{"id":"57001299-c710-4293-a616-4d6833e160e5","arxiv_id":"2411.14628","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A heat-equation-based loss for neural signed distance functions gives an asymptotically sufficient condition for convergence to the true distance, with better surface and distance accuracy on shape benchmarks.","lead":"HotSpot is a new training loss for neural signed distance functions, built from the screened Poisson equation, that converges to a true distance field as a parameter grows. It improves surface and distance accuracy over eikonal-based methods such as DiGS and StEik on 2D and 3D benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Eq. (9) bounds distance to the artificial point-cloud boundary, not to the interpolated surface, so the central asymptotic sufficiency claim is not established.","rationale":"The reader's weakest_assumption correctly identifies the boundary-interpolation premise. My stress test sharpens it: the proof's boundary is the discrete point set, not the interpolated surface, so the theorem as written is about the wrong distance target. This is not an external disagreement with consensus but an internal gap between the stated theorem and its proof: Eq. (38) uses d_Γ while the boundary in Eq. (32) is a union of balls around input points. The empirical evaluation is honest and strong, and the method may still work well on dense point clouds where the two distances nearly coincide, but the asymptotic guarantee as stated is not established. Because the reader already conditioned acceptance on exactly this premise, my read does not change the verdict: the paper should remain CONDITIONAL pending a proof or a clarifying statement that the guarantee is for distance to the point cloud, or an analysis of how the discrete-to-continuous boundary interpolation error vanishes.","tokens_in":25799,"tokens_out":4561,"duration_ms":46751,"concrete_test":"Numerical check on a 2D circle: take N=8 points on a circle of radius 1. Solve the screened Poisson equation (∇²-λ²)h=0 with h=1 on the true circle and, separately, with h=1 on the 8 points (or on small balls of radius ε < half-spacing). Measure E_λ = d_circle(x) + (1/λ)ln h_λ(x) at the arc midpoint between two adjacent points for λ=10,100,1000,10000. If the discrete-boundary solution's E_λ saturates near 1-cos(π/8)≈0.038 instead of decaying as ~ln(ε/d)/λ, then Eq. (38) cannot hold with d_Γ taken as the surface distance. Equivalently, train HotSpot with a very high boundary weight on this 8-point circle and inspect whether the midpoint error decreases with λ; a plateau would falsify the asymptotic sufficiency claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In supplement A.3, the boundary condition is imposed on balls B(x_i, ε) around each input point (Eqs. 30-32), and the solution is a linear combination of single-point heat kernels. The proven bound (Eq. 38) therefore controls d_{∂(∪B(x_i,ε))}(x) - |u_λ(x)|, i.e., distance to the point cloud with artificial balls. The paper's key claim, however, is that |u_λ| converges to d_Γ where Γ is the continuous surface interpolating the point cloud. These two distances are not the same: at the midpoint between two adjacent points, the distance to the point set is approximately half the spacing, while the distance to the interpolated surface is zero; this discrepancy does not vanish as λ→∞. The transition from the discrete boundary to the continuous surface rests on the assertion that 'if there are some boundaries connecting them, the boundary condition still holds' (Sec. 3) and the isolation assumption h_j(S_i)=h_j(x_i) (A.3). Neither is proven, and the isolation assumption fails when point spacing is comparable to ε or when balls overlap. Without an argument that the learned minimizer (or the true heat solution) enforces u=0 on a continuous interpolating surface, Eq. (9) does not establish the asymptotic sufficient condition claimed in the abstract. Fig. 6's collapse to unsigned distance is a visible symptom of this gap. The temporal-stability analysis being performed in h-space rather than u-space is secondary; the boundary mismatch is the load-bearing issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HotSpot, a loss for optimizing neural signed distance functions based on the screened Poisson equation. The authors derive a heat loss L_heat = (1/2)∫ e^{-2λ|u|}(||∇u||² + 1) dx, whose minimizer in the heat-field variable satisfies the screened Poisson equation, and they prove a bound (Eq. 9 / Eq. 38 in the supplement) showing that the recovered distance |u_λ| converges linearly to the true distance as λ→∞. The paper also claims spatial and temporal stability, natural surface-area penalization, and a connection to the PHASE method. Experiments on 2D and 3D datasets, including ShapeNet and high-genus shapes, report improved reconstruction and distance metrics over baselines such as DiGS and StEik.","tokens_in":26075,"tokens_out":10485,"duration_ms":92574,"significance":"If the theoretical claim holds, this is a valuable contribution to neural SDF optimization: it offers a principled alternative to the eikonal loss with a quantitative convergence guarantee, grounded in Varadhan's classical heat-kernel asymptotics. The paper is well-written, the derivations are mostly transparent, and the empirical gains are substantial and consistently presented across multiple datasets and metrics. The authors also honestly discuss the fragility of the boundary condition and disclose the relationship to PHASE. However, the central theoretical claim rests on a proof that currently has a concrete mathematical flaw and a gap between the discrete boundary used in the proof and the continuous surface claimed in the main text.","major_comments":[{"comment":"The bound in Eq. (9)/(38) is proven for a discrete boundary consisting of balls B(x_i, ε) around the input points, not for the continuous interpolated surface Γ. The proof explicitly assumes the isolation condition h_j(S_i)=h_j(x_i) for i≠j and ε<r_0, which only holds when the balls are well separated and inter-ball influence is negligible. The main text asserts that 'if there are some boundaries connecting them, the boundary condition of Eq. (4) still holds' (Section 3), but this is not proved. Consequently, the central claim that |u_λ| converges to d_Γ (the distance to the interpolated surface) is not established; at best the proof controls the distance to the point cloud with artificial balls, which differs from d_Γ by a non-vanishing amount in the limit λ→∞. This gap needs to be closed, or the claim should be reformulated to state what is actually proven.","section":"Section 3 and Supplement A.3, Eqs. (30)-(38)"},{"comment":"The lower-bound proof contains an incorrect inequality. The function h0(r)=ε/r e^{-λr} is positive and decreasing for r≥ε. Since Σc_i r_i/Σc_i ≥ d_Γ(x), the second inequality in Eq. (37) should read h0(Σc_i r_i/Σc_i) ≤ h0(d_Γ(x)), not ≥. As written, the chain does not imply h(x) ≥ ε e^{-λd_Γ}/d_Γ. The lower bound in Eq. (38) is therefore not proven by the given argument. The upper bound appears correct, but the lower bound is essential for the claimed two-sided convergence rate.","section":"Supplement A.3, Eq. (37)"},{"comment":"The temporal stability analysis is performed for the gradient flow in the heat-field variable h, ∂h/∂t = ∇²h - λ²h. However, the actual optimization is on the network output u with the combined loss of Eq. (8). The dynamics of u under gradient descent are not governed by this h-heat equation, and the assertion that |u| converges because the h-flow is stable and 'u is continuous' is not justified. The temporal stability claim for the proposed method is therefore unsupported as stated.","section":"Section 4.2.2, Eqs. (14)-(15)"}],"minor_comments":[{"comment":"The claimed Euler-Lagrange equation for the coarea loss is incorrect. A direct computation gives ∇·(e^{-λ|u|}∇u/||∇u||) + λ sgn(u)e^{-λ|u|}||∇u|| = 0, which for ||∇u||=1 reduces to a mean-curvature condition, not 1-∇²u=0. This does not invalidate the qualitative conclusion, but the stated derivation should be corrected or removed.","section":"Section 4.3, Eq. (17)"},{"comment":"The phrase 'our loss, when minimized, provides an asymptotically sufficient condition' should be qualified: the theoretical claim concerns the exact minimizer of the heat-loss energy functional, whereas the method minimizes a weighted combination of boundary, eikonal, and heat losses with a neural network. A brief discussion of the approximation gap would avoid overclaiming.","section":"Section 4.1, around Eq. (8)"},{"comment":"The λ scheduler used in the ShapeNet experiments is described in the supplement but not in the main text; a short description in the main text would improve reproducibility.","section":"Section 5.3"},{"comment":"There are minor typos: 'presist' in the paragraph before Proposition 1 and 'backproping' in Supplement C.","section":"Section 4.2.1 and Supplement C"}],"recommendation":"major_revision","confidential_remarks":"The paper has a clear mathematical error in the supplement's lower-bound proof and a substantive gap between the discrete boundary in the proof and the continuous surface in the claim. These issues directly affect the central theoretical contribution. However, the empirical results are strong and the core idea is promising; the errors appear fixable either by correcting the proof (if a valid lower bound exists) or by relaxing the claim. The temporal-stability argument also needs reframing to the actual u-space optimization. I recommend major revision rather than rejection, given the paper's potential value and the clarity of its presentation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the practical recipe is good: the heat loss works, the ShapeNet and 2D numbers are strong, and the authors are honest about its equivalence to PHASE's energy. Second, the advertised sufficiency proof has a hole. The linear bound in Eq. (38) is proven for the distance to the union of ε-balls around the input points, not for the distance to the interpolated surface, and the step connecting them is asserted rather than proven.\n\nWhat is genuinely new: the linear-in-1/λ bound (mod the gap), the spatial error-decay analysis in Prop. 2, the stability discussion, and the practical point that a firm boundary weight is needed—PHASE's recommended small w_b demonstrably collapses to an unsigned field. The experiments are extensive, honest, and state-of-the-art on several benchmarks. Near-surface distance accuracy is meaningfully better than the baselines, and the sphere-tracing results are convincing.\n\nThe stress-test note is on target. In supplement A.3, the boundary condition is imposed on balls B(x_i, ε), and the proof uses the isolation assumption h_j(S_i) = h_j(x_i). At the midpoint between two adjacent points, the distance to the point set is half the spacing while the distance to the interpolated surface is zero, and this discrepancy does not vanish as λ→∞. So Eq. (9) as proven controls d_{∂(∪B(x_i,ε))}, not d_Γ. The Section 3 remark about \"boundaries connecting them\" is not a proof; it is the load-bearing step. This is not a minor blemish because the abstract's central claim depends on it.\n\nSecondary issues: the temporal stability analysis is done for the h-gradient flow while training happens in u-space; the Euler-Lagrange derivation for the coarea loss (1 − ∇²u = 0) does not follow from the written loss; no code is released; and the λ scheduler is described qualitatively. None of these are fatal on their own.\n\nWho this is for: anyone working on neural SDF optimization. The heat loss is a practical, well-tested alternative to eikonal-style regularization, and the PHASE connection is worth understanding. The theoretical claim needs repair before it can be accepted as proven. I would bring this to a reading group: the gap between discrete boundary and continuous surface is a useful example of where a plausible assertion hides a real missing argument.\n\nRecommendation: yes, send it to peer review. A serious referee can pin the A.3 gap and ask the authors to either prove the interpolation step or soften the claim. The empirical contribution deserves to be in circulation, and the theory can be fixed or properly scoped in revision.","headline":"Strong empirical paper with a real proof gap: Eq. (38) bounds distance to the artificial ε-balls, not to the interpolated surface, so the headline 'asymptotically sufficient condition' is not established as stated.","tokens_in":26742,"tokens_out":2520,"would_cite":true,"duration_ms":25393,"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":"A heat-diffusion loss provably drives neural signed distances to the true distance function.","keywords":["signed distance functions","neural implicit surfaces","screened Poisson equation","heat–distance relation","point cloud reconstruction","eikonal regularization","distance approximation","surface area regularization"],"falsifier":"Solve the screened Poisson equation numerically for a known analytic surface with $h = 1$ on a densely sampled boundary, recover $u_\\lambda = -\\frac{1}{\\lambda}\\ln h_\\lambda$, and check whether $d_\\Gamma(x) - |u_\\lambda(x)|$ stays within the paper's bound and falls linearly with $1/\\lambda$ as $\\lambda$ grows. Then thin the samples until the proof's assumption that boundary balls do not influence each other clearly fails, and test whether the minimizer of the heat loss still converges to a true signed distance; if the error leaves the predicted envelope, or remains distant at those densities, the asymptotic sufficiency claim is limited to the dense regime.","tokens_in":25497,"feed_emoji":"🌡️","tokens_out":15879,"duration_ms":129380,"temperature":0.7,"pith_summary":"This paper argues that a neural signed distance function can be trained with a loss that is asymptotically sufficient: as the absorption parameter $\\lambda$ of a screened Poisson equation grows, exactly minimizing the loss drives the network output to the true signed distance to the surface, with an error that shrinks linearly in $1/\\lambda$. The loss is built from a classical heat–distance relation, and the proof shows the minimizer satisfies a quantitative bound involving the distance itself, the boundary sampling scale, and the number of input points, so the guarantee holds even at finite $\\lambda$. The same loss makes optimization stable, because a local perturbation decays exponentially and the gradient flow is the heat equation, and it penalizes large surface area as a byproduct rather than through distorting regularizers. If the claim holds, point-cloud reconstruction gains both accuracy and reliability: the reported ShapeNet IoU is 0.9796 against 0.9641 for the leading compared baseline [4], and rendered sphere tracing needs fewer steps because near-surface distances are more accurate.","feed_headline":"Heat-loss training provably converges to true signed distances","feed_subtitle":"Eikonal constraints can't guarantee a distance field; this heat loss bounds the error and stabilizes training.","key_machinery":"The machinery is the screened Poisson equation $\\nabla^2 h - \\lambda^2 h = 0$ with the Dirichlet condition $h = 1$ on the target surface, together with the classical heat–distance relation $\\lim_{\\lambda\\to\\infty} \\frac{1}{\\lambda}\\ln h_\\lambda(x) = -d_\\Gamma(x)$, which says the log of an absorbed heat field recovers the distance to the boundary as absorption grows. The paper turns this PDE into a trainable loss by substituting $h(x) = e^{-\\lambda|u(x)|}$ into the energy functional $\\frac{1}{2}\\int_\\Omega \\|\\nabla h\\|^2 + \\lambda^2 h^2\\,dx$, yielding the heat loss whose variation recovers the screened Poisson equation. The absorption coefficient $\\lambda$ is the accuracy knob: the supplement proves the recovered field $|u_\\lambda| = -\\frac{1}{\\lambda}\\ln h_\\lambda$ deviates from $d_\\Gamma$ by $1/\\lambda$-order terms that depend on $\\ln(\\varepsilon/d_\\Gamma)$ and $\\ln N$, where $N$ is the number of input points.","core_discovery":"The central claim is that minimizing the heat loss $L_{\\mathrm{heat}} = \\frac{1}{2}\\int_\\Omega e^{-2\\lambda|u(x)|}\\left(\\|\\nabla u(x)\\|^2 + 1\\right)\\,dx$, together with a boundary term that pins $u$ to zero at the input points, makes the recovered field converge to the true signed distance as $\\lambda \\to \\infty$. The paper proves that the solution of the screened Poisson equation $\\nabla^2 h - \\lambda^2 h = 0$, with $h = e^{-\\lambda\\varepsilon}$ on small balls of radius $\\varepsilon$ around the sample points, satisfies $\\frac{1}{\\lambda}\\ln\\frac{\\varepsilon}{d_\\Gamma(x)} \\le d_\\Gamma(x) - |u_\\lambda(x)| \\le \\frac{1}{\\lambda}\\left(\\ln\\frac{\\varepsilon}{d_\\Gamma(x)} + \\ln N\\right)$, so the error is first-order small in $1/\\lambda$. Because the heat loss is the variational form of that equation under the substitution $h = e^{-\\lambda|u|}$, its minimizer inherits this heat–distance behavior, in contrast to eikonal-type losses that are only necessary conditions and admit non-distance solutions even when minimized almost everywhere. Three structural consequences follow: spatial stability (a perturbation of the field decays exponentially instead of traveling along rays), temporal stability (the gradient flow is the heat equation and converges), and an automatic penalty on large surfaces through a coarea-type bound that does not distort the level sets.","pith_inferences":["The bound suggests a density-aware default for $\\lambda$ that the paper leaves implicit: with an estimate of the point spacing and a target error tolerance, Eq. 38 can be inverted to choose $\\lambda$ up front, turning the hand-tuned scheduler into a parameter-free recipe.","A testable prediction of the theory is a sharp phase transition in point-cloud density: reconstruction quality should degrade abruptly once the mean spacing approaches the scale of the boundary balls used in the proof, and the error bound should visibly break at that density.","Because the loss needs only the field value and its first derivative, and the boundary condition is the only geometric input, the same construction could attach the screened-Poisson boundary to silhouettes or depth maps rather than points, which would extend the asymptotically sufficient guarantee to image-driven reconstruction; the paper lists inverse rendering as future work without making this ","The stability analysis applies just as well to initialization noise as to injected errors, so the heat loss should make results less sensitive to random seeds than eikonal training, a comparison the paper does not run."],"forward_implications":["Raw point clouds yield better surfaces: on ShapeNet the method reports IoU 0.9796 against 0.9641 for the strongest compared baseline, with lower Chamfer and Hausdorff distances, and on high-genus shapes it recovers the correct topology where baselines produce extra boundaries.","Distance queries become more reliable exactly where they matter: near-surface errors drop by more than a third relative to the best baseline on the paper's ShapeNet metrics, and sphere tracing converges in fewer steps as a direct result.","Optimization is provably better behaved: the heat-loss gradient flow is the heat equation and converges, while a local perturbation of the field decays exponentially instead of persisting along characteristic lines as it does for the eikonal equation.","Area shrinkage no longer distorts the field: because the heat loss upper-bounds a coarea term, large surfaces are penalized automatically, and the separate area regularizers that bend the level sets of prior methods become unnecessary.","Accuracy is predictable in advance: the linear-in-$1/\\lambda$ bound means the residual error of a trained field is controlled by $\\lambda$, the sampling scale $\\varepsilon$, and the point count $N$, which the paper exploits through a $\\lambda$ scheduler."],"supporting_citations":[{"why":"Supplies the classical heat–distance limit (Eq. 5) that the whole loss is built on, relating the log of the screened-Poisson solution to the distance function.","marker":"[6]"},{"why":"The closely related PHASE model; the paper derives the log-occupancy connection, contrasts boundary-weight prescriptions, and reproduces PHASE's collapse in Fig. 6.","marker":"[49]"},{"why":"The StEik baseline the method must beat, and the source of the von Neumann stability analysis for the eikonal loss that the paper extends to the heat loss.","marker":"[4]"},{"why":"The DiGS baseline and the source of the divergence-loss ablation components and the 2D/ShapeNet comparison protocol.","marker":"[3]"},{"why":"Establishes the necessary-but-insufficient status of the eikonal condition; the paper's 1D experiment also rules out its CP loss as a sufficient constraint.","marker":"[5]"},{"why":"Spectral bias of neural networks, used to justify that the network interpolates a continuous boundary between sparse input points so the Dirichlet condition approximately holds.","marker":"[58]"},{"why":"The geodesics-in-heat method that popularized approximating distance from heat flow, giving the practical lineage the paper extends to neural SDF optimization.","marker":"[7]"},{"why":"Defines the coarea loss whose surface-area interpretation the heat loss bounds through inequality (16), replacing explicit area regularization.","marker":"[30]"},{"why":"Provides the high-genus benchmark shapes (bunny, genus-6, Voronoi sphere, dino, kangaroo) where the paper demonstrates correct topology reconstruction.","marker":"[1]"}],"fun_headline_variants":["HotSpot: heat loss guarantees distance fields asymptotically","Screened Poisson loss yields true signed distances","Eikonal fails, heat loss converges to distance","Stable neural SDFs via screened Poisson loss","Asymptotic sufficiency: heat loss beats eikonal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes the input points are dense enough that a continuous surface connecting them can act as a firm boundary where the heat field is held at a constant value, whereas the optimizer enforces that boundary only softly with a weighted loss; if the points are too sparse or the boundary weight too small, the signed distance can collapse into an unsigned one.","fun_headline_variants_meta":{"raw":{"variants":["HotSpot: heat loss guarantees distance fields asymptotically","Screened Poisson loss yields true signed distances","Eikonal fails, heat loss converges to distance","Stable neural SDFs via screened Poisson loss","Asymptotic sufficiency: heat loss beats eikonal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000397,"raw_usage":{"total_tokens":2112,"prompt_tokens":1012,"completion_tokens":1100,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":1023}},"tokens_in":628,"tokens_out":1100,"duration_ms":8103,"temperature":1.0,"reasoning_tokens":1023,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:07:41.697227+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the screened Poisson equation numerically for a known analytic surface with $h = 1$ on a densely sampled boundary, recover $u_\\lambda = -\\frac{1}{\\lambda}\\ln h_\\lambda$, and check whether $d_\\Gamma(x) - |u_\\lambda(x)|$ stays within the paper's bound and falls linearly with $1/\\lambda$ as $\\lambda$ grows. Then thin the samples until the proof's assumption that boundary balls do not influence each other clearly fails, and test whether the minimizer of the heat loss still converges to a true signed distance; if the error leaves the predicted envelope, or remains distant at those densities, the asymptotic sufficiency claim is limited to the dense regime.","supporting_citations":[{"cited_title":"On the behavior of the fundamental solution of the heat equation with variable coef- ficients","cited_arxiv_id":null,"evidence_quote":"Supplies the classical heat–distance limit (Eq. 5) that the whole loss is built on, relating the log of the screened-Poisson solution to the distance function."},{"cited_title":"Phase transitions, distance functions, and im- plicit neural representations","cited_arxiv_id":null,"evidence_quote":"The closely related PHASE model; the paper derives the log-occupancy connection, contrasts boundary-weight prescriptions, and reproduces PHASE's collapse in Fig. 6."},{"cited_title":"StEik: Stabilizing the Optimization of Neural Signed Distance Functions and Finer Shape Representation","cited_arxiv_id":"2305.18414","evidence_quote":"The StEik baseline the method must beat, and the source of the von Neumann stability analysis for the eikonal loss that the paper extends to the heat loss."},{"cited_title":"DiGS: Divergence guided shape implicit neu- ral representation for unoriented point clouds","cited_arxiv_id":null,"evidence_quote":"The DiGS baseline and the source of the divergence-loss ablation components and the 2D/ShapeNet comparison protocol."},{"cited_title":"Constructive solid geometry on neural signed dis- tance fields","cited_arxiv_id":null,"evidence_quote":"Establishes the necessary-but-insufficient status of the eikonal condition; the paper's 1D experiment also rules out its CP loss as a sufficient constraint."},{"cited_title":"On the spectral bias of neural networks","cited_arxiv_id":null,"evidence_quote":"Spectral bias of neural networks, used to justify that the network interpolates a continuous boundary between sparse input points so the Dirichlet condition approximately holds."},{"cited_title":"Geodesics in heat: A new approach to computing distance based on heat flow","cited_arxiv_id":null,"evidence_quote":"The geodesics-in-heat method that popularized approximating distance from heat flow, giving the practical lineage the paper extends to neural SDF optimization."},{"cited_title":"Visco grids: Surface reconstruction with viscosity and coarea grids","cited_arxiv_id":null,"evidence_quote":"Defines the coarea loss whose surface-area interpretation the heat loss bounds through inequality (16), replacing explicit area regularization."},{"cited_title":"A level set theory for neural implicit evolution under explicit flows","cited_arxiv_id":null,"evidence_quote":"Provides the high-genus benchmark shapes (bunny, genus-6, Voronoi sphere, dino, kangaroo) where the paper demonstrates correct topology reconstruction."}],"review_version":1}