{"id":"f6119700-d4be-4c2a-a10a-3037b6f3ada7","arxiv_id":"2608.00249","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Habituation requires nonlinearity, and the minimal structure that produces it is a single linear fading-memory state composed with a static nonlinear readout.","lead":"This review formalizes the hallmarks of habituation as behavioral constraints on input–output dynamics and shows that no linear time-invariant system can satisfy them with nonnegative output, so nonlinearity is mandatory. It then assembles a minimal habituation motif—one linear fading-memory state feeding a static nonlinearity—and maps that motif onto biological, circuit, and machine-learning examples.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LTI barrier depends on whether U includes finite pulse trains; under a periodic-only reading of U, an explicit LTI counterexample satisfies H0+H1.","rationale":"The reader's verdict is CONDITIONAL, and my analysis supports keeping that condition. The identified issue is more specific than the reader's weakest assumption: it is not chiefly the lossy translation of verbal hallmarks into inequalities, but the under-specified domain of admissible inputs in H0 and H1, on which the LTI impossibility theorem hinges. If U includes finite pulse trains — which the paper's own proof sketch suggests — the theorem is true and the central claim stands. If U is read as periodic-only, the theorem is false, and an explicit LTI counterexample exists. This is a genuine formal gap, not a manufactured concern, and it is load-bearing because Sec. 2.2 is the foundation for the rest of the review. The paper otherwise has independent value: the Wiener motif construction is simple and convincing for H1/H2; the distinction between adaptation and habituation is well argued; the survey of realizations is broad and useful; and the limitations of the lossy formalization are acknowledged. The fix is local: state that U includes all finite truncations of periodic pulse trains (or equivalently that H0 applies to finite pulse trains), and replace the informal superposition argument with a rigorous proof, e.g., via the finite-window positivity argument. Because this is a clarifiable formal condition rather than evidence that the intended result is false, I do not change the reader's CONDITIONAL verdict.","tokens_in":25633,"tokens_out":25971,"duration_ms":264812,"concrete_test":"Simulate the LTI system with impulse response h(t) = p(t) - 0.4 p(t-T) - 0.3 p(t-2T), where p is a rectangular pulse of width dT, driven by a semi-infinite periodic pulse train of period T starting at t=0. Verify that the peak sequence is 1, 0.6, 0.3, 0.3, ... and that the output is nonnegative for every phase shift φ in [0,T); if so, the LTI barrier fails under the periodic-only reading of U. Then drive the same system with a finite train of, say, five pulses and observe that the output becomes negative after the last pulse, confirming that the barrier holds only when finite pulse trains are admissible. This single experiment settles which domain the theorem requires.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Sec. 2.2 — that any system satisfying H0 and H1 must be nonlinear — is true only for a precisely specified admissible input set U, and the paper leaves U ambiguous. H0 is stated as 'for all admissible inputs,' while Sec. 2.1 defines U as 'non-negative periodic pulse trains of period T, amplitude A, and duty cycle d,' and the proof of the barrier explicitly uses a finite pulse train. These are different domains. If U is read as semi-infinite periodic pulse trains only (the natural reading of 'periodic stimulus' in Table 1), the LTI barrier is false. Let p(t) be a rectangular pulse of width dT, and take an LTI system with impulse response h(t) = p(t) - c1 p(t-T) - c2 p(t-2T), with c1,c2>0 and c1+c2<1. For any semi-infinite T-periodic pulse train starting at any phase, the delayed negative copies align exactly with later positive pulses, so the output is nonnegative and its peak sequence is 1, 1-c1, 1-c1-c2, 1-c1-c2, ... — satisfying H1 with K=1 and H0 on that input class. This is a linear time-invariant system, contradicting the claimed structural necessity of nonlinearity. The barrier is rescued only if U includes finite pulse trains, as the proof assumes: then the negative tail of the final pulse appears after the last pulse with no canceling pulse, violating H0. The paper therefore must state explicitly whether U contains finite truncations of periodic trains; the cited Prop. 3.1 presumably does so, but the self-contained argument in this review does not. This is not a dispute about the biological interpretation; it is an internal under-specification of the formal framework on which the headline result rests.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This review-style manuscript formalizes the classical hallmarks of habituation as behavioral constraints on input–output behavior, argues that linear time-invariant (LTI) systems are structurally incapable of habituation under a nonnegative-output assumption, and constructs minimal nonlinear motifs—linear fading-memory dynamics with static nonlinear readouts—that satisfy the core hallmarks. It then surveys realizations across biological systems, analog circuits, memristive materials, and machine-learning architectures, and discusses extensions for frequency and intensity sensitivity. The central claims are: nonlinearity is necessary for habituation with nonnegative output; a single fading-memory state with a nonlinear readout is sufficient for the core hallmarks H1/H2; and simple structural extensions (series composition, input nonlinearity) account for H4(b) and H5. The paper is written as an Annual Reviews-style synthesis, with the main formal results deferred to the authors' prior publications (refs. 1 and 2).","tokens_in":26061,"tokens_out":11878,"duration_ms":136912,"significance":"If the structural claims hold, the paper provides a principled, domain-independent answer to what dynamical ingredients habituation requires, and it usefully organizes a scattered literature across biology, physics, and machine learning. The behavioral-constraints framing is clear, the minimal Wiener motif is genuinely simple and mechanistically interpretable, and the survey is broad and current. The paper's core derivation is not a data-fitting exercise; the motif is constructed from specification, and no parameters are fitted in the main argument. The main structural theorem (LTI impossibility) is elegant but, as written, suffers from an ambiguity in the admissible input set that is load-bearing: under one natural reading (semi-infinite periodic pulse trains only), the theorem is false. The paper also leans heavily on the authors' own CDC/PNAS results for the formal propositions; for a review this is acceptable, but the self-contained argument needs to be precise about its domain.","major_comments":[{"comment":"The admissible input set U is not fixed precisely. Section 2.1 defines stimuli as \"non-negative periodic pulse trains of period T, amplitude A, and duty cycle d,\" while the proof sketch in Section 2.2 uses a finite pulse train. Under the semi-infinite-periodic reading, the claimed LTI barrier is false: the LTI system with impulse response h(t)=δ(t)−c1δ(t−T)−c2δ(t−2T), with c1,c2>0 and c1+c2<1, maps any semi-infinite T-periodic pulse train to a nonnegative output with peak sequence 1, 1−c1, 1−c1−c2, ..., so H0 and H1 are both satisfied. This counterexample fails only if U includes finite truncations of periodic trains (because the negative tail after the final pulse violates H0), or if H2 is appended to the definition of habituation so that withholding is an admissible operation. Since the central conclusion \"nonlinearity is structurally necessary\" depends entirely on this, the paper must","section":"Secs. 2.1–2.2, Table 1"},{"comment":"Even accepting finite pulse trains as admissible, the proof sketch is not self-contained. It invokes superposition and time invariance but does not specify how H1 applies to the summed input ũ=u+u_shift, which is not necessarily an admissible periodic pulse train, nor exactly how the peak sequence of the summed system contradicts the superposition identity. Since the formal proposition is deferred to ref. 2, the review should at least state the precise proposition, including the class of inputs and the notion of habituation used, so that the reader can verify the argument without consulting the CDC paper.","section":"Sec. 2.2"},{"comment":"The claim that a series connection of two Wiener units satisfies H4(b) (faster recovery under more frequent stimulation) is supported only by a verbal timescale argument and by the examples shown in Fig. 5(a,b). No formal sufficient condition is given (e.g., explicit timescale separation α1≫α2 with parameter bounds), and it is not stated whether H4(b) holds robustly or only for the plotted parameters. Since H4(b) is one of the advertised \"structurally distinct extensions,\" the paper should either supply a proof or a precise parameter regime, or explicitly label the claim as a numerical demonstration.","section":"Sec. 4.2.1, Fig. 5"},{"comment":"The intensity-sensitivity claim H5 is verified through the asymptotic ratio ρ=y[∞]/y[0], but the Table 1 criterion for H5 is an inequality on the normalized response sequence y1[k]≤y2[k] for all k. A phase diagram of the asymptotic ratio alone does not establish the full-sequence inequality. Please show the normalized peak sequences for representative A1<A2, or explicitly restrict the claim to the asymptotic regime and state that the full hallmark is not demonstrated.","section":"Sec. 4.2.2, Fig. 5(e,f)"}],"minor_comments":[{"comment":"The index in the H3 criterion \"y[K(L+L′)+k]<y[k]\" is ambiguous; k is said to range over \"some subsequent stimuli,\" but the bounds on k should be made explicit to be mathematically precise.","section":"Table 1, H3"},{"comment":"The notation ρ=y[∞]/y[0] should be defined as the asymptotic peak-response ratio, since under periodic stimulation the system does not converge to a constant output and y[∞] is not a steady-state value.","section":"Sec. 4.2.2"},{"comment":"The phrase \"time-varying receptivity\" for σ(t)=e^{−αt}H(t) is slightly misleading because σ depends on absolute time rather than on stimulus history; the text immediately explains this, but rewording would improve clarity.","section":"Sec. 4.1, Step 1"},{"comment":"The discussion of ideal memristors lacking spontaneous recovery is useful, but the claim that habituation \"requires a leaky memory\" would benefit from a precise pointer back to H2 and the definition of fading memory in Sec. 4.4.","section":"Sec. 5.1.2"},{"comment":"The statement that the minimal motif's steady-state attenuation \"can be made small in appropriate limits\" is vague; giving one explicit parameter limit (e.g., α→0 or β→∞) would make the point concrete.","section":"Sec. 3.3"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the unstated domain of admissible inputs in the LTI barrier. The counterexample in my major comment is real under the natural semi-infinite-periodic reading, so the authors must either expand U to include finite pulse trains or adjust the claim. The paper is otherwise a solid review with a useful synthesis; the reliance on refs. 1 and 2 for formal proofs is acceptable for this venue, but the editor may wish to confirm that the CDC paper is already published or available. No concerns about novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis is worth a serious look, but the central claim as written is less settled than it sounds. The review restates the authors' own PNAS and CDC results rather than proving them fresh, and it says so plainly. The genuinely new material is the adaptation-vs-habituation independence argument and the treatment of the hallmarks as behavioral specifications. That part is good: the damped oscillator and AIC examples cleanly separate asymptotic adaptation from transient decrement, and the discussion of why neither implies the other is careful.\n\nWhere it gets soft is the LTI barrier. The sketch in Sec. 2.2 uses a finite pulse train, while Sec. 2.1 defines the admissible inputs as periodic pulse trains. Those are different domains. If U contains only semi-infinite periodic trains, the barrier is false: an LTI system with impulse response p(t) - c1 p(t-T) - c2 p(t-2T), with c1, c2 > 0 and c1 + c2 < 1, gives a nonnegative output with peak sequence 1, 1-c1, 1-c1-c2, ... on any periodic pulse train, satisfying H0 and H1. If U includes finite truncations, then the negative tail after the last pulse violates H0 and the barrier stands. So the theorem needs U pinned down explicitly. The authors themselves note that the verbal-to-mathematical formalization is lossy; that caveat is doing real work here. The minimality claim is also asserted rather than proved: the Wiener motif is plausibly minimal for the intended class, but no no-smaller-system argument is given. And the simulations are not accompanied by code or parameter ranges, which matters for a paper whose selling point is robustness across parameters.\n\nNone of this kills the paper. The authors are clearly serious, the literature synthesis is genuinely useful, and they explicitly defer to the formal proofs in their earlier papers. A reader who wants an organizing map of habituation across biology, circuits, and machine learning will get value. But a referee should send it back with specific demands: define U, prove or precisely state the LTI theorem, provide code or parameter sweeps, and soften or prove the minimality claim.\n\nI would send it to peer review. It is important enough and the underlying work is real; it just needs the formal edges sealed.","headline":"Useful synthesis, but the headline 'nonlinearity is necessary' needs the admissible input set pinned down before it is a theorem.","tokens_in":26596,"tokens_out":4738,"would_cite":true,"duration_ms":49333,"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":"Any habituating system with nonnegative, bounded output must be nonlinear, and a single fading-memory state with a static nonlinear readout is enough to capture the core hallmarks.","keywords":["habituation","behavioral constraints","fading memory","nonlinear systems","Wiener model","transient response","state-space models","adaptation vs habituation"],"falsifier":"Take any linear time-invariant system with a nonnegative impulse response (e.g., an RC low-pass filter measuring capacitor voltage), drive it with a periodic positive pulse train, and record the peak output in each period followed by the peak after a stimulus-free pause. If the peaks strictly decrease over the first several periods and then recover toward baseline, the paper's central impossibility claim is wrong; the theorem predicts the peaks must instead converge monotonically upward to steady state.","tokens_in":25491,"feed_emoji":"🧠","tokens_out":7497,"duration_ms":77623,"temperature":0.7,"pith_summary":"This review argues that habituation—the progressive lessening of a response to repeated stimulation and its recovery when stimulation stops—is not a collection of substrate-specific tricks but a constraint on dynamical structure. Formalizing the classical hallmarks of habituation as inequalities on peak responses, the authors show that no linear time-invariant system with nonnegative output can habituate; nonlinearity is therefore necessary, not a modeling choice. They then construct the minimal structure that satisfies the core hallmarks: one linear fading-memory state (a leaky integrator of recent input) followed by a static nonlinearity that attenuates the response when the memory is large. This single motif captures the defining features of habituation, and simple extensions—two units in series, a static input nonlinearity—add frequency- and intensity-sensitivity. The review positions this motif as the shared 'normal form' underlying habituating systems from ciliates to circuits to machine-learning sequence models.","feed_headline":"No linear time-invariant system can habituate","feed_subtitle":"Nonlinearity is a structural requirement; one leaky memory plus a static nonlinear readout suffices for the core hallmarks.","key_machinery":"The key object is the Wiener model: linear fading-memory dynamics (a first-order leaky integrator x_dot = beta*u - alpha*x, so x is a discounted memory of recent input) followed by a static, monotone-decreasing nonlinear output map y = u*sigma(x) with sigma(0)=1. The memory state x accumulates during bursts of stimulation, pushing sigma(x) down and attenuating the response; when stimulation stops, x decays and the system recovers. This block structure is the minimal motif the review constructs from the hallmarks, and it is the common thread connecting RC-diode circuits, molecular memory models, and reservoir or state-space computing. A classical theorem on fading memory guarantees that such","core_discovery":"The paper's central claim is that any system that habituates with nonnegative, bounded output must be nonlinear, because linear time-invariant systems obey superposition and time invariance, which forbid the history-dependent attenuation that defines habituation. Conversely, habituation is not computationally demanding: a single leaky-integrator state that remembers recent stimulation, feeding a static nonlinear readout that suppresses output when the memory is full, satisfies the core hallmarks H1 (progressive decrement) and H2 (spontaneous recovery), along with H3 and frequency sensitivity H4(a). The authors derive this Wiener-type motif step-by-step from the hallmarks rather than assuming","pith_inferences":["A testable extension: because the impossibility argument uses only superposition, time invariance, and nonnegativity, it should extend to any periodic input family, not just the pulse trains used in the paper; checking monotone attenuation on a broader class of inputs would probe the robustness of the structural conclusion.","Boundary of the claim: the minimality result is tied to the 'there exists a stimulus' reading of H1; if one requires habituation for all stimulation frequencies, the single-unit motif may no longer be sufficient, and a stronger architecture—possibly involving an internal model of the stimulus—would be needed.","Design corollary for artificial sequence models: to obtain habituation-like filtering, keep the core recurrence linear and make the readout or decay rate input-dependent; this is the cheapest way to satisfy the behavioral constraints and could be tested directly in model ablations.","Empirical prediction: single-trial behavioral data from any habituating system should be describable by a one-dimensional hidden state with exponential forgetting; if two timescales are required, frequency-dependent recovery should be observed."],"forward_implications":["Any habituating system with nonnegative, bounded output must be nonlinear: linear time-invariant dynamics with a linear readout cannot produce monotone attenuation and recovery.","A single fading-memory state with a static nonlinear readout is sufficient for the core hallmarks H1, H2, H3, and H4(a); no more structural complexity is needed at the core.","Frequency-dependent recovery (H4b) requires at least two timescales, realized by connecting two Wiener units in series, and intensity sensitivity (H5) requires a static input nonlinearity.","Adaptation and habituation are logically independent: each can occur without the other, and both require nonlinearity once outputs are required to stay nonnegative.","Across biological, physical, and algorithmic systems, the recurring architectural principle is a fading memory of recent input coupled to a nonlinear readout."],"fun_headline_variants":["Habituation demands nonlinearity","Linear systems can't habituate","A simple nonlinear motif explains habituation","Nonlinearity is the missing ingredient in habituation","One leaky integrator plus nonlinearity yields habituation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The argument stands or falls on the chosen mathematical encoding of the verbal hallmarks—in particular, H1 is formalized as 'there exists some periodic stimulus whose peak responses decrease monotonically,' and H0 requires outputs to be nonnegative and bounded; a different, stricter formalization could defeat both the impossibility result and the claimed minimality.","fun_headline_variants_meta":{"raw":{"variants":["Habituation demands nonlinearity","Linear systems can't habituate","A simple nonlinear motif explains habituation","Nonlinearity is the missing ingredient in habituation","One leaky integrator plus nonlinearity yields habituation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1161,"prompt_tokens":671,"completion_tokens":490,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":415,"completion_tokens_details":{"reasoning_tokens":435}},"tokens_in":415,"tokens_out":490,"duration_ms":6028,"temperature":1.0,"reasoning_tokens":435,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T00:55:08.172196+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any linear time-invariant system with a nonnegative impulse response (e.g., an RC low-pass filter measuring capacitor voltage), drive it with a periodic positive pulse train, and record the peak output in each period followed by the peak after a stimulus-free pause. If the peaks strictly decrease over the first several periods and then recover toward baseline, the paper's central impossibility claim is wrong; the theorem predicts the peaks must instead converge monotonically upward to steady state.","supporting_citations":[],"review_version":1}