{"id":"cec97d2b-ba2a-4bdb-a56d-eea0a34bc5cb","arxiv_id":"2504.14546","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"TeMBM models a time-varying Hurst exponent with a smoothed telegraph process, yields a stationary beta distribution, and shows a decaying ensemble autocovariance signature that separates it from FBM and FBMRE.","lead":"Telegraphic multifractional Brownian motion lets the Hurst exponent, a random walk's memory strength, switch between two levels and relax smoothly over time, producing the beta distribution of exponents seen in experiments. The paper also proposes an autocovariance-based rule to distinguish it from fixed-H and realization-random-H processes, and apply it to weather, cell, and electricity-price data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FBM's zero-ACVF baseline is not structurally guaranteed: finite-sample TAMSD slope estimates for constant-H FBM can be correlated across segments (via overlap or long-range dependence), so the three-way classifier may mislabel persistent FBM as TeMBM.","rationale":"The analytical model contribution is credible and independently supported: the beta stationary distribution and the ACVF formula (Eqs. (6) and (8)) follow from standard smoothed-telegraph theory, and the authors provide reproducible simulation code. My concern targets the separate empirical-identification claim. The reader's weakest assumption already identified the segment-wise TAMSD estimates as load-bearing; I sharpen this by pointing to a concrete mechanism, namely that the estimation-error process itself can have a nonzero, decaying autocovariance even when the true H is constant. Overlapping windows create shared-data correlations by construction, and for persistent FBM the long-range dependence of increments can correlate even disjoint-segment estimates. Figure 3(a) only shows FBM with H = 0.1, so the universal zero baseline is not demonstrated. This does not overturn the model or the derivation, but it means the central methodology is conditional on an unstated calibration of w, o and on an unverified insensitivity of the FBM baseline to H. The reader's CONDITIONAL verdict captures this appropriately, so I recommend no change to the verdict, with the specific test above as the decisive next step.","tokens_in":19836,"tokens_out":10177,"duration_ms":107039,"concrete_test":"Run the Supplemental Sec. VI-VII estimation and distinguishing algorithm on 5,000 simulated FBM trajectories with H = 0.4 and H = 0.8 (and H = 0.1 as a control), for a grid of segment lengths w (e.g., 50, 100, 200) and overlaps o = 0, w/4, w/2, w - 1; compute the sample ACVF gamma(j,k) as in Eq. (20) of the Supplement. If for any persistent-H FBM the ACVF at small lags is significantly above zero and decays monotonically beyond the 95% confidence band shown in Fig. 3(a), the FBM baseline is not zero and the classifier cannot separate FBM from TeMBM without calibrated thresholds on w and o.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The distinguishing algorithm in Supplemental Sec. VII rests on the claim that FBM yields a zero sample ACVF of estimated Hurst exponents. This is not a structural property of FBM: the Hurst estimates in Supplemental Sec. VI are finite-sample TAMSD slopes (Eq. (9) of the main text), and for overlapping segments (o > 0, which the algorithm permits) adjacent estimates share data and are positively correlated. Even for disjoint segments, FBM increments for H > 1/2 are long-range correlated, so slope-estimation errors across segments need not vanish. Thus the sample ACVF of estimated H for a constant-H FBM can be nonzero and decaying at small lags, exactly the TeMBM signature. Figure 3(a) tests FBM only at H = 0.1 and with one unspecified (w, o); it does not establish the zero baseline for persistent FBM. Moreover, Eq. (8) is the ACVF of the underlying H(t) process, not of the estimated segment-wise Hurst exponents; the resemblance in Fig. 3(a) is a numerical demonstration for one parameter set, not a proven invariance. If H = 0.8 FBM yields a decaying ACVF that exceeds the 95% confidence band, the classification procedure cannot separate FBM from TeMBM without specifying and justifying w, o and a quantitative decision rule.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces telegraphic multifractional Brownian motion (TeMBM), in which the Hurst exponent H(t) follows a stationary smoothed telegraph process with a beta stationary density. The authors derive the conditional autocovariance of the process, the stationary beta density of H(t), and the autocovariance function of H(t) itself. They then propose an estimation and classification scheme: segment-wise TAMSD slopes yield estimates of H along a trajectory, and the ensemble sample ACVF of these estimates is claimed to be zero for FBM, constant nonzero for FBMRE, and decaying for TeMBM. The model is validated on simulations, and the classification procedure is applied to temperature-anomaly data, quantum-dot trajectories, mucin-gel bead trajectories, and electricity-price data, with the authors concluding FBM, FBMRE, FBMRE, and TeMBM, respectively.","tokens_in":20115,"tokens_out":9917,"duration_ms":91259,"significance":"If the classification methodology were fully validated, TeMBM would be a valuable analytically tractable model for within-trajectory Hurst-exponent fluctuations, with a stationary beta marginal that matches previously reported experimental fits. The analytical derivations of Eqs. (2)-(3), (6), and (8) are standard and internally consistent, and the simulation checks against the analytical stationary density and ACVF are encouraging. The provision of code on GitHub is also a strength. The main weakness is that the distinguishing methodology is not yet calibrated: the zero baseline for FBM is asserted from a property of H(t) rather than demonstrated for the finite-sample estimator, only one FBM parameter is tested, and no quantitative decision rule is provided. The model contribution is solid, but the paper's claim to provide 'a methodology to identify our model in experimental data' needs substantial additional validation.","major_comments":[{"comment":"The defining equation for H(t) is dimensionally inconsistent as printed. Eq. (5) reads dH/dt = -H(t) + H_TP(t)/tau; if tau is a relaxation time, the first term should be -H(t)/tau, i.e., dH/dt = (H_TP(t) - H(t))/tau. This is exactly the equation used in Supplement Sec. II (where f(x) = -gamma x with gamma = 1/tau) and in Supplement Sec. V, step 2(b), and it is the equation that yields the beta density in Eq. (6). Please correct Eq. (5) and make the main text consistent with the Supplement.","section":"Main text, Eq. (5)"},{"comment":"The claimed zero baseline for FBM is a property of the constant underlying H(t), not of the estimated segment-wise Hurst exponents that the algorithm actually analyzes. Finite-sample TAMSD slope estimates are correlated across overlapping segments (o > 0 is permitted by Supplement Sec. VI), and for persistent FBM with H > 1/2 the long-range dependence of increments can induce correlated estimation errors even for disjoint segments. Consequently, a constant-H FBM can produce a nonzero, decaying sample ACVF at small lags, which is precisely the TeMBM signature. Figure 3(a) tests FBM only at H = 0.1 and with one unstated choice of w and o, so it does not establish the baseline. Please add simulations for persistent FBM (e.g., H = 0.6 and H = 0.8), for both disjoint and overlapping segmentation, and report the resulting ACVFs with confidence bands.","section":"Fig. 3(a) and Supplement Secs. VI-VII"},{"comment":"The classification step is performed by visual inspection of whether the sample ACVF 'stabilizes at zero,' 'stabilizes at a non-zero level,' or 'decays.' No quantitative decision rule is specified, and the key algorithm parameters w (segment length) and o (overlap) are not given numerical values anywhere in the main text or the Supplement. Without these values and a calibrated decision criterion, the method cannot be applied objectively to new data. Please specify w and o, justify them through a sensitivity analysis over w, o, trajectory length N, and ensemble size M, and define the decision rule used for the classifications in Fig. 3(b).","section":"Supplement Sec. VII and Fig. 3(b)"}],"minor_comments":[{"comment":"Equation (20) defines gamma(j,k) as an unnormalized sum over trajectories; for a sample ACVF it should be divided by M (or M-1).","section":"Supplement Sec. VII, Eq. (20)"},{"comment":"The 95% confidence bands in Fig. 3 are not defined; please state how they are computed, for example by bootstrap, analytically, or from across-trajectory variation.","section":"Fig. 3"},{"comment":"The Supplemental Material author line contains corrupted names ('Micha/suppress l Balcerek' and 'Wy/suppress loma´ nska'); the published PDF must render the author names correctly.","section":"Supplemental Material title page"},{"comment":"The statement that the diffusion coefficient can be ignored 'under certain conditions' should specify those conditions, since short-window log-log slope estimates can be affected by the noise level and the fitting range.","section":"Supplement Sec. VII"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision rather than rejection because the issues are fixable within the manuscript's scope: correct Eq. (5), add persistent-FBM and sensitivity simulations, and state a decision rule. I do not see circularity in the derivations; the reliance on the authors' earlier Ref. [74] for experimental beta distributions is self-referential but not logically circular. The real-data classification would be substantially more convincing if the FBM zero baseline were demonstrated for persistent H values and for the exact segmentation parameters used."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The model is the real contribution here. Putting a smoothed telegraph process inside the MBM spectral representation gives you a tractable process where H(t) fluctuates within a single trajectory and still has a clean stationary beta distribution. That is genuinely new relative to the earlier FBMRE work, where H was constant per realization. The derivations of the ACVF, the beta density, and the H-ACVF are standard but careful, and the simulations match the analytics. The paper also ships code and data, which is good practice and makes the simulation part reproducible.\n\nThe soft spot is the classification methodology. The claim that FBM gives a zero sample ACVF of estimated Hurst exponents is not structurally guaranteed. The estimates are finite-sample TAMSD slopes from overlapping segments. Adjacent segments share data, so their estimation errors are correlated. For persistent FBM with H > 1/2, the increments themselves are long-range dependent, so even disjoint segments can produce correlated slope errors. The paper only tests FBM at H = 0.1 and with one unspecified segment length and overlap. If an H = 0.8 FBM produces a decaying sample ACVF that looks like the TeMBM signature, the three-way classifier is not separating the two processes. This is not a minor gap; it is load-bearing for the empirical identification claim.\n\nThe real-data section is also more visual than quantitative. There is no fitted TeMBM model for dataset 4, no quantitative decision rule for calling a decay \"decaying,\" and no comparison with alternative time-varying-H models. The reader's weakest-assumption note nails it: the whole procedure hinges on the estimator preserving the shape of the underlying H(t) autocovariance, and that is demonstrated numerically for one parameter set, not proven.\n\nThat said, the core model is sound and the classification idea is worth developing. The gaps are addressable with a calibration study: vary H, segment length, and overlap; show the false-positive rate for persistent FBM; give a threshold rule for calling a decay. I would send this to peer review, because the constructive model deserves scrutiny and the empirical method needs to be pushed before it becomes a standard tool. A referee should focus on the statistical baseline, not on the analytics.","headline":"The TeMBM model is a genuinely useful addition, but the classification method needs a firmer statistical baseline before I would trust the empirical claims.","tokens_in":20704,"tokens_out":1406,"would_cite":true,"duration_ms":14468,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G22","60G18","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a smoothed telegraph process for the Hurst exponent yields a tractable multifractional Brownian motion, and a single autocovariance measurement can tell it apart from fixed- and random-exponent fractional Brownian…","keywords":["telegraphic multifractional Brownian motion","Hurst exponent","smoothed telegraph process","beta distribution","autocovariance classification","fractional Brownian motion","single-particle tracking","anomalous diffusion"],"falsifier":"Apply the estimation algorithm to simulated TeMBM where the telegraph switching rate is much faster than the estimation segment length, so that smoothing hides most switches within segments; if the sample autocovariance of estimated Hurst exponents shows a constant plateau instead of the exponential decay of Eq. (8), the classifier would mislabel TeMBM as FBMRE in a resolvable and physically plausible parameter regime.","tokens_in":19641,"feed_emoji":"📊","tokens_out":10187,"duration_ms":82510,"temperature":0.7,"pith_summary":"Multifractional Brownian motion has lacked a simple analytical model that lets the Hurst exponent $H(t)$ wander along a single trajectory while remaining bounded and tractable. This paper supplies one: model $H(t)$ as a smoothed telegraph process, a two-level random switch filtered through a relaxation time. The smoothed process has a stationary beta distribution, matching exponent histograms seen in biological single-particle tracking, and the paper derives its autocovariance. It then gives a practical procedure that separates this telegraphic multifractional Brownian motion from plain fractional Brownian motion and from fractional Brownian motion with a random exponent: the autocovariance of estimated Hurst exponents is decaying for TeMBM, constant for FBMRE, and zero for FBM. The paper demonstrates the procedure on simulated trajectories and on real data from biology, climate, and finance.","feed_headline":"One autocovariance curve separates three fractional-motion classes","feed_subtitle":"For TeMBM, the estimated Hurst exponent's autocovariance decays; for FBMRE it plateaus, and for FBM it is zero.","key_machinery":"The load-bearing object is the smoothed telegraph process $H(t)$, generated by $dH/dt=-H+H_{TP}(t)/\\tau$, where $H_{TP}$ is a two-state telegraph switching between $H_1$ and $H_2$ with rates $\\lambda_{12}$ and $\\lambda_{21}$. The smoothing filter $\\tau$ produces continuous, bounded paths and a stationary $\\beta$ distribution with unimodal or bimodal shapes. Its autocovariance, Eq. (8), is $\\langle(H(t)-\\langle H\\rangle)(H(s)-\\langle H\\rangle)\\rangle = \\frac{\\lambda_{12}\\lambda_{21}(H_2-H_1)^2}{4\\lambda^2(4\\lambda^2\\tau^2-1)}\\left(2\\lambda\\tau e^{-|t-s|/\\tau}-e^{-2\\lambda|t-s|}\\right)$. The classification protocol rests on comparing the sample autocovariance of estimated Hurst exponents against the three ideal shapes that this formula predicts.","core_discovery":"On the paper's own terms, telegraphic multifractional Brownian motion (TeMBM) is defined through the spectral representation $B_{H(t)}(t)=C(H(t))\\int_{-\\infty}^{\\infty}\\frac{e^{i\\omega t}-1}{|\\omega|^{H(t)+1/2}}dB(\\omega)$, with $H(t)$ a stationary smoothed telegraph process independent of $B(t)$. This renders the Hurst exponent bounded, smooth, and random along a single trajectory, and its stationary law is the $\\beta$ distribution $p(h)\\propto(h-H_1)^{\\lambda_{12}\\tau-1}(H_2-h)^{\\lambda_{21}\\tau-1}$. The central discovery for data analysis is that the ensemble autocovariance of segment-wise estimated Hurst exponents reproduces the autocovariance of $H(t)$: zero for fixed-exponent FBM, constant for FBMRE, and the decaying exponential combination of Eq. (8) for TeMBM. The paper uses these three signatures to classify trajectories and reports TeMBM-like decay for electricity prices, FBMRE-like plateaus for quantum dots in cells and beads in mucin gels, and FBM-like zero for temperature anomalies.","pith_inferences":["Beyond the paper, the three-way autocovariance diagnostic could be turned into a formal model-selection test by deriving the finite-sample distribution of the estimated autocovariance under each model; that would supply the missing quantitative threshold for segment length and overlap.","The same smoothed-telegraph construction could be applied to stochastic diffusivity instead of the Hurst exponent, coupling time-varying mobility and time-varying memory in one analytically tractable process; the paper notes the need for such combined models but does not construct one.","Because the beta-distribution match in earlier data is between the model and histograms of estimated exponents, the fitted parameters may partly absorb the estimator's smoothing; a direct check would compare the switching times implied by fitted rates with switches detected in individual trajectories.","If the electricity-price result holds up, the time-dependent Hurst exponent becomes a quantitative, continuously varying measure of market efficiency, giving the qualitative adaptive-market narrative a concrete observable; the paper hints at this interpretation but does not develop it."],"forward_implications":["Any data set whose exponent autocovariance decays as in Eq. (8) can be represented by TeMBM, and the fit gives its switching levels, rates, and relaxation time.","In single-particle tracking, a decaying exponent autocovariance indicates the medium or the particle's state is changing in time, whereas a constant plateau indicates static heterogeneity between trajectories.","The distinguishing procedure does not need a priori knowledge of the model and applies to any multifractional Brownian motion, because the three shape categories are defined by the autocovariance of the estimated exponent.","Fitting the beta distribution to exponent histograms and Eq. (8) to their autocovariance yields a complete parametrisation of TeMBM from experimental trajectories."],"supporting_citations":[{"why":"Defines fractional Brownian motion by the spectral integral representation that TeMBM generalizes, supplying the baseline covariance structure.","marker":"[7]"},{"why":"Establishes the theory of multifractional processes with random exponent, including the self-similarity property TeMBM inherits when the Hurst exponent is stationary.","marker":"[30]"},{"why":"Gives the stationary distribution and autocovariance of the asymmetric random telegraph signal, from which the beta density and Eq. (8) follow.","marker":"[73]"},{"why":"Introduces FBM with a beta-distributed random Hurst exponent, the FBMRE class that TeMBM must be distinguished from, together with earlier biological fits.","marker":"[74]"},{"why":"Supplies the quantum-dot cytoplasm data set whose exponent autocovariance plateau identifies FBMRE.","marker":"[47]"},{"why":"Supplies the mucin-hydrogel bead data set whose exponent autocovariance plateau identifies FBMRE.","marker":"[42]"},{"why":"Provides the dichotomous Markov noise formalism used to derive the stationary beta distribution of the smoothed telegraph process.","marker":"[56]"},{"why":"Provides the differentiation formulas used in the derivation of the autocovariance of the smoothed telegraph process.","marker":"[57]"},{"why":"Reports bimodal Hurst-exponent distributions in biological single-particle tracking, motivating the beta family's flexibility.","marker":"[48]"}],"fun_headline_variants":["Hurst autocovariance: zero, plateau, decay separates models","Telegraphic Hurst exponent: autocovariance reveals the model","Beta-distributed Hurst exponent: telegraphic Brownian motion","Autocovariance shape tells FBM, FBMRE, TeMBM apart"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on the assumption that Hurst exponents estimated from short overlapping TAMSD segments are accurate enough for the ensemble autocovariance of those estimates to preserve the true autocovariance shape (zero, constant, or decaying) of $H(t)$, yet the paper specifies no quantitative rule for choosing the segment length $w$ and overlap $o$ that guarantee this.","fun_headline_variants_meta":{"raw":{"variants":["Hurst autocovariance: zero, plateau, decay separates models","Telegraphic Hurst exponent: autocovariance reveals the model","Beta-distributed Hurst exponent: telegraphic Brownian motion","Autocovariance shape tells FBM, FBMRE, TeMBM apart"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000422,"raw_usage":{"total_tokens":2143,"prompt_tokens":898,"completion_tokens":1245,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":1167}},"tokens_in":514,"tokens_out":1245,"duration_ms":10323,"temperature":1.0,"reasoning_tokens":1167,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:46:17.548781+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the estimation algorithm to simulated TeMBM where the telegraph switching rate is much faster than the estimation segment length, so that smoothing hides most switches within segments; if the sample autocovariance of estimated Hurst exponents shows a constant plateau instead of the exponential decay of Eq. (8), the classifier would mislabel TeMBM as FBMRE in a resolvable and physically plausible parameter regime.","supporting_citations":[{"cited_title":"FitzHugh, Statistical properties of the asymmetric random telegraph signal, with applications to single- channel analysis, Mathematical Biosciences 64, 75 (1983)","cited_arxiv_id":null,"evidence_quote":"Gives the stationary distribution and autocovariance of the asymmetric random telegraph signal, from which the beta density and Eq. (8) follow."},{"cited_title":"Balcerek, K","cited_arxiv_id":null,"evidence_quote":"Introduces FBM with a beta-distributed random Hurst exponent, the FBMRE class that TeMBM must be distinguished from, together with earlier biological fits."}],"review_version":1}