{"id":"ce1c8771-4d45-4c7b-a9c1-cfc2a3328574","arxiv_id":"2502.02566","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Tail bounds for Dyson series of random Schrödinger equations on Z^d and R^d give new eigenfunction localization and delocalization estimates, including Floquet states.","lead":"This math paper proves new bounds on how much a random potential can change quantum time evolution, showing the Dyson series can be truncated at times as long as λ^{-2+ε}. The same elementary proof yields localization and delocalization estimates for eigenfunctions and periodic (Floquet) systems, but the continuous-space part relies on a spatial cutoff that is not justified for the full time range.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"R^d proof of Theorem 1.1 fails at Lemma 3.5: the fixed cutoff χ_{100L} is used for times where the low-frequency propagator has spread well beyond it, so the finite-speed bound (3.7) is applied outside its validity.","rationale":"The reader's weakest assumption identifies exactly the step I would flag. Re-reading Lemma 3.5 and its use of (3.7), the mismatch between the fixed spatial cutoff B_{KL} and the time-dependent spreading scale Ls is real and quantitative: the finite-speed estimate is simply not applicable for s larger than a constant depending on K. This is an internal validity gap in Proposition 3.7, not a disagreement with external consensus. The deterministic simplex lemma and the Z^d argument are coherent and provide partial independent support, so the method is promising and likely repairable with a time-dependent cutoff, but the R^d theorem as stated is not proved. I agree with the reader's REJECT verdict, and no adjustment is needed.","tokens_in":22280,"tokens_out":23114,"duration_ms":242467,"concrete_test":"Verify Lemma 3.5 in a one-dimensional analogue: take L=10^3, K=101, R=1, s=60, x=KL, y=0, and compute the kernel ⟨x|e^{isH0}P_{≤L}|y⟩ from (3.5). Since |x−y|/(sL) ≈ 1.68 < 2, the phase (x−y)ξ + s|ξ|^2/2 has a stationary point inside supp ρ, so the kernel is ≍ s^{−1/2}, not (KL)^{−10}. Re-running the Schur test in Lemma 3.5 with this estimate gives an O(t) contribution at the level of (3.3), destroying the tL^{−10d} error term.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The R^d half of Theorem 1.1 rests on Proposition 3.7, whose finite-rank reduction uses Proposition 3.2. The load-bearing step is the spatial-cutoff estimate Lemma 3.5, which claims ∥T1(t)P_{≤L}(1−χ_{KL})∥_{op} ≤ Ct(KL)^{−10d} by invoking the low-frequency finite-speed bound (3.7). That bound is a stationary-phase estimate and holds only when |x−y| > 100Ls. For x outside B_{KL} and y inside B_R the separation is only ≈ KL, so (3.7) is available only for s ≲ K/100; even with the sharp non-stationary condition it is available only for s ≲ K/2. In the regime 1 < t < R^{5d} required by Proposition 3.7, almost the entire integration interval [0,t] lies beyond this range, and there the low-frequency propagator has already spread to radius ≈ Ls ≫ KL, so the kernel is of order s^{−d/2} rather than (KL)^{−N}. Thus the tL^{−10d} term in (3.3) is unjustified; the available estimates give no better than an O(t) error, which is far too large for the √t target. Consequently the R^d part of Theorem 1.1 and the R^d corollaries are unproven as stated. The Z^d argument and the deterministic T1→Tk reduction in Section 4 do not share this defect, and a time-dependent cutoff χ_{KLt} would likely repair the proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies random Schrödinger equations on Z^d and R^d (d ≥ 2) with a compactly supported random potential of strength λ, and derives high-probability tail bounds for the operator norms of the Dyson-series terms T_j(t). The main theorem asserts that, with probability at least 1 − 2 exp(−cK^2), one has ||T_j(t)||_op ≤ (K^2 σ_d(t)/j)^{j/2} for all j and t, where σ_d(t) is t log(2+t) for d>2 and t log^2(2+t) for d=2. The proof combines three ingredients: a bound on T_1(t) using the noncommutative Khintchine inequality, a deterministic reduction from T_j to powers of T_1 (Lemma 4.1), and pointwise dispersive estimates for the free propagator. On R^d, the argument introduces a frequency cutoff P_≤L and a spatial cutoff χ_{100L} to pass to a finite-rank approximation. From the main theorem the paper derives corollaries on frequency localization and spatial delocalization of approximate eigenfunctions, as well as an analogue for periodically driven (Floquet) systems.","tokens_in":22425,"tokens_out":4957,"duration_ms":47653,"significance":"If the main theorem were established in the stated generality, this would be a significant contribution: it replaces restriction theory and diagrammatic expansions with a self-contained random-matrix argument, removes energy restrictions (including the flat level set at E=0), and yields new results for Floquet states. The paper is also commendable for being fully self-contained: the noncommutative Khintchine inequality is proved in Appendix A, the dispersive estimates are proved in Lemmas 2.2 and 3.3, and the T1-to-Tk reduction is deterministic and parameter-free. No fitted parameters or external benchmarks are used. However, the proof on R^d contains a load-bearing gap in the spatial-cutoff step, so the R^d and Floquet corollaries are not currently supported.","major_comments":[{"comment":"The bound ||T_1(t)P_≤L(1−χ_{KL})||_op ≤ C t (KL)^{−10d} is obtained by writing the integral over [0,t] and applying the low-frequency finite-speed estimate (3.7) pointwise in s. However, (3.7) is valid only when |x−y| > 100 L s. For x in the support of 1−χ_{KL} and y in B_R, the separation is only about KL, so the condition holds only for s ≲ K/100 (or s ≲ K/2 under a sharp non-stationary phase condition), independent of L. In Proposition 3.7 the estimate is needed for all 1 < t < R^{5d}, so almost the entire integration interval lies beyond the range of validity of (3.7). In that range the low-frequency propagator has already spread to scale L s ≫ KL and the kernel is of order s^{−d/2}, giving no (KL)^{−N} gain. Thus the error term t L^{−10d} in (3.3) is unjustified; the available estimates give only an O(t) error, which is far too large for the √t target. This is a load-bearing gap that invalidates the R^d part of Theorem 1.1 as stated. A time-dependent cutoff of the form χ_{KL s} inside the integral would likely repair the proof, but as written the R^d theorem is unproven.","section":"§3, Lemma 3.5"},{"comment":"The time-dependent (Floquet) theorem and corollaries on R^d inherit the same defect. The proof of Theorem 1.4 in Section 5 states that the proof of Proposition 3.7 (which itself relies on Proposition 3.2 and Lemma 3.5) goes through verbatim. Since the spatial-cutoff step fails for intermediate times, the high-probability bound for sup_{a,t} ||T_1(a+t,a)|| is not established on R^d, and consequently Corollary 1.5 on R^d is also unsupported. The Z^d part of the Floquet statement does not share this issue because the discrete finite-speed estimate (2.2) is valid globally in time.","section":"§5 and §6"},{"comment":"Even setting aside the validity of Lemma 3.5, the decomposition of the error into (t^{1/2}L^{−1/2}R^{1/2} + tL^{−10d} + tε) is necessary for the later choice L=R^{100d} and ε=L^{−1}. With those choices the claimed error is small for 1 < t < R^{5d} only if the t L^{−10d} term is correct. Because that term is not justified, Proposition 3.2 cannot be used to approximate T_1(t) by a finite-rank operator on the required time interval. This comment is not an independent objection but rather clarifies that the flaw in Lemma 3.5 is not localized to a single estimate: it permeates the entire R^d reduction.","section":"§3, Proposition 3.2"}],"minor_comments":[{"comment":"The statement of Lemma 3.5 reads 't < Rwe have that ...'; there is a missing space and the intended condition is unclear. More importantly, the stated hypothesis 't < R' does not address the issue that (3.7) is used only for s ≲ K/100, so the proof of the displayed inequality would still be invalid for t larger than K/100.","section":"§3, Lemma 3.5"},{"comment":"The spelling 'Khinthine' appears in the appendix heading; it should be 'Khintchine' to match the main text and references.","section":"Appendix A"},{"comment":"In display (2.8), the expression contains a minor notational slip: after writing 'χ_L ∑_{|n|≤R} e^{is_1 H0}|n⟩⟨n|e^{−is_1 H0}χ_L e^{is_0 H0}|n⟩⟨n|e^{−is_0 H0}χ_L' the second copy of |n⟩⟨n| is not explicitly displayed, and the same line mixes operator and scalar notation. The intended meaning is clear, but the presentation would benefit from a cleaner derivation.","section":"§2, proof of Proposition 2.4"},{"comment":"The statement uses λ for both the coupling constant and the large-deviation parameter, writing 'for all λ ≥ √log(R)' where a different letter (e.g., K) would avoid confusion. This is consistent with the authors' convention but can be misread in the context of the main theorem.","section":"§5, Proposition 5.1"}],"recommendation":"major_revision","confidential_remarks":"The reader's stress-test concern about Lemma 3.5 is, on close reading, correct and load-bearing: the R^d proof applies the low-frequency finite-speed estimate outside its time range of validity. The Z^d half and the deterministic T1-to-Tk reduction appear sound, and the stated result is likely repairable by using a time-dependent cutoff. The paper is not rejectable on circularity or self-consistency grounds; it is simply incomplete in its current R^d part. I recommend major revision rather than rejection because the flaw is localized to the spatial-cutoff argument and there is a plausible repair path within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the Z^d half is a clean, honest piece of work, and Lemma 4.1 is a genuinely neat deterministic reduction. But the R^d half has a load-bearing gap in the cutoff argument. I checked the stress-test concern and it lands: Lemma 3.5 uses the finite-speed bound (3.7) outside its validity, so Theorem 1.1 on R^d is not proved as written.\n\nWhat is actually new: the reduction from ||T_k|| to ||T_1|| via the simplex decomposition is a real idea and is deterministic. Applying the noncommutative Khintchine inequality to the Dyson series in this setting is a sensible unification, and the payoff — all energies including E=0, all d≥2, plus the Floquet analogue — is a genuine improvement over [SSW02] and [Che05]. The paper is also self-contained: the NCK inequality is proved in an appendix, the dispersive estimates are proved, and I see no fitted parameters or circularities.\n\nThe soft spot is the R^d finite-rank approximation. Lemma 3.5 claims ||T1(t)P≤L(1-χ_KL)|| ≤ Ct(KL)^{-10d}, and the proof applies (3.7) for every s in [0,t]. But (3.7) holds only when |x-y|>100Ls. For x outside B_KL and y inside B_R, that condition forces s ≲ K/100. Once s exceeds that, the low-frequency propagator has spread past B_KL and the kernel is of order s^{-d/2}, not (KL)^{-N}. The integral then gives an O(t) error, which is far too large for the target √t bound. So Proposition 3.2 and Proposition 3.7 do not justify the R^d part of Theorem 1.1 or the R^d corollaries. This is a genuine proof gap, not a cosmetic issue.\n\nThe Z^d argument does not share the defect: on Z^d there is true finite speed, and the approximation Lemma 2.5 is valid. Lemma 4.1 and the corollary derivations are also fine as deterministic steps. The only other issue is minor: the NCK tail bound requires K > 4√log d, and with the finite-rank dimension used here that becomes K ≳ C_d √log R, not just K > √log R. That is adjustable.\n\nI think the method is promising and the Z^d results alone look publishable. But the R^d half needs a repair — likely a time-dependent cutoff or a sharper low-frequency kernel estimate — and the current version should not be cited for the R^d theorem. A serious referee should see it, mainly to check the cutoff and to judge whether the repair is cheap. I would not desk-reject it.","headline":"Z^d argument is solid and the deterministic T1-to-Tk reduction is a genuinely nice trick, but the R^d half of Theorem 1.1 has a real gap in the spatial-cutoff step, so the main theorem is unproved as stated.","tokens_in":23197,"tokens_out":9563,"would_cite":false,"duration_ms":93132,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81Q10","60B20","35Q41","82B44"],"pacs":[],"model":"deepseek-v4-flash","headline":"A random-matrix inequality gives square-root cancellation for every term of the Dyson series of a random Schrödinger equation, yielding localization and delocalization bounds in all dimensions d≥2.","keywords":["random Schrödinger operator","Dyson series","noncommutative Khintchine inequality","tail bounds","frequency localization","spatial delocalization","Floquet states","dispersive estimates"],"falsifier":"Evaluate the low-frequency error term $\\|\\chi_R U_0(s) P_{\\le L}(1-\\chi_{100L})\\|_{\\mathrm{op}}$ for s with $L^{-1}\\ll s\\ll R^{5d}$ on $\\mathbb{R}^d$, using the paper's choice L=$R^{{100d}}$; if this norm is not bounded by (KL)^{-10d} but instead grows with s, then Proposition 3.2 fails for intermediate times and Theorem 1.1 on R^d does not follow from the given argument.","tokens_in":21856,"feed_emoji":"🎲","tokens_out":9751,"duration_ms":91267,"temperature":0.7,"pith_summary":"This paper claims that for random Schrödinger operators H=H0+λV on $\\mathbb{Z}^d$ or $\\mathbb{R}^d$ with d≥2 and a compactly supported potential of radius R>$λ^{{-1}}$, the terms T_j(t) of the Dyson series obey Gaussian tail bounds with square-root cancellation: with probability at least 1−$2e^{{-cK^2}}$, $\\|T_j(t)\\|_{\\mathrm{op}} \\leq (K^2\\sigma_d(t)/j)^{j/2}$ for all t>0 and j≥1. Summing these bounds shows that the true evolution is within Kλ√σ_d(t) of the free evolution for times up to order $λ^{{-2+ε}}$, with high probability. From that comparison the paper derives frequency localization and spatial delocalization of approximate eigenfunctions, as well as the same localization statements for Floquet states of time-periodic potentials. The proof is elementary: a deterministic reduction of T_j to powers of T_1, then a random-matrix bound applied to T_1, whose variance is computed from dispersive decay of the free propagator.","feed_headline":"Random potentials tame the Dyson series to square-root size","feed_subtitle":"At the weak-coupling time scale, the evolution stays near free motion while eigenfunctions localize in frequency and delocalize in space.","key_machinery":"The load-bearing object is the first Dyson term $T_1(t)=\\int_0^t e^{isH_0}Ve^{-isH_0}\\,ds$, which is a structured random matrix of the form $\\sum_n g_n A_n(t)$ with independent random coefficients g_n. A concentration theorem for such matrices bounds $\\|\\sum_n g_n A_n\\|_{\\mathrm{op}}$ by roughly $\\|(\\sum_n A_n^2)^{1/2}\\|_{\\mathrm{op}}$ up to a logarithmic factor and a Gaussian tail, and the paper computes this variance from the diagonal dispersive decay $|\\langle n|e^{-isH_0}|n\\rangle| \\lesssim |s|^{-d/2}$. To pass from T_1 to all T_j, a deterministic lemma shows $\\|T_k(t)\\|_{\\mathrm{op}} \\leq (Ck^{-1/2}(\\|V\\|_{L^\\infty}+M_t)\\sqrt{t\\log(2+t)})^k$, where $M_t=\\sup_{s\\le t}\\|T_1(s)\\|_{\\mathrm{op}}/\\sqrt{s}$, using an approximate product structure of the simplex of integration times. On $\\mathbb{R}^d$, a phase-space cutoff $\\Pi=P_{\\le L}\\chi_{100L}$ is introduced to make T_1 finite-rank so the random matrix theorem applies, with high and low frequencies controlled separately.","core_discovery":"The central claim is that the Dyson series of the random Schrödinger evolution has square-root cancellation, uniformly in t and in the order j. Specifically, with H=H0+λV on $\\mathbb{Z}^d$ or $\\mathbb{R}^d$, d≥2, there is c>0 such that for all K>√(log R), P(∥T_j(t)∥op ≥ ($K^{2}$σ_d(t)/j)^{j/2} for some t>0, j∈ℕ) ≤ 2 exp(−$cK^{2}$), where σ_d(t)=|t| log(2+|t|) for d>2 and |t| log²(2+|t|) for d=2. Summing over j gives $\\|e^{-itH}-e^{-itH_0}\\|_{\\mathrm{op}} \\leq CK\\lambda\\sqrt{\\sigma_d(t)}$ with probability at least 1−$2e^{{−cK^2}}$, so the perturbed evolution stays close to the free evolution at times of order $λ^{{−2+ε}}$. From that comparison, the paper derives frequency localization and spatial delocalization of approximate eigenfunctions, and the analogous statements for Floquet states of time-periodic potentials. The authors emphasize that the proof uses neither sophisticated harmonic analysis nor diagrammatic expansions, only the noncommutative Khintchine inequality plus pointwise dispersive estimates for the free Schrödinger propagator.","pith_inferences":["If the tail bound could be upgraded to a second-moment control of $\\|e^{-itH}-\\sum_{j\\le M}T_j\\|_{\\mathrm{op}}^2$, the high-probability truncation would turn into the averaged observable estimates needed for quantum diffusion; the paper notes this gap and the present method is a natural starting point.","On $\\mathbb{R}^d$, a time-dependent cutoff $\\Pi(t)=P_{\\le L}\\chi_{100L+Lt}$ that follows the low-frequency wavefront would likely extend the argument beyond the range where the fixed spatial cutoff is reliable.","The deterministic reduction from T_j to T_1 is independent of the randomness and should apply to any background Hamiltonian whose free evolution has enough dispersive decay, provided the finite-rank approximation step can be replaced."],"forward_implications":["For times t up to order λ^{-2+ε}, the full evolution $e^{-itH}$ differs from the free evolution $e^{-itH_0}$ by at most CKλ√σ_d(t) in operator norm, with probability at least 1−2e^{−cK^2}.","The Dyson series can be truncated at roughly λ²σ_d(t) terms with exponentially small error, so the interaction picture of the evolution is effectively finite-dimensional at that scale.","Approximate eigenfunctions satisfying $\\|(H-E)\\psi\\|_{L^2}\\le\\lambda^2$ have Fourier mass concentrated near the level sets of H0 up to an error $CK\\lambda|\\log\\lambda|^2\\delta^{-1/2}$, and are spatially delocalized on scales of order $K\\lambda|\\log\\lambda|^3\\sqrt{\\ell}/E^{1/4}$.","The same localization statements hold for Floquet states of τ-periodic potentials: a quasienergy state stays Fourier-localized under the full evolution for any number of periods.","No curvature restriction on the dispersion relation is needed, so the flat level set at E=0 and all dimensions d≥2 are covered."],"supporting_citations":[{"why":"Supplies the structured random matrix concentration theorem (Theorem A.1) that gives the tail bound for T1.","marker":"[Van17]"},{"why":"Provides the original noncommutative Khintchine inequality used to control moments of T1.","marker":"[Lus86]"},{"why":"Extends the Khintchine inequality to the noncommutative setting, underpinning the matrix concentration argument.","marker":"[LP91]"},{"why":"The frequency-concentration results that Corollary 1.3 extends to all d≥2 and all energies.","marker":"[SSW02]"},{"why":"The diagrammatic localization-length analysis whose time-periodic analogue is proved here.","marker":"[Che05]"}],"fun_headline_variants":["Dyson series of random Schrödinger has square-root tails","Random potentials: square-root cancellation in Dyson series","Weak-coupling Dyson series tamed to square-root size","Dyson series tail bounds: square-root cancellation","Khintchine plus dispersive estimates tame Dyson series"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The R^d argument requires that the low-frequency part of the free evolution stays inside a fixed spatial ball of radius 100L for all times up to $R^{{5d}}$; the claim fails if the wavefront travels a distance comparable to the cutoff before that time.","fun_headline_variants_meta":{"raw":{"variants":["Dyson series of random Schrödinger has square-root tails","Random potentials: square-root cancellation in Dyson series","Weak-coupling Dyson series tamed to square-root size","Dyson series tail bounds: square-root cancellation","Khintchine plus dispersive estimates tame Dyson series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000765,"raw_usage":{"total_tokens":3417,"prompt_tokens":994,"completion_tokens":2423,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":2344}},"tokens_in":610,"tokens_out":2423,"duration_ms":18839,"temperature":1.0,"reasoning_tokens":2344,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T11:46:58.563446+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the low-frequency error term $\\|\\chi_R U_0(s) P_{\\le L}(1-\\chi_{100L})\\|_{\\mathrm{op}}$ for s with $L^{-1}\\ll s\\ll R^{5d}$ on $\\mathbb{R}^d$, using the paper's choice L=$R^{{100d}}$; if this norm is not bounded by (KL)^{-10d} but instead grows with s, then Proposition 3.2 fails for intermediate times and Theorem 1.1 on R^d does not follow from the given argument.","supporting_citations":[],"review_version":1}