{"id":"7ca32a46-edee-411e-a966-d83ba8e66c41","arxiv_id":"2603.13546","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PGH deforms Boltzmann densities of nonconvex objectives into a soft-min homotopy whose Monte Carlo gradients bias descent toward low-energy regions and recovers a posterior-mean Moreau envelope.","lead":"The paper introduces Probabilistic Gaussian Homotopy (PGH), which smooths a nonconvex objective by deforming its Boltzmann distribution and aggregates gradients with Boltzmann weights that favor low-energy regions. This probability-space continuation can help optimizers escape poor basins on hard nonconvex problems where classical gradient methods and objective-space Gaussian smoothing fail.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The practical superiority claim rests on an unproved path-tracking assumption plus heavy per-problem tuning of the annealing schedule and K, so the reported gains may not generalize beyond the tuned settings.","rationale":"The Reader correctly isolates the weakest link: the unproved regularity assumption that lets the discrete annealing schedule track a useful global path. Theorems 1–2 are standard and correct; the soft-min / posterior-mean interpretation is clean. The only material risk to the strongest claim is therefore empirical: whether the reported gains survive a fixed, non-tuned schedule. Because that risk is already flagged by the Reader and does not invalidate the theoretical construction, the verdict remains CONDITIONAL; no stronger rejection is warranted.","tokens_in":12605,"tokens_out":517,"duration_ms":4747,"concrete_test":"Fix a single, untuned schedule (e.g., T=100, K=8, cosine LR=1.0, tmin=0.1) and re-run the four n=10 benchmarks of Table 1 for 30 independent seeds; if the mean function-evaluation counts rise above those of CMA-ES or Implicit Filtering on any two functions, or if success rate drops below 80 %, the practical-superiority claim is schedule-dependent and the path-tracking assumption is not load-bearing in practice.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that PGHO reliably tracks a useful global path (and thereby outperforms classical GH and black-box global optimizers) depends on Section 3.3: minimizers of Ft evolve smoothly under the non-autonomous flow dx/dt = −∇x Ft under only “suitable regularity conditions.” No theorem, Lipschitz/coercivity hypothesis, or counter-example analysis is supplied. In practice the continuous path is replaced by a discrete schedule (Algorithm 1) with very small Monte Carlo size K (typically 4, occasionally 16) and problem-specific choices of homotopy length (50–200 steps), initial learning rates (0.1–50), and tmin. Because the Boltzmann weights become extremely peaked as λ(t) decreases, the Monte-Carlo estimator can collapse onto a single sample and the discrete trajectory can jump basins or stall; the paper never quantifies how often this occurs. Consequently the strong empirical numbers in Table 1 and Figure 3 may be artifacts of extensive per-function hyper-parameter search rather than a robust consequence of the probability-space construction.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Probabilistic Gaussian Homotopy (PGH), a continuation method that smooths the Boltzmann density of a nonconvex objective rather than the objective itself. The resulting energy Ft is a log-sum-exp (soft-min) aggregation of Gaussian-perturbed function values; its gradient is a Boltzmann-weighted average of local gradients (Eq. 14). Under the scaling α=1, β=√λ the construction coincides with a soft Moreau envelope (Theorem 1), and gradient steps on Ft are posterior-mean updates under a Gaussian denoising model (Theorem 2 / Tweedie). A practical Monte-Carlo algorithm (PGHO, Algorithm 1) is derived from the non-autonomous gradient flow of Ft and is shown to outperform classical Gaussian homotopy and several global optimizers on standard multimodal benchmarks (Table 1, Fig. 3) and on a smooth-ℓ0 sparse-recovery problem (Fig. 4).","tokens_in":12920,"tokens_out":1002,"duration_ms":8878,"significance":"If the empirical gains hold under modest hyper-parameter budgets, the work supplies a clean conceptual bridge among classical Gaussian homotopy, Moreau envelopes, Bayesian denoising, and diffusion-style smoothing, together with a simple, parallelizable first-order algorithm. The two short theorems are standard but correctly placed; the soft-min / posterior-mean view of continuation is a useful organizing principle for the community. The main practical contribution is the demonstration that Boltzmann-weighted gradient aggregation can be competitive with (and often faster than) both objective-space GH and black-box global methods on the chosen test suite.","major_comments":[{"comment":"Section 3.3 asserts that minimizers of Ft evolve smoothly under the non-autonomous flow dx/dt = −∇Ft under only “suitable regularity conditions,” yet no Lipschitz, coercivity, or isolation hypotheses are stated and no counter-example analysis is given. Because the discrete PGHO schedule (Algorithm 1) with small K (typically 4) is justified by this continuous path-tracking claim, the central algorithmic guarantee remains incomplete; at minimum a precise statement of the conditions under which the discrete trajectory stays near a continuous minimizer path is needed.","section":null},{"comment":"Appendix B.1 and the experimental protocol reveal extensive per-problem tuning of homotopy length (50–200 steps), initial learning rates (0.1–50), K, and tmin. Table 1 and Figure 3 report mean function-evaluation counts and success rates without any sensitivity study or fixed-hyper-parameter ablation. Given that Boltzmann weights become extremely peaked as λ(t)\to0, the Monte-Carlo estimator can collapse onto a single sample; without evidence that the reported gains survive modest hyper-parameter budgets, the practical superiority claim over classical GH and the black-box baselines is only weakly supported.","section":null}],"minor_comments":[{"comment":"Notation for the temperature schedule is inconsistent: λ(t) appears both as the Boltzmann temperature and (implicitly) as the Moreau parameter; a single, explicit schedule table would help.","section":null},{"comment":"Figure 1 caption refers to a “nearly flat” surface at large smoothing, but the plotted range is not quantified; adding the corresponding λ or β values would make the illustration more informative.","section":null},{"comment":"In Algorithm 1 the Monte-Carlo objective (20) uses λ(tk)−1 while the surrounding text uses 1/λ(t); the two expressions should be aligned.","section":null},{"comment":"Related-work discussion of single-loop GH (Iwakiri et al., 2022) and of Laplace–Moreau approximations (Tibshirani et al., 2025) is present but brief; a short paragraph clarifying the precise algorithmic difference from SLGH would strengthen the novelty claim.","section":null},{"comment":"Sparse-recovery experiments fix τ=0.05 and a single sensing ratio; a brief note on sensitivity to these choices would improve reproducibility.","section":null}],"recommendation":"major_revision","confidential_remarks":"The theoretical core is sound but thin; the paper’s value is primarily empirical and conceptual. If the authors can supply either a rigorous path-tracking result under stated hypotheses or a convincing fixed-hyper-parameter ablation, the contribution is suitable for a solid ML/optimization venue. Without that, the risk that the gains are largely the product of per-function schedule search remains material."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful thing here is the geometric switch: instead of convolving f and averaging gradients uniformly, they smooth the Boltzmann density and get a soft-min (Boltzmann-weighted) gradient. That is not just a re-labeling of classical Gaussian homotopy. Under the scaling β=√λ the energy is exactly a soft Moreau envelope (Theorem 1), and the gradient step is a posterior-mean update via Tweedie (Theorem 2). Both proofs are short change-of-variable / score calculations and look correct. The Monte-Carlo estimator and Algorithm 1 are simple enough to implement, and the connection to diffusion schedules is made explicit without overclaiming.\n\nWhat the paper does well is keep the math honest and the experiments focused. Table 1 (n=10, 30 runs, fixed budget) shows PGH-GD/Adam reaching the 5e-2 threshold with far fewer evaluations than GH, CMA-ES, DE, annealing, etc. Success-rate vs dimension on Ackley stays high while several baselines collapse. The sparse-recovery L-curves and objective-vs-λ plots are consistent with the same story: better sparsity-fidelity trade-off than plain GD/Adam on the smooth-ℓ0 problem. Citations cover the right classical GH, Moreau, and diffusion literature; no obvious gaps or self-citation games.\n\nSoft spots, in proportion. Section 3.3 asserts that minimizers of Ft evolve smoothly under dx/dt = −∇Ft under “suitable regularity conditions,” then immediately discretizes with tiny K (usually 4) and problem-specific schedules (T=50–200, learning rates 0.1–50, tmin, etc.). No Lipschitz/coercivity hypotheses, no counter-example analysis, no quantification of how often the peaked weights collapse the estimator or the discrete path jumps basins. That is a genuine gap for the “reliably tracks a global path” claim, and the stress-test note is right that the reported gains could partly be tuning artifacts. Code is not released, statistical reporting is only means, and free parameters are numerous. These are real but do not break the central construction or the two theorems.\n\nThis is for people who already care about continuation, Moreau envelopes, or nonconvex inverse problems and want a first-order method that systematically biases toward low-energy basins. It is not a foundational breakthrough, but it is a coherent algorithmic result with correct short theory and usable empirical signal. I would send it to referees; the gaps are fixable with more analysis or broader ablation, not desk-reject material. Worth reading and, if the path-tracking story holds up under scrutiny, worth citing.","headline":"Clean probability-space reworking of Gaussian homotopy with a correct soft-Moreau link and solid empirical gains; the path-tracking theory is thin and the numbers are heavily tuned, but the construction itself is real and worth referee time.","tokens_in":13524,"tokens_out":647,"would_cite":true,"duration_ms":5724,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Smoothing the Boltzmann density, not the objective, gives descent directions that favor low-energy basins and recovers a soft Moreau envelope.","keywords":["homotopy continuation","Gaussian smoothing","Moreau envelope","Boltzmann distribution","nonconvex optimization","score-based diffusion","sparse recovery"],"falsifier":"On a multimodal test function of moderate dimension, replace the Boltzmann weights by uniform weights (recovering classical Gaussian homotopy) while keeping every other hyper-parameter identical; if success rate and evaluation count do not degrade, the claimed advantage of probability-space weighting disappears.","tokens_in":13498,"feed_emoji":"📉","tokens_out":581,"duration_ms":5052,"temperature":0.7,"pith_summary":"Classical Gaussian homotopy smooths a nonconvex objective by convolution and averages gradients uniformly, so competing basins pull equally. This paper instead smooths the Boltzmann distribution associated with the objective and takes a log-sum-exp (soft-min) average of the same perturbed gradients. The resulting weights are exponential in energy, so low-energy neighborhoods dominate the descent direction. Under a simple scaling the construction is exactly a finite-temperature, posterior-mean version of the Moreau envelope, and its minimizers obey a non-autonomous gradient flow that can be tracked by Monte Carlo annealing. The practical algorithm, PGHO, is shown to reach target accuracy with far fewer function evaluations than classical homotopy and several global optimizers on standard high-dimensional benchmarks and on sparse recovery.","feed_headline":"Weight gradients by energy, not uniformly, to escape bad basins","feed_subtitle":"A soft-min homotopy on the Boltzmann density beats classical Gaussian smoothing on hard nonconvex tests","key_machinery":"The soft-min energy Ft(x) = −λ(t) log pt(x), where pt is the Gaussian-smoothed Boltzmann density; its Monte-Carlo gradient is the Boltzmann-weighted aggregation that biases steps toward low-energy regions and recovers the posterior-mean identity of the Moreau envelope when β² = λ.","core_discovery":"Probabilistic Gaussian Homotopy deforms the Boltzmann density of a nonconvex objective rather than the objective itself; the induced gradient is a Boltzmann-weighted average of Gaussian-perturbed gradients, which coincides with the gradient of a soft (log-sum-exp) Moreau envelope and yields a tractable continuation path from a nearly flat landscape to the original problem.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Boltzmann weights bias gradients to low-energy basins","Deform the Boltzmann density, not the objective itself","Soft-min homotopy recovers objectives via energy-weighted grads","PGH induces soft-min Moreau envelopes for nonconvex paths","Annealed Boltzmann path yields energy-biased descent directions"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That the minimizers of the smoothed energies move smoothly enough under ordinary gradient flow that a simple annealing schedule with only a handful of Monte-Carlo samples can track a useful global path.","fun_headline_variants_meta":{"raw":{"variants":["Boltzmann weights bias gradients to low-energy basins","Deform the Boltzmann density, not the objective itself","Soft-min homotopy recovers objectives via energy-weighted grads","PGH induces soft-min Moreau envelopes for nonconvex paths","Annealed Boltzmann path yields energy-biased descent directions"]},"model":"grok-4.5","effort":"low","cost_usd":0.00368,"raw_usage":{"total_tokens":1164,"prompt_tokens":730,"num_sources_used":0,"completion_tokens":62,"cost_in_usd_ticks":36800000,"prompt_tokens_details":{"text_tokens":730,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":372,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":730,"tokens_out":62,"duration_ms":3389,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T21:45:45.276696+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a multimodal test function of moderate dimension, replace the Boltzmann weights by uniform weights (recovering classical Gaussian homotopy) while keeping every other hyper-parameter identical; if success rate and evaluation count do not degrade, the claimed advantage of probability-space weighting disappears.","supporting_citations":[],"review_version":1}