acceptodds
Under review as a conference paper at ICLR 2027

Restarting Heavy Ball: Sharp Intervals for Momentum Retention

Abstract

Resetting momentum too frequently can destabilize a convergent heavy-ball method. We characterize restart intervals over the complete scalar class with spectral radius at most . Building on the known repeated-root peak effect, we show that full reset is uniformly stable exactly when . Retaining a fixed fraction of momentum changes the exact answer: the optimal retention tends to , and we determine the restart interval through its constant-order asymptotic term. Exact rational exclusions and independently checked interval certificates establish exact results at finite margins: at , partial retention permits shorter optimal common intervals than full reset. A secondary gradient-only checkpoint construction preserves an accelerated exponent on fixed quadratics. The experiments separate rejected work from delays in checking the stopping target: observed-target stopping removes all gradient-query overhead when the proposal equals the fallback. A delayed-gradient handoff illustrates a distinct use for velocity control: endpoint loss can be small while a subsequent position limit is violated. Controlled optimization experiments find no general speed advantage from joint-energy certification. The main contribution is a sharp restart design rule, with the scope and costs of additional certificates made explicit.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
heavy-ball method, momentum restart, stability analysis, partial momentum retention, accelerated optimization

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_XTqhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies heavy-ball momentum on quadratics when the step size and momentum may change every hh steps. Those changes are chosen by an adversary among all settings that converge at rate at least ss. At every change, a fixed fraction rr of the velocity is kept. The question the authors address is, what is the smallest hh for which the method converges for every sequence of changes. For a full reset (r=0r = 0), Theorem 1 gives an exact condition, sh[1+h(1+s)]<1s^h[1 + h(1+s)] < 1. Theorem 2 shows that keeping part of the velocity allows a shorter hh, and that as s→1s \to 1 the best fraction rr tends to (3−5)/2(3-\sqrt{5})/2.

Critical strengths

1. The result on how much velocity to keep is interesting.

Erasing the velocity at every switch and keeping all of it fail for opposite reasons. The first makes the method overshoot again in every block, while the second lets alternating rules build up the error. The paper shows that in the limit these are the only two ways the method can fail, and that keeping the fixed fraction (3−5)/2(3-\sqrt{5})/2 balances them for every margin. It is a simple and elegant answer to a problem that looks complicated at first.

2. The results appear correct.

I checked the proofs of the main results and found no errors.

Critical weaknesses

1. The paper is very hard to read

  • Title and abstract. They present the work as a study of restarting, but the main result depends on the optimizer's settings changing adversarially at every block boundary. The abstract never says this, and it only becomes clear in the introduction and Section 2.
  • Writing. Many sentences are hard to follow, for example "Exact rational exclusions and independently checked interval certificates establish exact results at finite margins." Such repeated disclaimers keep interrupting the argument and make it necessary to read the same sentence over and over again to understand it.
  • Unexplained terms and experiments. Terms such as "commit length," "contract," and "handoff" are not explained when first used. Section 7.3 squeezes four studies into a few paragraphs, relies on objects defined only in Appendix C, and includes a "two-scale proposal-control" experiment that is not described anywhere.
  • No figure of the mechanism. The idea is easy to picture, so I was expecting some accompanying figures. Erasing the velocity makes the fast oscillating rule overshoot again in every block, while keeping it lets a fast and a slow rule take turns building up the error. A plot of a single mode in these cases would make Sections 3–4 much easier to follow than the stability inequalities alone.

The paper needs a major rewrite. These are only a few examples of writing issues.

2. Limited scope.

As mentioned in the first point, the paper's setting is different than classical restarts in accelerated optimization, so I failed to understand why the shortest stable block length matters or when this setting arises. With a fixed rule, keeping all the velocity is already stable, so the trade-off exists only because the rule changes. The worst case also needs the step size to jump by a factor of κ\kappa (under the paper's Polyak parameters) between blocks. If all rules must converge on the same range of curvatures [μ,L][\mu, L], this worst case cannot happen, so it is unclear how much of the result carries over to changing the optimizer on a given problem.

Within the model, keeping part of the velocity only improves the constant in the formula. Block lengths shrink by about 10% and that saving gets smaller as κ\kappa grows. While the goal of the paper is to establish stability, I would still have liked some intuition on what happens to the convergence rate. Right now, the scope of the paper is limited.

Finally, the paper should compare more concretely with dwell-time methods for switched systems. They are only mentioned in the third paragraph of the intro. Perhaps, a better positioning against that literature would also allow to effectively show why this setting is significant.

Questions for authors
  1. Is there a practical setting where the optimizer's settings change as much as the worst case requires?
  2. How do the results change if the kept fraction can depend on the rules before and after the switch, does that extra information buy any extra gains?
  3. Does the derived block length and kept fraction lead to faster convergence than a full reset under any switching schedule? Not necessarily a theoretical argument, even an empirical experiment would be interesting.
2 comments
AuthorsOpenReview ↗

Response to Reviewer XTqh (1/2)

Thank you for checking the proofs and for the concrete suggestions. We agree with the criticism of the presentation.

Title and abstract. The abstract now opens with the switching model: the rule (step size and momentum) may change every hh steps, a fraction rr of the velocity is kept at each change, and the guarantee covers every rule sequence with spectral radius at most s<1s<1. Because the retention result needs rules that change between blocks, we propose the title "Restarting Heavy Ball under Changing Hyperparameters: Sharp Intervals for Momentum Retention". The revised PDF keeps the submitted title until the change is approved.

Writing. The sentence you quote is now a plain statement of the exact intervals. We removed the framed scope box and many repeated caveats, gathered the main scope statements in one paragraph of Section 2, and made Sections 3, 4, 5 and 7 open by saying what they show.

Terms. "Commit length" is now "block length", the number hh of steps between rule changes. "Contract" (after Theorem 3) is the guaranteed decrease W+≤qhWW^+\le q^hW at each completed checkpoint, where W=∥g∥2+L2∥v∥2W=\lVert g\rVert^2+L^2\lVert v\rVert^2 and q<1q<1 is fixed. "Handoff" (contributions) means that an executor keeps moving with the last velocity for KK steps until the next gradient arrives.

Section 7.3. It now opens with its question and answer: certifying a joint position-velocity energy did not help the optimizer in any of four controls. Three labelled paragraphs then define the objects in place, including the two-scale proposal control, an evolution strategy whose two mutation scales are frozen or adapted. Appendix O specifies these experiments.

Mechanism figure. Figure 1 now shows one eigenmode at 1−s=0.051-s=0.05 and h=93h=93. In panel (a), the fastest allowed rule A−A_- runs in every block, and each full reset re-excites its transient. In panel (b), A−A_- alternates with the slowest rule A+A_+, and full retention builds up the error. The certified r=0.3727r=0.3727 contracts in both.

Q1: does the worst case arise in practice? For full reset, it needs no rule change at all. With Polyak's parameters, the heavy-ball matrix of the top eigenmode is exactly A−A_-, so a single well-tuned rule restarted every hh steps diverges whenever h<hreseth<h_{\rm reset}, the full-reset threshold. On a 200-dimensional quadratic, the measured growth per block matches Ds(h)=sh[1+h(1+s)]D_s(h)=s^h[1+h(1+s)] with relative error below 10−1010^{-10} (Section 7.5). For partial retention, the worst case needs the step in one mode to change by the factor κs=((1+s)/(1−s))2\kappa_s=((1+s)/(1-s))^2, which is 1521 at s=0.95s=0.95. With a scalar step and momentum on a quadratic with condition number κ=L/μ\kappa=L/\mu, this requires κ=1\kappa=1, whereas steps that differ between eigendirections, such as coordinate-wise steps in an eigenbasis, also allow it (Proposition 4(iii), Appendix H.1). Schedules, warm restarts, population-based search and learned optimizers change rules, but we do not claim that they select the worst case (Appendix H.3).

Weakness 2: rules that converge on the same [μ,L][\mu,L]. You are right for retention when κ>1\kappa>1, because its worst case needs the step ratio κs\kappa_s within one mode, but the full-reset worst case needs only one rule and still occurs. The retention designs of Table 1 therefore stay valid for every κ\kappa and optimal for κ\kappa near 1, but may be conservative for large κ\kappa. In uncertified grid estimates at s=0.95s=0.95 (Table 6), full reset still needs 104 steps for every κ\kappa, while the best of six tested retentions needs about 72 steps at κ=10\kappa=10 (with r=0.8r=0.8), and 54 at κ=100\kappa=100 and 40 at κ=1000\kappa=1000, both with full retention.

AuthorsOpenReview ↗

Response to Reviewer XTqh (2/2)

Size of the gain and the rate. We agree that the interval gain is a constant-order term. Retention shortens the interval from 104 to 93 steps at 1−s=0.051-s=0.05, about 10%, and from 320 to 292 and from 724 to 669 steps at 1−s=0.021-s=0.02 and 0.010.01 (Table 1). At a fixed block length, it also reduces the worst-case per-block factor by a constant ratio. As s→1s\to1 with hsh→chs^h\to c, this factor tends to 2cmax⁡{1−r,r}2c\max\lbrace1-r,\sqrt r\rbrace (Corollary 5, Appendix I). Hence r⋆=(3−5)/2r_\star=(3-\sqrt5)/2 improves on full reset by the factor 1/φ≈0.621/\varphi\approx0.62, where φ\varphi is the golden ratio. At s=0.95s=0.95, the certified per-block factor for r⋆r_\star is 0.64 times that of full reset at every tested hh from 104 to 300 (Table 8). Per step, the gain is only φ−1/h\varphi^{-1/h}. To reduce the worst-case bound by 10610^6 at h=120h=120, r⋆r_\star needs about 1,459 steps and full reset 2,383, but at h=300h=300 they need 438 and 460. Retention therefore matters most near the stability threshold and when hh is fixed externally.

Dwell-time methods. In the new Appendix K, a dwell-time bound in the Euclidean norm, which needs ∥Ah∥2<1\lVert A^h\rVert_2<1 for every allowed rule AA, requires at least 126, 373 and 829 steps at the three margins. At our retentions, a numerical search finds no common quadratic Lyapunov function for blocks shorter than 96, 300 and 686 steps. These are 3, 8 and 17 steps more than our exact optima of 93, 292 and 669, although this search is not certified. Our analysis also proves that blocks below these optima fail, and it gives the exact no-reset intervals 105, 320 and 725. Mode-dependent conditions in the style of Geromel & Colaneri (2006) use one Lyapunov function per rule. Our certificate for rule-dependent retention (Appendix J) is of this kind, with one norm per sign of the rule's trace, but we did not compute one for constant retention. Our certified designs also contract every longer block.

Q2: retention that depends on both rules. This information helps, but it can at most halve the leading term of the interval (Proposition 6, Appendix J), because every such policy needs Ds(h)⋅sh[1+h(1−s)]<1D_s(h)\cdot s^h[1+h(1-s)]<1, which means h≥61h\ge61, 185 and 416 at the three margins. On the complete class of rules, we certify the following policy at h=68h=68, 201 and 461: keep the velocity when consecutive rules have traces of the same sign, and reset it otherwise. This is 27 to 31% shorter than the best constant retention and within 12% of the bound. It acts on each mode separately, so it needs the eigenbasis and is an oracle policy.

Q3: is it faster than full reset? Only in the worst case, where it has a better rate and is stable at shorter blocks: the certified r=0.3727r=0.3727 converges for every h≥93h\ge93, while full reset can diverge for h≤103h\le103. Under random switching with five rule distributions, every retention converged, including r=0r=0 and r=1r=1 at h=93h=93 (Section 7.5, Appendix M). Retention barely mattered on three distributions, and full retention was fastest on the two with extreme rules. For example, when A−A_- and A+A_+ are drawn with equal probability, the per-step rates (lower is faster) at h=93h=93 were 0.980 for r=1r=1, 0.986 for r=0r=0 and 0.991 for r=0.3727r=0.3727, so the certified retention was the slowest. The static design thus protects against bad sequences but does not improve average speed (Section 7.5).

Reviewer_ddE7OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This article analyzes the restart effect in the heavy-ball method / gradient-descent with momentum. Convergence of the algorithm is analyzed in one-dimensional quadratic functions, based on a uniform control of the spectral radius of Jacobian matrix across a scalar class of functions. Numerical results in high dimension seem to be performed to validate the insights obtained from the analysis.

Critical strengths

The idea of analyzing a restart effect is an important question for momentum-based methods, and the analysis in quadratic functions makes it possible to have a clear understanding of the effect.

Critical weaknesses

The article is not easy to follow. Some of the mathematical notations and technical concepts are not clearly defined, making both the algorithm analysis and numerical results hard to understand.

The analysis is limited to one-dimensional functions. It is not clear why it cannot be done for higher-dimensional functions.

Questions for authors
  • Please clarify your notation in equation 2: how are they related to the H in f(x)?

  • A boundary first replaces …, line 69; what does the boundary stand for?

  • In Theorem 1, what does (A^h)_11 mean in eq. 4? It is very hard to follow the discussion on lines 104- 106 regarding (A^h)_11.

  • In Section 7, what is Z_t in eq 18? It is not clearly defined.

  • What is d in Section 7.1? Is it the dimension of the variable x to be optimized? Also, some technical terms such as ‘commit length ’ and ‘Polyak parameters’ are not easy to understand.

1 comment
AuthorsOpenReview ↗

Response to Reviewer ddE7

Thank you for the specific questions. We agree that the notation was unclear. Section 2 now defines every model object before its use, and Appendix G adds a notation table and a glossary. Section 7 now says what each experiment tests, and Appendices H, L, M and O specify them.

Is the analysis limited to one-dimensional functions? No. In every dimension dd, heavy ball splits exactly into dd independent two-dimensional recursions, one per eigenvector of the Hessian HH, as Section 2 now shows. Along the eigenvector qiq_i with eigenvalue λi\lambda_i, the error and velocity components evolve as (ei,vi)↦A(αλi,β)(ei,vi)(e_i,v_i)\mapsto A(\alpha\lambda_i,\beta)(e_i,v_i), where AA is the 2×22\times2 heavy-ball matrix for step α\alpha and momentum β\beta. All our stability results therefore hold in every dimension, and in a 200-dimensional experiment the measured growth per block matches Theorem 1 with relative error below 10−1010^{-10} (Appendix H.2). Dimension matters in one respect. To show that shorter intervals fail, Theorem 2 alternates the fastest and the slowest admissible step and momentum on one eigenvector, which one scalar step and momentum cannot do when the eigenvalues of HH lie in [μ,L][\mu,L] with κ=L/μ>1\kappa=L/\mu>1 (Proposition 4(iii)). Our block length hh and retention rr then remain valid but may be conservative. Preconditioners whose eigenvectors differ from those of HH are outside the theory.

Q1: how does eq. (2) relate to HH? Here λ\lambda is an eigenvalue of HH, a=αλa=\alpha\lambda, and (e,v)(e,v) are the components of the position error x−x⋆x-x_\star and of the velocity along its eigenvector. So A(a,β)A(a,\beta) is one heavy-ball step in that mode, as Section 2 now says before eq. (2).

Q2: what is a boundary? Time is divided into blocks of hh steps, and a boundary is the start of a block. There the rule, meaning the step size and momentum, may change, and J(r)=diag(1,r)J(r)=\mathrm{diag}(1,r) multiplies the velocity by rr, so r=0r=0 is a full reset and r=1r=1 means no restart. Section 2 defines it in the paragraph "Blocks, boundaries, and retention".

Q3: the (1,1) entry of AhA^h in eq. (4). It is the position after hh steps from position 1 and velocity 0. After a full reset the velocity is 0, so the position at the next boundary is this entry times the current position. By Theorem 1, its largest absolute value over all rules with spectral radius at most s<1s<1 is Ds(h)=sh[1+h(1+s)]D_s(h)=s^h[1+h(1+s)], attained when both eigenvalues of AA equal −s-s. Section 3 now opens with this explanation and derives the bound after Theorem 1, and Figure 1(a) shows the transient.

Q4: what is ZtZ_t in eq. (18)? It is the error measure at step tt. In Section 7.3, ZtZ_t is f(xt)−f⋆f(x_t)-f_\star in the three controls that use the exact Hessian, and ∥gt∥2+τ∥vt∥2\lVert g_t\rVert^2+\tau\lVert v_t\rVert^2, with gradient gtg_t, velocity vtv_t and τ=μL/2\tau=\mu L/2, in the guard tests on non-quadratic tasks (paragraph "Nonlinear observable guards"). Section 7 now defines both cases next to eq. (18).

Q5: dd, "commit length" and "Polyak parameters". The symbol dd is the dimension of xx. "Commit length", now "block length", is the number hh of steps between two boundaries, and in the checkpoint experiments of Section 7.1 it is the number of gradient queries per checkpoint, 14, 58 and 228 for κ=10\kappa=10, 100 and 1000. Polyak's parameters, now defined in Section 2, are the classical optimal tuning α⋆=4/(L+μ)2\alpha_\star=4/(\sqrt L+\sqrt\mu)^2 and β⋆=((κ−1)/(κ+1))2\beta_\star=((\sqrt\kappa-1)/(\sqrt\kappa+1))^2.

We would be grateful if you would reconsider your assessment.

Reviewer_C321OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper analyzes how often heavy ball should be restarted when minimizing a quadratic objective $f(x) = \frac{1}{2} (x-x\ast)T H (x x\ast)$. The authors consider two restart strategies. The first is a full reset, in which the momentum term is reset to 0 after hh heavy ball updates. In Theorem 1, they prove their exact design rule for the number of heavy ball updates before restarting. It is given by the smallest natural number hh that satisfying hlog⁡(1/s)>log⁡(1+h(1+s))h \log (1/s) > \log (1 + h (1+s)). For the second case, they consider partial retention, in which the momentum term is scaled by a factor r∈[0,1]r \in [0,1] after hh heavy ball updates. They provide the smallest uniformly stable hh in Theorem 2. The authors also propose a safeguarded checkpoint scheme. It tries a proposed block of at most h gradient queries, and accepts it only if the gradient-based energy shrinks by a factor qh. Otherwise it rolls back and runs hh heavy ball steps.

Critical strengths

The paper provides the smallest restart interval hh for heavy ball to stay uniformly stable. This gives the practitioner a helpful insight for the heavy ball method.

Critical weaknesses
  1. There are many undefined or ambiguous objects, so the setup cannot be followed or checked without guessing. For example in Section 2 Model and Scope:
  2. (064 - 065) The initial conditions are not given (x0x_0 and v0v_0, for instance whether v0=0v_0=0).
  3. (066) λ\lambda is not defined
  4. (066) The decomposition H=QΛQTH=Q \Lambda QT is missing, and so is the statement ∇f(x)=He\nabla f(x) = H e. Without them, “scalar Hessian mode” is unexplained.
  5. (062 + 066) The symbols e and v denote both the vectors and the scalar modal coordinates.
  6. (070) The “Frozen Rule” and “Schur stable” are not defined and the claim “With β≥0\beta \geq 0, this means β<1\beta < 1 and $0 < a < 2 (1 + \beta)$ is stated without a reference or derivation.
  7. (081) η\eta is not defined.
  8. (082) ρ\rho is not defined. Is it the spectral radius?
  9. The paper never explains the crucial restart operator JJ. It does not describe, that r=0r = 0 is a full reset and r=1r = 1 means no restart or that it acts identically in every eigen-coordinate.
Questions for authors
  1. Your update rule in (066-070) uses e=x−x∗e = x-x_\ast, which is not observable. A practical rule needs observable quantities. How do you use this algorithm in practice?

Suggestions for future work - Is it possible to extend it beyond quadratics?

1 comment
AuthorsOpenReview ↗

Response to Reviewer C321

Thank you for the detailed list. Section 2 now defines each of these objects where it is first used, and Appendix G adds a notation table.

  • Initial conditions. The initial state (x0,v0)(x_0,v_0) is arbitrary, because every guarantee holds uniformly over initial states, and the experiments use v0=0v_0=0.
  • The symbol λ\lambda denotes an eigenvalue of the Hessian HH, with μ≤λ≤L\mu\le\lambda\le L. The eigenvalues of the 2×22\times2 matrix AA are now written ζ1,ζ2\zeta_1,\zeta_2.
  • Modes. Section 2 now writes H=QΛQTH=Q\Lambda Q^T with eigenvectors qiq_i, e=x−x⋆e=x-x_\star and ∇f(x)=He\nabla f(x)=He. In the coordinates ei=qiTee_i=q_i^Te and vi=qiTvv_i=q_i^Tv, heavy ball becomes dd independent recursions (ei,vi)↦A(αλi,β)(ei,vi)(e_i,v_i)\mapsto A(\alpha\lambda_i,\beta)(e_i,v_i), which is what "scalar Hessian mode" meant. Sections 3 and 4 treat one mode and write e,ve,v for its scalars ei,vie_i,v_i.
  • Frozen rule and Schur stability. A frozen rule is one step size and momentum pair (α,β)(\alpha,\beta) used forever, and it is Schur stable if its spectral radius is below one. Section 2 now gives the characteristic polynomial ζ2−(1−a+β)ζ+β\zeta^2-(1-a+\beta)\zeta+\beta of A(a,β)A(a,\beta), with a=αλa=\alpha\lambda, whose roots lie inside the unit disk exactly when 1−β1-\beta, aa and 2(1+β)−a2(1+\beta)-a are all positive. Appendix A.1 derives this from the Jury criterion, eq. (21).
  • The symbols η\eta and ρ\rho. The margin is η=1−s\eta=1-s, where ss bounds the spectral radius ρ\rho of every allowed rule.
  • The restart operator. At each boundary, which starts a block of hh steps, J(r)=diag(1,r)J(r)=\mathrm{diag}(1,r) multiplies the velocity by rr in every eigen-coordinate, so r=0r=0 is a full reset and r=1r=1 means no restart.

Q1: how is the method used without x⋆x_\star? The policy is static and needs only gradients, while ee and x⋆x_\star appear only in the analysis. Section 2 now states its three steps.

  1. Choose ss to bound the spectral radius of every rule that may be used. For a known [μ,L][\mu,L], Proposition 4(i) gives the test β≤s2\beta\le s^2 and (1−s)(1−β/s)/μ≤α≤(1+s)(1+β/s)/L(1-s)(1-\beta/s)/\mu\le\alpha\le(1+s)(1+\beta/s)/L.
  2. For full reset, take the smallest hh with Ds(h)=sh[1+h(1+s)]<1D_s(h)=s^h[1+h(1+s)]<1, which is hreseth_{\rm reset} and exists for every ss. For partial retention, Table 1 gives hh and rr at three certified margins, and other margins need a new certificate computation. These designs are valid for every condition number κ=L/μ\kappa=L/\mu and may be conservative when κ>1\kappa>1.
  3. Change the rule only every hh steps, and multiply the velocity by rr at each boundary.

Q2: beyond quadratics. Proposition 7 (Appendix L) extends the restart rules locally, with ∇2f(x⋆)\nabla^2f(x_\star) in the role of HH. If the Hessian is Lipschitz near x⋆x_\star, full reset with Ds(h)<1D_s(h)<1 is locally uniformly exponentially stable, and so is every certified design of Table 1. Conversely, if the largest Hessian eigenvalue at x⋆x_\star equals LL, the rule α=(1+s)2/L\alpha=(1+s)^2/L, β=s2\beta=s^2 restarted every hh steps is not Lyapunov stable when Ds(h)>1D_s(h)>1. On two logistic-regression problems, every h≥hreseth\ge h_{\rm reset} converged, with a per-block factor over the last ten blocks within 0.04% of Ds(h)D_s(h). Runs started at x0=0x_0=0 with blocks shorter than hreseth_{\rm reset} stalled at a distance of 1.7⋅10−41.7\cdot10^{-4} to 6.7⋅10−46.7\cdot10^{-4} from x⋆x_\star (Table 10). Global results remain open, because Polyak's parameters are not globally convergent for all strongly convex functions (Lessard et al., 2016).

We hope the setup can now be checked without guessing, and would be grateful if you reconsidered your scores.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

All three reviewers found the paper hard to read and asked about the scope of the model, and the revision addresses both. The main text still fits in 10 pages, and no original result or numbering changed. The new results are in Appendices H to M.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.