acceptodds
Under review as a conference paper at ICLR 2027

What a Point Set Knows About a Riemannian Minimizer: A Scaling Law at the Cut Locus

Abstract

Quasi-Monte Carlo (QMC) point sets come with integration certificates: a discrepancy bounds the error of every integrand in a function class. Riemannian optimization asks where the minimizer of the sampled loss lies. One object connects the two. The point set minus the target law is a signed measure, the discrepancy charge; through the loss it generates a potential whose gradient, the field, is the error of the sampled gradient. The minimizer error is the field at the minimizer times an inverse Hessian, and reusing a point set, refreshing it, or solving once reads the field in three norms, which gives their sample complexities. What a certificate says about the field passes through one integrand, the gradient of the loss, and on a manifold this integrand is singular at the cut locus. For positive rules, dilation predicts that a singularity of order on a set of codimension leaves the exponent of the discrepancy, where is the dilation exponent of the certificate norm; we prove this scaling law, and that it cannot be improved, for point singularities on spheres in Sobolev norms and for point and polygonal singularities in the square in Hardy–Krause variation. On spheres the Fr\'echet problem is electrostatics: the cap discrepancy is the energy of the charge, the value error its potential at the antipode, the gradient error its field there. This gives the sharp exponent for the cap discrepancy on ; a floor of order on the worst-case gradient error, which spherical designs meet up to ; and a chart that removes the singularity when it blows up the cut point. Experiments on positive definite matrices, Grassmannians and spheres test the main predictions.

Primary area:
optimization
Keywords:
quasi-Monte Carlo, Riemannian optimization, discrepancy, sample average approximation, stochastic gradient descent, cut locus, Fréchet mean, spherical designs, Koksma–Hlawka inequality, scaling law

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1DcmOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Please see "questions for authors"

Critical strengths

Please see "questions for authors"

Critical weaknesses

Please see "questions for authors"

Questions for authors

I want to start by saying that I found the abstract very hard to follow. I strongly encourage the authors to re-write the abstract in a way that states upfront : what is the problem at hand and what is the proposed solution and contribution.

..Moreover, the submission stands at 80 pages or so : conference papers are supposed to be short and to the point, and a conference review process is too short for anyone to go over so many pages of appendices.

My second recommendation is please shorten your submission to a readable length. I think it is perfectly fine if the proofs are not included, but only sketched, and if background material is replaced by references to litterature (this is where it came from in the first place).

The introduction does a better job of stating the main problem, which I find quite interesting, and also relevant to applied problems. However, I think the auhots attempt something that is way too difficult : to express the problem and its intended solution in written language (in English).

I recommend the introduction use more mathematics (equations/symbols) to write down the main problem in a clear fashion (roughly, one clearly explained equation is much easier to follow that several dense paragraphs of text).

The "relation to prior work" paragraph is extremely confusing and hard to follow. You should redo completely !! Also, I noticed the use of "We", while the submission is supposed to be anonymous.

In Section 2 "Setting" is quite clear to me, and I like the fact that it is concise, but I also ask myself whether non-experts will be able to make any sense out of it. I do not ask you to change anything here, but just share with you this issue.

I agree that Proposition 2.2. is elementary, and I think there is almost no point in discussing its proof. It would be much better to explain how it is going to be used.

Recommendation : what is a "Sobol’ or Halton sequence" .. I have never heard of that. Please explain or provide a reference.

In the paragraph "Sturm's contraction" :

(a) is Log_x(y) 1-Lipschitz ? Its covariant derivative is the Hessian of squared distance, and that is not bounded !

(b) the remainder of the paragraph does not need the above statement and moreover perhaps does not belong here in the first place (it is supposed to explain Thm E.3).

Section 3 :

The opening sentence is very problematic !! You claim that the norm of G_x is infinite at the cut-locus, but G_x is not defined there. Moreover, its norm remains bounded near the cut locus. You are talking about a problem that does not exist !!

Do Propositions 3.3 and 3.4 even address the main issue of the paper (behavior near singularity..) ? if you impose that g is supported away from the singularity (at a fixed distance), you are not dealing with behavior near the singularity.

Conclusion :

Strengths :

(1) the paper raises an important issue, and attempts to address it in a rigorous mathematical way

Weaknesses :

(2) the paper is incomprehensible to non-experts (..I was able to follow it, but I believe many if not most ICLR attendees will not find it cogent/relevant), and often uses obscure language.

(3) the main results in Section 3 are based on strong hypotheses that do not hold in practice and (mathematically) make everything a bit too easy

4 comments

Response to Reviewer 1Dcm (1/4)

Thank you for the paragraph-level reading. Each of the three technical objections (Q9, Q10 with Q13, Q8) is answered by a statement that is in the submission and is proved there. For each we give the statement, its location, and a short derivation. A revised PDF is uploaded, with changes in blue; line numbers below refer to the submitted PDF.

Q9 (Section 3, opening sentence).

You claim that the norm of G_x is infinite at the cut-locus [...] its norm remains bounded near the cut locus.

The problem exists, and the sentence is about a different norm from the one in the objection. It reads: "The three readings need ∥Gx∥H<∞\Vert G_x\Vert_H\lt\infty. For F=12d2F=\frac12d^2 the field integrand is Gx=−LogxG_x=-\mathrm{Log}_x, and where the data can reach the cut locus of xx this norm is infinite." "This norm" is the function-space norm ∥Gx∥H=sup⁡∥v∥=1∥⟨Gx,v⟩∥H\Vert G_x\Vert_H=\sup_{\Vert v\Vert=1}\Vert\langle G_x,v\rangle\Vert_H of the map y↦Gx(y)y\mapsto G_x(y) (Definition 2.1), not the pointwise length ∥Gx(y)∥\Vert G_x(y)\Vert. The submission nowhere says that the pointwise length is unbounded: ∥Logxy∥=d(x,y)≤π\Vert\mathrm{Log}_xy\Vert=d(x,y)\le\pi on Sd\mathbb S^d, and Theorem 3.3 assumes that gg is bounded. A QMC error bound ∣QNg−Pg∣≤∥g∥HDH(QN)\vert Q_Ng-Pg\vert\le\Vert g\Vert_H D_H(Q_N) needs the function-space norm, and on the sphere that norm is infinite for the Sobolev classes that carry QMC rates, the classes Hτ(Sd)H^\tau(\mathbb S^d) with τ>d/2\tau\gt d/2: their members are continuous, and GxG_x has no continuous extension to −x-x. Section 4 treats the polar chart in which the norm is finite (Proposition 4.2(a)). In normal coordinates zz centred at the antipode −x-x (Eq. (4)),

Logx(Exp−xz)=πz∥z∥−z.\mathrm{Log}_x(\mathrm{Exp}_{-x}z)=\pi\frac{z}{\Vert z\Vert}-z .

The first term is bounded, homogeneous of degree 00, and discontinuous at z=0z=0. The gradient of its component along vv has length π∣sin⁡∠(z,v)∣/∥z∥\pi\vert\sin\angle(z,v)\vert/\Vert z\Vert, whose square is not integrable near z=0z=0 in two dimensions, so ⟨Logx(⋅),v⟩∉H1(S2)\langle\mathrm{Log}_x(\cdot),v\rangle\notin H^1(\mathbb S^2); Proposition E.8 proves this. The L2L_2 cap discrepancy bounds integration errors only for functions in H3/2(S2)H^{3/2}(\mathbb S^2). The introduction states the same fact (lines 080-083).

What is lost is rate, not convergence. On S2\mathbb S^2, point sets whose cap discrepancy has the optimal order N−3/4N^{-3/4} integrate H3/2H^{3/2} functions at that rate. For the Fréchet gradient Theorem 3.3 gives N−3/5N^{-3/5} for the same point sets, and Theorem 3.4 shows that this exponent of the discrepancy cannot be improved and that no equal-weight rule with distinct nodes yiy_i does better than πmax⁡ih(yi)/N\pi\max_ih(y_i)/N uniformly in xx, which is of order 1/N1/N for a positive density hh; for smooth integrands spherical tt-designs with NN of order t2t^2 reach N−τ/2N^{-\tau/2} for every τ\tau.

On "GxG_x is not defined there": it is undefined at the single point y=−xy=-x only, a set of measure zero. Theorem D.6 allows any bounded value at that point, and the gradient-error statements are for x∉−supp QNx\notin-\mathrm{supp}\thinspace Q_N (Theorem 3.4(c), Corollary D.7), so no statement depends on it. So that the two norms cannot be confused, the opening of Section 3 in the revised PDF now names the norm in words: the pointwise length is at most π\pi, and the function-space norm ∥Gx∥H\Vert G_x\Vert_H of Definition 2.1 is infinite for every Sobolev class Hτ(Sd)H^\tau(\mathbb S^d) with τ>d/2\tau\gt d/2, the classes that carry QMC rates.

Response to Reviewer 1Dcm (2/4)

Q10, Q13 (hypotheses of Theorems 3.3 and 3.4).

if you impose that g is supported away from the singularity (at a fixed distance), you are not dealing with behavior near the singularity.

This is not what the theorems assume. In Theorem 3.3(i), zz are normal coordinates centred at the singular point pp, and for the Fréchet gradient at xx the singular point is p=−xp=-x, the cut point itself, not xx. The integrand is b(z)∥z∥mω(z/∥z∥)b(z)\Vert z\Vert^m\omega(z/\Vert z\Vert) plus an HτH^\tau function, with "bb supported in {∥z∥<π/4}\lbrace\Vert z\Vert\lt\pi/4\rbrace": a ball around the singular point. b(0)≠0b(0)\ne0 is allowed and is the case of interest; the lower bound (Theorem D.11) assumes it. The proof of Theorem D.6 removes a ball of radius rr around pp and shrinks it with the discrepancy, r=D1/(τ+d/2)→0r=D^{1/(\tau+d/2)}\to0. Up to constants, the part outside the ball costs rm+d/2−τDr^{m+d/2-\tau}D, and the part inside is at most rmr^m times the mass the rule puts in the ball, which positivity of the weights bounds by rd+rd/2−τDr^d+r^{d/2-\tau}D. No fixed distance enters. Corollary D.7 applies the theorem to GxG_x, and Table 1 lists the resulting exponents; the revised PDF now states its gradient bound in the main text, in the paragraph "The Fréchet gradient on spheres" after Theorem 3.3. Theorem 3.3(i) now names the ball as the ball around the singular point and lists the dependence of its constant as Theorem D.6 does.

On "strong hypotheses that do not hold in practice": Theorem 3.3(i) has four hypotheses, and the Fréchet mean problem on the sphere with a smooth data density satisfies all of them. (1) The manifold is a round sphere, the space of mean directions and normalized embeddings. (2) The error is measured by the worst-case error in Hτ(Sd)H^\tau(\mathbb S^d), τ>d/2\tau\gt d/2, for the uniform measure, and the data density hh enters the loss as a weight (Setting paragraph of Section 2). A known sampling law is the premise of every QMC method, because the points are constructed for it. The theorem asks in addition that its density relative to the uniform measure be smooth, as it is for the von Mises-Fisher laws of Section 5. This covers injected noise and model laws; for data given only as a sample, QMC does not apply at all, with or without a cut locus. (3) The rule has nonnegative weights, as every equal-weight average does. (4) The integrand is bounded with a point singularity of the displayed form and 0≤m<τ−d/20\le m\lt\tau-d/2. The Fréchet-mean gradient satisfies (4) with no further assumption: by Eq. (4), −h(y)⟨Logxy,v⟩-h(y)\langle\mathrm{Log}_xy,v\rangle has m=0m=0, ω(θ)=⟨θ,v⟩\omega(\theta)=\langle\theta,v\rangle and b(0)=−πh(−x)b(0)=-\pi h(-x), uniformly in xx and vv (Corollary D.7). The submission states where the exponent is proved and where it is a prediction (lines 091-093 and 247-249): it is proved for point singularities on spheres and for point and polygonal singularities in the unit square, and it is a prediction for cut loci of codimension one on curved manifolds.

On "too easy": the upper bound has a short proof, and Theorem 3.4 shows that it is nevertheless the right one. Parts (a,b) construct node sets for which no larger exponent holds, so no bound in terms of the discrepancy alone improves the exponents (in the square, the logarithm) of Theorem 3.3; these node sets are built for the purpose and are not sets of optimal order in NN. Part (c) shows that every equal-weight rule with NN distinct nodes has worst-case gradient error at least πmax⁡ih(yi)/N\pi\max_ih(y_i)/N.

Response to Reviewer 1Dcm (3/4)

Q8 (Sturm's contraction).

(a) is Log_x(y) 1-Lipschitz ? Its covariant derivative is the Hessian of squared distance, and that is not bounded !

(a) Yes, in the variable the sentence names. The statement is that y↦⟨Logxy,v⟩y\mapsto\langle\mathrm{Log}_xy,v\rangle is 1-Lipschitz for fixed xx: the variable is the data point yy. The Hessian of the squared distance is the derivative in the other variable, ∇x(−Logxy)=Hessx12d(x,y)2\nabla_x(-\mathrm{Log}_xy)=\mathrm{Hess}_x\tfrac12d(x,y)^2. On a Hadamard manifold it is at least the identity, and it is unbounded when the curvature is bounded above by a negative constant (in hyperbolic space its largest eigenvalue is ℓcoth⁡ℓ\ell\coth\ell with ℓ=d(x,y)\ell=d(x,y)). That concerns the derivative in xx and does not bear on the statement. The derivative in yy is (d Expx)−1(d\thinspace\mathrm{Exp}_x)^{-1}, of norm at most one, because Expx\mathrm{Exp}_x does not decrease distances in nonpositive curvature. Proposition C.21 gives the proof: in the geodesic triangle (x,y,y′)(x,y,y') the angle at xx is at most its Euclidean comparison angle, hence

d(y,y′)2 ≥ ∥Logxy∥2+∥Logxy′∥2−2⟨Logxy,Logxy′⟩=∥Logxy−Logxy′∥2.d(y,y')^2\ \ge\ \Vert\mathrm{Log}_xy\Vert^2+\Vert\mathrm{Log}_xy'\Vert^2-2\langle\mathrm{Log}_xy,\mathrm{Log}_xy'\rangle=\Vert\mathrm{Log}_xy-\mathrm{Log}_xy'\Vert^2 .

(b) The rest of the paragraph uses this constant. With x⋆=b(ν)x^\star=b(\nu), the one-point bound (Proposition 2.2(c), in its appendix form C.7(c) for general probability measures) and 1-strong convexity give d(b(ν),b(ν′))≤∥∫Logx⋆ d(ν′−ν)∥d(b(\nu),b(\nu'))\le\Vert\int\mathrm{Log}_{x^\star}\thinspace d(\nu'-\nu)\Vert. Kantorovich-Rubinstein duality bounds the integral of a 1-Lipschitz function against ν′−ν\nu'-\nu by W1(ν,ν′)W_1(\nu,\nu'); a Lipschitz constant LL would give L W1L\thinspace W_1. The pointer to Theorem E.3 belongs to the last clause of the paragraph: the same inequality with the Hardy-Krause class in place of the Lipschitz class gives the QMC rate of Karcher means. The revised sentence now has the words "for fixed xx" and the pointer to Proposition C.21.

Q1, Q3, Q12 (abstract, introduction, vocabulary). The revised PDF now states the problem in standard terminology in the introduction: the text after Eq. (1) and the paragraph that follows define the point set, the rule and the certificate, display the QMC error bound and the gradient error, and name the three bounds of Section 2. The abstract is kept. It states the problem (what a QMC error bound says about a Riemannian minimizer), the main result (the exponent (m+c)/(c+β)(m+c)/(c+\beta), proved and sharp in the settings it lists) and the scope; one clause now says that the singular integrand is that of squared distances. The paper's own terms ("charge", "potential", "field") are defined in Definition 2.1 and are kept, because on the sphere they are literal: the cap discrepancy is the energy of the charge, and the value and gradient errors are its chordal potential and field at the antipode up to smoother remainders (Proposition 4.1). Each now appears next to its standard name (Section 1; Table 3 in Appendix A). The equations are also in the general response above.

Q4 (prior-work paragraph; "We"). The paragraph is now set out as three labelled lists in the revised PDF: what we use, related work, and what is new. On "We": the first-person plural identifies no one. The PDF lists "Anonymous authors" and contains no names, affiliations or links.

Q6 (Proposition 2.2 is elementary). The submission says the same: "Each part is one line" (line 162) and "The readings of the field are known arguments organized around one object" (lines 118-119). None of it is claimed as new. Its use: part (c) reduces the minimizer error to the integration error of one integrand, Gx⋆G_{x^\star}, which is the quantity Theorems 3.3 and 3.4 bound when it is singular; parts (a) and (b) give the sample complexities of Corollary 2.3. The revised text now states this use before the proposition.

Q7 (Sobol' and Halton sequences). They are the two standard low-discrepancy sequences in [0,1]s[0,1]^s: their first NN points have star discrepancy O(N−1(log⁡N)s)O(N^{-1}(\log N)^s), against order N−1/2N^{-1/2} for random points (Halton, 1960; Sobol', 1967; Dick, Kuo and Sloan, 2013). The submission cites them and states this bound in Appendix B.2. The revised introduction now describes them at first use, with the references, and Corollary 2.3 cites them.

Response to Reviewer 1Dcm (4/4)

Q2 (length). The main text is 9 pages in the submission and 10 in the revision. The other pages are proofs, experimental protocols and a notation table. ICLR sets no limit on appendices and does not require reviewers to read them. We keep the complete proofs, because the contribution consists of theorems and their soundness can be checked only against proofs. Table 2, at the start of the appendix ("How the appendix is organized"), names for each statement of the main text the appendix result that proves it, so one theorem can be checked without reading the rest.

Q5 (Setting). No change was requested; for readers from the optimization side, the revised introduction now contains a short QMC primer (the text after Eq. (1)).

If the pointers under Q8, Q9 and Q10 resolve the soundness concerns, we would appreciate it if you would reconsider the assessment.

Reviewer_N67HOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper studied theoretical guarantees on Riemannian QMC.

Critical strengths

None

Critical weaknesses

My main concern: why is this an important question. For example, nowadays do people still do QMC on Riemannian manifold (in particular for the problem studied in this submitted paper)? I asked because the cited papers are mostly before year 2021. Also I see there is a discussion on related works around line 4580, perhaps they should be in main text.

A following question is, what is the expected reader community? For example, for people familiar with Riemannian optimization but not QMC, they might find it very difficult to read this paper.

Questions for authors

I feel this paper difficult to understand. The difficulty could be due to my lack of knowledge in QMC community. But I feel some writing style of this paper is causing difficulty in understanding. For example, the word "reading" appears very frequent, but I don't know why the author use this word. Also "discrepancy charge", I don't know if this is a commonly used word.

The author can correct me if my understanding is wrong. As I mentioned I am not familiar with QMC and find it difficult to understand this paper. I think it might be better to write in a way that is easier to understand. For example can consider to directly introduce what is QMC through some math equations, and what could be the practical settings that this will he used, and how to formulate the objective under practical settings. Then what are the gradient estimator (mathematically), why such estimation could cause problem.

I know the cut point would introduce some singularity issues, and this might be the core of this paper. But I feel it's difficult to see this point from the main text.

What is a "point set"? A set of a single point? Or a set of many points? I think this should be defined.

I don't understand the sentence "The answer passes through one object." (line 71). Then in this paragraph, many notation come from nowhere: what is a "rule"? a probabilit distribution (maybe it is not)? I think this should be defined.

Maybe Corollary D7, after revise to make it more readable, could be more intuitive for readers.

Again, the difficulty in understanding this submitted paper could be due to my ignorance in the field of QMC, but I think the presentation of this paper itself doesn't meet the standard to be published. I believe it's the author's responsibility to make the presentation clear and in the meanwhile easy to follow. If other reviewers doesn't raise concerns on presentation, then it's likely that the above mentioned presentation issue is my own problem, but if other reviewers also raise such concern, then I believe the paper should be revised for others to understand. Therefore I incline to reject this paper. If the author are able to revise this paper so that presentation become good, then i would reconsider my decision accordingly.

3 comments

Response to Reviewer N67H (1/2)

Thank you for the specific questions. Each is answered below with the definition or equation it asks for, and the corresponding text is now in the revised PDF, in blue.

Q7 (presentation). The introduction of the revised PDF now states the problem in standard terminology, with the QMC background in the main text (the text after Eq. (1) in Section 1) and a table of terms (Table 3, Appendix A). The answers to Q2, Q4 and Q5 below follow the revised text.

Q2 (QMC, the setting, the objective, the gradient estimator, and why there is a problem). In equations:

min⁡x∈Mf(x),f(x)=∫ΩF(x,ω) dP(ω),fN(x)=1N∑i=1NF(x,ωi).\min_{x\in\mathcal M}f(x),\qquad f(x)=\int_\Omega F(x,\omega)\thinspace dP(\omega),\qquad f_N(x)=\frac1N\sum_{i=1}^NF(x,\omega_i).

Here M\mathcal M is a manifold (a sphere for directions, a Grassmannian for subspaces, positive definite matrices for covariances), FF is a loss and ω\omega is a data point or injected noise. One cannot compute ff, so one minimizes fNf_N. Monte Carlo draws ω1,…,ωN\omega_1,\dots,\omega_N independently, and the error of an average is of order N−1/2N^{-1/2}. QMC chooses the points to be more evenly spread than random points, and then for every function gg in a class HH

∣1N∑i=1Ng(ωi)−∫g dP∣ ≤ ∥g∥H⋅DH(ω1,…,ωN),\Big\vert\frac1N\sum_{i=1}^Ng(\omega_i)-\int g\thinspace dP\Big\vert\ \le\ \Vert g\Vert_H\cdot D_H(\omega_1,\dots,\omega_N),

where the discrepancy DHD_H measures how evenly the points are spread. It decays faster than N−1/2N^{-1/2}: like N−1N^{-1} up to logarithms for the star discrepancy in the cube, where this is the Koksma-Hlawka inequality and ∥g∥H\Vert g\Vert_H is the Hardy-Krause variation of gg, and like N−3/4N^{-3/4} for the cap discrepancy on S2\mathbb S^2. The gradient estimator and its error are

grad fN(x)=1N∑i=1NGx(ωi),eN(x)=grad fN(x)−grad f(x),Gx(ω)=gradxF(x,ω).\mathrm{grad}\thinspace f_N(x)=\frac1N\sum_{i=1}^NG_x(\omega_i),\qquad e_N(x)=\mathrm{grad}\thinspace f_N(x)-\mathrm{grad}\thinspace f(x),\qquad G_x(\omega)=\mathrm{grad}_xF(x,\omega).

If fNf_N is μ\mu-strongly geodesically convex, its minimizer x^N\hat x_N and the minimizer x⋆x^\star of ff satisfy μ d(x^N,x⋆)≤∥eN(x⋆)∥\mu\thinspace d(\hat x_N,x^\star)\le\Vert e_N(x^\star)\Vert (Proposition 2.2(c)). So the minimizer inherits the QMC rate if ∥Gx⋆∥H\Vert G_{x^\star}\Vert_H is finite. The problem: for the squared-distance loss F(x,y)=12d(x,y)2F(x,y)=\frac12d(x,y)^2, whose minimizer is the Fréchet mean, Gx(y)=−LogxyG_x(y)=-\mathrm{Log}_xy is discontinuous in yy at the cut locus of xx, and on a sphere ∥Gx∥H=∞\Vert G_x\Vert_H=\infty for the Sobolev classes HH in which QMC rates are known there. The inequality above then gives nothing, and the paper determines what rate remains.

Q3 (the cut point). It is the core of the paper, and the submission shows it in Figure 1(b), in lines 080-083 of the introduction and in Eq. (4). The opening paragraph of Section 3 in the revised PDF now gives the following example, and the paragraph "The Fréchet gradient on spheres" after Theorem 3.3 gives the resulting bound. On a sphere the cut locus of xx is the antipode −x-x. Near it, in normal coordinates zz centred at −x-x, Logx=πz/∥z∥−z\mathrm{Log}_x=\pi z/\Vert z\Vert-z (Eq. (4)): the vector has length close to π\pi, and its direction z/∥z∥z/\Vert z\Vert depends on the side from which the data point approaches −x-x. The integrand is bounded but jumps at one point. For such an integrand the integration error of every equal-weight rule is at most C D2d/(2d+1)C\thinspace D^{2d/(2d+1)} in terms of its L2L_2 cap discrepancy DD on Sd\mathbb S^d, in place of C DC\thinspace D for smooth integrands (Theorem 3.3 with Corollary D.7), and Theorem 3.4 shows that this exponent cannot be improved.

Q4 ("point set"). A set of NN points ω1,…,ωN\omega_1,\dots,\omega_N in Ω\Omega, the sample points at which the loss is averaged. It is the standard QMC term for a deterministic or randomized sample of size NN; the introduction of the submission writes "an average over NN points". The revised text now gives the definition in words after Eq. (1).

Response to Reviewer N67H (2/2)

Q5 ("rule"; "The answer passes through one object"). A (quadrature) rule is the averaging operation QNQ_N that sends a function gg to 1N∑ig(ωi)\frac1N\sum_ig(\omega_i). It can be identified with the probability measure 1N∑iδωi\frac1N\sum_i\delta_{\omega_i}, the empirical distribution of the point set, so your interpretation as a probability distribution is right. The "one object" is the signed measure ρN=QN−P\rho_N=Q_N-P: the integration error of any gg is ρNg\rho_Ng, the objective error is fN(x)−f(x)=ρNF(x,⋅)f_N(x)-f(x)=\rho_NF(x,\cdot), and the gradient error is eN(x)=ρNGxe_N(x)=\rho_NG_x. "Rule" is the standard term of numerical integration (quadrature rule), and the Setting paragraph of Section 2 gives it as QN=1N∑iδωiQ_N=\frac1N\sum_i\delta_{\omega_i}. The revised introduction now states these definitions as well.

Q1 ("reading", "discrepancy charge"). Both are this paper's terms, not standard ones. "Discrepancy charge" names the signed measure ρN=QN−P\rho_N=Q_N-P and is defined in Definition 2.1. The name records a fact: on the sphere the error of the Fréchet objective is −π-\pi times the chordal-distance potential of ρN\rho_N at the antipode plus a smoother remainder, and the cap discrepancy is exactly the energy of ρN\rho_N (Proposition 4.1). "Reading" names each of the three bounds on the gradient error in Proposition 2.2. In the revised PDF Proposition 2.2 is now titled "three bounds on the gradient error, or readings of the field", the sentence before it now reads "three known bounds on the gradient error, which we call readings of the field", and Table 3 (Appendix A) lists each of the paper's names next to its standard meaning.

Q6 (Corollary D.7). It is the most concrete statement of the main result, and the revised PDF now states its gradient bound in the main text (paragraph "The Fréchet gradient on spheres" after Theorem 3.3). In words: For the Fréchet objective on Sd\mathbb S^d with a smooth data density hh, every equal-weight rule satisfies sup⁡x∥eN(x)∥≤C D2d/(2d+1)\sup_x\Vert e_N(x)\Vert\le C\thinspace D^{2d/(2d+1)}, where DD is its L2L_2 cap discrepancy and the supremum is over xx whose antipode is not a sample point. The exponent cannot be improved (Theorem 3.4(a)), and for every equal-weight rule with NN distinct nodes yiy_i this supremum is at least πmax⁡ih(yi)/N\pi\max_ih(y_i)/N (Theorem 3.4(c)).

W1 (importance; dates of the references; related work). The question is current. 89 of the 101 references predate 2021 because the tools are classical (Koksma-Hlawka, Stolarsky's invariance principle, Sturm's barycentres), not because the problem is old. QMC inside stochastic optimization and inference: Buchholz et al. (ICML 2018), Liu and Owen (JMLR 2021), Lu et al. (ICLR 2022), Liu (NeurIPS 2023), Melnikov and Milz (JOTA 2025), all cited. QMC point sets on spheres are used in recent ICLR papers, which the revision now cites: Nguyen, Bariletto and Ho (ICLR 2024) for sliced Wasserstein distances and Hertrich, Jahn and Quellmalz (ICLR 2025) for kernel summation. QMC samples for an objective whose parameter lives on a manifold are used by Luu et al. (ICLR 2025), cited in the first paragraph: scrambled Sobol' minibatches for Gaussian variational inference on the Bures-Wasserstein manifold, as an empirical baseline. To our knowledge that use has not been analysed, and the submission supplies the analysis. A concrete case is a mean direction on a sphere estimated from data with full support, such as a von Mises-Fisher law. There the paper's prescription is to centre the polar chart of the point set at an estimate of the minimizer, and in the experiment of Section 5 this changes the fitted slope of the Fréchet-mean error from −1.01-1.01 to −1.55-1.55; the separation is measured, and the proved bounds for the two charts differ by a logarithm. On placement: the introduction has a "Relation to prior work" paragraph and Appendix H the longer discussion; the comparison with the works above is now in the main text of the revised PDF ("Relation to prior work", Section 1).

W2 (expected readers). Readers who optimize sampled objectives on manifolds (Riemannian optimization, geometric statistics, variational inference) and want to know what replacing random samples by QMC points buys; and QMC readers interested in integrands with singularities. For the first group the QMC background must be in the main text, and the primer under Q2 is now there (the text after Eq. (1) in Section 1).

If the revised text resolves the presentation concern, we would appreciate it if you would reconsider the rating.

Again by reading the paper itself I can't understand it.

There are too many abstract definitions and too many notations. For example the author use eNe_N to denote gradient of Φ\Phi. I think this way of introducing too many notations would make it very difficult for human reader to follow.

And the way theorems are stated makes it impossible for non-experts to understand, as another reviewer pointed out.

As the main text is of limited length, I think it's better to highlight some of the main results and try to explain it in plain English so that people can understand what it is doing. The rest can go to Appendix.

Reviewer_XVA6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper purports to study quasi-Monte Carlo (QMC) methods, when the function being estimated FF takes as input (p,x)(p, x) where pp lives on a manifold and xx is in a probability space. Their main results appear to relate to the optimization landscape over pp of the gap between the QMC rule and the full integral, and in particular case where FF is the squared distance on the sphere.

Critical strengths

None.

Critical weaknesses
  • The paper is written in a way that is nearly impossible to read. Not only is the prose frequently strangely formed, but it is not clear from reading the introduction (or indeed anywhere else that I can tell) what the actual goal of the paper is. For example, in the introduction, when discussing Section 2, it says "reusing one point set along an optimization path, refreshing it at every step, and solving once read the field in three norms, which is where their sample complexities come from; and that whatever a certificate says about the field, it says through one tangent-vector-valued integrand, the gradient Gx=gradxF(x,⋅)G_x = \mathrm{grad}_x F(x, \cdot) of the loss in the decision variable". I looked through section 2 but I still cannot be sure what the authors are saying.
  • To the limited extent to which I have been able to interpret them, their results do not appear to be significantly novel: for example their Proposition 2.2(b) is immediate from their definitions, and part (c) seems to be a standard consequence of strong convexity. For some reason, they call their Theorem 3.3 a "scaling law", but it is nonetheless only studying results for very particular functions on the sphere or unit square.
Questions for authors
  • Could you please write, in a short paragraph and using standard mathematical terminology the precise contribution of this paper and how it relates to previous work?
2 comments

Response to Reviewer XVA6 (1/2)

Thank you for the review. The paragraph requested in Q1 comes first, since it also answers W1; W2 follows. A revised PDF is uploaded, with changes in blue; line numbers below refer to the submitted PDF.

Q1 (contribution and relation to previous work, in standard terminology).

Problem. Let f(x)=∫F(x,ω) dP(ω)f(x)=\int F(x,\omega)\thinspace dP(\omega) on a complete Riemannian manifold and fN(x)=1N∑i≤NF(x,ωi)f_N(x)=\frac1N\sum_{i\le N}F(x,\omega_i) for a QMC point set, with minimizers x⋆x^\star and x^N\hat x_N. QMC theory bounds integration errors by ∥g∥HDH\Vert g\Vert_H D_H for gg in a function space HH, where DH=O(N−1(log⁡N)s)D_H=O(N^{-1}(\log N)^s) for the star discrepancy of low-discrepancy sequences in [0,1]s[0,1]^s and, at best, DH=O(N−τ/d)D_H=O(N^{-\tau/d}) for the worst-case error in Hτ(Sd)H^\tau(\mathbb S^d). Which rate does d(x^N,x⋆)d(\hat x_N,x^\star) inherit?

Results. (1) If fNf_N is μ\mu-strongly geodesically convex, then μ d(x^N,x⋆)≤∥grad fN(x⋆)−grad f(x⋆)∥\mu\thinspace d(\hat x_N,x^\star)\le\Vert\mathrm{grad}\thinspace f_N(x^\star)-\mathrm{grad}\thinspace f(x^\star)\Vert, the integration error of one integrand, Gx⋆=gradxF(x⋆,⋅)G_{x^\star}=\mathrm{grad}_xF(x^\star,\cdot), with a matching lower bound when fNf_N is also smooth near the minimizer (Proposition 2.2; known arguments, not claimed as new). (2) For Fréchet means on Sd\mathbb S^d, Gx=−LogxG_x=-\mathrm{Log}_x jumps at the antipode of xx and lies in no Hτ(Sd)H^\tau(\mathbb S^d) with τ>d/2\tau\gt d/2, so (1) gives no rate. Theorem 3.3: let τ>d/2\tau\gt d/2 and 0≤m<τ−d/20\le m\lt\tau-d/2; for rules with nonnegative weights and bounded integrands that near one point equal a smooth cutoff times a function homogeneous of degree mm, plus an HτH^\tau function, the error is O(D(m+d)/(τ+d/2))O(D^{(m+d)/(\tau+d/2)}) in the worst-case error DD for Hτ(Sd)H^\tau(\mathbb S^d). For the Fréchet gradient and the L2L_2 cap discrepancy this is O(D2d/(2d+1))O(D^{2d/(2d+1)}). In the unit square with the star discrepancy D∗D^\ast the error is O(D∗log⁡(1/D∗))O(D^\ast\log(1/D^\ast)) at a point singularity of degree 00 and O((D∗)1/2)O((D^\ast)^{1/2}) for jumps across edges of convex polygons. Theorem 3.4 gives node sets for which no larger exponent holds, and shows that every equal-weight rule with NN distinct nodes yiy_i has worst-case gradient error at least πmax⁡ih(yi)/N\pi\max_ih(y_i)/N for a data density hh. Corollary 3.5: for d≥2d\ge2, spherical tt-designs with NN of order tdt^d reach N−1+ϵN^{-1+\epsilon}. (3) On Sd\mathbb S^d, 12d(x,y)2=π22−π∥x+y∥\frac12d(x,y)^2=\frac{\pi^2}2-\pi\Vert x+y\Vert plus a smoother term, which expresses the Fréchet value error through the chordal-distance potential of Stolarsky's invariance principle (Proposition 4.1). For QMC points mapped to S2\mathbb S^2 by a polar chart, the singularity disappears when the pole of the chart is at xx (Proposition 4.2).

Previous work. QMC sample averages and QMC-SGD have been analysed for Euclidean parameters and smooth integrands (Koivu, 2005; Homem-de-Mello, 2008; Buchholz et al., 2018; Liu and Owen, 2021). QMC on spheres bounds the error for integrands in HτH^\tau (Brauchart, Saff, Sloan and Womersley, 2014), a class that excludes Logx\mathrm{Log}_x. The nearest bounds we know of are three. Owen (2006) treats unbounded point singularities ∥x−z∥−A\Vert x-z\Vert^{-A} in the cube: the expected error of randomized QMC is O(n(−1+ϵ)(d−A)/d)O(n^{(-1+\epsilon)(d-A)/d}), obtained by cutting the integrand at a radius balanced against the discrepancy. This is the exponent (m+c)/(c+β)(m+c)/(c+\beta) at m=−Am=-A, c=dc=d, β=0\beta=0. Theorem 3.3 treats bounded singularities, where positivity of the weights gives a bound for every rule and not in expectation, in Sobolev norms on spheres, with lower bounds. Owen (2006) is cited in the submission, and the revised PDF now gives this comparison in the related-work paragraph of Section 1 and in Appendix H. Second, the interpolation inequality wce(Hs′)≤c wce(Hs)s′/s\mathrm{wce}(H^{s'})\le c\thinspace\mathrm{wce}(H^{s})^{s'/s} for d/2<s′<sd/2\lt s'\lt s (Lemma 26 of Brauchart et al., after Brandolini et al.) needs continuous integrands, so it does not reach the gradient (m=0m=0); for m>0m\gt0 it gives exponents below (m+d/2)/τ(m+d/2)/\tau, which is smaller than (m+d)/(τ+d/2)(m+d)/(\tau+d/2) whenever m<τ−d/2m\lt\tau-d/2. The difference is that Theorem 3.3 approximates the integrand from above and below within HτH^\tau and does not interpolate between Sobolev norms. The second comparison is now there as well. Third, the exponent 1/21/2 for polygonal jumps in Theorem 3.3(ii) is the classical bound of the isotropic discrepancy by the star discrepancy (Kuipers and Niederreiter, 1974), as the submission notes.

Response to Reviewer XVA6 (2/2)

W1 (readability; the goal of the paper). The goal is stated after Eq. (1): to determine what a QMC error bound implies for the minimizer of the sampled objective. The sentence quoted in the review summarizes Section 2 in the terms that Section 2 introduces (Setting, Definition 2.1, Proposition 2.2). In standard terms it says: Section 2 gives three bounds on the gradient error eN(x)=grad fN(x)−grad f(x)e_N(x)=\mathrm{grad}\thinspace f_N(x)-\mathrm{grad}\thinspace f(x). A bound uniform in xx, sup⁡x∥eN(x)∥≤sup⁡x∥Gx∥HDH\sup_x\Vert e_N(x)\Vert\le\sup_x\Vert G_x\Vert_H D_H, is what gradient descent on one fixed point set needs, because its iterates depend on the points. A mean-square bound at fixed xx is what SGD with a fresh randomized point set per step needs. A bound at x⋆x^\star alone is what the sample minimizer needs. In all three the only function that must be integrated accurately is ω↦Gx(ω)\omega\mapsto G_x(\omega). The introduction of the revised PDF now says this in formulas, after Eq. (1) and in the paragraph that follows; the general response gives the equations and a table of terms.

W2 (novelty; "scaling law").

their Proposition 2.2(b) is immediate from their definitions, and part (c) seems to be a standard consequence of strong convexity

The submission says the same: "Each part is one line" (line 162) and "The readings of the field are known arguments organized around one object. New is the analysis at the cut locus" (lines 118-119). Proposition 2.2 is the reduction that tells which integrand must be controlled. The results claimed as new are Theorem 3.3 except the exponent 1/21/2 for polygonal jumps, Theorem 3.4, Corollary 3.5 and Propositions 4.1 and 4.2. The 1/N1/N bound and the identity of Proposition 4.1 have elementary proofs; what is claimed is what they imply, that no point set escapes 1/N1/N and that the Fréchet loss on a sphere has no singularity beyond the chordal distance to the antipode. Theorem 3.4(a) is the least standard part: the lower-bound measure is a multipole of width rr whose moments vanish below an order above τ−d/2\tau-d/2, so that its discrepancy is of order rτ+d/2r^{\tau+d/2} while it moves the integral by order rm+dr^{m+d}.

only studying results for very particular functions on the sphere or unit square

The domains are as the review says, and the submission states the limit: "We prove the law for point singularities on spheres [...] and for point singularities and polygonal jumps in the square [...]; for cut loci of other codimensions on other manifolds it is a prediction" (lines 091-093). Within these domains the theorem covers a class of integrands, not particular functions: on Sd\mathbb S^d, every bounded gg that near a point equals b(z)∥z∥mω(z/∥z∥)b(z)\Vert z\Vert^m\omega(z/\Vert z\Vert) plus an HτH^\tau function, for every degree 0≤m<τ−d/20\le m\lt\tau-d/2, every smooth angular profile ω\omega and cutoff bb, every τ>d/2\tau\gt d/2 and every dimension d≥1d\ge1, with a constant independent of the location of the singularity. This class contains the gradient of the Fréchet objective at every xx for every smooth data density (Corollary D.7), and its value when τ>d/2+1\tau\gt d/2+1. The name "scaling law" refers to the exponent min⁡{1,(m+c)/(c+β)}\min\lbrace1,(m+c)/(c+\beta)\rbrace, which every row of Table 1 with a singular set instantiates and which comes from a dilation argument (the paragraph "Why this exponent"); the abstract says in which settings it is proved.

Soundness. No statement or proof step is named as incorrect in the review. Every statement of the main text has a complete proof in the appendix, and Table 2 ("How the appendix is organized", at the start of the appendix) lists where; if a specific step is in doubt, we will address it.

Comments on the paper

2 comments

General Response (1/2)

We thank the three reviewers. This comment gives the problem and the results in standard terminology, answers the three technical objections, and states what is new. A revised PDF is uploaded, with changed text in blue; line numbers below refer to the submitted PDF.

Strengths stated in the reviews. Reviewer 1Dcm: the paper "raises an important issue, and attempts to address it in a rigorous mathematical way", and the main problem is "quite interesting, and also relevant to applied problems".

1. The problem and the results in standard terminology; the introduction of the revised PDF now states them this way. (Reviewer 1Dcm Q1, Reviewer 1Dcm Q3, Reviewer 1Dcm Q12, Reviewer N67H Q7, Reviewer N67H W2, Reviewer N67H Q1, Reviewer N67H Q2, Reviewer N67H Q4, Reviewer N67H Q5, Reviewer XVA6 W1, Reviewer XVA6 Q1) All three reviews ask for the problem and the results in standard terms. The problem is

min⁡x∈Mf(x),f(x)=∫ΩF(x,ω) dP(ω),fN(x)=1N∑i=1NF(x,ωi),\min_{x\in\mathcal M}f(x),\quad f(x)=\int_\Omega F(x,\omega)\thinspace dP(\omega),\quad f_N(x)=\frac1N\sum_{i=1}^NF(x,\omega_i),

∣1N∑i=1Ng(ωi)−∫g dP∣≤∥g∥H DH(ω1,…,ωN)for g∈H,\Big\vert\frac1N\sum_{i=1}^Ng(\omega_i)-\int g\thinspace dP\Big\vert\le\Vert g\Vert_H\thinspace D_H(\omega_1,\dots,\omega_N)\quad\text{for }g\in H,

eN(x)=grad fN(x)−grad f(x)=1N∑i=1NGx(ωi)−∫Gx dP,Gx(ω)=gradxF(x,ω).e_N(x)=\mathrm{grad}\thinspace f_N(x)-\mathrm{grad}\thinspace f(x)=\frac1N\sum_{i=1}^NG_x(\omega_i)-\int G_x\thinspace dP,\qquad G_x(\omega)=\mathrm{grad}_xF(x,\omega).

The second line is the QMC error bound. Its discrepancy DHD_H decays faster than N−1/2N^{-1/2}: like N−1N^{-1} up to logarithms for the star discrepancy in the cube (the Koksma-Hlawka inequality), and at best like N−τ/dN^{-\tau/d} for the worst-case error in Hτ(Sd)H^\tau(\mathbb S^d). If fNf_N is μ\mu-strongly geodesically convex, then μ d(x^N,x⋆)≤∥eN(x⋆)∥\mu\thinspace d(\hat x_N,x^\star)\le\Vert e_N(x^\star)\Vert for the minimizers x^N\hat x_N of fNf_N and x⋆x^\star of ff, so the minimizer inherits the QMC rate when ∥Gx⋆∥H\Vert G_{x^\star}\Vert_H is finite. For the Fréchet mean on a sphere, F=12d2F=\frac12d^2 and Gx=−LogxG_x=-\mathrm{Log}_x jumps at the antipode of xx; this norm is infinite for the Sobolev classes that carry QMC rates, and the paper determines the rate that remains.

The terms of the submission are introduced in Sections 2 and 3 (Setting, Definition 2.1, Proposition 2.2, Lemma 3.1, Theorem 3.4); "point set" and "rule" are the standard QMC terms. They correspond to standard ones as follows.

Term in the submissionStandard meaning
point setthe NN sample points ω1,…,ωN\omega_1,\dots,\omega_N
rule QNQ_Nthe equal-weight average g↦1N∑ig(ωi)g\mapsto\frac1N\sum_ig(\omega_i), i.e. the empirical measure
certificatea function class HH with the error bound of the second display
discrepancy charge ρN\rho_Nthe signed measure QN−PQ_N-P
potential ΦN\Phi_Nthe objective error fN−ff_N-f
field eNe_Nthe gradient error grad fN−grad f\mathrm{grad}\thinspace f_N-\mathrm{grad}\thinspace f
field integrand GxG_xthe loss gradient ω↦gradxF(x,ω)\omega\mapsto\mathrm{grad}_xF(x,\omega)
three readingsthree bounds on eNe_N: uniform in xx, mean-square at fixed xx, at x⋆x^\star only
sandwichone-sided approximation by a majorant and a minorant in HH
floora lower bound valid for every equal-weight NN-point rule

2. The singularity at the cut locus is real and is what Section 3 treats; each of the three technical objections is answered by a statement proved in the submission. (Reviewer 1Dcm Q9, Reviewer 1Dcm Q10, Reviewer 1Dcm Q13, Reviewer 1Dcm Q8, Reviewer N67H Q3) No proof, rate, exponent or hypothesis changes; the main-text statement of Theorem 3.3(i) now lists the dependence of its constant as the appendix statement (Theorem D.6) does. The reply to Reviewer 1Dcm gives each statement, its location and a derivation; in short:

  • "its norm remains bounded near the cut locus". The pointwise length is bounded, ∥Logxy∥≤π\Vert\mathrm{Log}_xy\Vert\le\pi. The norm in the paper's sentence is the function-space norm ∥Gx∥H\Vert G_x\Vert_H of Definition 2.1, which the error bound requires, and on the sphere it is infinite: the components of Logx\mathrm{Log}_x are not in H1(S2)H^1(\mathbb S^2) (Proposition E.8), and the cap discrepancy controls H3/2H^{3/2}.
  • "gg is supported away from the singularity". In Theorem 3.3 the cutoff bb is supported in a ball around the singular point, with b(0)≠0b(0)\ne0 allowed; for the Fréchet gradient at xx the singular point is −x-x (Corollary D.7), and the proof shrinks the excised ball with the discrepancy.
  • "is Log_x(y) 1-Lipschitz ?" Yes, as a function of yy for fixed xx on a Hadamard manifold (Proposition C.21, by angle comparison). The Hessian of the squared distance is the derivative in xx.

General Response (2/2)

3. What is claimed as new, and its scope. (Reviewer XVA6 W2, Reviewer 1Dcm Q6, Reviewer N67H W1) Proposition 2.2 collects known arguments, and Corollary 2.3 follows from it by standard descent and stochastic-approximation arguments; the submission says so (lines 118-119, 162 and 171). They identify the integrand that must be controlled. The results claimed as new are: the error bound O(D(m+d)/(τ+d/2))O(D^{(m+d)/(\tau+d/2)}) for rules with nonnegative weights and integrands with a point singularity of degree mm on Sd\mathbb S^d, and its analogue for point singularities in the unit square (Theorem 3.3; the exponent 1/21/2 for polygonal jumps is classical); node sets for which no larger exponent holds, and the lower bound πmax⁡ih(yi)/N\pi\max_ih(y_i)/N on the worst-case gradient error of every equal-weight rule with NN distinct nodes (Theorem 3.4); for d≥2d\ge2, the rate N−1+ϵN^{-1+\epsilon} for spherical tt-designs with NN of order tdt^d (Corollary 3.5); the reduction of the Fréchet value error on spheres to the chordal-distance potential of Stolarsky's invariance principle (Proposition 4.1); and the removal of the singularity by a polar chart centred at the minimizer (Proposition 4.2). The nearest earlier bounds we know of are Owen (2006), for unbounded point singularities in the cube and randomized QMC in expectation, and the interpolation inequality between Sobolev worst-case errors on the sphere (Brauchart et al., 2014, Lemma 26); the reply to Reviewer XVA6 gives the comparison with both, and it is now in the revised PDF (Section 1 and Appendix H).

Scope. The exponent (m+c)/(c+β)(m+c)/(c+\beta) is proved for point singularities on spheres in Sobolev norms, for a data law with a smooth density relative to the uniform measure, and for point and polygonal singularities in the unit square in Hardy-Krause variation. For cut loci of codimension one on curved manifolds it is a prediction, and the separation between charts measured in Section 5 exceeds what is proved. The submission states both (lines 091-093, 110-111).

Revisions (in blue in the revised PDF):

ChangeWhereRaised by
Abstract: one clause made precise ("for squared distances")abstractReviewer 1Dcm Q1
Point set, rule and certificate defined; QMC error bound and gradient error displayed; the three bounds named; QMC primerSection 1, after Eq. (1)Reviewer 1Dcm Q1, Reviewer 1Dcm Q3, Reviewer N67H Q2, Reviewer N67H Q7, Reviewer N67H W2, Reviewer XVA6 W1, Reviewer XVA6 Q1
"point set" and "rule" defined in the introduction; each term of the paper next to its standard name; table of terms; Proposition 2.2 titled "three bounds on the gradient error"Sections 1-2; Table 3 (Appendix A)Reviewer 1Dcm Q12, Reviewer N67H Q1, Reviewer N67H Q4, Reviewer N67H Q5
Prior work as three lists (used; related; new), with the comparison to Owen (2006) and to the interpolation inequality, and recent work on QMC in optimization and on spheres (two references added)Section 1, Appendices B.2 and HReviewer 1Dcm Q4, Reviewer N67H W1, Reviewer XVA6 W2
Use of Proposition 2.2 stated before it; Sobol' and Halton described and cited at first useSections 1-2Reviewer 1Dcm Q6, Reviewer 1Dcm Q7
Sturm paragraph: "for fixed xx" and the pointer to Proposition C.21Section 2Reviewer 1Dcm Q8
Opening of Section 3: the norm named in words (function-space norm, infinite in Hτ(Sd)H^\tau(\mathbb S^d) for τ>d/2\tau\gt d/2; pointwise length at most π\pi, also stated in the introduction); the value at the cut point; that the joint continuity assumed in Section 2 is not used in Section 3 (Appendix D); in "Why this exponent", that the radius is chosen at the endSections 1 and 3Reviewer 1Dcm Q9, Reviewer N67H Q3
Theorem 3.3(i): the ball named as the ball around the singular point, b(0)≠0b(0)\ne0 allowed; the dependence of the constant listed as in Theorem D.6Section 3Reviewer 1Dcm Q10, Reviewer 1Dcm Q13
Gradient bound of Corollary D.7 stated in the main text (paragraph "The Fréchet gradient on spheres")Section 3Reviewer N67H Q6
Main result stated for the Fréchet mean on Sd\mathbb S^d (upper bound C D2d/(2d+1)C\thinspace D^{2d/(2d+1)}, sharpness, lower bound πmax⁡ih(yi)/N\pi\max_ih(y_i)/N)Section 1Reviewer 1Dcm Q1, Reviewer N67H Q6, Reviewer XVA6 Q1

Pages. The main text of the revised PDF is 10 pages, the limit for the revision; the submission had 9.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.