{"id":"480b4ba2-472d-4c4d-9bb2-c1b25419d9a3","arxiv_id":"2412.16189","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"In an asymmetrically coupled pair of nonchaotic Rulkov neurons, coexisting spiking and chaotic bursting dynamics form a short-lived quasimultistability with fractal basin boundaries and extreme final-state sensitivity.","lead":"This paper studies two electrically coupled model neurons and finds that the system briefly behaves as if it has two possible states, calm spiking or chaotic bursting, before eventually settling into one. It maps the fractal boundaries between these short-lived behaviors, showing that tiny changes in starting conditions can produce very different short-term activity.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'extreme final state sensitivity' rests on a finite-time Lyapunov cutoff that defines the pseudo-attractor basin; no test shows u4 ≈ 0.037 is independent of that cutoff.","rationale":"Agree with the reader: the weakest assumption is the finite-time Lyapunov cutoff. The paper deserves credit for explicitly labeling the chaotic object a 'pseudo-attractor' and acknowledging that all orbits eventually reach the spiking attractor, but the quantitative conclusions in the abstract — fractal dimension 3.963, factor 10^27 — rely on a classification rule that is not shown to be robust. The uncertainty-exponent method is intended for genuine multiattractor systems; applying it to transient chaos requires at least a demonstration that the measured exponent converges as the classification horizon is varied. The proposed cutoff-sweep is the minimal check: it directly tests whether u4 is a property of the dynamics or of the horizon. If u4 remains near 0.037 across a decade of cutoffs, the concern is largely resolved; if it drifts toward 0.3–1, the extreme sensitivity becomes a finite-horizon artifact. No change to the reader's conditional verdict is needed unless the check fails badly.","tokens_in":21005,"tokens_out":5040,"duration_ms":43482,"concrete_test":"Recompute the four-dimensional uncertainty statistic ϱ4(ε) and the fitted exponent u4 using basin-classification cutoffs of 5,000, 10,000, 20,000, 40,000, and 100,000 iterations, holding the domain S'_4 and the perturbation protocol (Eq. 42) fixed; also record the fraction of the 20,000-iteration 'chaotic' sample whose λ1 remains positive after 100,000 iterations. If u4 changes by more than ~0.01 or the 'chaotic' fraction decays substantially with cutoff, the extreme-sensitivity claim is an artifact of the finite-time classifier. A complementary check is to re-estimate u4 after labeling states by the actual settling time to the spiking attractor rather than by λ1 sign.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim — an uncertainty exponent u4 ≈ 0.037, a basin-boundary dimension d4 ≈ 3.963, and a 10^27 improvement factor — is derived from the uncertainty-exponent method of McDonald et al., which presupposes two coexisting attractors with invariant basins (Sec. V, Eqs. 38-39). Here the only true attractor is the nonchaotic spiking attractor (Sec. II); the 'chaotic pseudo-attractor' is a long transient, and 'basin membership' is assigned by the sign of the finite-time maximal Lyapunov exponent after 5,000 iterations (Sec. IV A and Sec. V) or 20,000 iterations (Sec. IV B). Because all orbits eventually converge to the spiking attractor, the measured 'basin boundary' is not a boundary between different final states but a level set of a finite-time observable, and it is not invariant under a change of the iteration cutoff. The paper neither reports how u4, the basin fractions, or the fractal dimensions vary with this cutoff, nor provides evidence that the 5,000/20,000-iteration classification is a robust proxy. Since Eq. (39) and the factor 10^27 follow directly from u4, the headline 'extreme final state sensitivity' is currently conditional on an arbitrary numerical choice. The transient dynamics are interesting, but the quantitative exponents — the main new results — describe the boundary of a finite-time classifier rather than an invariant basin structure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies two asymmetrically electrically coupled nonchaotic Rulkov neurons and reports the coexistence of a true nonchaotic spiking attractor with a chaotic spiking-bursting 'pseudo-attractor' that captures orbits for long transient times (on the order of 10^4 iterations) before they eventually converge to the true spiking attractor. The authors call this quasimultistability. They compute box-counting and Lyapunov dimensions of the pseudo-attractor, classify basins of attraction in a two-dimensional slice and in four-dimensional state space using the Sprott-Xiong method, and use the uncertainty-exponent method to claim an extreme final-state sensitivity with u4 ≈ 0.037, a basin-boundary dimension d4 ≈ 3.963, and a required initial-uncertainty reduction of about 10^27 to improve final-state prediction by a factor of 10.","tokens_in":21321,"tokens_out":6144,"duration_ms":54858,"significance":"If the quantitative claims were robust, this would be a useful case study of long chaotic transients mimicking multistability in a discrete-time neuronal map, and the reported fractal basin geometry would be of genuine interest to the nlin.CD community. The paper is clearly written, gives explicit model equations, and honestly acknowledges that the chaotic object is not a true attractor and that all orbits eventually reach the nonchaotic spiking attractor. The central issue is that the headline quantitative results — the uncertainty exponent, the fractal dimension of the basin boundary, and the 10^27 factor — are defined with respect to a finite-time Lyapunov-sign classifier rather than invariant basins, and the paper does not test how these quantities depend on the classifier's time horizon.","major_comments":[{"comment":"The uncertainty-exponent method of McDonald et al. presupposes two coexisting attractors with invariant basins, but the paper itself concludes in Sec. II that the system is not multistable and that all orbits eventually converge to the nonchaotic spiking attractor. Basin membership is assigned by the sign of the finite-time maximal Lyapunov exponent after 5,000 iterations (Sec. IV A) or 20,000 iterations (Sec. IV B). Because the paper does not test how u4, the basin fractions, or the fractal dimensions vary with the classification horizon, the claim of 'extreme final state sensitivity' is conditional on an arbitrary numerical choice. Please add a systematic study of ϱ4(ϵ) and u4 for several horizons (e.g., T = 10^3, 5×10^3, 10^4, 2×10^4, 5×10^4) and, ideally, for an alternative classifier such as a threshold on a finite-time average of a voltage variable, and report whether u4 and d4 are robust or how they trend with the horizon.","section":"Sec. V, Eqs. (38)-(46)"},{"comment":"The box-counting dimension d ≈ 1.84 is computed from only three box sizes ε = 1/20, 1/30, 1/40, which span a factor of only 2 in length scale; this is an insufficient range to establish a scaling exponent, and for a finite orbit segment the box-count N(ε) necessarily saturates for sufficiently small ε. Please extend the estimate to at least two decades of ε, report the scaling range, and state clearly that the result applies to the finite-time pseudo-attractor sample rather than to an invariant set.","section":"Sec. III"},{"comment":"The reported values of ϱ4(ϵ) show a plateau at small ϵ (from 0.349 at ϵ = 1/32 to 0.302 at ϵ = 1/2048), and the fitted exponent u4 ≈ 0.037 is strongly influenced by this flat tail. The fit uses all tabulated points without error bars, so the inferred dimension d4 ≈ 3.963 has no quantified uncertainty. Please provide bootstrap or Monte Carlo confidence intervals for u2 and u4, and report how the fitted exponents depend on the range of ϵ included (e.g., using only ϵ ≤ 1/8 versus all points).","section":"Sec. V, Table III"}],"minor_comments":[{"comment":"The regression equations are written as if Pw(ξ4) and Pb(ξ4) were linear in ξ4, but the text states that the regressions were performed on log-log plots; please rewrite these equations in terms of log2 P and log2 ξ4 to avoid confusion.","section":"Sec. IV B, Eqs. (35)-(36)"},{"comment":"The text says 'we use an indirect method of determining which basin an initial state is in' based on a 5,000-iteration Lyapunov exponent; this is an explicit admission that the 'basin of the pseudo-attractor' is a finite-time basin. Please state this distinction more prominently in the abstract and introduction so that readers do not mistake the pseudo-attractor basin for an invariant basin.","section":"Sec. IV A"},{"comment":"The number of Monte Carlo samples used for the basin classification (Table I and Table II) is not reported; please include sample sizes so that the statistical significance of the P(ξ) estimates can be assessed.","section":"Sec. IV B"},{"comment":"The conclusion that the results 'could have important applications in neurobiology' is speculative given that the central final-state sensitivity result is tied to a finite-time classifier; please temper this claim or condition it on the transient interpretation.","section":"Sec. VI"},{"comment":"The Jacobian matrix rendering is deferred to the author's preprint [49]; if this matrix is important for reproducibility, consider including the explicit expression in the appendix rather than citing an unpublished source.","section":"Sec. III"},{"comment":"The color maps of the maximal Lyapunov exponent would be easier to interpret with a color bar, and the basin plots would benefit from a legend explicitly labeling white and black regions.","section":"Figs. 6 and 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is a single computational case study. The authors are honest that the system lacks true multistability and that the chaotic object is a pseudo-attractor, but the central quantitative claim of extreme final-state sensitivity is currently a property of a finite-time classifier. The manuscript can be made publishable if the authors demonstrate that the reported exponents are robust (or at least correctly characterize the cutoff dependence) across the classification horizon. If the exponent u4 turns out to be highly cutoff-dependent, the paper would be more appropriately framed as a study of finite-time basin boundaries rather than as a claim about extreme final-state sensitivity in the usual sense."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely new numerical case study, and the quasimultistability idea is worth having. But the central number, u4 ≈ 0.037 and the 10^27 factor, is computed from a basin boundary defined by the sign of a finite-time Lyapunov exponent after 5,000 or 20,000 iterations. Since the authors themselves state that all orbits eventually converge to the nonchaotic spiking attractor, that boundary is not a boundary between final states; it is a level set of a finite-time observable. The uncertainty exponent method of McDonald et al. presupposes two coexisting attractors with invariant basins. Applying it to a long transient is a category error unless you show the results are insensitive to the cutoff. The paper does not do that. So the 'extreme final state sensitivity' is conditional on an arbitrary numerical choice, and the 10^27 improvement factor is a direct consequence of that choice.\n\nWhat is good: the paper is clearly written, the authors are honest about the pseudo-attractor being a transient, and the geometric plots (Figs. 6, 8, 9) show a rich structure. The Sprott-Xiong classification of the two-dimensional slice is a reasonable application, and the distinction between Class 3 in the slice and Class 2 in the full space is interesting. The claim of novelty for nonchaotic Rulkov basins appears to hold against the cited literature. Also, they give the Jacobian explicitly, which helps reproducibility.\n\nSoft spots in proportion: (1) the cutoff dependence is untested; this is the load-bearing issue. (2) No error bars or sample sizes anywhere; the ϱ(ϵ) values in Table III have no statistical uncertainties, and the linear fit for u4 has R2=0.967 but 12 points with no error bars. (3) The box-counting dimension uses only three box sizes, which is thin for a fractal dimension estimate. (4) The biological relevance framing in the abstract and conclusions overshoots; the map is phenomenological and only one parameter set is studied. (5) Minor: Eq. (35)-(36) are presented confusingly with low R2, and they interpret that as γ=0; that is defensible but needs clearer reporting.\n\nWho is this for: readers interested in basins of transient chaos in discrete-time neuron maps, or in the Sprott-Xiong/uncertainty-exponent toolbox. It deserves a serious referee, but the referee should push for a cutoff-robustness analysis and a reframing of 'final state sensitivity' as 'finite-time sensitivity' (or 'transient sensitivity').","headline":"A well-written numerical exploration of transient chaos in coupled Rulkov neurons, but the headline uncertainty exponent is a property of the chosen finite-time Lyapunov cutoff, not an invariant basin boundary.","tokens_in":21851,"tokens_out":3907,"would_cite":false,"duration_ms":32174,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37D45","37N25","37C70"],"pacs":[],"model":"deepseek-v4-flash","headline":"Two asymmetrically coupled nonchaotic Rulkov neurons produce a chaotic pseudo-attractor and a fractal basin boundary of dimension about 3.96, giving extreme sensitivity to initial conditions.","keywords":["Rulkov map","quasimultistability","pseudo-attractor","fractal basin boundary","uncertainty exponent","final state sensitivity","slow-fast dynamics","coupled neurons"],"falsifier":"Compute the basin fractions and the four-dimensional uncertainty exponent using Lyapunov-evaluation horizons of $10^3$, $10^4$, $10^5$, and $10^6$ iterations while holding everything else fixed. If the uncertainty exponent rises toward 1 or the measured fraction of the chaotic pseudo-attractor basin falls toward zero as the horizon grows, the reported extreme sensitivity is a cutoff artifact; if the exponent stays near 0.037 and the fraction holds, the claim stands.","tokens_in":20757,"feed_emoji":"🧠","tokens_out":8299,"duration_ms":73614,"temperature":0.7,"pith_summary":"This paper studies what happens when two identical nonchaotic Rulkov neuron maps are coupled asymmetrically. It claims that the coupled system has only one true attractor, a nonchaotic spiking cycle, but that a chaotic spiking-bursting \"pseudo-attractor\" governs orbits for long transients (order $10^4$ iterations), producing what the author calls quasimultistability. The paper classifies the basins of these two behaviors: in a two-dimensional slice near the attractors the chaotic basin is almost everything (Class 1), while in the full four-dimensional state space both basins occupy fixed fractions (Class 2). It then quantifies the basin boundary via uncertainty exponents, finding extreme final-state sensitivity: the boundary has dimension about 3.963 in four dimensions, so reducing initial uncertainty by a factor of $10^{27}$ is needed to reduce short-term final-state uncertainty by a factor of 10. The paper argues these geometrical features could matter for how small perturbations of biological neurons lead to different functional outcomes.","feed_headline":"Neuron pair's fate needs 10^27 precision","feed_subtitle":"Tiny errors in a coupled neuron pair's state are amplified: a 10x prediction gain costs 10^27 precision.","key_machinery":"The carrying object is the two-neuron slow-fast Rulkov map with a piecewise fast-variable function and asymmetric electrical coupling; its Jacobian is partitioned into five piecewise blocks so that Lyapunov spectra can be computed along orbits. The argument then runs on the uncertainty-exponent relation $u = n - d$, which converts the measured scaling of final-state uncertainty with initial uncertainty into a fractal dimension for the basin boundary. Basin membership is assigned by the sign of the finite-time maximal Lyapunov exponent after a fixed number of iterations, and the normalized-distance basin classification scheme $P(\\xi) = P_0 \\xi^{-\\gamma}$ sorts the basins into classes by how their occupancy scales with distance from the attractor.","core_discovery":"The system's true asymptotic behavior is a single nonchaotic spiking attractor, yet a chaotic spiking-bursting pseudo-attractor traps orbits for roughly ten thousand iterations, so the system behaves as if it were multistable. Treating this pseudo-attractor as a second attractor reveals two basins whose four-dimensional boundary is extremely fractal: the uncertainty exponent is $u_4 \\approx 0.037$, giving a boundary dimension $d = 4 - u_4 \\approx 3.963$ and requiring an initial-uncertainty reduction on the order of $10^{27}$ to shrink final-state uncertainty by a factor of 10. In a two-dimensional slice, the nonchaotic basin is a finite-measure Class 3 set concentrated near synchronization, while the chaotic pseudo-attractor basin fills the slice; in all of four-dimensional space both basins are Class 2, occupying fixed fractions of state space. The paper also reports that the pseudo-attractor's box-counting dimension ($\\approx 1.84$) does not match its Lyapunov dimension ($\\approx 2.07$), which it attributes to the transient, non-invariant nature of the pseudo-attractor.","pith_inferences":["Beyond the paper: re-running the uncertainty-exponent measurement with longer Lyapunov-evaluation horizons (e.g., $10^5$ and $10^6$ iterations) would show whether the four-dimensional uncertainty exponent is stable or an artifact of the 20,000-iteration cutoff; because every orbit eventually reaches the nonchaotic attractor, the pseudo-attractor basin is a transient construct.","Beyond the paper: varying the coupling asymmetry and the parameter mismatch between the two neurons would test whether the fractal boundary dimension and the $10^{27}$ sensitivity figure tune continuously with coupling, which would make the effect a tunable feature rather than a single-parameter accident.","Beyond the paper: the same random-sampling pipeline could be applied to other slow-fast neuron models, including continuous-time ones, to see whether extreme final-state sensitivity is a general feature of small coupled neuron systems rather than specific to this parameter set."],"forward_implications":["If the central claim holds, a coupled pair of nonchaotic neurons can display chaos-like dynamics for tens of thousands of iterations even though its asymptotic state is a simple periodic spike.","Final-state sensitivity of this degree means that short-term predictions of which firing pattern a neuron pair enters are practically impossible from initial-condition measurements alone.","The basin classification implies that near the attractor slice the nonchaotic basin has finite measure, whereas in full four-dimensional space both basins occupy fixed fractions of state space.","The mismatch between box-counting dimension and Lyapunov dimension for the pseudo-attractor is presented as a signature of transient chaos rather than a genuine strange attractor.","Quasimultistability is proposed as a general phenomenon for small sets of coupled identical nonchaotic systems that can temporarily synchronize into periodic orbits."],"supporting_citations":[{"why":"Introduces the nonchaotic Rulkov map and the spiking-bursting neural behaviors the paper builds on.","marker":"[1]"},{"why":"Provides the modified form of the Rulkov map and the map-based neuronal dynamics context used by the model.","marker":"[14]"},{"why":"Supplies the normalized-distance basin classification method with the power-law form for basin occupancy.","marker":"[26]"},{"why":"Defines final-state sensitivity and motivates measuring how initial-condition uncertainty transfers to final-state uncertainty.","marker":"[27]"},{"why":"Establishes the uncertainty-exponent method and the relation between uncertainty exponent and basin-boundary dimension.","marker":"[46]"},{"why":"Gives the explicit Jacobian rendering and Lyapunov-exponent computation pipeline used to classify orbits as chaotic or not.","marker":"[49]"}],"fun_headline_variants":["Fractal neuron basin needs 10^27 precision to predict","Neuron pair's chaotic mirage hides true stability","Asymmetric coupling yields extreme final state sensitivity","10^27 state accuracy required for coupled neuron fate","Quasimultistability: neurons trapped in transient chaos"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole analysis depends on labeling an initial state by whether its short-term chaos indicator is positive after a fixed number of steps (5,000 in the two-dimensional slice, 20,000 in four dimensions), even though every orbit eventually settles into the same nonchaotic spiking pattern; the reported basins and sensitivity numbers are tied to that chosen time cutoff.","fun_headline_variants_meta":{"raw":{"variants":["Fractal neuron basin needs 10^27 precision to predict","Neuron pair's chaotic mirage hides true stability","Asymmetric coupling yields extreme final state sensitivity","10^27 state accuracy required for coupled neuron fate","Quasimultistability: neurons trapped in transient chaos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1468,"prompt_tokens":938,"completion_tokens":530,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":452}},"tokens_in":554,"tokens_out":530,"duration_ms":5558,"temperature":1.0,"reasoning_tokens":452,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:51:00.731104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the basin fractions and the four-dimensional uncertainty exponent using Lyapunov-evaluation horizons of $10^3$, $10^4$, $10^5$, and $10^6$ iterations while holding everything else fixed. If the uncertainty exponent rises toward 1 or the measured fraction of the chaotic pseudo-attractor basin falls toward zero as the horizon grows, the reported extreme sensitivity is a cutoff artifact; if the exponent stays near 0.037 and the fraction holds, the claim stands.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the modified form of the Rulkov map and the map-based neuronal dynamics context used by the model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the normalized-distance basin classification method with the power-law form for basin occupancy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines final-state sensitivity and motivates measuring how initial-condition uncertainty transfers to final-state uncertainty."},{"cited_title":"Bashkirtseva, A","cited_arxiv_id":null,"evidence_quote":"Establishes the uncertainty-exponent method and the relation between uncertainty exponent and basin-boundary dimension."},{"cited_title":"Gotthans, J","cited_arxiv_id":null,"evidence_quote":"Gives the explicit Jacobian rendering and Lyapunov-exponent computation pipeline used to classify orbits as chaotic or not."}],"review_version":1}