A Variational Perspective on High-Resolution ODEs

We consider unconstrained minimization of smooth convex functions. We propose a novel variational perspective using forced Euler-Lagrange equation that allows for studying high-resolution ODEs. Through this, we obtain a faster convergence rate for gradient norm minimization using Nesterov's accelerated gradient method. Additionally, we show that Nesterov's method can be interpreted as a rate-matching discretization of an appropriately chosen high-resolution ODE. Finally, using the results from the new variational perspective, we propose a stochastic method for noisy gradients. Several numerical experiments compare and illustrate our stochastic algorithm with state of the art methods.

Paper

References (34)

Scroll for more · 22 remaining

Similar papers

Peer review

Reviewer rs2m7/10 · confidence 3/52023-07-03

Summary

This paper introduces forced Lagrangians in order to better understand discrete schemes for accelerate convex numerical optimization from a continuous-time ODE perspective.

Strengths

Two choices of F enable to re-derive continuous-time ODEs which were introduced recently in the literature (Section 2). Likewise, accelerated discrete schemes derive subsequently (Sections 3, 4) recover known schemes or in slightly modified form with slightly superior convergence rates. Empirically, it is demonstrated that even for non-convex network training, applying stochastic versions of accelerated schemes may perform well. This paper significantly contributes to our understanding what unconstrained convex optimization with optimal convergence rates really means.

Weaknesses

Authors use their jargon right from the beginning which makes the paper hard to read for readers who do not work on similar topics. For example, what “LR-ODE”, “HR-ODE” and the “rate matching technique” means should be briefly explained to a broader educated readership in the introduction.

Questions

none

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

see "weaknesses" on how to improve the presentation

Reviewer A9z27/10 · confidence 2/52023-07-04

Summary

This work combines variational perspective approach with high-resolution ODE functions to investigate the Nesterov accelerated gradient descent algorithm (NAG). With variational perspective, the authors reconstruct various high-resolution ODEs derived in previous research using alternative methods. Moreover, through this approach, they propose a special representation of NAG that exhibits an improved convergence rate in terms of gradient norm minimization. The authors also discuss some new properties of rate-matching technique. Finally, the authors analyze the stochastic setting both theoretically and empirically.

Strengths

1. The idea of combining variational perspective and high-resolution ODEs by including external forces is very interesting. 2. The authors show several important theoretical results in their manuscript. By carefully checking their proofs of theorems (except for Section 5 because of time), I think overall these results are correct. 3. The numerical results indicate the potential of having better optimization algorithms based on the theoretical results in this manuscript. 4. Overall, the manuscript is well written. Overall, I have a positive impression of this work, however, I also admit that I am not an expert in this specific field, which may affect my confidence in assessing its accuracy and significance.

Weaknesses

1. It would be better to have a more detailed introduction to low-resolution ODEs and high-resolution ODEs. Without reading some previous work, it is difficult to understand the differences between low-resolution ODEs and high-resolution ODEs. Are high-resolution ODEs the ODE functions that contain the learning rate $s$? Why do people need to care about high-resolution ODEs? 2. I feel that the proof of Proposition 4.1 is more of an intuition rather than a rigorous demonstration. The treatment of the condition $s\to 0$ seems imprecise, as sometimes the authors will directly consider this condition as $s=0$ (e.g. line 60), while at other times the authors maintain $s$ to be a non-zero value (e.g. line 59). The use of the word “approximately” in Prop 4.1 is also vague. 3. Some typos/ unclear parts in the proofs that the authors may need to double-check. (a) line 84, $\frac{\partial L}{\partial X}(X_t,...)$ should be $\frac{\partial L}{\partial X_t}(X_t,...)$. (b) equation (7), left hand side, $\bigtriangledown f$ should be $\bigtriangledown f(X_t)$. (c) Theorem 2.1, $\dot{\gamma}=e^{\alpha t}$ should be $\dot{\gamma}=e^{\alpha_t}$. (d) line 138, ODE(14) should be ODE(12). (e) equation (25) $(3/t+\sqrt{s}\bigtriangledown f(X_t))$ should be $(3/t+\sqrt{s}\bigtriangledown^2 f(X_t))$. (f) line 196, $\sigma$ not introduced, is it the variance of the noise? (g) equation (38) (appendix), second line $\sqrt(s)e^{-\alpha_t}\ddot{\beta_t}$ should be $\sqrt(s)e^{-2\alpha_t}\ddot{\beta_t}$. (h) line 392, (7) should be (11). (i) line 406, first equality $+\frac{1}{2}||v_k-x^*||^2$ should be $-\frac{1}{2}||v_k-x^*||^2$. (j) line 408, the term $+\frac{s^2(k+2)}{4}||\bigtriangledown f(x_{k+1})||^2$ in the second inequality is left out, therefore the authors need to double-check whether the results still hold after considering this term. (k) line 411, the first equaltiy redundant, (l) line 413 "not ethat" typo (m) line 421, the term $\frac{ks}{2}(\bigtriangledown f(x_k-\bigtriangledown f(x_k))$ miss ")". (n) equation (62), $(3/t+\sqrt{s}\bigtriangledown f(X_t))$ should be $(3/t+\sqrt{s}\bigtriangledown^2 f(X_t))$. 4. Figure 1 is hard to read. (font size too small) 5. It would be better to describe what "NAG" means (Nesterov accelerated gradient?) the first time this abbreviation is used.

Questions

1. It is unclear to me that equation (21) and equation (27) (by replacing one term) is equal to NAG (line 76), why is that? 2. In line 211, the authors say that practically $k_0$ is lower than the term $(\cdot)^{1/\alpha}$. However, one of the conditions in Theorem 5.1 and Theorem 5.2 is that $k_0\geq (\cdot)^{1/\alpha}$, does that mean this condition will not be satisfied in practice?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The authors don't specifically discuss the limitations of this work. The authors may consider adding a paragraph in their manuscript to discuss the limitations of their work based on the reviewer's feedback. I don't think there will be a significant negative societal impact of this work.

Reviewer ZUpb6/10 · confidence 1/52023-07-04

Summary

They generalize the Lagrangian formulation of known first order optimization methods [Wibisono et al., 2016, Wilson et al., 2021] by introducing the notion of external forces, and show theorems on convergence of convex/strongly convex functions. In addition, they show the variational analysis leads a special representation of NAG, and it enjoys to superior convergence rates than [Shi et al., 2019]. Based on the special representation of NAG, they generalize it to stochastic variation in Section 5, called NNAG, and show its convergent theorem. They check the practical behavior of the proposed NNAG by using - binary classification, - classification on CIFAR10.

Strengths

originality - Introducing the external force term to the Lagrangian formulation and show convergent theorems with it. quality/clarity - They show their idea and mathematical statements clearly, and show their proofs. significance - Their formulation can lead known dynamics also. In this sense, their proposal can be regarded as a unification of various optimization dynamics.

Weaknesses

- The first part of the paper sounds natural, but to me, there seems no theoretical reason to introduce i.i.d. noise to the gradients in section 5. - In my opinion, they should conduct more numerical experiments. For example, the experiments in classification on CIFAR10, NNAG and SVRG+NNAG are competitive to SGD. This result, itself is good, but it means there is no big incentive to use the proposed algorithm, but SGD.

Questions

- Is there any natural/theoretical motivation to introduce i.i.d. noise to the gradients in section 5? - Is there any strong incentive to use NNAG or its variants compared to SGD in practice? - Besides it, the authors wrote `As the figure depicts, the SVRG+NNAG performs faster than the other methods in terms of minimizing the training error.`, but minimizing training error itself sounds possibility of overfitting in machine learning context. If not, please correct me.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

n/a

Reviewer CWtz7/10 · confidence 2/52023-08-01

Summary

This paper did four works about High-Resolution ODEs and first-order optimization algorithms. The first part uses forced Euler-Lagrange equations to generalize the analysis of Low-Resolution ODEs to High-Resolution ODEs. The second part is a refined result for bound estimation of gradient norm. The third part is an interpretation of Nesterov’s acceleration by rate-matching discretization. In the last part, the authors propose a stochastic version of Nesterov’s accelerated gradient method, named NNAG in this paper, and compared it to several known stochastic methods on binary classification and training CNN.

Strengths

This paper proposes a new idea of external force and uses the idea of forced Euler-Lagrange equations to generalize the analysis of Low-Resolution ODEs to High-Resolution ODEs. This leads to a new formulation of NAG. Combined with rate-match discretization, the authors discover a new connection of NAG and the continuous ODE, i.e., it is understood as a perturbation of the Low-Resolution ODE. The new formulation of NAG further inspires the framework of Noisy NAG. The idea of forced Euler-Lagrange equations may bring new insight to fast algorithm design.

Weaknesses

Although this paper provides a new reformulation of NAG, as the author suggested, the refined convergence rate of gradient norm for NAG is already appeared in [S. Chen, B. Shi, and Y.-x. Yuan. Gradient norm minimization of nesterov acceleration: o(1/k^3). arXiv preprint arXiv:2209.08862, 2022]. The Lyapunov function is essentially the same using the implicit-velocity form and the explicit-velocity form in [B. Shi, S. S. Du, M. I. Jordan, and W. J. Su. Understanding the acceleration phenomenon via high-resolution differential equations. ArXiv, abs/1810.08907, 2021].

Questions

The Lagrange function is kinetic energy minus potential energy (though multiplied by some decreasing coefficient). The Lagrange function and external force is considered separately in this paper. Is it possible to give physical explanation about the acceleration by forced Euler-Lagrange equation and use different energy function and external force to inspire better algorithm design?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The authors have adequately addressed the limitations and have indicated problems for further study. This work appears to have little negative societal impact.

Reviewer A9z22023-08-13

Thank you for the response

I would like to thank the authors for their response. I have one follow up question. I agree that equation (21) can be transferred to the form of the one-line representation above, However, according to line 76, the one-line representation of NAG should be $$x_{k+2}=x_{k+1}+\frac{k+1}{k+4}(x_{k+1}-x_{k})-\frac{s(k+1)}{k+4}(\bigtriangledown f(x_{k+1})-\bigtriangledown f(x_{k}))-s\bigtriangledown f(x_{k+1})$$. Do these differences matter?

Authorsrebuttal2023-08-15

Response 1

Thank you for your comments. The relation of the NAG (the one-line update) holds for a 3-point sequence, it does not matter if the sequence is $x_k,x_{k+1},x_{k+2}$ or $x_{k-1},x_{k},x_{k+1}$. This is just a matter of notation/convention; in fact, we can define a new sequence $z_k:=x_{k+1}$ and get exactly the same one-line update as the NAG.

Reviewer A9z22023-08-15

Thank you for your response. The difference I notice here is not whether the sequence is represented by $x_k,x_{k+1},x_{k+2}$ or $x_{k-1},x_{k},x_{k+1}$, the difference is the coefficient $(k+1)/(k+4)$ vs. $k/(k+3)$ given the same representation, this difference cannot be removed by a simple transformation $z_{k}:=x_{k+1}$ (because then the coefficients containing $k$ would also need to be changed). Although I guess this difference should be fine?

Authorsrebuttal2023-08-15

Thanks for your comment. The equation in line 76 is equivalent to $$x_{k+1}=x_k+\frac{k}{k+3}(x_k-x_{k-1})-\frac{s(k)}{k+3}(\nabla f(x_k)-\nabla f(x_{k-1}))-s\nabla f(x_k)$$ with an initial point $x_{-1}$. And the one-line representation of (21) is $$\hat x_{k+2}= \hat x_{k+1}+\frac{k}{k+3}(\hat x_{k+1}-\hat x_{k})-\frac{s(k)}{k+3}(\nabla f(\hat x_{k+1})-\nabla f(\hat x_{k}))-s\nabla f(\hat x_{k+1})$$ with an initial point $\hat x_0$. Now, if we introduce $z_k:=\hat x_{k+1}$, we get $$z_{k+1}=z_k+\frac{k}{k+3}(z_k-z_{k-1})-\frac{s(k)}{k+3}(\nabla f(z_k)-\nabla f(z_{k-1}))-s\nabla f(z_k),$$ with initial point $z_{-1}$; which indicates that the updates for the sequence $z_k$ are exactly the same as the ones for sequence $x_k$.

Reviewer A9z22023-08-15

Thanks, I got this point. I decide to increase the score by one based on the responses from the reviewers.

Reviewer ZUpb2023-08-16

I would like to thank the authors for their response. Their response to my questions: - motivation to introduce i.i.d. noise - The authors answered that it is a one step closer to the real-world practical scenarios like SGD in training of neural network. This makes sense from practical perspective. - incentive to use NNAG - The authors answered that "SVRG+NNAG outperforms SGD in terms of convergence rate in our experiments". In addition, I understand that the contribution of this paper is more theoretical, a new interpretation of the NAG method by the forced Euler-Lagrange equation. - on overfitting problem - The author explained why "faster convergence rates of NNAG+SVRG do not pose problems of overfitting" in the rebuttal, and I agree with the author's point. In summary, I can say that the author's explanation has, for the most part, addressed my concerns. So I would like to raise my score to weak accept.

Reviewer CWtz2023-08-20

I would like to thank for the author's reply. I agree that the forced Euler-Lagrangian is anothor way to help simplifying the proof in [B. Shi, S. S. Du, M. I. Jordan, and W. J. Su. "Understanding the acceleration phenomenon via high-resolution differential equations"(2021)]. Although the level of simplification compared to [S. Chen, B. Shi, and Y.-x. Yuan. "Gradient norm minimization of nesterov acceleration: o(1/k^3)"(2021)] is still not very clear to me, the authors have addressed my question. On the other hand, the physical explanation of the non-potential external force and its potential usage seem interesting to me. In summary, I decide to raise my score by one based on the responses from the authors.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC