{"id":"bbf287ca-015f-4a65-9f20-bdb2cea4fc42","arxiv_id":"2507.01854","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"With C1 convergence plus a positive lower bound on critical point spacing, maxima, minima, and saddle counts converge; with C2 convergence to a Morse function, Morse index counts converge.","lead":"This paper identifies regularity conditions under which a sequence of functions approaching a limit function has the same number of peaks, pits, and saddles as the limit. The results give theoretical backing to a common assumption in brain imaging and other fields that count topological features from finite samples.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's Step 2 applies Theorem 10 at the fixed level c=f(p), but the stated mountain pass theorem returns a point at its own level c_n<f_n(p_n); no argument gives c_n→c, so the transversality contradiction does not follow as written.","rationale":"I read the paper's goal as providing minimal sufficient conditions for convergence of critical-point counts. The C^2/Morse strand is substantially better supported: Theorem 4 tracks explicit constants from the parametric Morse lemma, and Theorem 5's uniqueness argument is plausible. The C^1 strand, which carries the reader's strongest_claim, depends on Theorem 3. The weakest point is Step 2's use of Theorem 10: the theorem as stated does not produce a point at a preassigned level, and the proof neither states which endpoints are used nor proves that the produced level converges to f(p). This is an internal gap, not a disagreement with external consensus. Assumption 3 is indeed load-bearing and explicitly acknowledged, and Example 3 shows it cannot be dropped; but it is an assumption, whereas the Step 2 gap affects the theorem's proof under its own hypotheses. I do not conclude that Theorem 3 is false; the gap appears repairable with a level-convergence lemma or a correctly stated fixed-level mountain-pass variant. Therefore the reader's CONDITIONAL verdict remains appropriate.","tokens_in":33928,"tokens_out":23546,"duration_ms":288391,"concrete_test":"Analytical check: let f(x,y)=x^2−y^2 on B_1, take p'=(ε,0), and consider any C^1 sequence f_n satisfying the hypotheses of Theorem 3 with a unique local maximum p_n→0 and no other critical point in B_1. For each n, apply Theorem 10 with p_1=p_n and p_2=p', and record the level c_n=f_n(p'_n) of the resulting boundary-tangential or critical point. If liminf_n c_n < f(p)=0, then the claimed contradiction at level f(p) fails and Theorem 3 is false as stated; if lim_n c_n=0, the proof needs an explicit missing lemma establishing this level convergence. Re-running Step 2 with the actual c_n in place of c will show which repair is needed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing soft spot is in the proof of Theorem 3, Step 2 (§3.2). The proof fixes c=f(p), chooses p' with f(p')>c, and after showing f_n(p')>f_n(p_n) invokes the mountain pass theorem (Theorem 10, Appendix D) to obtain p'_n that is either a critical point or lies on f_n^{-1}(c) tangent to ∂B_δ(p). But Theorem 10 does not fix a level: given p_1,p_2 it returns p_3 with f(p_3)=c_mp<f(p_1) (or, under the reversed ordering, a different inequality). Applying it with p_1=p_n, p_2=p' gives a level below f_n(p_n), not equal to f(p); applying it with reversed endpoints gives a level below f_n(p'). The proof supplies no estimate showing c_mp→f(p), and none follows from C^1 convergence alone because the Hessian of f_n at the local maximum is unconstrained. Without c_mp→c, the limiting equation f(p^*)=c and the transversality contradiction at level c do not follow. This gap is internal to the central C^1 theorem and is independent of the explicitly assumed bounded-resolution condition; it is the main reason the paper's headline convergence result is not fully proven as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies when C^k convergence of functions on a compact manifold with boundary implies convergence of the numbers of critical points of various types. In the C^1 setting the authors introduce a bounded critical-point-resolution assumption, then use a generalized Poincaré-Hopf theorem and a mountain-pass theorem to claim convergence of the numbers of local maxima, minima, and saddles. In the C^2 setting they prove convergence of Morse-indexed critical point counts via a homotopy-based Morse lemma with explicit constants. The paper closes with a probabilistic reformulation in the language of empirical processes.","tokens_in":34183,"tokens_out":17916,"duration_ms":204597,"significance":"If the central theorems held, they would provide useful and fairly minimal regularity conditions under which topological summaries of finite-sample approximations stabilize, and the paper is honest about failure modes (Examples 1–3). The C^2/Morse strand, Theorems 4 and 5, is carefully argued with explicit constant tracking and appears sound. However, the C^1 strand, which is the paper's headline contribution, has a central gap in the proof of Theorem 3 and an overclaim in Theorem 2, so the paper is not yet in a publishable state.","major_comments":[{"comment":"The application of Theorem 10 does not justify the stated conclusion. Theorem 10 returns a point p'_n with f_n(p'_n) = c_n < f_n(p_n); it does not allow one to prescribe the level c = f(p). The proof then treats p'_n as lying on f_n^{-1}(c), but no estimate shows c_n = c or c_n → f(p). Consequently the limiting identity c = f(p*) and the transversality contradiction at level c do not follow. This is the load-bearing step for convergence of local maxima.","section":"§3.2, Theorem 3, Step 2"},{"comment":"The displayed limiting argument `0 = ∇f_n(p'_n)·(p'_n - p)` is inconsistent with the definition of tangency. If f_n^{-1}(c_n) is tangential to ∂B_δ(p) at p'_n, then ∇f_n(p'_n) is normal to ∂B_δ(p), hence collinear with the radial vector p'_n - p; the scalar product is ±|∇f_n(p'_n)||p'_n - p|, not 0. The contradiction with Condition (2) should be obtained by passing to the limit of the vanishing tangential component, not by the displayed equality. As written, the limiting step is invalid.","section":"§3.2, Theorem 3, Step 2"},{"comment":"The statement `lim sup_{n→∞} N_0^H(f_n) = N_0^H(f)` is stronger than what the proof establishes and is false as stated. The proof only shows that near an undulation point of f there is at most one critical point of f_n, and if present it has index zero; it gives no existence. For S = [-1,1], f(x) = x^3, and f_n(x) = x^3 + x/n, we have f_n → f in C^1, Assumptions 1 and 3 hold, f has one undulation point at 0, and N_0^H(f_n) = 0 for all n while N_0^H(f) = 1. The statement should be changed to `lim sup ≤ N_0^H(f)`, and downstream uses, including Theorem 3's final line and the proof of Theorem 6, should be re-checked; the weaker inequality suffices when undulations are assumed absent.","section":"§3.2, Theorem 2"},{"comment":"The proof does not verify condition (iii) on the annulus B_{2γ} \\ B_γ for the adjusted functions \\tilde f_n. After defining \\tilde f_n, the gradient lower bound is shown only on B_ϵ \\ B_{2γ}; condition (iii) requires absence of critical points on all of B_ϵ \\ Int(B_γ). The missing estimate can likely be obtained by the same modulus-of-continuity argument applied to f_n, but it should be written out.","section":"§3.2, Lemma 3"}],"minor_comments":[{"comment":"Assumption (M2) is written as `P*[L = 0]` and appears to be missing the required `= 0`; as stated it is not a probability condition.","section":"§4, Theorem 7, assumption (M2)"},{"comment":"The statement uses `Γ_n : B_{r_n}(p_n) → R^D` while writing `(x - p_n)` in the displayed identity, which treats p_n as both a point of S and a coordinate origin. This is reconciled after the translations in the proof, but the statement should say so explicitly.","section":"§3.3, Theorem 4"},{"comment":"The sentence `Noting the final line of Theorem 2` is ambiguous once Theorem 2 is corrected to a limsup inequality; the proof should state explicitly that H5 together with the limsup bound yields the needed convergence of N_0^H.","section":"§4, Theorem 6 proof"}],"recommendation":"major_revision","confidential_remarks":"The C^2/Morse strand (Theorems 4 and 5) is the strongest part and appears sound. The C^1 strand is not ready as written: the mountain-pass application in Theorem 3 has both a level-mismatch problem and an incorrect tangency limiting equation, and Theorem 2 overclaims an equality that the proof does not support. The probabilistic theorems inherit these issues through their reliance on the C^1 results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the paper carefully, and the reader's report is close to the mark, though I'd sharpen one thing: the gap in Theorem 3's Step 2 is real and load-bearing, not a minor technicality.\n\nThe good news first. The paper does something genuinely useful: it moves beyond the local-maxima-specific results of Cheng & Schwartzman and Davenport et al., and gives convergence statements for minima, saddles, and Morse-indexed counts, on manifolds with boundary. The C2 strand (Theorems 4 and 5) is carefully argued, with explicit constants in the Morse-neighborhood construction and a clean one-to-one correspondence argument. That part is, as far as I can tell, correct and a real contribution. The examples are well chosen and illustrate the failure modes nicely.\n\nThe soft spot is the C1 strand. In Step 2 of Theorem 3, the proof fixes a level c = f(p), then applies the mountain pass theorem (Theorem 10) with p1 = p_n, p2 = p'. That theorem returns a point p'_n at a level c_n < f_n(p_n), not at the prescribed level c. The proof then asserts f_n^{-1}(c) intersects ∂B_δ(p) tangentially at p'_n, but this is simply not what the theorem gives. No estimate shows c_n → f(p), and C^1 convergence alone doesn't provide one. So the limiting tangential intersection at level c doesn't follow, and the contradiction collapses. This is exactly the stress-test concern, and I don't see a way around it as written.\n\nA secondary note: the mountain pass condition (every path from p_n to p' has min < f_n(p_n)) is not verified for a possibly non-strict local maximum. That's a smaller issue; one might handle it by perturbing the ball, but it's still unaddressed.\n\nThe reader's worry about Lemma 1's finiteness claim, by contrast, doesn't land: isolated improper maxima in a compact set are finite, so that argument is fine.\n\nThe probabilistic results in Section 4 inherit the C1 gap for Theorem 6 because they lean on Theorem 3. Theorem 7, resting on the C2 strand, should be fine.\n\nBottom line: this paper deserves a serious referee, but not acceptance as-is. The C1 gap is repairable—one needs a mountain-pass variant that can be pushed to a prescribed level, or a continuity argument for the level—but it needs to be written. I'd send it out with a request for major revision.","headline":"The C2/Morse half is solid, but the C1 half has a genuine mountain-pass gap that breaks Theorem 3 as written.","tokens_in":34693,"tokens_out":5585,"would_cite":true,"duration_ms":60075,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["58K05","60G60"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under bounded critical point resolution, $C^1$ convergence makes counts of maxima, minima, and saddles converge; $C^2$ convergence with a Morse limit makes Morse-indexed counts converge.","keywords":["critical point convergence","homological index","Morse theory","C^k convergence","empirical processes","mountain pass theorem","Poincaré-Hopf theorem","bounded critical point resolution"],"falsifier":"Run a $C^1$-convergent sequence of smooth functions on a compact domain whose two local maxima move toward each other and merge into a single maximum of the limit, while keeping all critical points away from the boundary; this sequence satisfies the paper's other hypotheses but violates bounded critical point resolution and exhibits $N_M(f_n)=2$ converging to $N_M(f)=1$. To test the positive claim directly, one would need a sequence satisfying Assumptions 1 and 3 with $C^1$ convergence where maxima, minima, or saddle counts fail to converge; the paper's Theorem 3 asserts that none exists.","tokens_in":33677,"feed_emoji":"📐","tokens_out":5615,"duration_ms":58586,"temperature":0.7,"pith_summary":"This paper asks when the number of local maxima, minima, and saddles of a sequence of functions $f_n$ converges to those of a limit $f$, under convergence in the $C^1$ or $C^2$ metric. The authors show that $C^1$ convergence alone is not enough: two peaks of $f_n$ can merge into one peak of $f$. They identify a condition called bounded critical point resolution — the distance between distinct critical points of $f_n$ is bounded below by a positive constant — under which the counts of maxima, minima, and saddles do converge, even when the limit has degenerate critical points. When the limit is Morse and convergence is $C^2$, they show the Morse-indexed counts converge as well. The paper matters because practitioners routinely read topological features off finite-sample approximations, and these theorems state the regularity conditions that make that practice valid.","feed_headline":"Bounded peak separation forces critical-point counts to converge","feed_subtitle":"When approximating functions keep their peaks apart, maxima, minima, and saddles converge to the limit's.","key_machinery":"The load-bearing device is the critical point resolution $R(\\{f_n\\}) := \\liminf_{n\\to\\infty} \\inf_{p_1\\ne p_2 \\in Z(\\nabla f_n)} |p_1-p_2|$, the limiting minimal distance between distinct critical points of the approximating functions; Assumption 3 requires this to be positive. It prevents the collision of peaks and saddles that breaks convergence. The arguments also rely on two named tools: the homological index of a vector field on a manifold with boundary, used through the generalized Poincaré–Hopf theorem, and a mountain pass theorem for convex domains that detects whether a critical point is a maximum, minimum, or saddle. In the $C^2$ setting, the Morse lemma is proved via a homotopy whose flows give explicit constants for the size of Morse neighborhoods, allowing the neighborhoods of $f_n$ and $f$ to be matched.","core_discovery":"The central claim is that two regularity conditions on the approximating sequence — no critical points of the limit on the boundary, and bounded critical point resolution — are sufficient for the number of local maxima, minima, and saddles of $f_n$ to converge to those of $f$ under $C^1$ convergence (Theorem 3). The proof works by pairing a generalized Poincaré–Hopf theorem for manifolds with boundary, which forces critical points of $f_n$ to appear near those of $f$ with matching homological index, with a mountain-pass theorem on convex domains, which rules out a local maximum of $f_n$ converging to a saddle or undulation of $f$. Under $C^2$ convergence with a Morse limit, the authors prove a stronger statement (Theorem 5): the number of critical points of each Morse index converges, via a homotopy-based Morse lemma that provides explicit bounds on the Morse neighborhoods and a controlled correspondence between critical points. Probabilistic versions of both theorems are obtained for random processes by importing these deterministic results onto almost-sure representations.","pith_inferences":["A testable consequence for practice: estimators built on fixed lattices plausibly satisfy bounded critical point resolution, whereas kernel smoothers whose bandwidth shrinks to zero may not, and one can check whether $R(\\{\\hat G_n\\})$ has positive liminf in probability for a given estimator.","The same Poincaré–Hopf machinery that prevents critical-point collisions may also control the convergence of the Euler characteristic of excursion sets, since it already counts critical points with sign; the authors flag this direction as future work.","The theorems suggest a practical diagnostic: monitor the minimal distance between critical points of successive approximations; if it collapses, topological summaries such as peak counts are unstable and should not be trusted.","For smooth Gaussian random fields, the condition that peaks are almost surely separated is plausible in generic settings, and verifying it would let the probabilistic theorems apply directly to common neuroimaging peak-counting pipelines."],"forward_implications":["If $f$ has no undulation points, then under $C^1$ convergence plus the two assumptions the total number of critical points $N_C(f_n)$ converges to $N_C(f)$; otherwise it can only be bounded above in the limit.","The convergence statements hold for maxima, minima, and saddles separately, not just for homological-index counts, which is what applications that count peaks and clusters need.","Under $C^2$ convergence with $f$ Morse, every Morse-indexed count $N^M_\\lambda(f_n)$ converges to $N^M_\\lambda(f)$, giving the sharp guarantee that no extra critical points of any index appear in the limit.","For random processes, weak convergence of the process and its derivatives, plus mild probability conditions, imply weak convergence of the critical-point counts; in particular, when the limiting process almost surely has no undulation points, the total count converges.","The assumptions can be checked mostly in terms of the approximating functions $f_n$ themselves, so the theory gives a template for verifying convergence in concrete estimators."],"supporting_citations":[{"why":"Supplies the generalized Poincaré–Hopf theorem for manifolds with boundary that forces critical points of $f_n$ to appear near those of $f$ with matching homological index.","marker":"[11]"},{"why":"Provides the mountain pass theorem on convex domains used to distinguish maxima from saddles in the $C^1$ convergence proof.","marker":"[12]"},{"why":"Provides the homotopy-based Morse lemma with explicit neighborhood constants that anchors the $C^2$ convergence proof and the one-to-one Morse point correspondence.","marker":"[14]"},{"why":"Supplies the empirical process weak convergence framework used to transform the deterministic theorems into probabilistic convergence statements.","marker":"[15]"}],"fun_headline_variants":["Critical counts converge when boundary is clean and peaks separate","Two regularity rules force critical point count convergence","No boundary criticals, bounded resolution: counts converge","Sufficient regularity conditions for stable critical point counts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole convergence result rests on the assumption that the critical points of the approximating functions stay at least some fixed positive distance apart as the sample grows; if two critical points drift together and merge in the limit, the counts can fail to converge even when derivatives converge.","fun_headline_variants_meta":{"raw":{"variants":["Critical counts converge when boundary is clean and peaks separate","Two regularity rules force critical point count convergence","No boundary criticals, bounded resolution: counts converge","Sufficient regularity conditions for stable critical point counts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000263,"raw_usage":{"total_tokens":1588,"prompt_tokens":920,"completion_tokens":668,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":608}},"tokens_in":536,"tokens_out":668,"duration_ms":8487,"temperature":1.0,"reasoning_tokens":608,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:45:31.266522+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a $C^1$-convergent sequence of smooth functions on a compact domain whose two local maxima move toward each other and merge into a single maximum of the limit, while keeping all critical points away from the boundary; this sequence satisfies the paper's other hypotheses but violates bounded critical point resolution and exhibits $N_M(f_n)=2$ converging to $N_M(f)=1$. To test the positive claim directly, one would need a sequence satisfying Assumptions 1 and 3 with $C^1$ convergence where maxima, minima, or saddle counts fail to converge; the paper's Theorem 3 asserts that none exists.","supporting_citations":[{"cited_title":"A generalized Poincar\\'e-Hopf index theorem","cited_arxiv_id":"0903.0697","evidence_quote":"Supplies the generalized Poincaré–Hopf theorem for manifolds with boundary that forces critical points of $f_n$ to appear near those of $f$ with matching homological index."},{"cited_title":"The Mountain Pass Theorem: Variants, Generalizations and Some Appli- cations","cited_arxiv_id":null,"evidence_quote":"Provides the mountain pass theorem on convex domains used to distinguish maxima from saddles in the $C^1$ convergence proof."},{"cited_title":"Parametric Morse lemmas for C1,1-functions","cited_arxiv_id":null,"evidence_quote":"Provides the homotopy-based Morse lemma with explicit neighborhood constants that anchors the $C^2$ convergence proof and the one-to-one Morse point correspondence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the empirical process weak convergence framework used to transform the deterministic theorems into probabilistic convergence statements."}],"review_version":1}