{"id":"df414751-0e00-4465-9f7a-4072a3fca779","arxiv_id":"1908.09217","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For elastic manifolds in high-dimensional random landscapes, the annealed complexity of stationary points and of minima is derived explicitly; both vanish at the Larkin mass, with quadratic and cubic scaling, and a depinning bound follows.","lead":"This paper derives explicit formulas for the number of stationary points and local minima in the high-dimensional random energy landscape of an elastic manifold. The formulas vanish at a critical curvature, the Larkin mass, and yield an upper bound on the force needed to unpin the manifold.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unproven spectral-rigidity conjecture (13)/(39) is load-bearing in the infinite-block limit; the supplementary q→0 resolvent check does not test the fluctuations that the conjecture requires to vanish.","rationale":"After reading the paper in good faith, the central claim is exactly what the authors state: explicit annealed complexities and a depinning bound, conditional on Eq. (13). All subsequent manipulations (saddle point in ξ, elimination of y, expansions near μ_c) are internally consistent; I checked the algebra of the supplementary expansions, including the relations leading to (34) and the d=2 parametric solution, and found no internal inconsistency. The depinning bound from Markov's inequality is logically sound given the annealed complexity. The only place where the argument can break is the determinant-averaging step, because it is the only step that is not proven and that carries the full weight of the large-deviation estimate. The provided q→0 resolvent check is real evidence but only constrains the first moment of log det; the failure mode that matters is the second and higher cumulants in the L^d → ∞ limit. A numerical test of Var and R is well-defined and feasible for modest N and L, and would settle whether the conjecture holds or whether the formulas need correction. I therefore agree with the reader's CONDITIONAL verdict; no verdict change is needed.","tokens_in":15018,"tokens_out":16210,"duration_ms":166596,"concrete_test":"Monte Carlo sampling of the block banded model: take d=1, t=1, μ=0.5 μ_c, choose N ∈ {8,16,32,64} and L ∈ {1,2,4,8,16}; for each (N,L) generate independent GOE blocks H(x) and independent Gaussians ξ(x), compute log|det(K+X+μI)|, and estimate m = ⟨log|det|⟩, v = Var(log|det|), and R = log⟨|det|⟩ - m. The conjecture (13) requires v = o(NL) and R = o(NL). Plot R/(NL) and v/(NL) versus N and L; if the curves extrapolate to a positive constant, the annealed complexity differs from Eq. (32) by that constant and the vanishing/scaling statements need revision. Repeating the same check with the θ(K+X+μ) restriction tests Eq. (39) and the uniform-boundary assumption used for Σ_st.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The derivation of Eqs. (32) and (43) hinges on replacing the disorder average of |det| by the exponential of the averaged log, conjectured in Eq. (13) and extended to minima in Eq. (39). For L^d=1 this is a theorem; for block banded matrices with finite block number the authors cite spectral rigidity, but the paper needs the same rigidity after the number of blocks L^d is sent to infinity. The consistency check in the Supplementary Material (Section 'LIMIT q→0 AND RESOLVENT') only verifies the q→0 derivative of log|det|, i.e. the first moment of the log-determinant; it says nothing about the variance or higher cumulants. If Var(log|det(K+X+μI)|) grows as c N L^d, or R = log⟨|det|⟩ - ⟨log|det|⟩ saturates to c N L^d, then the true annealed complexity acquires an additive correction c/2, changing the vanishing exponents and the depinning bound f_st^c in Eq. (49). A related gap is the assumption in the minima calculation that the constrained saddle on the boundary of D is the uniform configuration ξ_e; this follows only if the low-energy eigenmodes are delocalized, which is part of the same unproven rigidity. The authors are explicit that (13) is a conjecture, so this is a known limitation, but it is load-bearing: without it the explicit formulas and the bound collapse.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies the energy landscape of an elastic manifold with internal dimension d<4 embedded in a high-dimensional random medium, with energy functional (1) and Gaussian disorder covariance (2). Using the Kac-Rice formula, the authors define annealed complexities for the total number of stationary points and for local minima (Eq. (3)). They decompose the Hessian into a block random matrix and, under the spectral-rigidity conjecture stated in Eq. (13) (extended to minima in Eq. (39)), derive explicit formulas for the two complexities: Eqs. (32) and (43). These formulas vanish at the Larkin mass μ_c, with quadratic behavior for stationary points and cubic behavior for minima near μ_c. The massless limit μ→0 is used to obtain an upper bound f_st^c for the depinning threshold, Eq. (49). The supplementary material contains the near-transition expansions, explicit results for the continuum model in d=1,2,3, and a consistency check for the mean resolvent.","tokens_in":15364,"tokens_out":3920,"duration_ms":42760,"significance":"If the central claims hold, the paper provides one of the first explicit high-dimensional landscape-complexity calculations for disordered elastic manifolds with internal dimension d<4, extending the d=0 toy model and connecting landscape topology to depinning. The derivation is transparent and systematic: the saddle-point equations, the self-consistent resolvent equation, and the explicit continuum-model formulas are carefully presented. The authors also give machine-checkable algebraic details in the supplementary material, including exact results in d=2, a numerical constant in d=1, and a universal ratio Σ_st(0)/Σ(0). The conditional nature of the central step is openly acknowledged, which is a strength. However, the main quantitative formulas and the depinning bound are not rigorously established because they depend on an unproved spectral-rigidity conjecture in the infinite-block limit; the consistency check provided tests only the first moment of the log-determinant, not the fluctuations that the conjecture requires to vanish.","major_comments":[{"comment":"The replacement ⟨|det(K+X+μI)|⟩ ≈ exp⟨Tr log|K+X+μI|⟩ in Eq. (13) is load-bearing for the central formulas (32) and (43). It is a theorem for L^d=1, but here it must hold after the number of blocks L^d is sent to infinity, not only for finite block number. The supplementary consistency check computes only the q→0 derivative of log|det|, i.e., the first moment of the log-determinant or the mean resolvent. It does not test the variance or higher cumulants. If Var(log|det(K+X+μI)|) grows as c N L^d, or if log⟨|det|⟩−⟨log|det|⟩ is of order N L^d, then the exponent in the annealed complexity acquires an additive correction c/2, which would change the vanishing exponents near μ_c and the value of f_st^c in Eq. (49). The authors should either provide a proof of the required rigidity in the infinite-block limit, or present numerical evidence directly for the fluctuations of the log-determinant, or explicitly restrict the claims in the abstract and conclusions to the conjectural status of this step.","section":"Eq. (13) and Supplementary Material, section 'LIMIT q→0 AND RESOLVENT'"},{"comment":"The complexity of local minima additionally relies on the claim that the integral over the domain D is dominated by the uniform boundary configuration ξ_e = −λ_e^− − μ. This is not automatic: it requires that the constrained saddle point of the action on the boundary of D is the constant configuration, which in turn relies on delocalization of the low-energy eigenmodes of the block banded matrix K. That delocalization is part of the same spectral-rigidity assumption used in Eq. (13). If the boundary saddle were nonuniform, Eq. (43) and the cubic law (45) would not follow. This step should be derived explicitly or stated as a separate conjecture with a numerical test, since the cubic vanishing at μ_c is one of the headline results.","section":"Eq. (39) and paragraph after Eq. (42)"},{"comment":"The depinning bound f_c ≤ f_st^c = sqrt(4B'(0) Σ_st(μ=0)) is derived from the annealed complexity of minima. As the authors note, the quenched complexity may differ from the annealed one, so the bound is conditional not only on the rigidity conjecture but also on the relation between annealed and quenched counts. This is not an error, but because the bound is one of the two headline results, the conditional status should be stated in the abstract and in the concluding discussion, not only in the body near Eq. (5).","section":"Eq. (5) and Eq. (49)"}],"minor_comments":[{"comment":"There is a parenthesis typo in the displayed formula: the integrand should read (I_1(μ̃) − I_1(μ_c)) dμ̃, not (I_1(μ̃) − I_1(μ_c)dμ̃ as printed.","section":"Eq. (43)"},{"comment":"The notation I_p(μ) is defined after Eq. (34) but used earlier in Eq. (43) before that definition; please define I_1(μ) before its first use.","section":"After Eq. (34)"},{"comment":"Reference [33] is listed as 'in preparation'; if a published or preprint version is available by the time of revision, it should be updated.","section":"References"},{"comment":"The numerical constant Σ(μ=0)|_{d=1} ≈ 0.375 t^{-2/3} and the related constants C_{∞,d} in Eq. (91) would benefit from a brief statement of the numerical integration method or an independent check, since these are quoted as decimal values.","section":"Supplementary Material, Eq. (87)"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly written and the authors are transparent that the central step is a conjecture. The main question for the editor is whether the journal is willing to publish results whose primary formulas rest on an unproved spectral-rigidity assertion in the infinite-block limit. If the authors can supply a proof or, failing that, strong numerical evidence for the fluctuations of the log-determinant, I would support publication. At present, the load-bearing nature of the conjecture makes the central claims conditional rather than established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nWhat you should know: this paper actually delivers what it promises. It generalizes the known annealed complexity calculations for the d=0 toy model and the d=1, N=1 directed polymer to elastic manifolds with internal dimension 1≤d<4. The explicit formulas for the total complexity (32) and the minima complexity (43), the quadratic/cubic vanishing near the Larkin mass, and the depinning bound (5)/(49) are new and cleanly derived. I checked the near-transition expansions in the supplementary material and they are correct; the d=2 parametric result is a nice touch.\n\nThe derivations are laid out with care. The self-consistent equation for the resolvent, the saddle-point calculation for ξ(x), and the handling of the minima via the domain D are all presented clearly. The authors are also honest: they label the key replacement (13), and its extension (39), as conjectures rather than theorems. That is refreshing.\n\nThe soft spot is real and load-bearing. The replacement of the disorder average of |det| by the exponential of the averaged log requires spectral rigidity in the limit where the number of blocks L^d goes to infinity. The supplementary q→0 resolvent check only verifies the first moment of log|det|—it says nothing about the variance or higher cumulants. If Var(log|det|) grows as c N L^d, the annealed complexity acquires an additive correction that would change the critical exponents and shift the depinning bound. The stress-test note is right about this. It is not a hidden flaw, but it is exactly the kind of gap that a careful referee should dig into.\n\nThe minima calculation also assumes that the constrained saddle point sits on the boundary of D at the uniform configuration ξ_e. That depends on delocalization of low-energy eigenmodes, which is tied to the same unproven rigidity. So the paper is complete conditional on a plausible but unproved assumption. I don't think the conjecture is obviously false—the finite-block result of Shcherbina is suggestive—but it is not established.\n\nWho this is for: people working on random landscape complexity, disordered elastic systems, and replica mean-field transitions. The paper deserves serious refereeing, not a desk reject. A referee should push on (13) and (39), and the authors should either prove them in a tractable case or provide numerical evidence for the variance. If the conjecture holds, this is a strong result; if it fails, the formulas collapse, so the ball is in the authors' court.\n\nI would accept it for peer review and cite it if I worked in the area.","headline":"A clean, genuinely new extension of the d=0 and d=1,N=1 complexity results to 1≤d<4, but the central determinant-averaging conjecture is explicitly unproved and the consistency check does not test the variance it would need.","tokens_in":15865,"tokens_out":1862,"would_cite":true,"duration_ms":21316,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"For a $d<4$ elastic manifold in a high-dimensional random medium, the annealed complexities of stationary points and of minima are explicit functions of the curvature $\\mu$, and both vanish at the Larkin mass $\\mu_c$.","keywords":["annealed complexity","random energy landscape","elastic manifold","Kac-Rice formula","Larkin mass","depinning threshold","random matrix spectral determinant","replica symmetry breaking"],"falsifier":"Numerically evaluate, for the block-banded Hessian with block size $N$ and $L$ blocks, the ratio $\\langle|\\det(K+X+\\mu I)|\\rangle / \\exp(\\langle\\operatorname{Tr}\\log|K+X+\\mu I|\\rangle)$ for increasing $N$ at fixed $L$ and then for increasing $L$. If the ratio does not approach $1$, or if it deviates once $L$ grows before $N\\to\\infty$, the conjecture (13) fails and the formulas (32) and (43) are not justified.","tokens_in":14803,"feed_emoji":"📉","tokens_out":10501,"duration_ms":99187,"temperature":0.7,"pith_summary":"This paper counts the stationary points and local minima of the energy landscape of an elastic manifold dragged through a high-dimensional random medium. It claims that for internal dimension $d<4$ and large embedding dimension $N$, the annealed complexities—the log of the mean number of equilibria per degree of freedom—are given by explicit formulas: Eq. (32) for all stationary points and Eq. (43) for minima. Both complexities vanish exactly at the curvature $\\mu_c$ known as the Larkin mass, so the landscape switches from simple (no exponentially many equilibria) to complex (exponentially many) as $\\mu$ drops below $\\mu_c$. The vanishing is quadratic in $\\mu_c-\\mu$ for stationary points and cubic for minima. The same formulas have a finite massless limit, which yields a bound on the force needed to depin the manifold.","feed_headline":"Random-manifold landscape counts collapse at a critical curvature","feed_subtitle":"Closed formulas count random-manifold equilibria; both counts vanish at the critical curvature, bounding depinning.","key_machinery":"The load-bearing machinery is the Kac-Rice formula, which writes the mean number of stationary points as an integral over configurations of $|\\det K_0|$, where $K_0$ is the Hessian. The Hessian has a block structure: off-diagonal blocks come from the discrete Laplacian, while diagonal blocks are independent GOE matrices plus Gaussian shifts $\\xi(x)$, so its mean spectral density obeys the self-consistent resolvent equation (19). The decisive step is the conjecture (13), that the GOE average of the absolute determinant can be replaced by the exponential of the average log-determinant; spectral rigidity makes the log-determinant self-averaging. That replacement turns a difficult determinant average into a saddle-point integral over the shifts, whose solution reduces the complexity to integrals involving a single parameter $y(\\mu)$ defined by Eq. (33).","core_discovery":"On the paper's own terms, the central discovery is that the annealed complexity of the high-dimensional random-manifold landscape is controlled by the mean spectral density of a block random-matrix Hessian. For curvature $\\mu>\\mu_c$ the complexity of stationary points and of minima is identically zero. For $\\mu<\\mu_c$, with $\\mu_c$ set by $1=\\int_k (\\mu_c-t\\Delta(k))^{-2}$, the total complexity is $\\Sigma(\\mu)=\\int_{\\mu_c}^{\\mu} d\\tilde\\mu \\int_k \\frac{y(\\tilde\\mu)^2}{(\\tilde\\mu-t\\Delta(k))((\\tilde\\mu-t\\Delta(k))^2+y(\\tilde\\mu)^2)}$ and the minima complexity is $\\Sigma_{\\rm st}(\\mu)=-\\frac12(\\mu_c-\\mu)^2+\\int_\\mu^{\\mu_c}(I_1(\\tilde\\mu)-I_1(\\mu_c))\\,d\\tilde\\mu$, with $y(\\mu)$ determined by $1=\\int_k ((\\mu-t\\Delta(k))^2+y(\\mu)^2)^{-1}$. Near $\\mu_c$, $\\Sigma\\propto(\\mu_c-\\mu)^2$ while $\\Sigma_{\\rm st}\\propto(\\mu_c-\\mu)^3$, and the massless limits give the depinning bound $f_c\\le f_{\\rm st}^c=\\sqrt{4B'(0)\\Sigma_{\\rm st}(0)}$.","pith_inferences":["If the spectral rigidity conjecture survives the infinite-block limit, the same saddle-point calculation should extend to the quenched complexity, since the log-determinant fluctuations that distinguish annealed from quenched averages are subextensive; the paper leaves that equality open.","The universal dimension-dependent ratio $\\Sigma_{\\rm st}(0)/\\Sigma(0)$ (about $0.63$ in $d=1$ and $0.405$ in $d=2$ for the continuum model) predicts a quantitative relation between metastable-state count and the depinning bound that numerical interface simulations could test.","A direct testable extension is to count stationary points in finite-size discrete manifolds and check the predicted quadratic and cubic vanishing laws, with finite-$N$ corrections expected from random-matrix edge fluctuations as in the $d=0$ case.","The bound $f_c\\le f_{\\rm st}^c$ may become an equality in models where annealed and quenched complexities coincide, while the random-field-type models cited in the paper suggest settings where the inequality should be strict."],"forward_implications":["For $\\mu>\\mu_c$ the complexity is zero, so a typical landscape has no exponentially large family of stationary points or minima.","For $\\mu<\\mu_c$ the number of equilibria grows exponentially with $N L^d$ at rate $\\Sigma(\\mu)$, and the rate for minima vanishes cubically at the transition while the total rate vanishes quadratically.","The massless limit yields $N_{\\rm tot}\\sim e^{C N (L/L_c)^d}$, so the Larkin length $L_c$ controls the exponential count of equilibria in the critical limit.","The depinning threshold under a uniform force obeys $f_c\\le \\sqrt{4B'(0)\\Sigma_{\\rm st}(0)}$, a bound sharper than the bound obtained from the total stationary-point complexity.","In $d=0$ the formulas reproduce the earlier toy-model complexity, and in the continuum the complexity of minima has an elementary closed form for $d<4$, including a finite limit at $d=2$."],"supporting_citations":[{"why":"Establishes the d=0 toy-model complexity and the absolute-determinant method that the present calculation generalizes.","marker":"[18]"},{"why":"Shows in d=0 that the onset of exponential complexity coincides with replica symmetry breaking, the property extended here to d<4.","marker":"[21]"},{"why":"Provides the d=0 critical behavior of the number of minima, the result whose cubic scaling is generalized.","marker":"[22]"},{"why":"Develops the Kac-Rice counting and force/width method for an elastic string, including the depinning bound that is lifted to general N and d.","marker":"[26]"},{"why":"Supplies the self-consistent resolvent formalism for the Hessian spectrum used to obtain the mean density rho_K.","marker":"[27]"},{"why":"Introduces the block Hessian model and the Larkin mass criterion for the manifold, giving the spectral input for the saddle point.","marker":"[28]"},{"why":"Gives the self-consistent equation (19) for the resolvent of random operators with off-diagonal randomness.","marker":"[30]"},{"why":"Proves spectral rigidity for block band matrices with finitely many blocks, the main support for the conjecture (13).","marker":"[31]"}],"fun_headline_variants":["Landscape complexity collapses at Larkin curvature","Random-manifold complexity vanishes at critical mass","Counts of stationary points die at critical curvature","Depinning threshold bounded by vanishing complexity","Zero complexity beyond Larkin mass in random landscapes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every explicit formula rests on the conjecture that the logarithm of the absolute determinant of the random block Hessian is self-averaging at large $N$, so the average of the determinant is the exponential of the average log-determinant, even when the number of blocks tends to infinity.","fun_headline_variants_meta":{"raw":{"variants":["Landscape complexity collapses at Larkin curvature","Random-manifold complexity vanishes at critical mass","Counts of stationary points die at critical curvature","Depinning threshold bounded by vanishing complexity","Zero complexity beyond Larkin mass in random landscapes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000384,"raw_usage":{"total_tokens":2058,"prompt_tokens":999,"completion_tokens":1059,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":991}},"tokens_in":615,"tokens_out":1059,"duration_ms":9022,"temperature":1.0,"reasoning_tokens":991,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:17:34.836570+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Numerically evaluate, for the block-banded Hessian with block size $N$ and $L$ blocks, the ratio $\\langle|\\det(K+X+\\mu I)|\\rangle / \\exp(\\langle\\operatorname{Tr}\\log|K+X+\\mu I|\\rangle)$ for increasing $N$ at fixed $L$ and then for increasing $L$. If the ratio does not approach $1$, or if it deviates once $L$ grows before $N\\to\\infty$, the conjecture (13) fails and the formulas (32) and (43) are not justified.","supporting_citations":[{"cited_title":"Fyodorov","cited_arxiv_id":null,"evidence_quote":"Establishes the d=0 toy-model complexity and the absolute-determinant method that the present calculation generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows in d=0 that the onset of exponential complexity coincides with replica symmetry breaking, the property extended here to d<4."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the d=0 critical behavior of the number of minima, the result whose cubic scaling is generalized."},{"cited_title":"Fyodorov, P","cited_arxiv_id":null,"evidence_quote":"Develops the Kac-Rice counting and force/width method for an elastic string, including the depinning bound that is lifted to general N and d."},{"cited_title":"V Fyodorov and P","cited_arxiv_id":null,"evidence_quote":"Supplies the self-consistent resolvent formalism for the Hessian spectrum used to obtain the mean density rho_K."},{"cited_title":"Manifolds pinned by a high-dimensional random landscape: Hessian at the global energy minimum","cited_arxiv_id":"1903.07159","evidence_quote":"Introduces the block Hessian model and the Larkin mass criterion for the manifold, giving the spectral input for the saddle point."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the self-consistent equation (19) for the resolvent of random operators with off-diagonal randomness."},{"cited_title":"Shcherbina","cited_arxiv_id":null,"evidence_quote":"Proves spectral rigidity for block band matrices with finitely many blocks, the main support for the conjecture (13)."}],"review_version":1}