{"id":"704f9c46-23f5-4c53-be43-6a8cd0ed20c6","arxiv_id":"1908.11388","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"On a random regular graph with intermediate disorder, an initially localized wave packet spreads subdiffusively, with width growing as t^beta where beta = 1 - W/W_AT, for a disorder range where earlier work expected diffusion.","lead":"This paper studies how a quantum particle spreads on a random regular graph, a tree-like network used as a simplified model of many-body localization. It reports that, for intermediate disorder, the wave packet expands much slower than ordinary diffusion, moving as a subdiffusive front, and connects this to slow equilibration in disordered quantum systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central phase claim inherits the stretched-exponential exponent beta(W)=1-W/W_AT from Ref.","rationale":"The paper contains a genuinely useful numerical dataset: the four-regime decomposition of Pi(x,t), the front collapse at fixed beta, and the f/L robustness checks in Appendix A are real evidence that something slower than diffusion happens on the accessible RRG. I do not fault the authors for using their own previous result self-consistently; citing prior work is normal. The load-bearing weakness is that the phase statement is quantitatively anchored to Eq. (5) while the new observables are analyzed only through that anchor. The self-consistent semiclassical relation in Eqs. (7)-(9) connects the front exponent to the return-probability exponent, so if Eq. (5) is accepted, the consistency of the collapses is a meaningful check; but it is not an independent measurement. The paper's own final paragraph acknowledges the thermodynamic-limit caveat, and that caveat is central rather than cosmetic: the RRG diameter grows only logarithmically with L, so the graph cannot separate a genuine subdiffusive phase from a long crossover. On the strength of the data, a conditional accept is appropriate, with the condition that the exponent be measured freely and that the time-window stability be quantified. This matches the reader's verdict; no verdict change is needed.","tokens_in":17202,"tokens_out":4942,"duration_ms":51268,"concrete_test":"Take the same L=2^20 exact-evolution data (or the publicly deposited data) at W=8, 10, and 12 and re-analyze without imposing Eq. (5): (i) fit X(t)=A t^beta + B over a sliding time window before saturation and record the free beta; (ii) compute the log-log derivative d ln X/d ln t as a function of t; (iii) do the same two-parameter fit for ln[Pi(0,t)-Pi(0,infty)] = ln C - Gamma t^beta. If the free beta is within ~0.1 of 1-W/W_AT and does not drift toward 1 as the window moves to later times, the imported exponent is validated; if it drifts or differs, Eqs. (10)-(11) are an imposed collapse and the subdiffusive-phase claim should be downgraded to a transient crossover. This settles the concern without relying on the prior fit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations (10) and (11) assert X_front(t) ~ Gamma t^beta and X(t) ~ t^beta, but beta(W)=1-W/W_AT is not measured in this paper; it is taken from the same group's return-probability study (Ref. [73], quoted as Eq. (5)). The collapses in Figs. 4(b), 5(b), and 11 use this value as the time-rescaling input, so the data collapse demonstrates consistency with the assumed exponent, not a free extraction of it. The only independent input is the qualitative separation between four dynamical regimes and the observation that X(t) grows slower than linearly before the front reaches the graph diameter. Since the RRG diameter D ~ ln L / ln K is about 14 at L=2^20, the pre-saturation time window is t < t_Th ~ (ln L)^{1/beta} (roughly 10^2 at W=8-12), and the paper explicitly concedes in the Conclusions that a crossover to beta=1 in the thermodynamic limit cannot be ruled out. If the stretched-exponential fit in Ref. [73] is biased or finite-time, or if the true asymptotic beta at fixed W is different, the claimed subdiffusive phase has no independent quantitative support. No error bars or collapse-quality metrics are given, and the absence of code prevents machine-checking or reproduction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript studies the finite-time dynamics of an initially localized wave packet in the Anderson model on a random regular graph (RRG), using the full probability distribution Π(x,t) as a function of distance from the initial site. The authors identify four space-time regimes controlled by a propagating wave front X_front(t) and present numerical evidence that both the front and the mean distance X(t) grow sub-diffusively as t^β with β(W) ≈ 1 − W/W_AT for 0.4 W_AT ≲ W ≲ 0.7 W_AT. They also demonstrate a space-time factorization Π(x,t) − Π(x,∞) = g(x)[Π(0,t) − Π(0,∞)] after the front passes, connect the return probability to the front position through an exponential relation, and argue that the RRG can serve as a proxy for many-body localization dynamics. The numerical results are for L up to 2^20, with the energy-shell fraction f = 1/8, and are supported by appendices on finite-size and energy-window dependence.","tokens_in":17570,"tokens_out":5841,"duration_ms":58574,"significance":"If the claims hold, this paper would establish that the RRG Anderson model hosts a genuine dynamical subdiffusive phase over a finite disorder interval, in contrast to Anderson models on Z^d with d>2 where subdiffusion is restricted to the critical point. The four-regime description of Π(x,t) and the factorization property are potentially useful organizing principles for slow dynamics near the MBL transition, and the connection to the Thouless time scaling is suggestive. The paper's strengths include the use of a well-defined observable (the radial distribution), consistency checks on the energy window and system size in Appendix A, and an explicit statement of the residual uncertainty about a thermodynamic-limit crossover to diffusion. However, the central exponent β(W) is taken from the same authors' previous return-probability work (Ref. [73]) rather than measured here, the numerical collapses use that value as an input, and the paper provides no error bars, realization counts, or code/data release. These issues make the quantitative phase claim weaker than the qualitative observation of sublinear spreading.","major_comments":[{"comment":"The exponent β(W) is not measured in this paper. It is imported from Ref. [73] as Eq. (5), and then used as the time-rescaling exponent in the collapses of Figs. 4(b), 5(b), and 11. Consequently, those collapses demonstrate consistency with the assumed β(W), not an independent extraction of β from X(t) or X_front(t). Since the central quantitative claim is X_front(t) ~ Γ(W)t^β and X(t) ~ t^β, the authors should fit β directly from the new data, report confidence intervals and collapse-quality metrics, and compare the fitted values with Eq. (5). Without this, the subdiffusive-phase exponent remains an inherited assumption rather than a result of the present analysis.","section":"§Model and methods, Eq. (5); §Results, Eqs. (10)–(11); Figs. 4(b), 5(b), 11"},{"comment":"No statistical information is reported: the numbers of disorder realizations, graph samples, and initial states are not given, and no error bars appear on Π(x,t), X(t), or the collapse curves. For a numerical claim about a scaling exponent and a collapse, this is load-bearing: the apparent agreement with the proposed functional forms cannot be distinguished from finite-sampling scatter or systematic drift. Please provide uncertainties (for example, bootstrap over samples) and state the number of realizations for each W and L. Releasing the data-processing code would also allow the collapses to be checked.","section":"§Results, Figs. 2–5 and Appendix A"},{"comment":"The paper explicitly concedes that a crossover to β = 1 in the thermodynamic limit cannot be ruled out. Because the RRG diameter grows only as ln L and the pre-saturation time window is t < t_Th ~ (ln L)^{1/β}, all data shown are intrinsically finite-time. The abstract's claim of an \"entire subdiffusive phase\" is therefore stronger than the evidence presented. Either soften the claim to finite-time subdiffusive dynamics or add a scaling analysis in L showing that the exponent extracted in the available window is stable as L increases.","section":"Conclusions, final paragraph before Acknowledgments"},{"comment":"The abstract states that the numerical results are supported by a \"self-consistent semiclassical picture\" relating β to the relaxation rate of the return probability, but no derivation of β is given. Appendix D shows only that the two collapse forms, Eqs. (7) and (8), are mutually consistent if the return probability decays exponentially with the front position, δR(x) ~ e^{−λx}; it does not determine β(W). The wording should be revised to describe this as a consistency condition or an ansatz, not a self-consistent derivation.","section":"Abstract and Appendix D"}],"minor_comments":[{"comment":"The term \"subdiffusive\" is used with a nonstandard definition (X(t) ~ t^β with β < 1 on a hierarchical graph, where diffusive motion would be X(t) ~ t). This is explained in the text, but the abstract and Fig. 1 could benefit from an early clarification to avoid confusion with the standard mean-square-displacement terminology.","section":"§Introduction, Fig. 1"},{"comment":"The figure axis is labeled X^{−1}(0) while the text discusses X(0); please check the label or define the plotted quantity explicitly.","section":"Appendix A, Fig. 6"},{"comment":"The sentence \"Xfront(t) it is given by the value of x for which Π(x,t) has a maximum\" contains an extra \"it is\"; please correct to \"Xfront(t) is given by...\".","section":"Footnote 79"},{"comment":"The inset is said to show all four stages of the evolution, but the caption does not explicitly mark the four stages; adding labels or arrows would make the claimed four-regime structure easier to verify.","section":"§Results, Fig. 4(a) inset"}],"recommendation":"major_revision","confidential_remarks":"The central exponent is taken from the same group's previous paper (Ref. [73]), and the present collapses are not independent of that input. The editor may wish to have the underlying Ref. [73] analysis checked with the same care as this manuscript, since the phase claim rests on it. The absence of error bars and realization counts is also a significant reproducibility concern for a numerical paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a solid numerical study of wave-packet spreading on the random regular graph, and the genuinely new piece is the full spatial analysis of Pi(x,t): the four-regime decomposition, the sublinear motion of the wave front, and the space-time factorization connecting front propagation to the return probability. The data in Fig. 5 do show X(t) growing slower than linearly on the available time scales, and the factorization is demonstrated consistently across W = 8–13. That is a real contribution and worth reading.\n\nThe soft spot is the load-bearing exponent. beta(W) = 1 - W/W_AT is not measured in this paper; it is imported from the same group's earlier return-probability fit (Ref. [73], Eq. 5). The collapses in Figs. 4, 5, and 11 rescale time with that beta, so the agreement establishes consistency with an assumed exponent, not an independent extraction. Since the 'subdiffusive phase' claim is quantitatively tied to beta, the strongest statement this paper can support on its own is that spreading is sublinear on finite times. The authors are honest about this, explicitly conceding in the Conclusions that a crossover to beta = 1 in the thermodynamic limit cannot be ruled out. The pre-saturation time window is also short—t_Th is on the order of 10^2 at L = 2^20—so the asymptotic regime is not probed. The dispute with Refs. [61,85], which claim diffusive or ballistic propagation, is left unresolved rather than engaged quantitatively.\n\nAlso, no error bars, realization counts, or code/data are given, so the collapse quality can't be checked independently. That is a minor-to-moderate reproducibility issue, not a fatal one.\n\nOverall: the qualitative subdiffusive dynamics and the four-regime picture are plausible and likely correct at these finite sizes. The phase interpretation is conditional. A serious referee should engage; I would ask for an independent fit of beta from X(t) (or a clear statement that the exponent is being assumed), error estimates, and a more quantitative confrontation with the ballistic/diffusive claims before accepting the phase language.\n\nRecommendation: send to peer review, with an expectation of revision. I would cite this work for the four-regime analysis and the numerical observation, with a caveat on the exponent.","headline":"Plausible and well-organized numerical evidence for subdiffusive spreading on the RRG, but the phase claim leans on an inherited exponent that the collapses do not independently test.","tokens_in":18000,"tokens_out":2545,"would_cite":true,"duration_ms":25997,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An initially localized wave packet on a random regular graph spreads subdiffusively: both the wave front and the mean distance grow as $t^{\\beta}$ with $\\beta(W) \\approx 1 - W/W_{\\mathrm{AT}}$, evidence of a subdiffusive phase rather than…","keywords":["Anderson localization","random regular graph","subdiffusion","wave-packet dynamics","return probability","stretched exponential","many-body localization","non-ergodic phase"],"falsifier":"A direct test would be to evolve wave packets on random regular graphs larger than $L = 2^{20}$ and longer than the current maximum times, extracting $\\beta(W)$ separately from the wave-front position, the first moment $X(t)$, and the return probability. If the three fitted exponents disagree, or if $X(t)$ visibly crosses over to linear growth at any fixed $W$ in $8 \\le W \\le 14$, the subdiffusive-phase claim as stated would be falsified.","tokens_in":17016,"feed_emoji":"🌀","tokens_out":7644,"duration_ms":67789,"temperature":0.7,"pith_summary":"The paper studies how an initially localized wave packet spreads on a random regular graph, a tractable stand-in for the Fock space of a disordered interacting system. It claims that for a range of disorder strengths below the Anderson transition the dynamics is genuinely subdiffusive: both the wave front and the mean distance grow as $t^{\\beta}$ with $\\beta < 1$ over the entire range, not merely at a critical point. The exponent is the same $\\beta(W) \\approx 1 - W/W_{\\mathrm{AT}}$ that controls the stretched-exponential decay of the return probability, connecting spatial propagation to the relaxation of the initial site. If correct, this establishes a full subdiffusive phase on the random regular graph and, through the many-body-localization proxy, suggests a mechanism for slow relaxation of local observables in disordered interacting systems.","feed_headline":"Wave fronts on random graphs spread as t^β, with β<1","feed_subtitle":"For disorder between 40% and 70% of the Anderson transition, transport is subdiffusive—a proxy for many-body localization.","key_machinery":"The central object is the distance-resolved probability distribution $\\Pi(x,t)$, the probability that the particle is at graph distance $x$ from its initial site at time $t$, restricted to a microcanonical energy shell around mid-spectrum. The wave front $X_{\\mathrm{front}}(t)$, defined as the distance at which $\\Pi(x,t)$ has its maximum, is the moving boundary that separates frozen, front, and relaxed regions in spacetime. The argument is carried by the relation between spatial and temporal relaxation: the stretched-exponential return probability $\\Pi(0,t) \\sim \\exp(-\\Gamma t^{\\beta(W)})$ fixes the exponent $\\beta(W) = 1 - W/W_{\\mathrm{AT}}$, and the same exponent collapses both the wave-front motion and the growth of the first moment $X(t)$. The space-time factorization behind the front, $\\Pi(x,t)-\\Pi(x,\\infty) = g(x)[\\Pi(0,t)-\\Pi(0,\\infty)]$, is what ties local relaxation to the front motion.","core_discovery":"On the Anderson model on a random regular graph, the full radial probability distribution $\\Pi(x,t)$ of an initially localized wave packet is shown to organize into four spacetime regimes determined by a moving wave front. Before the front arrives, the distribution is frozen; near the front, $\\Pi$ minus its infinite-time value collapses as a function of $X_{\\mathrm{front}}(t) - x$; behind the front, it factorizes as $g(x)[\\Pi(0,t)-\\Pi(0,\\infty)]$ and relaxes with the return probability; at very long times it saturates to the uniform-over-sites distribution. The central quantitative claim is that for disorder strengths $0.4W_{\\mathrm{AT}} \\lesssim W \\lesssim 0.7W_{\\mathrm{AT}}$, both the front $X_{\\mathrm{front}}(t) \\approx \\Gamma(W) t^{\\beta(W)}$ and the mean distance $X(t) \\sim t^{\\beta(W)}$ grow subdiffusively with $\\beta(W) \\approx 1 - W/W_{\\mathrm{AT}}$. Because the same exponent also controls the stretched-exponential return probability from earlier work, the paper presents the spatial spreading and the temporal relaxation as two faces of one subdiffusive phase, in contrast to $\\mathbb{Z}^d$ lattices with $d>2$ where subdiffusion appears only at the critical point.","pith_inferences":["If the stretched-exponential exponent survives at larger sizes, one could test whether the same $\\beta$ controls entanglement growth in the corresponding many-body system, since the graph distance maps to the Hamming distance in Fock space.","The four-regime structure is likely generic to locally tree-like graphs with exponential growth of the number of sites with distance; a possible extension is to derive $\\beta(W)$ from statistics of rare resonances along the boundary rather than importing it from the return probability.","A finite-time to diffusive crossover at larger graph sizes cannot be excluded from these data; the paper's own claim should be read as establishing the subdiffusive phase for the accessible time and size window, with the thermodynamic extrapolation as the open question."],"forward_implications":["If the claim holds, the random regular graph hosts a genuine subdiffusive phase for a finite interval of disorder, not just a critical point, with the exponent decreasing linearly to zero at the Anderson transition.","The Thouless time, when the front reaches the graph diameter, scales as $(\\ln L / \\ln K)^{1/\\beta(W)}$, which diverges as a power of the logarithm near the transition; this is the same scaling found in subdiffusive many-body-localized systems.","Slow spreading on the random regular graph implies slow relaxation of local observables in systems that map to it, offering a route to subdiffusion in many-body systems that does not invoke Griffiths effects.","The observed slow dynamics rules out full ergodicity in the random-matrix sense for $W \\ge 8$ even in parameter regions where eigenfunction-based measures such as the inverse participation ratio may look ergodic."],"supporting_citations":[{"why":"Supplies the stretched-exponential return probability $\\Pi(0,t) \\sim \\exp(-\\Gamma t^\\beta)$ with $\\beta(W) = 1 - W/W_{\\mathrm{AT}}$ that the paper imports and uses for all spatial collapses.","marker":"[73]"},{"why":"Provides the RRG Anderson localization transition value $W_{\\mathrm{AT}} \\approx 18.1$ and the phase-diagram context that the subdiffusive phase sits below it.","marker":"[34]"},{"why":"Documents the debated non-ergodic extended phase on regular graphs that motivates distinguishing dynamical subdiffusion from eigenfunction ergodicity.","marker":"[44]"},{"why":"Establishes the random regular graph as a proxy for many-body localization dynamics on Fock space, the setting the paper's conclusions are transferred to.","marker":"[33]"},{"why":"Provides the $\\mathbb{Z}^d$ comparison: subdiffusion on finite-dimensional lattices is believed to occur only at the Anderson critical point, making the RRG subdiffusive phase a singular large-dimension effect.","marker":"[67,68]"},{"why":"Gives the Bethe-lattice return-probability decay that underlines the difference between diffusive spreading on trees and on Euclidean lattices.","marker":"[77]"},{"why":"Provide the many-body Thouless-time scaling with which the paper compares its $(\\ln L)^{1/\\beta}$ estimate.","marker":"[82,83]"}],"fun_headline_variants":["Subdiffusive wave-front on random graphs: four regimes, one exponent","Subdiffusion in Anderson RRG: wave-front moves as t^β, β<1","Four spacetime regimes from a subdiffusive front in Anderson RRG","Wave-packet spread on random graphs: subdiffusive front, four regimes","Subdiffusive front exponent β≈1−W/W_AT on Anderson RRG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's spatial exponents are not fitted independently: it assumes that the stretched-exponential return probability with $\\beta(W) = 1 - W/W_{\\mathrm{AT}}$, taken from its earlier study, is accurate and not a finite-time artifact, and pins the wave-front and mean-distance growth to this same $\\beta$.","fun_headline_variants_meta":{"raw":{"variants":["Subdiffusive wave-front on random graphs: four regimes, one exponent","Subdiffusion in Anderson RRG: wave-front moves as t^β, β<1","Four spacetime regimes from a subdiffusive front in Anderson RRG","Wave-packet spread on random graphs: subdiffusive front, four regimes","Subdiffusive front exponent β≈1−W/W_AT on Anderson RRG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001117,"raw_usage":{"total_tokens":4721,"prompt_tokens":1089,"completion_tokens":3632,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":705,"completion_tokens_details":{"reasoning_tokens":3538}},"tokens_in":705,"tokens_out":3632,"duration_ms":25030,"temperature":1.0,"reasoning_tokens":3538,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:16:50.196077+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to evolve wave packets on random regular graphs larger than $L = 2^{20}$ and longer than the current maximum times, extracting $\\beta(W)$ separately from the wave-front position, the first moment $X(t)$, and the return probability. If the three fitted exponents disagree, or if $X(t)$ visibly crosses over to linear growth at any fixed $W$ in $8 \\le W \\le 14$, the subdiffusive-phase claim as stated would be falsified.","supporting_citations":[{"cited_title":"Return probability for the anderson model on the random regular graph,","cited_arxiv_id":null,"evidence_quote":"Supplies the stretched-exponential return probability $\\Pi(0,t) \\sim \\exp(-\\Gamma t^\\beta)$ with $\\beta(W) = 1 - W/W_{\\mathrm{AT}}$ that the paper imports and uses for all spatial collapses."},{"cited_title":"Critical behavior at the localization transition on random regular graphs,","cited_arxiv_id":null,"evidence_quote":"Provides the RRG Anderson localization transition value $W_{\\mathrm{AT}} \\approx 18.1$ and the phase-diagram context that the subdiffusive phase sits below it."},{"cited_title":"Non- ergodic delocalized phase in anderson model on bethe lattice and regular graph,","cited_arxiv_id":null,"evidence_quote":"Documents the debated non-ergodic extended phase on regular graphs that motivates distinguishing dynamical subdiffusion from eigenfunction ergodicity."},{"cited_title":"Delocalized glassy dynam- ics and many-body localization,","cited_arxiv_id":null,"evidence_quote":"Establishes the random regular graph as a proxy for many-body localization dynamics on Fock space, the setting the paper's conclusions are transferred to."},{"cited_title":"Exact closed form of the return prob- ability on the bethe lattice,","cited_arxiv_id":null,"evidence_quote":"Gives the Bethe-lattice return-probability decay that underlines the difference between diffusive spreading on trees and on Euclidean lattices."}],"review_version":1}