Directional Smoothness and Gradient Methods: Convergence and Adaptivity

We develop new sub-optimality bounds for gradient descent (GD) that depend on the conditioning of the objective along the path of optimization rather than on global, worst-case constants. Key to our proofs is directional smoothness, a measure of gradient variation that we use to develop upper-bounds on the objective. Minimizing these upper-bounds requires solving implicit equations to obtain a sequence of strongly adapted step-sizes; we show that these equations are straightforward to solve for convex quadratics and lead to new guarantees for two classical step-sizes. For general functions, we prove that the Polyak step-size and normalized GD obtain fast, path-dependent rates despite using no knowledge of the directional smoothness. Experiments on logistic regression show our convergence guarantees are tighter than the classical theory based on $L$-smoothness.

Paper

References (47)

Scroll for more · 35 remaining

Similar papers

Peer review

Reviewer Z7fB5/10 · confidence 3/52024-06-25

Summary

The work propose different directional smoothness function and use them to establish sub-optimality bounds that are adaptive to the optimization trajectory. This approach can be explicitly used in quadratics or with an exponential search technique for convex objectives to use adaptive stepsizes that enjoy better guarantees compared to constant stepsize GD tuned according to the global smoothness. Furthermore, the paper shows that Polyak and normalized GD obtain convergence guarantees that depend on the directional smoothness without modifying the algorithms. The theoretical claims are empirically evaluated using two logistic regression problems.

Strengths

1. The technique naturally yields results for several existing methods. 2. Tighter, adaptive guarantees are appealing due to the observed success beyond the theoretical convergence bounds.

Weaknesses

3. The paper discusses only deterministic optimization, while modern large-scale optimization is largely stochastic. It begs the question, which is not discussed in the paper, whether such approaches are even valid for stochastic optimization (e.g. [1]). 4. The novelty of the approach is unclear for some of the observed results. Consider especially Theorem 4.4 with the Polyak stepsizes, which achieve the best empirical performance in the experiments. It is straightforward to translate the decrease mentioned in [2] (page 2, end of left column), $\lVert x_{t+1} - x^* \rVert^2 - \lVert x_{t} - x^* \rVert^2 \leq - \frac{(f(x_t)-f(x^*))^2}{\lVert \nabla f(x_t) \rVert^2}$ to a bound of the weighted $\overline{x}$ of $f(\overline{x_t})-f(x^*) \leq \frac{\lVert x_0 - x^*\rVert^2}{\sum_{k=0}^{t} \eta_t}$. Theorem 4.4 can be considered as replacing $\eta_t$ with a possibly worse $M(x_{t+1},x_t)$. In this case, the simple adaptive result is only framed for convenience in the directional smoothness framework. [1] N. Loizou, S. Vaswani, I. H. Laradji, and S. Lacoste-Julien. Stochastic Polyak step-size for SGD: An adaptive learning rate for fast convergence. In International Conference on Artificial Intelligence and Statistics, pages 1306–1314, 2021. [2] Hazan, E. and Kakade, S., 2019. Revisiting the Polyak step size. arXiv preprint arXiv:1905.00313.

Questions

5. Can the framework be of use for stochastic optimization? 6. Can the authors elaborate on the technical novelty of the results? In particular, it is arguably known that the standard analysis technique of GD exploits smoothness only between consecutive points. What is the main difficulty when we replace global smoothness with an adaptive variant? Overall, the paper presents an appealing framework for adaptive bounds which encompasses a long list of stepsize strategies, while the novelty and deterministic setting are somewhat in question.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

See weaknesses.

Reviewer vUdK6/10 · confidence 4/52024-07-11

Summary

The paper proposes a new type of non-uniform smoothness, which the authors label _directional smoothness_. Directional smoothness replaces $L$ in the typical $L$-smoothness inequality by a function $M(x,y)$, which describes the smoothness along the line between $x$ and $y$. The authors proof multiple basic properties of this notion as well as sub-optimality bounds for GD with general stepsizes in the deterministic convex setting. Additionally, they provide sub-optimality bounds for three specific stepsizes in the same setting.

Strengths

- The proposed smoothness notion – directional smoothness – is generalising many previous notions of smoothness while still allowing for a meaningful analysis of algorithms. - Convergence analysis with directional smoothness can potentially give some intuition which algorithms automatically adapt to (potentially beneficial) local smoothness properties. - The authors provide a wide range of properties for directional smoothness and different ways to compute upper bounds on the directional smoothness. - The sub-optimality gap for Polyak’s step-size, showing it automatically adapts to the local smoothness in a meaningful way, could potentially explain empirical success of the step-size schedule in practise.

Weaknesses

- Many of the results until Section 4.2 are direct consequences of the (directional) descent inequality. In particular, it was to be expected that the global smoothness inequality can be replaced with a localized version, and that results along the lines of Proposition 3.1 – 3.3 should hold. That being said, these results can be seen as preliminary results for the propositions / theorems in the later portion of the paper, which provide new insights. - ~~There might be some questions regarding the soundness of Proposition 3.2, see Question 1. However, this result is not used in the consequent results and does hence not have any further implications.~~ - Given this submission introduces a new way to analyse algorithms, seeing its application in different settings and for different algorithms would be highly beneficial to showcase the benefit of the proposed approach. All results in the work are however limited to the deterministic convex setting, and only Gradient Descent with different stepsizes is considered. Showcasing the approach for the stochastic setting or for different algorithms (such as Nesterov AGD) could give a more wholistic picture of the approach. --- **Edit after Author Rebuttal** The authors did fix the proof of Proposition 3.2.

Questions

# Questions 1. In the proof of Proposition 3.2, why does the last inequality (above line 438) hold? More specifically, in the case $\eta_k > \mu_k^{-1}$, the recursive argument cannot be applied as the sign is flipped. How do the authors deal with this case? 2. Theorem 4.3 is very unclear to me. What is one “iteration” of GD? In order to reach a point $\hat{x}_T$ with $f(\hat{x}_T) - f(x^*) \leq \varepsilon$, do you require $\widetilde{O}(\varepsilon^{-1})$ or $\widetilde{O}(\varepsilon^{-2})$ gradient evaluations? Maybe these unclear points could be addressed by including the full algorithm, producing the final output $\hat{x}_T$, in Algorithm 1. 3. The authors first derive suboptimality bounds for general stepsizes, before proving results for specific stepsizes. However, the proofs for Theorems 4.4 and 4.5 are not based on the general results, rather they are standalone proofs. Why did the authors not plug in these stepsizes into the general results? This would be a great opportunity to showcase the application of Propositions 3.1 – 3.3. 4. Given directional smoothness generalises other non-uniform smoothness notions, did the authors examine whether their sub-optimality bounds reconstruct / improve previous results for notions of smoothness that can be modelled by directional smoothness? For example, can results from $(L_0, L_1)$-smoothness [1] be reconstructed? The corresponding $M(x,y)$ is specified in [2, Lemma A.3] / [3, Lemma 8]. 5. Is there hope for any lower bound results that show strongly adapted stepsizes are indeed the best (non-accelerated) stepsize one can choose in your generalized setting? Given such result, the Polyak-stepsize results would get even stronger. 6. Are there any difficulties extending the notion of directional smoothness to the typical nonconvex setting and results, i.e. $|| \nabla F(x) - \nabla F(y) || \leq M(x,y) || x-y||$ and the corresponding GD guarantees? [1]: Li et al., Convex and Non-convex Optimization Under Generalized Smoothness, NeurIPS 2023 [2]: Zhang et al., Improved Analysis of Clipping Algorithms for Non-convex Optimization, NeurIPS 2020 [3]: Hübler et al., Parameter-Agnostic Optimization under Relaxed Smoothness, AISTATS 2024 # Remarks - There are multiple potential inconsistencies throughout the work. None of them have an impact on the soundness of the work, but it would be great if the authors could address them. a) Most results use $k$, some use $T$ while others use $t$ as the last iteration index. b) The averaged iterate is sometimes denoted using hats while other times using widebars. c) There are inconsistencies whether differentiability assumptions are explicitly stated or not. In sections 2 and 4 they are consistently mentioned, section 3 and its proofs in the appendix do not mention it even-though it is required. d) Usage of $f(x^*)$ vs. $f^*$. e) Appendix C denotes the Euclidian norm by $|| \cdot ||$, while all other sections use $|| \cdot ||_2$. f) All results in the appendix besides the three Theorems are numbered relative to the appendix, the theorems relative to the main section. - Lemma A.4 might better fit in Appendix B. - Proposition 3.2 could potentially be improved by using $2 \eta_k [f(x^*) – f(x_{k+1})] \leq -\eta_k \mu_k \Delta_{k+1}$ instead of removing the former term in the last inequality after line 435. This divides each term in the product by $(1+\eta_k\mu_k)$. - In it’s current version, Equation (41) does not provide additional information. The authors could either remove it or modify Equation (40) to include this step. - As mentioned in Line 562, the convergence results in Theorem 4.5 require some smoothness assumption such as $L$-smoothness. This assumption is missing in the statement of Theorem 4.5. Below some potential typos. - Proposition 3.1: $\delta_k$ -> $\delta_{k+1}$ - Proposition 3.2: $\Delta_k$ -> $\Delta_{k+1}$ - Theorem 3.4, Case 1: Missing parenthesis - Equation (23): $\bar x$ -> $\bar x_t$ - Last inequality in equation after Line 232: $M(x_{k+1},x_{k})^{-1}$ -> $M(x_{k+1},x_{k})$ - Equation after Line 393: $f(x)$ -> $f(y) - f(x)$ - Second inequality in equation after Line 435: $|| \nabla f(x) ||$ -> $|| \nabla f(x) ||^2$ - Equation after Line 468: $\Rightarrow$ -> $\Leftarrow$ (or $\Leftrightarrow$) - Equations after line 475: $M$ -> $D$ - Algorithm 1, Line 1: $L_0$ -> $\eta_0$ - Algorithm 1, Line 3: Wrong parentheses - Line 527: follow -> following

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The introduction could more clearly mention that the paper (besides Lemma 2.4) exclusively focuses on the deterministic convex setting.

Authorsrebuttal2024-08-09

Thank you for engaging with our response! Please let us know if the following answers address your concerns. > Does the result hinge on this weighted average? Thank you for this relevant question. As we note in Line 228 and after Eq. 25, the weighted average of iterates using the (unknown) smoothness function is always lower-bounded (in $f$) by the best iterate, so we can replace $\bar x$ with $\text{argmin}_{k} f(x_k)$ without changing the guarantee. Thus, the best iterate is adaptive to any directional smoothness function without any knowledge of $M$. If we have access to function value evaluations, calculating the argmin is no more difficult than calculating the weighted average. > Applicability to other settings. We understand your concerns. We can extend our results to Nesterov acceleration to obtain optimal rates which only depend on the directional smoothness. Nesterov [N18] shows that any gradient method making progress like, $$ f(x_{k+1}) \leq f(x_k) - \frac{\eta_k}{2} \\|\nabla f(x_k)\\|_2^2. $$ can be generically accelerated using an estimating sequences argument. This condition is satisfied by gradient descent with step-sizes adapted to a directional smoothness function $M$. Thus, GD with adapted step-sizes can be accelerated to obtain optimal linear/sub-linear rates with a dependence on the maximum directional smoothness observed over the optimization trajectory. Such a result can only improve on the classical accelerated rate. We will add a formal proof of this result to the paper. We'd like to reiterate that it has not been established in prior work that adapted stepsizes exist, or that they can be found easily. While our provided counterexample means that in general non-convex optimization might not be amenable to an easy analysis, this does not rule out that it can be useful under extra assumptions (e.g. that the Hessian is locally Lipschitz, or with directional smoothness functions that can potentially depend on more than two points). [N18] Nesterov, Yurii. Lectures on convex optimization. Vol. 137. Berlin: Springer, 2018. > Proof of Proposition 3.2 Thanks again for catching this issue. > $(L_0, L_1)$ smoothness and Lemma [1, Lemma 8] Thank you for pointing us to this more modern result and reference. We now agree that indeed that $(L_0, L_1)$ smoothness is a valid directional smoothness function according to our definition. This is actually a very interesting observation, which has several consequences including expanding the applications of directional smoothness, and explaining recent work on $(L_0, L_1)$ smoothness and the Polyak step-size. 1. [Polyak step size and $(L_0, L_1)$ smoothness] During the NeurIPS deadline (May 2024), two works appeared online showing that the Polyak step-size can be analysed under $(L_0, L_1)$ smoothness [3], and a related local curvature condition [4]. In particular Theorem 4 in [4] establishes that if the objective $f(x)$ is convex, smooth and $(L_0, L_1)$ smooth, that GD with a Polyak step size converges at a $O(1/T)$ rate with favourable constants. But now that it is clear that $(L_0, L_1)$ smoothness is a directional smoothness, we can immediately improve up on Theorem 4 in [2] in two ways: (a) we now know that assuming L smoothness (in addition to $(L_0, L_1)$ smooth) is not needed, and (b) this $O(1/T)$ result holds not just for $(L_0, L_1)$ smoothness, but for *every* directional smoothness constants as we have shown in Theorem 4.4. We also believe this will have consequences on the related definition of local curvature and the results in [4]. In other words, directional smoothness is a generalization of $(L_0,L_1)$ smoothness that also guarantees convergence at the same rate. 2. [Non-convex path forward] Both [3] and [4] consider the non-convex setting and analyse variants of normalized gradient descent with momentum. Now that we know that $(L_0, L_1)$ smooth is a particular type of directional smoothness, we can consider generalizing their result to any type of directional smoothness function, thus giving us a clear way forward for the non-convex setting. To do this, we believe we will be able to leverage our Lemma D.4. Yet it is not immediately obvious how to do this, so we cannot promise any substantial result in this paper. We sincerely thank the reviewer for bringing this up. We will include now an extended remark on this connection and consequences to [3] and [4]. We will also add an acknowledgement to Reviewer vUdK for pointing this out to us, and for a insightful review.

Reviewer ETZ96/10 · confidence 3/52024-07-13

Summary

This paper develops refined sub-optimality bounds for gradient descent in the convex setting. The authors consider directional smoothness, a local and path dependent smoothness condition, instead of assuming globally bounded smoothness constants in classical analyses. They discussed several interesting examples of directional smoothness functions, and derived sub-optimality bounds with them. For convex quadratic problems, they show that the stepsizes that minimize the upper bounds can be easily computed, which leads to new guarantees for two classical stepsizes. For general convex functions, they show that ideal stepsizes exist and can be computed using Newton's method or by exponential search. They also show that Polyak stepsize and normalized GD can achieve fast and path-dependent rates without knowing the directional smoothness.

Strengths

1. The sub-optimality bounds in terms of directional smoothness and the stepsizes and convergence results following the upper bounds look novel. The perspective is quite interesting. 2. The proofs look correct and the results look reasonable. The authors have thoroughly studied relevant aspects and details of their theory. 3. The authors did a good job connecting their results with existing methods and theories, which is quite insightful. 3. The writing and presentation are also very good. I enjoyed reading the paper.

Weaknesses

1. The final convergence rates are path-dependent. Although they are more refined and should be better than classical rates, it is not clear how large the gap is. I wonder if there are any interesting applications for which their rates can be more explicit and show clear improvement over classical ones.

Questions

1. Have the authors considered non-convex settings? In fact, since the authors mentioned neural networks in the first paragraph of the introduction, I wonder if this directional smoothness can be used to explain certain phenomenons in deep neural network training, like edge of stability or benefits of adaptivity. Some potential typos: a. Equation (3): missing "+" b. Line 194: it should be the inverse of Hessian? c. Theorem 4.3. Case 1: missing ")"

Rating

6

Confidence

3

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors have adequately addressed limitations.

Reviewer vUdK2024-08-09

Response to Rebuttal by Authors

I would like to thank the authors for taking the time to respond to all reviewers. Some of my questions were answered completely, for some others I have a few follow-up questions to ensure a correct understanding on my side. > Direct consequence of Descent Inequality I see the authors' perspective and agree that formalising a long-standing insight can be valuable if it leads to new, previously unknown insights. At the time of my review, I believed that the Polyak stepsize result could be one such insight. However, the comment from reviewer Z7fB raises some questions. Specifically, while I agree that your result has a different flavour by demonstrating adaptivity to an unknown smoothness function, the final result only holds for the weighted average $\bar{x}_t$. The weights in this average require knowledge of the (unknown) smoothness function. Does the result of Theorem 4.4 hinge on this weighted average? Asked differently, can the authors extend the result to show that Polyak stepsizes lead to the same convergence without requiring knowledge of $M$ to output a final iterate? > Stochastic setting or different algorithms (such as Nesterov AGD) could give a more wholistic picture of the approach I appreciate the authors for providing sub-optimality bounds for four different algorithms using their introduced concept. However, all four algorithms adhere to the classical GD paradigm $x_{t+1} \gets x_t - \eta_t \nabla f(x_t)$. On a high level, all convergence guarantees demonstrate that this new concept can be applied to such algorithms. Since the work introduces a new concept, it might arguably be more important to convince readers of its significance and applicability. By limiting all results to such type of algorithms and the deterministic convex setting, it raises the question of whether this is the only setting that can be addressed using this concept. This might in turn explain the lack of previous publications on this straightforward generalisation. The authors response regarding the applicability to the non-convex setting underscores this concern. Could the authors address this concern? > Proof of Proposition 3.2 I would like to thank the authors for their correction of the proof, with which I agree. I will increase my Soundness score accordingly and remove the corresponding weakness. Please ensure the fix is included in future versions of the work. > $(L_0, L_1)$-Smoothness I fear that Lemma A.3 – while being the first to derive such result – does not state the most general version of the result. The result has been simplified for presentation purposes. For a more general version, see [1, Lemma 8]. Specifically, the function $M$ is given by $$ M(x,y) = B_0 (L_1 ||x-y||)L_0 + B_1 (L_1||x-y||) L_1 ||\nabla f (x)||, $$ where $$ B_0(c) = 1 + 2\frac{e^c - 1} c - 4 \frac{e^c - 1 - c}{c^2}, \qquad B_1(c) = 2\frac{e^c - 1 - c}{c^2}. $$ It would be great if the authors could revisit the question using the appropriate Definition / Lemma. --- [1] Hübler et al., Parameter-Agnostic Optimization under Relaxed Smoothness, AISTATS 2024

Reviewer Z7fB2024-08-10

I thank the authors for their response. After reading the other reviews and rebuttal responses, I tend toward the acceptance of the paper and I raise my score accordingly. In any case, extending the framework to the stochastic case, even under stronger than usual noise assumption, would give the framework more credibility and make the case for directional smoothness much stronger.

Authorsrebuttal2024-08-12

Thank you for reading our response and for updating your score! We greatly appreciate your help in improving our paper. We will make sure to carefully investigate the stochastic case as part of future work.

Reviewer ETZ92024-08-11

Thank you for the response! I have also read other reviewers and would like to keep my score.

Reviewer vUdK2024-08-12

I would like to express my gratitude to the authors for addressing my questions. While I can believe that Nesterov’s result allows to extend the result to accelerated regimes, it seems to be more of an existence result rather than a showcase of the applicability of Directional Smoothness to different algorithms. I hence still believe that this is a concern and it prevents me from giving a score higher than a weak accept. I am pleased to hear that the authors were able to reconstruct and even strengthen results that are a special case of their notion of smoothness. Regarding the first point, could the authors brievly outline how they show that $\sum_{k=0}^{T-1} M(x_{k+1},x_k) \in \tilde{O}(T)$ in the giving setting to reconstruct the cited result? It does not seem to be trivial to me. I will temporarily increase my score to a weak accept, finalising the score after a response on the above technical question.

Authorsrebuttal2024-08-12

Remaining technical question

Thank you again for engaging, and also thank you for raising your score. With regards to your remaining technical question. [Bounded average directional smoothness] Your questions is “when will $\bar{M}\_T := \sum\_{k=0}^{T-1} M(x_{k+1},x_k)/T = O(1)$“. The answer depends on what we assume. If we consider the same assumptions as Theorem 4 in [2], then the objective function is globally $L$ smooth and $\sum_{k=0}^{T-1} M(x_{k+1},x_k)/T \leq L = O(1)$, in which case our Theorem 4.4 generalizes Theorem 4 in [2], as it proves a $O(1/T)$ convergence for every type of directional smoothness. Alternatively, if we drop the global L smoothness assumption, then we need to assume that $\bar{M}_t$ is bounded. This is weaker than assuming global L smoothness, thus we would have both generalized this $O(1/T)$ in Theorem 4 in [2] to all directional smoothness functions, and also extended the class of functions to which is applies (to potentially non-globally L smooth functions). Please let us know if this is clear.

Reviewer vUdK2024-08-13

I appreciate the authors efforts in answering the technical question. My excitement stemmed from their claim that the results in their submission improved upon [1, Theorem 4] by removing the L-smoothness assumption. However, the authors' technical response did not meet my initial expectations. Instead of completely removing the $L$-smoothness assumption, the authors only manage to weaken it to a uniform bound on the smoothness between consecutive points. As reviewer Z7fB pointed out, it is well-known in the community that smoothness along the trajectory is sufficient for the convergence of Gradient Descent type algorithms in the deterministic setting, see e.g. the lecture notes [2, Lemma 3.7]. I do however want to mention that reconstructing [1, Theorem 4] is still an exciting result, as [1] is a concurrent work. While it is a difficult decision, I will maintain my current weak acceptance, with a tendency toward a borderline acceptance. --- [1] Yuki Takezawa, Han Bao, Ryoma Sato, Kenta Niwa, Makoto Yamada. Polyak Meets Parameter-free Clipped Gradient Descent. arXiv. May 2024. [2] Bernd Gärtner, Niao He, Martin Jaggi. Optimization for Data Science Lecture Notes, FS 23. June 2023.

Authorsrebuttal2024-08-13

Thank you for following-up. **Smoothness along the trajectory is sufficient for the convergence of Gradient Descent type algorithms** Thank you for pointing to the specific reference of Lemma 3.7 [2]. Lemma 3.7 mentions that the typical descent criteria "already holds if $f$ is smooth with parameter $L$ over the line segment connecting $x_t$ and $x_{t+1}$." But their proof technique for the convergence of gradient descent (Theorem 3.8) does not hold for local/path-wise (here depends on the specific definition) or directional smoothness. More broadly, we are unaware of any such work that establishes proofs for GD with path-wise directional smoothness. In the optimization community, there is an understanding the GD is a local algorithm, and should depend on local smoothness. But that not the same as determining the proper definition of local/directional smoothness, and writing down the proofs of convergence. **It seems to be more of an existence result rather than a showcase of the applicability of Directional Smoothness to different algorithms** We outline below the proofs for two accelerated results and the specific guarantees they obtain. We believe these two new theorems illustrate the power of directional smoothness for refining the analyses of existing algorithms and go beyond an existence result. In Equation 2.2.7 (General Scheme of Optimal Method), Nesterov [N18] shows that the only property required of the primal update for acceleration is that the descent lemma holds (Eq. (c) in 2.2.7). If we have an adapted step-size $\eta_k \leq 1/M(x_{k+1}, x_k)$, then the directional descent lemma implies, $$ f(x_{k+1}) \leq f(x_k) - \frac{\eta_k}{2} \\|\nabla f(x_k)\\|_2^2. $$ By replacing Eq. (c) with this inequality and changing Eq. (a) to be $$ \alpha_k^2/\eta_k = (1 - \alpha_k)\gamma_k + \alpha_k \mu, $$ we obtain the same guarantee as Theorem 2.2.1 [N18], but with a modified $\lambda_k$. In particular, $\lambda_k$ now depends on the directional smoothness (rather than on the global smoothness $L$) through the $\alpha_k$ sequence. If $\mu > 0$ and $\gamma_0 = \mu$, then $\gamma_{k} = \mu$ for every $k$, $\alpha_k = \sqrt{\mu \eta_k}$, and thus $\lambda_k = \prod_{i=0}^k (1 - \sqrt{\mu \eta_i})$. As a result, we obtain the following guarantee: **Theorem**: (Strongly-convex acceleration) If $f$ is $\mu > 0$ strongly-convex and $\eta_k \leq 1/M(x_{k+1}, x_k)$ for every $k$, then GD with Nesterov acceleration converges at the following rate: $$ f(x_k) - f(x^*) \leq \left\[\prod_{i=0}^k (1 - \sqrt{\mu \eta_i})\right\]\left(f(x_0) - f(x^*) + \frac{\mu}{2}\\|x_0 - x^*\\|_2^2\right) $$ This is at least as fast the standard strongly-convex accelerated convergence rate. On the other hand, if $\gamma_0 \leq 3 \max_i M_i$, then by slightly modifying Nesterov's Lemma 2.2.4 allows us to get $\lambda_k \leq \frac{4 \max_i M_i}{\gamma_0 (k+1)^2}$, which let's us prove the following rate. **Theorem**: (Convex acceleration) If $f$ is convex, $\gamma_0 \leq 3 \max_i M_i$, and $\eta_k \leq 1/M(x_{k+1}, x_k)$ for every $k$, then GD with Nesterov acceleration converges at the following rate: $$ f(x_k) - f(x^*) \leq \frac{4 \max_i M_i}{\gamma_0 (k+1)^2}\left(f(x_0) - f(x^*) + \frac{\mu}{2}\\|x_0 - x^*\\|_2^2\right). $$ Once again, this rate is at last as fast as the standard convex accelerated convergence rate.

Reviewer vUdK2024-08-14

Dear Authors I acknowledge your response and will take it into consideration for the discussion phase.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC