_{t+1} = theta_t + omega + u_t. For two input histories u and u' with u'_0 = u_0 + 2pi and u'_t = u_t for t >= 1, the observable state trajectories are identical for all t >= 1 because cos and sin are 2pi-periodic, while the memory target u_0 differs by 2pi. Hence no function of the state history, linear or nonlinear, with or without memory, can produce a time-invariant output that correctly distinguishes these two inputs. This is not an exotic case: phase/frequency modulation is a standard physical input pathway, and the paper's analytical examples instead use input acting on non-wrapping coordinates (radius in the periodic oscillator, the Z-coordinate in the Lissajous knot). The paper needs either an explicit observability/input-range condition (for example, phase unwrapping, input range narrower than the wrapping period, or a non-wrapping coordinate) or a reformulation of the central claim as holding for a restricted class of TV systems.","section":"TI Transformation"},{"comment":"The numerical TI transformation via a trained MLP is presented as covering 'the majority of physical systems,' but no condition is given under which a finite trained readout is guaranteed to find a TI output. The paper's own examples have positive conditional Lyapunov exponents, meaning the reservoir state depends on initial conditions as well as on input history; the demonstrations check a small number of initial conditions and show that the trained readouts are approximately initial-condition independent in those cases, but this does not establish a general mechanism. To make the central claim load-bearing, the authors should either state a sufficient condition for the existence of a TI readout (for example, an output-relevant subspace that is asymptotically independent of initial conditions) or explicitly narrow the claim to systems for which such a readout is demonstrated. As it stands, the abstract's statement that the framework 'removes the limitation' is stronger than the established results.","section":"TI Transformation and Applications"}],"minor_comments":[{"comment":"The averaged node-wise ESP index for the reservoir layer is given as 0.96 in the first paragraph of S3 and as 1.7 later in the same section; with the green threshold <0.3, neither value is consistent with the coloring rule described in the caption of Fig. 2c. Please reconcile the numbers and clarify the coloring rule.","section":"Supplementary S3"},{"comment":"The displayed definition of the global ESP index writes MSE(y_i^{(1)}, y_i^{(2)}) in the numerator but then expands it using \\hat y^{(1)}_{i,t} and \\hat y^{(2)}_i; the first argument should be \\hat y^{(1)} and \\hat y^{(2)}, and the subscript t is missing on \\hat y^{(2)}_i.","section":"Supplementary S4"},{"comment":"The general framework is introduced with readout f(xt, xt-1, ...), but the formal GRC definition two paragraphs later uses only \\hat y_t = f(xt). Please state explicitly in the definition that the readout may include delayed states.","section":"TI Transformation"},{"comment":"The formula for C(\\tau) has a mismatched normalization: the NMSE denominator is written as \\sum_{t=Ttrain}^{Ttrain+Ttest} y_t^2 while the numerator starts at Ttrain+1; please unify the summation range.","section":"Materials and Methods, Memory Capacity Task"}],"recommendation":"major_revision","confidential_remarks":"The empirical core of the paper is solid and the authors are transparent about the limits of their attractor-embedding results. The main problem is the unqualified generality claim in the abstract and in the TI Transformation section, which is falsified by a phase-modulated oscillator example. I think this is fixable within the scope of a revision: the authors can add explicit conditions under which a TI transformation exists, or reframe the contribution as a demonstration for a class of TV reservoirs. I would not reject, but the current abstract overstates the reach of the framework."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nI read the GRC paper with a skeptical but open mind. The core idea is simple and worth taking seriously: instead of demanding the echo state property from the reservoir, demand it from the output after a nonlinear readout, so time-variant reservoirs become usable. The paper does real work: clean analytic examples (periodic oscillator, Lissajous knot), a real spin-torque oscillator experiment, and a Lorenz 96 reservoir that emulates Rössler, periodic Lissajous targets, and KS dynamics for seven Lyapunov times. The global ESP index in the SI is a good check—nonlinear readout makes outputs nearly initial-condition-independent while linear readout fails. That part is solid.\n\nBut the paper's headline claim—that GRC 'removes the limitation' and that TV states can be transformed into TI outputs 'in general'—is not supported and, as stated, is false. The stress-test example is a phase-modulated oscillator: θ_{t+1}=θ_t+ω+u_t. Inputs differing by 2π produce observationally identical state histories, so no function of the state history can output both correct values. The analytic examples avoid this because the input acts on a non-wrapping coordinate (radius or an additive coordinate). No input-range restriction or observability condition is stated. So the 'in general' must go, or be replaced by a precise condition.\n\nOther soft spots: no public code or data (data only 'on reasonable request'), no repeated-seed statistics on MLP training, and no baseline comparison against a standard ESP reservoir. The embedding results are also approximate: Rössler output has a different Lyapunov exponent than the target, and the Lissajous output is not perfectly periodic. These are acknowledged in the SI, but they mean 'embedding' should be read as 'nearby model', not exact reproduction.\n\nNone of this kills the paper. The framework is useful for a nontrivial class of systems, and the demonstrations are genuine. But the abstract and TI-transformation section overclaim. A careful revision should state conditions under which a nonlinear readout can erase time dependence (e.g., modulo-2π ambiguities excluded), add baseline comparisons, and release code/data.\n\nMy verdict: worth a serious referee. I'd send it to review with a request for major revision on the claims. I'd also bring it to the reading group—the counterexample makes for a good discussion.","headline":"The output-ESP idea is real and the demonstrations are convincing, but the paper's universal claim is too strong and is false for phase-modulated reservoirs.","tokens_in":27577,"tokens_out":3790,"would_cite":true,"duration_ms":33688,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Generalized reservoir computing relaxes the echo state property by demanding it of the output instead of the reservoir, letting chaotic and oscillatory systems serve as computational resources.","keywords":["reservoir computing","echo state property","time-invariant transformation","physical reservoir computing","spin-torque oscillator","spatiotemporal chaos","attractor embedding","temporal information processing capacity"],"falsifier":"Repeat the Kuramoto-Sivashinsky embedding task with the Lorenz 96 reservoir, training the readout on one initial condition and then running the closed loop from a second initial condition; if the two generated outputs diverge by more than the target's own variance within one Lyapunov time, the claimed TI transformation has not removed the reservoir's time dependence.","tokens_in":26487,"feed_emoji":"🌀","tokens_out":8733,"duration_ms":77626,"temperature":0.7,"pith_summary":"Reservoir computing normally requires the physical reservoir to be stable: the same input history must produce the same reservoir state (the echo state property), which rules out oscillatory and chaotic materials. This paper proposes generalized reservoir computing (GRC), which moves that requirement from the reservoir state to the final output: as long as a trained nonlinear readout can turn the time-varying state into a time-invariant function of input history, any dynamical system can serve as a computational resource. The authors demonstrate the idea analytically with a periodic oscillator and a Lissajous knot, numerically with an echo state network whose conditional Lyapunov exponent is positive, and physically with a real spin-torque oscillator. They then show that a chaotic Lorenz 96 system, which fails the echo state property, can emulate targets including the Rossler attractor and the Kuramoto-Sivashinsky spatiotemporal chaos. If correct, GRC removes the main constraint on physical reservoir substrates.","feed_headline":"Reservoir computing works without the echo state property","feed_subtitle":"By requiring stability only at the output, oscillators and chaos become usable reservoirs.","key_machinery":"The central object is the time-invariant (TI) transformation: a nonlinear readout with memory, f(x_t, x_{t-1}, ...), that maps a time-variant reservoir state into a time-invariant output. The paper defines GRC by requiring the echo state property only on the output. The temporal information processing capacity (TIPC) expansion into TI and TV bases (Legendre polynomial chaos plus Fourier/time bases, or Volterra-Wiener-Korenberg series) is used to quantify how much of the state and output is time-invariant and how much depends on time; it is the diagnostic that shows TV states contain latent input memory that a nonlinear readout can expose.","core_discovery":"The paper's central claim is that reliable information processing does not require the reservoir itself to be reproducible; it requires the output to be reproducible. Formally, the reservoir state is allowed to be time-variant, x_t = h(t, u_{t-1}, u_{t-2}, ...), and GRC asks only that a readout f (possibly with memory) produce a time-invariant output \\hat{y}_t = f(x_t, x_{t-1}, ...) = F(u_{t-1}, u_{t-2}, ...), a function of input history alone. Conventional reservoir computing appears as the special case where the reservoir state is already time-invariant. The paper identifies two mechanisms: removing the time dependence from TV terms (periodic oscillator: $X_t^{2}$ + $Y_t^{2}$ = $r_t^{2}$) and amplifying small TI terms by canceling TV terms (Lissajous knot: -2X_t Y_t + Z_t = \\epsilon u_t). It then shows numerically that an ESN without ESP plus a multilayer perceptron can solve NARMA10, and that nonlinear readouts recover memory from Lorenz, Rossler, and a real spin-torque oscillator, where linear readouts recover none. Finally, a chaotic Lorenz 96 reservoir without ESP embeds several target attractors, including the Kuramoto-Sivashinsky model with errors staying small for more than seven Lyapunov times.","pith_inferences":["If the TI transformation is as general as claimed, the practical bottleneck in physical reservoir computing moves from the substrate to the readout; one testable extension is whether fixed, untrained nonlinear post-processing such as envelope extraction or detrending can substitute for trained MLPs in more time-variant systems.","The GRC perspective may apply to biological neural circuits that show trial-to-trial variability: the nonlinear response of downstream circuits could act as the readout, a possibility the paper raises and one could test by recording readout-neuron reliability across repeated sensory stimuli.","The paper leaves open the existence conditions for the TI transformation; characterizing the class of TV systems for which a finite-memory readout exists would turn the numerical demonstrations into a general theorem.","A practical caution follows implicitly: because the readout must cancel time-dependence, the readout's memory length and nonlinearity likely need to scale with the complexity of the reservoir's unforced dynamics."],"forward_implications":["Any physical system that fails the echo state property, such as oscillators, chaotic circuits, or biological networks, becomes a candidate reservoir provided a nonlinear readout with sufficient memory is trained.","Real devices such as the spin-torque oscillator can be used without forced synchronization; a trained nonlinear readout retrieves delayed inputs that a linear readout cannot.","Spatiotemporally chaotic systems, previously considered unusable, can be embedded as computational substrates: the Lorenz 96 reservoir emulates the Kuramoto-Sivashinsky dynamics with small prediction error for more than seven Lyapunov times.","The practical requirement of reservoir computing shifts from finding materials with stable reproducible dynamics to designing or training readouts, including physical readouts such as deep physical neural networks, that perform the TI transformation.","Conventional reservoir computing remains valid as the special case where the reservoir state already has ESP; GRC strictly enlarges the class of usable dynamics."],"supporting_citations":[{"why":"Defines the echo state network and the echo state property; the conventional-RC condition GRC relaxes and recovers as a special case.","marker":"[4]"},{"why":"Defines the liquid state machine and its Volterra-series prerequisite; the second conventional RC condition that requires time-invariant states.","marker":"[5]"},{"why":"Introduces the temporal information processing capacity (TIPC) that decomposes states and outputs into TI and TV terms; used to verify the transformation.","marker":"[33]"},{"why":"Supplies the NARMA10 benchmark task solved with the non-ESP ESN plus MLP readout.","marker":"[34]"},{"why":"Provides the memory-capacity task used to quantify latent memory recovered by nonlinear readouts.","marker":"[35]"},{"why":"Establishes the conditional Lyapunov exponent criterion for generalized synchronization and ESP, used to show the reservoirs lack ESP.","marker":"[47]"},{"why":"Describes the spin-torque oscillator setup whose real experimental data the GRC readout processes.","marker":"[25]"},{"why":"Defines the Kuramoto-Sivashinsky equation and the reservoir-computing approach to model-free prediction; the spatiotemporal-chaos benchmark that GRC embeds.","marker":"[37, 38]"}],"fun_headline_variants":["Generalized reservoir computing drops the echo state requirement","Reservoir computing without reservoir reproducibility","Chaos becomes a usable reservoir with GRC","No echo state property needed for reservoir computing","Generalized reservoir computing needs only stable output"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central claim stands on the assumption that for an arbitrary time-variant reservoir there is always a nonlinear readout with memory that makes the output a function only of the input history, and that a finite trained network can find it from finite data.","fun_headline_variants_meta":{"raw":{"variants":["Generalized reservoir computing drops the echo state requirement","Reservoir computing without reservoir reproducibility","Chaos becomes a usable reservoir with GRC","No echo state property needed for reservoir computing","Generalized reservoir computing needs only stable output"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000556,"raw_usage":{"total_tokens":2693,"prompt_tokens":1037,"completion_tokens":1656,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":1589}},"tokens_in":653,"tokens_out":1656,"duration_ms":11910,"temperature":1.0,"reasoning_tokens":1589,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:17:06.247869+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the Kuramoto-Sivashinsky embedding task with the Lorenz 96 reservoir, training the readout on one initial condition and then running the closed loop from a second initial condition; if the two generated outputs diverge by more than the target's own variance within one Lyapunov time, the claimed TI transformation has not removed the reservoir's time dependence.","supporting_citations":[{"cited_title":"Kubota, H","cited_arxiv_id":null,"evidence_quote":"Introduces the temporal information processing capacity (TIPC) that decomposes states and outputs into TI and TV terms; used to verify the transformation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the NARMA10 benchmark task solved with the non-ESP ESN plus MLP readout."},{"cited_title":"Jaeger, Short term memory in echo state networks , Vol","cited_arxiv_id":null,"evidence_quote":"Provides the memory-capacity task used to quantify latent memory recovered by nonlinear readouts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the conditional Lyapunov exponent criterion for generalized synchronization and ESP, used to show the reservoirs lack ESP."},{"cited_title":"Tsunegi, T","cited_arxiv_id":null,"evidence_quote":"Describes the spin-torque oscillator setup whose real experimental data the GRC readout processes."}],"review_version":1}