Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Dimension-adapted Momentum Outscales SGD

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The authors show that on power-law random feature models, a dimension- and data-adapted momentum schedule (DANA) improves the loss exponents of SGD in all training regimes $t=d^{\ell}$ with $0<\ell<2\alpha$ when $2\alpha>1$, while…

desk verdict DANA provably changes scaling exponents in a solvable model, but the headline theorem is proven one level removed from the actual algorithm; worth refereeing with the gap addressed. read the letter →

arxiv 2505.16098 v1 pith:BXOA6LOW submitted 2025-05-22 stat.ML cs.LGmath.OC

classification stat.MLcs.LGmath.OC MSC 68Q3260B2068W40
keywords scalinglawsstochasticmomentumNesterovaccelerationrandomfeaturespowerlawdatacompute-optimallossexponentshyperparameter
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's central claim is that the scaling-law exponents of stochastic gradient descent are not fixed by the problem alone: a momentum schedule tuned to the model dimension and data complexity can provably change them, whereas classical constant momentum cannot. Working on the power-law random features (PLRF) model, the authors prove that SGD with a Nesterov-type schedule $\Delta(t)=\delta(1+t)^{-1}$ and a momentum learning rate $\gamma_3$ scaled as a power of $d$ and $t$, the DANA rule, outscales SGD whenever $2\alpha>1$ and $\beta\le\alpha+1$, with four distinct loss-curve shapes across the $(\alpha,\beta)$ phase plane. DANA-decaying, with $\kappa_3=1/(2\alpha)$, improves the loss exponent for every training regime $t=d^{\ell}$ with $0<\ell<2\alpha$; DANA-constant improves it for $1<\ell<2\alpha$; and SGD-M yields exactly the SGD exponents. If true, this means optimizer choice can enter neural scaling laws as an exponent, not just a constant, and compute-optimal parameter counts need not follow the $d\asymp f^{1/2}$ Chinchilla rule.

What carries the argument

The load-bearing object is the pair of deterministic-equivalent measures $\mu_F$ and $\mu_K$, pure-point measures built from the fixed-point equation (33) for the resolvent of the projected covariance, together with the simplified ODE system (43) for the per-eigenmode moments $(\rho_j^2,\xi_j^2,\chi_j)$. Solving these ODEs turns the expected loss into a Volterra equation $P(t)=F(t)+\int_0^t K_t(s)P(s)\,ds$, whose forcing and kernel functions have known power-law asymptotics. The decisive mechanism is the DANA time change $\vartheta(t)=1+2\gamma_2 Bt+(\int_0^t\sqrt{\gamma_3(s)B}\,ds)^2$: it reparameterizes the SGD loss curve so that momentum changes the argument at which the bias and variance terms are evaluated. DANA-decaying's schedule $\gamma_3(t)\asymp(1+t)^{-1/(2\alpha)}$ is the concrete expression of the effective dimension $(tB)^{-1/(2\alpha)}$, the number of feature-space directions that have been substantially updated by time $t$.

What would settle it

On PLRF with fixed $(\alpha,\beta)$ such that $2\alpha>1$, for example $\alpha=1.4$, $\beta=1.0$, run DANA-decaying with $\kappa_3=1/(2\alpha)$, $\gamma_2=1/(2\mathrm{tr})$, and $\delta$ large against SGD with batch size $1$ across $d=100\cdot 2^i$, and fit the empirical loss exponent in the regime $t=d$. If the DANA exponent is not strictly larger than SGD's beyond the reported fitting error, the central claim fails; alternatively, large-$d$ divergence between discrete runs and the simplified ODE system (43) would invalidate the model.

Watch

Extended reading notes

Core claim

On the PLRF model, a random-features least-squares problem whose data-coordinate and target-coordinate decays are parameterized by $\alpha$ and $\beta$, the expected loss of any algorithm in the general momentum class (Gen-Mom-SGD) is shown to follow a Volterra equation whose forcing and kernel terms are controlled by two explicit measures, $\mu_F$ and $\mu_K$, arising from a deterministic equivalent of the random feature covariance. The main theorem states that under $2\alpha+2\beta>1$, $\alpha>1/4$, $\beta\le\alpha+1$, fixed batch size, and stability conditions, the loss obeys $P(t)\asymp \hat F(\vartheta(t))+\gamma \hat K_{pp}(\vartheta(t))$ with the time change $\vartheta(t)=1+2(\gamma_2+\gamma_3/\delta)Bt$ for SGD-M and $\vartheta(t)=1+2\gamma_2Bt+(\int_0^t\sqrt{\gamma_3(s)B}\,ds)^2$ for DANA. Consequently SGD-M has identical scaling-law exponents to SGD, since its effective learning rate is $\gamma_2+\gamma_3/\delta$, while the DANA time change accelerates the argument of the forcing function. With $\gamma_3(t)\asymp(1+t)^{-1/(2\alpha)}$, DANA-decaying outscales SGD for all regimes $t=d^{\ell}$, $0<\ell<2\alpha$, in the $2\alpha>1$ region, and the authors derive explicit compute-optimal parameter, loss, and data exponents showing that DANA's compute-optimal curve is not $d\asymp f^{1/2}$.

Load-bearing premise

The scaling-law exponents are proved for a simplified differential-equation model with a deterministic stand-in for the random covariance, not for the exact discrete stochastic momentum algorithm; if that reduction is unfaithful in some training regime, the predicted exponents may not describe the actual algorithm.

Editorial extensions

If this is right

  • DANA-decaying with $\kappa_3=1/(2\alpha)$ gives strictly better loss exponents than SGD in every training regime $t=d^{\ell}$, $0<\ell<2\alpha$, for $2\alpha>1$, and is more sample-efficient and compute-efficient than both SGD and DANA-constant.
  • SGD-M with any fixed momentum $\delta$ reproduces the SGD exponents exactly; its only effect is to replace $\gamma_2$ by the larger effective learning rate $\gamma_2+\gamma_3/\delta$.
  • DANA-constant, with $\kappa_3=0$ and $\gamma_3\asymp\gamma_2/d$, behaves like SGD for $t\le d$ and outscales SGD for $1<\ell<2\alpha$, a regime that exists only when $2\alpha>1$.
  • At compute-optimality for $2\alpha>1$, DANA-decaying never uses $d_\star(f)\asymp f^{1/2}$; depending on the phase it undertrains or overtrains relative to the Chinchilla ratio, and it lowers the compute-optimal loss exponent.
  • On LSTMs trained on text, DANA-decaying's empirical loss exponents vary smoothly with $\kappa_3$, recover the predicted divergent, outscaling, and SGD-like regimes, and improve on SGD near $\kappa_3\approx0.7$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The effective-dimension interpretation suggests a test outside quadratics: estimate $\alpha$ from the spectral decay of a real dataset's covariance, set $\kappa_3=1/(2\alpha)$, and check whether the loss-exponent improvement persists; the paper treats this only as an open question.
  • If DANA's gain carries over to preconditioned or adaptive optimizers, then Adam plus a DANA-style decaying momentum term could change exponents on real networks; the paper lists this as future work, so it is an extension, not a claim.
  • The result implies empirical compute-optimal frontier fits should report the optimizer: two algorithms that differ only in momentum can have different $d_\star(f)$ exponents, so a single Chinchilla exponent is not an algorithm-independent constant.
  • For $2\alpha<1$, all analyzed momentum variants collapse to SGD scaling, so the practical window for outscaling is precisely the regime where data power-law decay is strong enough; whether another algorithm can cross the high-dimensional line remains open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This paper studies scaling laws for stochastic momentum methods on the power law random features (PLRF) model. The authors derive a Poissonized ODE system (22) and an associated Volterra equation for the expected loss, then introduce a simplified ODE system (43) and replace the random spectral measures by deterministic equivalents (33). The main result, Theorem 3.1, states scaling laws of the form P(t) ≍ F(ϑ(t)) + γ Kpp(ϑ(t)) for SGD, classic momentum (SGD-M), and two DANA variants, with explicit time changes ϑ(t) and stability conditions. It is claimed that SGD-M matches SGD exponents while DANA-constant and DANA-decaying improve the loss exponents for 2α > 1, including in compute-optimal scaling. The theoretical predictions are tested on PLRF simulations and on LSTM language modeling (Fig. 2).

Significance. If the theorem applied to the exact discrete algorithms, this would be a significant contribution: it would show that dimension- and data-dependent momentum schedules can change the exponent of scaling laws, not merely the constant, and would provide principled compute-optimal training regimes that differ from the Chinchilla d⋆ ≍ f^{1/2} law. The derivation has a notable strength: the time change ϑ(t) in (9) contains no fitted constants, and the loss exponents are derived from the model rather than fitted to simulation, so the theory makes falsifiable predictions. The paper is also transparent about its two main reductions, the simplified ODE and the deterministic equivalent, and supports them with extensive numerical studies, including a large-scale LSTM experiment. The main weakness is that the proven scaling laws apply to the simplified ODE with the deterministic equivalent, not to the exact discrete algorithm as stated in the introduction, and some theorem hypotheses appear inconsistent with the broad claims made in the main text.

major comments (4)
  1. [Theorem 3.1, Remark C.3, §3.1] The scaling-law statements in Theorem 3.1 are proven for the solution of the simplified Volterra equation (58), which combines the simplified ODEs (43) with the deterministic equivalent measures (33)/(36), rather than for the exact discrete Gen-Mom-SGD algorithm introduced in (13). The exact ODE (22) contains additional terms of order Δ², γ3², γ1γ3, and λ², and Lemma C.1 justifies the simplified ODE only as a non-degenerate high-dimensional limit in which learning rates are scaled inversely with dimension; Remark C.3 explicitly states that the link between dropping those terms and the finite learning-rate regime used here (e.g., γ2 on the order of 1 for 2α > 1) is left for future work. The deterministic equivalent is also supported only numerically (Fig. 3c). Because the central outscaling claim is about the algorithm class Gen-Mom-SGD, the theorem as stated does not directly prove the headline claim; the PLRF experiments (e.g., Fig. 8) validate the simplified ODE against direct simulation, but do not compare the exact ODE (22) with the simplified ODE (43), so this key reduction is not tested. I recommend that the authors either prove the reduction, add a direct numerical comparison of (22) and (43) across the regimes used in the scaling laws, or explicitly restate the theorem's conclusions as being about the simplified model rather than the discrete algorithm.
  2. [Theorem 3.1, §4, Remark B.4] The recommended and headline choice for DANA-decaying is κ3 = 1/(2α) (Remark B.4, Section 4, Fig. 3), but the summarized Theorem 3.1 imposes the additional requirement (1/2α) < κ3 < 1 for DANA-decaying, which excludes κ3 = 1/(2α). The stability conditions listed just above the theorem do include a branch for κ3 ≥ 1/(2α), so the exclusion appears to be a mismatch rather than a deliberate boundary issue. If Theorem I.2 in the appendix covers the endpoint κ3 = 1/(2α), the main text needs to state that; otherwise the paper's optimal-schedule outscaling claim is not supported by the proved result. This is load-bearing because all of the DANA-decaying claims in the title and abstract rest on the endpoint schedule.
  3. [Theorem 3.1, §4, Fig. 6] Theorem 3.1 assumes β ≤ α + 1, yet the paper claims DANA-decaying outscales SGD for all (α, β) with 2α > 1 and presents compute-optimal experiments in Phase IIIa with α = 1.4, β = 3.0 (Fig. 6), which violates β ≤ α + 1 = 2.4. The authors note in Section C.4 that they believe α > 1/4 and α + 1 > β are proof artifacts, but as stated the theorem does not cover the regime shown in their own figure. Either the theorem should be extended to β > α + 1, or the claim of universal outscaling for 2α > 1 should be flagged as conjecture with a supporting argument.
  4. [Appendix B.2] The stability conditions in Theorem 3.1 are derived from the simplified ODEs, but the authors themselves warn that these can miss stability conditions of the original algorithm. In Appendix B.2, for SGD-M they observe that the simplified Volterra equation neglects Δ² terms, which makes any Δ > 0 appear convergent, and that the coin-flip ODEs (49) give an explicit condition 2δ(2−δ)/(2(2B+1)+δ(3B+1)) > γ3 > 0 with δ ∈ (0, 2). Since stability is a hypothesis of the scaling-law theorem, this discrepancy means the theorem may describe a regime that the actual algorithm does not enter. The LSTM sweep in Fig. 2a shows a divergence boundary near κ3 = 0.6, indicating the practical importance of getting the stability boundary right. The authors should reconcile the simplified-ODE stability conditions with the exact or coin-flip conditions, or at least verify by simulation that the stable set of the discrete algorithm contains the stable set claimed in the theorem.
minor comments (4)
  1. [Footnote 1] The definition of ≍ says the constants c and C are independent of d, but the scaling-law statements require comparison of functions of both t and d with constants uniform over the stated t-regimes; the definition should specify the full uniformity, e.g., independent of d and uniform in t over each scaling regime.
  2. [(5), (39), Theorem 3.1] The notation P(t) is used ambiguously: in (5) it is the expected loss of the discrete iterates, while in (39) and (58) it is the solution of the deterministic Volterra equation; Theorem 3.1 should state explicitly which object the ≍ bound refers to (Remark C.4 partially clarifies this, but not at the theorem statement).
  3. [Abstract, §3.1, Fig. 11] The abstract and Section 3 mention 'four distinct loss curve shapes', but the paper's phase diagrams (Fig. 11) contain subphases Ia, Ib, Ic, IIa, IIb, IIIa, IIIb, IVa, and IVb; the terminology should be aligned so that the reader can map the four shapes to the full phase structure.
  4. [Fig. 5] The colorbar label 'logd(time to reach irreducible loss)' is not informative about the base of the logarithm and whether it is per-algorithm or per-problem; please clarify the quantity being plotted.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the DANA exponents are derived from the PLRF Volterra/ODE model rather than fitted to the predicted loss exponents; the acknowledged model reductions are correctness risks, not circular steps.

full rationale

The paper's central derivation chain is self-contained in the sense required for a circularity audit: the scaling-law exponents and the outscaling claim are logical consequences of Theorem 3.1, which is proved from the simplified ODE system (43), the deterministic-equivalent measures (36)/(33), and the Volterra reduction (58). The DANA schedule is not fitted to the loss exponents it is said to predict. The hyperparameter κ3 = 1/(2α) is motivated by a spectral 'effective dimension' heuristic (Section 3, 'effective dimension at iteration t ≍ (tB)^{-1/(2α)}') and then independently proved, via Theorem 3.1 and the appendix, to produce better exponents; the paper's own sweeps in Fig. 3 are empirical confirmation, not the source of the theorem. Likewise, the constants γ2 = 1/(2 tr) and γ3 = (1/5)γ2 are chosen for stability and enter only as constants, not as fitted parameters in the exponent calculation. The comparison against LSTM loss exponents (Fig. 2) is a genuine external prediction: the theory fixes the relationship between κ3 and the expected improvement from the PLRF model, and the measured exponents are fit from runs, not fed back into the derivation. The only load-bearing imported result is the deterministic-equivalent analysis of the power-law random features covariance, attributed to [79] (whose authors overlap with the current paper). This is not circular: [79] is a separate published result with its own derivation, and the current paper's Appendix D strengthens and uses it to bound the measures µF and µK; the target result (DANA outscaling for the stable simplified dynamics) is not equivalent to that input by construction. The paper transparently flags two limitations: Remark C.3 states that the link between dropping higher-order ODE terms and the non-degenerate high-dimensional limit is left for future work, and the passage near Eq. (33) states that the PLRF deterministic-equivalent proof falls outside textbook RMT and is supported numerically (Fig. 3c). These are unproved modeling reductions that create correctness risk for the claim as a statement about the exact discrete algorithm, but they are not cases where a prediction reduces to a fitted input or to a self-citation by definition. No step in the derivation equates the predicted quantity with an input by construction, and therefore the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the PLRF model assumptions and on two modeling reductions: the simplified ODE system and the deterministic equivalent. The hand-chosen hyperparameters (κ3, κ2, δ, γ̃3/γ2) are part of the algorithm design rather than fitted from data, but they are free constants the user must set. No new physical entities are introduced.

free parameters (4)
  • k3 (DANA-decaying time exponent) = 1/(2α) for optimality; theorem holds for κ3 in (1/(2α),1)
    Chosen to match the effective dimension (tB)^(-1/(2α)); the optimal value is set by the model's spectral decay, not by external benchmarks. This is a hand-chosen schedule exponent.
  • k2 (DANA-constant dimension exponent) = 1 + κ1, with κ1 = max{0, 1-2α}
    DANA-constant uses γ3 ≍ γ2 * 1/d, the largest stable constant γ3 is d^{-1}; this choice is derived from stability considerations but is still a designed hyperparameter.
  • Ratio γ̃3/γ2 = 1/5 (recommended), c2 = 0.5/Tr(D) in experiments
    Hand-chosen small constants that satisfy the sufficient stability conditions. The exact values do not affect the exponents, but they are free constants in the algorithm recipe.
  • Momentum decay constant δ = δ = 8 for LSTM; δ > 4 max{(2α+2β-1)/α, 4-1/α} for PLRF
    A sufficient condition for stability and acceleration is derived, but the specific numerical value (e.g., 8) is chosen by hand and used in experiments.
assumptions (5)
  • domain assumption Power-law data and target (Assumption 1): data coordinates scale as j^{-α} and target coefficients as j^{-β}.
    Defines the PLRF model; all theoretical results are for this specific data model.
  • domain assumption Zero initialization (Assumption 2): θ0 = 0 and y_{-1} = 0.
    Used throughout the derivation of the ODE and Volterra equations; the scaling laws depend on this initialization.
  • ad hoc to paper Simplified ODEs (43) accurately model the risk curves of the discrete Gen-Mom-SGD algorithms.
    All theorems are proven for the simplified ODEs, and the exact-to-simplified link is only numerically validated; Remark C.3 leaves the link for future work.
  • ad hoc to paper Deterministic equivalent (33) with measures μF, μK correctly represents the random matrix spectral measures.
    Scaling laws are proven for the deterministic equivalent, not for the actual random covariance; the paper states this and gives numerical agreement in Fig. 3c.
  • domain assumption Theorem conditions: 2α+2β>1, α>1/4, β≤α+1, and parameters not on critical lines.
    These restrictions are stated in Theorem 3.1 and used in the proof of the scaling laws; they define the region where the theorem applies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dimension-adapted Momentum Outscales SGD." pith.science (2026). https://pith.science/paper/BXOA6LOW

@misc{pith2026250516098,
  author       = {Pith},
  title        = {Pith review of: Dimension-adapted Momentum Outscales SGD},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BXOA6LOW}},
  note         = {Machine review of arXiv:2505.16098}
}
read the original abstract

We investigate scaling laws for stochastic momentum algorithms with small batch on the power law random features model, parameterized by data complexity, target complexity, and model size. When trained with a stochastic momentum algorithm, our analysis reveals four distinct loss curve shapes determined by varying data-target complexities. While traditional stochastic gradient descent with momentum (SGD-M) yields identical scaling law exponents to SGD, dimension-adapted Nesterov acceleration (DANA) improves these exponents by scaling momentum hyperparameters based on model size and data complexity. This outscaling phenomenon, which also improves compute-optimal scaling behavior, is achieved by DANA across a broad range of data and target complexities, while traditional methods fall short. Extensive experiments on high-dimensional synthetic quadratics validate our theoretical predictions and large-scale text experiments with LSTMs show DANA's improved loss exponents over SGD hold in a practical setting.

Figures

Figures reproduced from arXiv: 2505.16098 by the authors.

Figure 1
Figure 1. For empirical runs and deterministic ODE (7) simulations of PLRF, DANA outscales SGD while Schedule-Free SGD and Adam do not. (left) SGD & DANA fixed d, single run. Gray dashed line indicates transition (t ≍ 1/γ3 ≍ d) where DANA-constant(κ2 = 1) shifts from SGD-like behavior to acceleration. Deterministic ODE predictions (bold curves) match single runs of the stochastic algorithms (faded curves). (middle) Determinis… view at source ↗
Figure 2
Figure 2. DANA-decaying improves the loss exponent on LSTM language models. (left) Sweeping DANA-decaying κ3 shows stability and divergence similar to PLRF (Fig. 3a). (center) DANA κ3 = 0.7 maximizes the compute-optimal loss exponent and outscales SGD. (right) Compute-optimal loss exponents (top) and validation loss for final iterate (bottom) vs DANA κ3. All loss exponents ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (left) & (center) DANA sweeps of κ2, κ3. A κ bigger than (stability) ⇒ divergence after some t; DANA-decaying w/ κ3 = 1/(2α) is the envelope of the divergent algorithms. κ3 = 1/(2α), κ2 = 0 is optimal in DANA. (right) Deterministic equivalent µF (8) vs empirical density estimate match. See Sec. M for details. 2. The power law random features model (PLRF) In this work, we analyze a four-parameter model called power l… view at source ↗
Figures from the paper (38 more)
Figure 4
Figure 4. Figure 4: Example of loss curves (Phase II, α > β > 0.5). Understanding stability. The stability conditions listed in Theorem 3.1 are essentially sharp, in that we can show the loss curves are unbounded in d if the hyperparameters have lower bounds that are on the same order as …
Figure 5
Figure 5. Figure 5: Full DANA class, time to reach irreducible loss. 4. Using Momentum to outscale SGD SGD-M fails to outscale. The typical approach to momentum is to use a constant momentum ∆(t) ≡ δ, in practice usually set to 1 − δ = 0.9. Regardless of the choice of fixed ∆, SGD-M produ…
Figure 6
Figure 6. Figure 6: Comparison of SGD, DANA-constant, and DANA-decaying with compute￾optimal curve predictions. Numerical set-up: d = 100 × 2 i , i = 1, . . . , 15; Simplified ODEs (43) plotted for the scaling-law-equivalent model. DANA-decaying outscales in all phases (Phase Ia, II, III)…
Figure 7
Figure 7. Figure 7: Compute-optimal scaling laws (Phase IIa, (α > (3 + √ 5)/4, α > β > 0.5)). We reparameterize [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Loss curves match almost exactly between empirical PLRF experiments and ODE solutions. Compute-optimal loss exponents match approximately. (left) Loss curves for simplified ODEs (7) closely match empirical PLRF across 13 d-values up to d = 12,800. Empirical loss expone…
Figure 9
Figure 9. Figure 9: Comparison of SGD, DANA with other known algorithms (Schedule￾free, AcSGD, Nesterov). Numerical setup: SGD (blue curves) learning rate γ2 = 0.5/Tr(D), DANA-constant (green) has γ1(t) = 1, γ2(t) = 0.5/Tr(D), γ3(t) = 0.1/d, ∆(t) = δ/(1 + t) where δ = max{2 − 1/α,(2α + 2β…
Figure 10
Figure 10. Figure 10: Deterministic equivalent of the forcing function measure µF. Numerical set-up: 100 randomly generated Kˆ = DWWT D and the ρj ’s computed; for empirical density µF, 500 bins equal spaced on log scale from 10−8 to 1 and counted the number of λj that fall into each bin w…
Figure 11
Figure 11. Figure 11: Phase diagrams for the various momentum algorithms in the compute￾optimal regime. See [PITH_FULL_IMAGE:figures/full_fig_p067_11.png]
Figure 12
Figure 12. Figure 12: Cartoon plots of the scaling laws for each algorithm above the high￾dimensional line (Phase Ia, IIa, IIb). Pictures for the scaling laws for all three algorithms in each of the different phases. When t < d, DANA-constant behaves like SGD/SGD-M. Observe that the trade …
Figure 13
Figure 13. Figure 13: Cartoon plots of the scaling laws for each algorithm above the high￾dimensional line (Phase IIIa, IIIb). Pictures for the scaling laws for all three algorithms in each of the different phases. When t < d, DANA-constant behaves like SGD/SGD-M. Observe that the trade of…
Figure 14
Figure 14. Figure 14: Cartoon plots of the scaling laws for each algorithm below the high-dimensional line (Phase Ib/c, IVa/b). Pictures for the scaling laws for all three algorithms in each of the different phases. Observe that the scaling laws are all the same for every algorithm; Tradeo…
Figure 15
Figure 15. Figure 15: Comparison between SGD and SGD-M. Numerical set-up: Run SGD with constant γ SGD 2 (14) and SGD-M with ∆ ≡ 0.1 and constant γ SGD-M 3 (15) on the PLRF model with batch size 1. SGD learning rate γ SGD 2 ranges 0.1/Tr(Kˆ ) to 1.0/Tr(Kˆ ) (100 equally spaced points where …
Figure 16
Figure 16. Figure 16: SGD-M (and SGD) cartoon plots and phase diagram. The loss curve is the sum of P(t) ≍ Fpp(t)+Fac(t)+ 1 γB Kpp(t)+F0(t), where γ def = γ2 for SGD and γ def = γ2+ γ3 δ for SGD-M. Each of these terms are explicit and take the form t −σd −τ (See [PITH_FULL_IMAGE:figures/f…
Figure 17
Figure 17. Figure 17: Cartoon Plots DANA-constant. Pictures for the scaling laws for DANA￾constant in each of the different phases. When t < d, DANA-constant behaves like SGD/SGD￾M. Observe that the trade off point for compute-optimum changes across phases (see Sec. E for derivations and …
Figure 18
Figure 18. Figure 18: DANA-decaying shows similar stability and divergence behavior across κ3 on PLRF and LSTMs. (left) We repeat Figure 3a on PLRF where γ3(t) ≍ (1 + t) −κ3 . This shows DANA-decaying diverges when κ3 < 1/2α and degrades to SGD as κ3 goes towards 1.0. (right) We repeat Fig…
Figure 19
Figure 19. Figure 19: Sweeps of κ3 for DANA decaying show qualitatively similar behavior across model sizes. Small values of κ3 diverge, moderate values of κ3 outperform SGD, and large values of κ3 trend back towards SGD. Each panel shows one model size with SGD in black and DANA-decaying …
Figure 20
Figure 20. Figure 20: Final loss versus DANA-decaying κ3 showing the κ3 which optimizes final loss is fairly consistent across model sizes. The optimal value of κ3 for the final loss is between 0.55 to 0.75 on all models with a trend towards 0.75 for the largest model sizes. Color indicate…
Figure 21
Figure 21. Figure 21: Loss exponents for SGD and DANA-decaying across κ3. Loss curves across model sizes (black or light blue) and power law fits (orange) through the compute￾optimal frontiers showing the loss exponent measurements (with R2 values) for SGD and DANA-decaying. The first pane…
Figure 22
Figure 22. Figure 22: DANA-decaying loss exponents show improvement over SGD, and traverse the regimes seen in PLRF. (left) Loss exponent versus κ3 for DANA-decaying. The loss exponent for SGD is shown with a black dotted line. The loss exponents for DANA-decaying have higher magnitude ind…
Figure 23
Figure 23. Figure 23: The equivalence in risk dynamics between SGD and SGD-M holds approximately in LSTMs. We sweep different values of momentum (different line styles) and different effective learning rates where effective learning rate = learning rate / (1 - momentum). Color corresponds …
Figure 24
Figure 24. Figure 24: SGD loss curves and compute-optimal loss vs. theory on PLRF. For various (α, β), we plot the mean population risk for stochastic algorithm runs over 10-100 seeds (see individual figures) for SGD (solid lines) with γ2 = 0.375/Tr(D). Colors indicate dimensionality d ran…
Figure 25
Figure 25. Figure 25: SGD loss curves and compute-optimal loss vs. theory on PLRF. See [PITH_FULL_IMAGE:figures/full_fig_p171_25.png]
Figure 26
Figure 26. Figure 26: SGD loss curves and compute-optimal loss vs. theory on PLRF. See [PITH_FULL_IMAGE:figures/full_fig_p172_26.png]
Figure 27
Figure 27. Figure 27: DANA-constant loss curves and compute-optimal loss vs. theory on PLRF. For various (α, β), we plot the mean population risk for stochastic algorithm runs over 10-100 seeds (see individual figures) for DANA-constant (solid lines) with γ1 = 1, γ2 = 0.375 Tr(D) , γ3 = 0.…
Figure 28
Figure 28. Figure 28: DANA-constant loss curves and compute-optimal loss vs. theory on PLRF. See [PITH_FULL_IMAGE:figures/full_fig_p174_28.png]
Figure 29
Figure 29. Figure 29: DANA-constant loss curves and compute-optimal loss vs. theory on PLRF. See [PITH_FULL_IMAGE:figures/full_fig_p175_29.png]
Figure 30
Figure 30. Figure 30: DANA-decaying loss curves and compute-optimal loss vs. theory on PLRF. For various (α, β), we plot the mean population risk for stochastic algorithm runs over 10-100 seeds (see individual figures) for DANA-decaying (solid lines) with γ1 = 1, γ2 = 0.375 Tr(D) , γ3(t) =…
Figure 31
Figure 31. Figure 31: DANA-decaying loss curves and compute-optimal loss vs. theory on PLRF. See [PITH_FULL_IMAGE:figures/full_fig_p177_31.png]
Figure 32
Figure 32. Figure 32: DANA-decaying loss curves and compute-optimal loss vs. theory on PLRF. See [PITH_FULL_IMAGE:figures/full_fig_p178_32.png]
Figure 33
Figure 33. Figure 33: SGD Chinchilla Approach 1. Gray stars plot the parameter count of the model size that is optimal for each value of flops using the loss curves from [PITH_FULL_IMAGE:figures/full_fig_p179_33.png]
Figure 34
Figure 34. Figure 34: SGD Chinchilla Approach 1. Gray stars plot the parameter count of the model size that is optimal for each value of flops using the loss curves from [PITH_FULL_IMAGE:figures/full_fig_p180_34.png]
Figure 35
Figure 35. Figure 35: SGD Chinchilla Approach 1. Gray stars plot the parameter count of the model size that is optimal for each value of flops using the loss curves from [PITH_FULL_IMAGE:figures/full_fig_p181_35.png]
Figure 36
Figure 36. Figure 36: DANA-constant Chinchilla Approach 1. Gray stars plot the parameter count of the model size that is optimal for each value of flops using the loss curves from [PITH_FULL_IMAGE:figures/full_fig_p182_36.png]
Figure 37
Figure 37. Figure 37: DANA-constant Chinchilla Approach 1. Gray stars plot the parameter count of the model size that is optimal for each value of flops using the loss curves from [PITH_FULL_IMAGE:figures/full_fig_p183_37.png]
Figure 38
Figure 38. Figure 38: DANA-constant Chinchilla Approach 1. Gray stars plot the parameter count of the model size that is optimal for each value of flops using the loss curves from [PITH_FULL_IMAGE:figures/full_fig_p184_38.png]
Figure 39
Figure 39. Figure 39: DANA-decaying Chinchilla Approach 1. Gray stars plot the parameter count of the model size that is optimal for each value of flops using the loss curves from [PITH_FULL_IMAGE:figures/full_fig_p185_39.png]
Figure 40
Figure 40. Figure 40: DANA-decaying Chinchilla Approach 1. Gray stars plot the parameter count of the model size that is optimal for each value of flops using the loss curves from [PITH_FULL_IMAGE:figures/full_fig_p186_40.png]
Figure 41
Figure 41. Figure 41: DANA-decaying Chinchilla Approach 1. Gray stars plot the parameter count of the model size that is optimal for each value of flops using the loss curves from [PITH_FULL_IMAGE:figures/full_fig_p187_41.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. On the Effectiveness of the z-Transform Method in Quadratic Optimization

    cs.LG 2025-07 conditional novelty 6.0 of 10

    Using z-transforms and Tauberian theory, this paper derives sharp asymptotic rates, such as 1/k^omega for gradient descent and 1/k^{min(2omega, omega+1)} for Nesterov acceleration, for quadratic optimization in infini...

Reference graph

Works this paper leans on

138 extracted references · 57 canonical work pages · cited by 1 Pith paper

  1. [1]

    Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dandelion Mané, Rajat Monga, Sherry Moore, Derek Murra...

  2. [2]

    Katyusha: The first direct acceleration of stochastic gradient methods.The Journal of Machine Learning Research, 18(1):8194–8244, 2017

    Zeyuan Allen-Zhu. Katyusha: The first direct acceleration of stochastic gradient methods.The Journal of Machine Learning Research, 18(1):8194–8244, 2017

  3. [3]

    Linear coupling: An ultimate unification of gradient and mirror descent.arXiv preprint arXiv:1407.1537, 2014

    Zeyuan Allen-Zhu and Lorenzo Orecchia. Linear coupling: An ultimate unification of gradient and mirror descent.arXiv preprint arXiv:1407.1537, 2014

  4. [4]

    From high-dimensional & mean-field dynamics to dimensionless odes: A unifying approach to sgd in two-layers networks

    Luca Arnaboldi, Ludovic Stephan, Florent Krzakala, and Bruno Loureiro. From high-dimensional & mean-field dynamics to dimensionless odes: A unifying approach to sgd in two-layers networks. InThe Thirty Sixth Annual Conference on Learning Theory (COLT), pages 1199–1227. PMLR, 2023

  5. [5]

    Springer, 2003

    Søren Asmussen, Soren Asmussen, and Sren Asmussen.Applied probability and queues, volume 2. Springer, 2003

  6. [6]

    On the Convergence of Nesterov’s Accelerated Gradient Method in Stochastic Settings

    Mahmoud Assran and Michael Rabbat. On the Convergence of Nesterov’s Accelerated Gradient Method in Stochastic Settings. InProceedings of the 37th International Conference on Machine Learning (ICML), 2020

  7. [7]

    Two-Point Deterministic Equivalence for Stochastic Gradient Dynamics in Linear Models.arXiv preprint arXiv:2502.05074, 2025

    Alexander Atanasov, Blake Bordelon, Jacob A Zavatone-Veth, Courtney Paquette, and Cengiz Pehlevan. Two-Point Deterministic Equivalence for Stochastic Gradient Dynamics in Linear Models.arXiv preprint arXiv:2502.05074, 2025

  8. [8]

    Courier Corporation, 2004

    Krishna B Athreya, Peter E Ney, and PE Ney.Branching processes. Courier Corporation, 2004

Show all 138 references
  1. [9]

    Robust Accelerated Gradient Methods for Smooth Strongly Convex Functions.SIAM Journal on Optimization, 30(1):717– 751, 2020

    Necdet Serhat Aybat, Alireza Fallah, Mert Gürbüzbalaban, and Asuman Ozdaglar. Robust Accelerated Gradient Methods for Smooth Strongly Convex Functions.SIAM Journal on Optimization, 30(1):717– 751, 2020

  2. [10]

    High-dimensional analysis of double descent for linear regression with random projections

    Francis Bach. High-dimensional analysis of double descent for linear regression with random projections. SIAM Journal on Mathematics of Data Science, 6(1):26–50, 2024

  3. [11]

    Explaining neural scaling laws.Proc

    Yasaman Bahri, Ethan Dyer, Jared Kaplan, Jaehoon Lee, and Utkarsh Sharma. Explaining neural scaling laws.Proc. Natl. Acad. Sci. USA, 121(27):Paper No. e2311878121, 8, 2024

  4. [12]

    Springer, 2010

    Zhidong Bai and Jack W Silverstein.Spectral analysis of large dimensional random matrices, volume 20. Springer, 2010

  5. [13]

    Crowdsourcing with enhanced data quality assurance: An efficient approach to mitigate resource scarcity challenges in training large language models for healthcare

    Prosanta Barai, Gondy Leroy, Prakash Bisht, Joshua M Rothman, Sumi Lee, Jennifer Andrews, Sydney A Rice, and Arif Ahmed. Crowdsourcing with enhanced data quality assurance: An efficient approach to mitigate resource scarcity challenges in training large language models for hea...

  6. [14]

    Bartlett, Philip M

    Peter L. Bartlett, Philip M. Long, Gábor Lugosi, and Alexander Tsigler. Benign overfitting in linear regression.Proc. Natl. Acad. Sci. USA, 117(48):30063–30070, 2020. 12 DIMENSION-ADAPTED MOMENTUM OUTSCALES SGD

  7. [15]

    On the fast convergence of minibatch heavy ball momentum.IMA Journal of Numerical Analysis, page drae033, 08 2024

    Raghu Bollapragada, Tyler Chen, and Rachel Ward. On the fast convergence of minibatch heavy ball momentum.IMA Journal of Numerical Analysis, page drae033, 08 2024

  8. [16]

    A Dynamical Model of Neural Scaling Laws

    Blake Bordelon, Alexander Atanasov, and Cengiz Pehlevan. A Dynamical Model of Neural Scaling Laws. InProceedings of the 41st International Conference on Machine Learning (ICML), volume 235 ofProceedings of Machine Learning Research, pages 4345–4382. PMLR, 2024

  9. [17]

    How Feature Learning Can Improve Neural Scaling Laws .International Conference on Learning Representations (ICLR), 2025

    Blake Bordelon, Alexander Atanasov, and Cengiz Pehlevan. How Feature Learning Can Improve Neural Scaling Laws .International Conference on Learning Representations (ICLR), 2025

  10. [18]

    On ill-posed problems and the method of conjugate gradients

    Helmut Brakhage. On ill-posed problems and the method of conjugate gradients. InInverse and Ill-Posed Problems, pages 165–175. Academic Press, 1987

  11. [19]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

  12. [20]

    Accelerated Linear Convergence of Stochastic Momentum Methods in Wasserstein Distances

    Bugra Can, Mert Gurbuzbalaban, and Lingjiong Zhu. Accelerated Linear Convergence of Stochastic Momentum Methods in Wasserstein Distances. InProceedings of the 36th International Conference on Machine Learning (ICML), volume 97, pages 891–901. Proceedings of Machine Learning Re...

  13. [21]

    Optimal rates for the regularized least-squares algorithm

    Andrea Caponnetto and Ernesto De Vito. Optimal rates for the regularized least-squares algorithm. Foundations of Computational Mathematics, 7:331–368, 2007

  14. [22]

    Learning with sgd and random features

    Luigi Carratino, Alessandro Rudi, and Lorenzo Rosasco. Learning with sgd and random features. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018

  15. [23]

    Dimension free ridge regression.arXiv preprint arXiv:2210.08571, 2022

    Chen Cheng and Andrea Montanari. Dimension free ridge regression.arXiv preprint arXiv:2210.08571, 2022

  16. [24]

    Coddington and Norman Levinson.Theory of ordinary differential equations

    Earl A. Coddington and Norman Levinson.Theory of ordinary differential equations. McGraw-Hill Book Company, Inc., New York-Toronto-London, 1955

  17. [25]

    Hitting the high-dimensional notes: an ODE for SGD learning dynamics on GLMs and multi-index models.Inf

    Elizabeth Collins-Woodfin, Courtney Paquette, Elliot Paquette, and Inbar Seroussi. Hitting the high-dimensional notes: an ODE for SGD learning dynamics on GLMs and multi-index models.Inf. Inference, 13(4):Paper No. iaae028, 107, 2024

  18. [26]

    The High Line: Exact Risk and Learning Rate Curves of Stochastic Adaptive Learning Rate Algorithms

    Elizabeth Collins-Woodfin, Inbar Seroussi, Begoña García Malaxechebarría, Andrew W Mackenzie, Elliot Paquette, and Courtney Paquette. The High Line: Exact Risk and Learning Rate Curves of Stochastic Adaptive Learning Rate Algorithms. InAdvances in Neural Information Processing...

  19. [27]

    Cambridge University Press, 2022

    Romain Couillet and Zhenyu Liao.Random Matrix Methods for Machine Learning. Cambridge University Press, 2022

  20. [28]

    Generalization Error Rates in Kernel Regression: The Crossover from the Noiseless to Noisy Regime.Advances in Neural Information Processing Systems (NeurIPS), 34, 2021

    Hugo Cui, Bruno Loureiro, Florent Krzakala, and Lenka Zdeborová. Generalization Error Rates in Kernel Regression: The Crossover from the Noiseless to Noisy Regime.Advances in Neural Information Processing Systems (NeurIPS), 34, 2021

  21. [29]

    (Accelerated) Noise-adaptive Stochastic Heavy-Ball Momentum.Transactions on Machine Learning Research (TMLR), 2025

    Anh Quang Dang, Reza Babanezhad Harikandeh, and Sharan Vaswani. (Accelerated) Noise-adaptive Stochastic Heavy-Ball Momentum.Transactions on Machine Learning Research (TMLR), 2025

  22. [30]

    On the interplay between data structure and loss function in classification problems.Advances in Neural Information Processing Systems (NeurIPS), 34:8506–8517, 2021

    Stéphane d’Ascoli, Marylou Gabrié, Levent Sagun, and Giulio Biroli. On the interplay between data structure and loss function in classification problems.Advances in Neural Information Processing Systems (NeurIPS), 34:8506–8517, 2021

  23. [31]

    The DeepMind JAX Ecosystem.http://github.com/google-deepmind, 2020

    DeepMind, IgorBabuschkin, KateBaumli, AlisonBell, SuryaBhupatiraju, JakeBruce, PeterBuchlovsky, David Budden, Trevor Cai, Aidan Clark, Ivo Danihelka, Antoine Dedieu, Claudio Fantacci, Jonathan Godwin, Chris Jones, Ross Hemsley, Tom Hennigan, Matteo Hessel, Shaobo Hou, Steven K...

  24. [32]

    The Road Less Scheduled

    Aaron Defazio, Xingyu Yang, Harsh Mehta, Konstantin Mishchenko, Ahmed Khaled, and Ashok Cutkosky. The Road Less Scheduled. InAdvances in Neural Information Processing Systems (NeurIPS), volume 37, pages 9974–10007, 2024

  25. [33]

    Dimension-free deterministic equiva- lents and scaling laws for random feature regression

    Leonardo Defilippis, Bruno Loureiro, and Theodor Misiakiewicz. Dimension-free deterministic equiva- lents and scaling laws for random feature regression. InAdvances in Neural Information Processing DIMENSION-ADAPTED MOMENTUM OUTSCALES SGD 13 Systems (NeurIPS), volume 37, pages...

  26. [34]

    Optimal Distributed Online Prediction Using Mini-Batches.Journal of Machine Learning Research (JMLR), 13(1), 2012

    Ofer Dekel, Ran Gilad-Bachrach, Ohad Shamir, and Lin Xiao. Optimal Distributed Online Prediction Using Mini-Batches.Journal of Machine Learning Research (JMLR), 13(1), 2012

  27. [35]

    Nonparametric stochastic approximation with large step-sizes

    Aymeric Dieuleveut and Francis Bach. Nonparametric stochastic approximation with large step-sizes. The Annals of Statistics, 44(4):1363 – 1399, 2016

  28. [36]

    NIST Digital Library of Mathematical Functions.https://dlmf.nist.gov/, Release 1.2.4 of 2025-03-

  29. [37]

    F. W. J. Olver, A. B. Olde Daalhuis, D. W. Lozier, B. I. Schneider, R. F. Boisvert, C. W. Clark, B. R. Miller, B. V. Saunders, H. S. Cohl, and M. A. McClain, eds

  30. [38]

    Continuized Accelerations of Deterministic and Stochastic Gradient Descents, and of Gossip Algorithms

    Mathieu Even, Raphaël Berthier, Francis Bach, Nicolas Flammarion, Hadrien Hendrikx, Pierre Gaillard, Laurent Massoulié, and Adrien Taylor. Continuized Accelerations of Deterministic and Stochastic Gradient Descents, and of Gossip Algorithms. In M. Ranzato, A. Beygelzimer, Y. D...

  31. [39]

    Scaling Exponents Across Parameterizations and Optimizers

    Katie E Everett, Lechao Xiao, Mitchell Wortsman, Alexander A Alemi, Roman Novak, Peter J Liu, Izzeddin Gur, Jascha Sohl-Dickstein, Leslie Pack Kaelbling, Jaehoon Lee, et al. Scaling Exponents Across Parameterizations and Optimizers. InInternational Conference on Machine Learni...

  32. [40]

    From Averaging to Acceleration, There is Only a Step-size

    Nicolas Flammarion and Francis Bach. From Averaging to Acceleration, There is Only a Step-size. In Proceedings of The 28th Conference on Learning Theory (COLT), volume 40 ofProceedings of Machine Learning Research, pages 658–695. PMLR, 03–06 Jul 2015

  33. [41]

    Stochastic Heavy Ball.arXiv preprints arXiv:1609.04228, 2016

    Sèbastien Gadat, Fabien Panloup, and Sofiane Saadane. Stochastic Heavy Ball.arXiv preprints arXiv:1609.04228, 2016

  34. [42]

    Rigorous dynamical mean-field theory for stochastic gradient descent methods.SIAM Journal on Mathematics of Data Science, 6(2):400–427, 2024

    Cedric Gerbelot, Emanuele Troiani, Francesca Mignacco, Florent Krzakala, and Lenka Zdeborova. Rigorous dynamical mean-field theory for stochastic gradient descent methods.SIAM Journal on Mathematics of Data Science, 6(2):400–427, 2024

  35. [43]

    Optimal stochastic approximation algorithms for strongly convex stochastic composite optimization I: A generic algorithmic framework.SIAM J

    Saeed Ghadimi and Guanghui Lan. Optimal stochastic approximation algorithms for strongly convex stochastic composite optimization I: A generic algorithmic framework.SIAM J. Optim., 22(4):1469–1492, 2012

  36. [44]

    Optimal stochastic approximation algorithms for strongly convex stochastic composite optimization, II: Shrinking procedures and optimal algorithms.SIAM J

    Saeed Ghadimi and Guanghui Lan. Optimal stochastic approximation algorithms for strongly convex stochastic composite optimization, II: Shrinking procedures and optimal algorithms.SIAM J. Optim., 23(4):2061–2089, 2013

  37. [45]

    On the resolvents of nonconvolution Volterra kernels.Funkcial

    Gustaf Gripenberg. On the resolvents of nonconvolution Volterra kernels.Funkcial. Ekvac., 23(1):83–95, 1980

  38. [46]

    Siegel, and Stephan Wojtowytsch

    Kanan Gupta, Jonathan W. Siegel, and Stephan Wojtowytsch. Nesterov acceleration despite very noisy gradients. InAdvances in Neural Information Processing Systems (NeurIPS), volume 37, pages 20694–20744, 2024

  39. [47]

    Deterministic equivalents for certain functionals of large random matrices.Ann

    Walid Hachem, Philippe Loubaton, and Jamal Najim. Deterministic equivalents for certain functionals of large random matrices.Ann. Appl. Probab., 17(3):875–930, 2007

  40. [48]

    Deep learning scaling is predictable, empirically

    Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017

  41. [49]

    Long short-term memory.Neural computation, 9(8):1735– 1780, 1997

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory.Neural computation, 9(8):1735– 1780, 1997

  42. [50]

    An empirical analysis of compute-optimal large language model training

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Ruther- ford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Os...

  43. [51]

    Lu, and Theodor Misiakiewicz

    Hong Hu, Yue M. Lu, and Theodor Misiakiewicz. Asymptotics of Random Feature Regression Beyond the Linear Scaling Regime.arXiv preprint arXiv:2403.08160, 2024

  44. [52]

    Accelerating Stochastic Gradient Descent for Least Squares Regression

    Prateek Jain, Sham Kakade, Rahul Kidambi, Praneeth Netrapalli, and Aaron Sidford. Accelerating Stochastic Gradient Descent for Least Squares Regression. InProceedings of the 31st Conference On Learning Theory (COLT), volume 75, pages 545–604, 2018. 14 DIMENSION-ADAPTED MOMENTU...

  45. [53]

    Exploring the limits of language modeling.arXiv preprint arXiv:1602.02410, 2016

    Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the limits of language modeling.arXiv preprint arXiv:1602.02410, 2016

  46. [54]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alex Radford, Jeffrey Wu, and Dario Amodei

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alex Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020

  47. [55]

    On the Insufficiency of Existing Momentum Schemes for Stochastic Optimization

    Rahul Kidambi, Praneeth Netrapalli, Prateek Jain, and Sham Kakade. On the Insufficiency of Existing Momentum Schemes for Stochastic Optimization. In2018 Information Theory and Applications Workshop (ITA), pages 1–9, 2018

  48. [56]

    Adam: A method for stochastic optimization.International Conference on Learning Representations (ICLR), 2015

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.International Conference on Learning Representations (ICLR), 2015

  49. [57]

    Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing.arXiv preprint arXiv:1808.06226, 2018

    Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing.arXiv preprint arXiv:1808.06226, 2018

  50. [58]

    A Generic Acceleration Framework for Stochastic Composite Optimization

    Andrei Kulunchakov and Julien Mairal. A Generic Acceleration Framework for Stochastic Composite Optimization. InAdvances in Neural Information Processing Systems (NeurIPS), volume 32, 2019

  51. [59]

    A Lyapunov analysis for accelerated gradient methods: from deterministic to stochastic case

    Maxime Laborde and Adam Oberman. A Lyapunov analysis for accelerated gradient methods: from deterministic to stochastic case. InProceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics (AISTATS), volume 108, pages 602–612. Proceedings ...

  52. [60]

    Trajectory of mini-batch momentum: batch size saturation and convergence in high dimensions.Advances in Neural Information Processing Systems (NeurIPS), 35:36944–36957, 2022

    Kiwon Lee, Andrew Cheng, Elliot Paquette, and Courtney Paquette. Trajectory of mini-batch momentum: batch size saturation and convergence in high dimensions.Advances in Neural Information Processing Systems (NeurIPS), 35:36944–36957, 2022

  53. [61]

    Risk bounds of accelerated SGD for overparameterized linear regression.arXiv preprint arXiv:2311.14222, 2023

    Xuheng Li, Yihe Deng, Jingfeng Wu, Dongruo Zhou, and Quanquan Gu. Risk bounds of accelerated SGD for overparameterized linear regression.arXiv preprint arXiv:2311.14222, 2023

  54. [62]

    Zhenyu Liao, Romain Couillet, and Mahoney Mahoney. A Random Matrix Analysis of Random Fourier Features: Beyond the Gaussian Kernel, a Precise Phase Transition, and the Corresponding Double Descent.arXiv preprint arXiv:2006.05013, 2020

  55. [63]

    Kakade, Peter L

    Licong Lin, Jingfeng Wu, Sham M. Kakade, Peter L. Bartlett, and Jason D. Lee. Scaling Laws in Linear Regression: Compute, Parameters, and Data. InAdvances in Neural Information Processing Systems (NeurIPS), volume 37, pages 60556–60606. Curran Associates, Inc., 2024

  56. [64]

    Accelerating SGD with momentum for over-parameterized learning

    Chaoyue Liu and Mikhail Belkin. Accelerating SGD with momentum for over-parameterized learning. InProceedings of the 37th International Conference on Machine Learning (ICML), 2020

  57. [65]

    Liu, Roman Novak, Jaehoon Lee, Mitchell Wortsman, Lechao Xiao, Katie Everett, Alexander A

    Peter J. Liu, Roman Novak, Jaehoon Lee, Mitchell Wortsman, Lechao Xiao, Katie Everett, Alexander A. Alemi, Mark Kurzeja, Pierre Marcenac, Izzeddin Gur, Simon Kornblith, Kelvin Xu, Gamaleldin Elsayed, Ian Fischer, Jeffrey Pennington, Ben Adlam, and Jascha-Sohl Dickstein. Nanodo...

  58. [66]

    Optimal Algorithms in Linear Regression under Covariate Shift: On the Importance of Precondition.arXiv preprint arXiv:2502.09047, 2025

    Yuanshi Liu, Haihan Zhang, Qian Chen, and Cong Fang. Optimal Algorithms in Linear Regression under Covariate Shift: On the Importance of Precondition.arXiv preprint arXiv:2502.09047, 2025

  59. [67]

    Momentum and stochastic momentum for stochastic gradient, Newton, proximal point and subspace descent methods.Comput

    Nicolas Loizou and Peter Richtárik. Momentum and stochastic momentum for stochastic gradient, Newton, proximal point and subspace descent methods.Comput. Optim. Appl., 77(3):653–710, 2020

  60. [68]

    Learning curves of generic features maps for realistic datasets with a teacher-student model.Advances in Neural Information Processing Systems (NeurIPS), 34:18137–18151, 2021

    Bruno Loureiro, Cedric Gerbelot, Hugo Cui, Sebastian Goldt, Florent Krzakala, Marc Mezard, and Lenka Zdeborová. Learning curves of generic features maps for realistic datasets with a teacher-student model.Advances in Neural Information Processing Systems (NeurIPS), 34:18137–18...

  61. [69]

    Roberts, and James Sully

    Alexander Maloney, Daniel A. Roberts, and James Sully. A Solvable Model of Neural Scaling Laws. arXiv preprint arXiv:2210.16859, 2024

  62. [70]

    An empirical model of large-batch training.arXiv preprint arXiv:1812.06162, 2018

    Sam McCandlish, Jared Kaplan, Dario Amodei, and OpenAI Dota Team. An empirical model of large-batch training.arXiv preprint arXiv:1812.06162, 2018

  63. [71]

    The generalization error of random features regression: Precise asymptotics and the double descent curve.Communications on Pure and Applied Mathematics, 75(4):667–766, 2022

    Song Mei and Andrea Montanari. The generalization error of random features regression: Precise asymptotics and the double descent curve.Communications on Pure and Applied Mathematics, 75(4):667–766, 2022

  64. [72]

    Anisotropic random feature regression in high dimensions

    Gabriel Mel and Jeffrey Pennington. Anisotropic random feature regression in high dimensions. In International Conference on Learning Representations, 2021

  65. [73]

    Connections between Schedule-Free Optimizers, AdEMAMix, and Accelerated SGD Variants.arXiv preprint arXiv:2502.02431, 2025

    Depen Morwani, Nikhil Vyas, Hanlin Zhang, and Sham Kakade. Connections between Schedule-Free Optimizers, AdEMAMix, and Accelerated SGD Variants.arXiv preprint arXiv:2502.02431, 2025. DIMENSION-ADAPTED MOMENTUM OUTSCALES SGD 15

  66. [74]

    Yoonsoo Nam, Nayara Fonseca, Seok Hyeong Lee, Chris Mingard, and Ard A. Louis. An exactly solvable model for emergence and scaling laws in the multitask sparse parity problem. InAdvances in Neural Information Processing Systems (NeurIPS), volume 37, pages 39632–39693, 2024

  67. [75]

    Springer, 2004

    Yuri Nesterov.Introductory lectures on convex optimization. Springer, 2004

  68. [76]

    The Role of Memory in Stochastic Optimization

    Antonio Orvieto, Jonas Kohler, and Aurelien Lucchi. The Role of Memory in Stochastic Optimization. InProceedings of The 35th Uncertainty in Artificial Intelligence Conference, volume 115 ofProceedings of Machine Learning Research, pages 356–366, 2020

  69. [77]

    SGD in the Large: Average-case Analysis, Asymptotics, and Stepsize Criticality

    Courtney Paquette, Kiwon Lee, Fabian Pedregosa, and Elliot Paquette. SGD in the Large: Average-case Analysis, Asymptotics, and Stepsize Criticality. InProceedings of Thirty Fourth Conference on Learning Theory (COLT), volume 134, pages 3548–3626, 2021

  70. [78]

    Dynamics of stochastic momentum methods on large-scale, quadratic models.Advances in Neural Information Processing Systems (NeurIPS), 34:9229–9240, 2021

    Courtney Paquette and Elliot Paquette. Dynamics of stochastic momentum methods on large-scale, quadratic models.Advances in Neural Information Processing Systems (NeurIPS), 34:9229–9240, 2021

  71. [79]

    Homogenization of SGD in high-dimensions: exact dynamics and generalization properties.arXiv preprint arXiv:2205.07069, 2022

    Courtney Paquette, Elliot Paquette, Ben Adlam, and Jeffrey Pennington. Homogenization of SGD in high-dimensions: exact dynamics and generalization properties.arXiv preprint arXiv:2205.07069, 2022

  72. [80]

    4+3 Phases of Compute- Optimal Neural Scaling Laws

    Elliot Paquette, Courtney Paquette, Lechao Xiao, and Jeffrey Pennington. 4+3 Phases of Compute- Optimal Neural Scaling Laws. InAdvances in Neural Information Processing Systems (NeurIPS), volume 37, 2024

  73. [81]

    Statistical optimality of stochastic gradient descent on hard learning problems through multiple passes.Advances in Neural Information Processing Systems (NeurIPS), 31, 2018

    Loucas Pillaud-Vivien, Alessandro Rudi, and Francis Bach. Statistical optimality of stochastic gradient descent on hard learning problems through multiple passes.Advances in Neural Information Processing Systems (NeurIPS), 31, 2018

  74. [82]

    B.T. Polyak. Some methods of speeding up the convergence of iteration methods.USSR Computational Mathematics and Mathematical Physics, 04, 1964

  75. [83]

    Resolving Discrepancies in Compute-Optimal Scaling of Language Models

    Tomer Porian, Mitchell Wortsman, Jenia Jitsev, Ludwig Schmidt, and Yair Carmon. Resolving Discrepancies in Compute-Optimal Scaling of Language Models. InAdvances in Neural Information Processing Systems (NeurIPS), volume 37, pages 100535–100570, 2024

  76. [84]

    Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research, 21(140):1–67, 2020

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research, 21(140):1–67, 2020

  77. [85]

    Generalization properties of learning with random features

    Alessandro Rudi and Lorenzo Rosasco. Generalization properties of learning with random features. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017

  78. [86]

    Almost sure convergence rates for Stochastic Gradient Descent and Stochastic Heavy Ball

    Othmane Sebbouh, Robert M Gower, and Aaron Defazio. Almost sure convergence rates for Stochastic Gradient Descent and Stochastic Heavy Ball. InProceedings of Thirty Fourth Conference on Learning Theory, volume 134, pages 3935–3971. Proceedings of Machine Learning Research (PMLR), 2021

  79. [87]

    Shallue, J

    C.J. Shallue, J. Lee, J. Antognini, J. Sohl-Dickstein, R. Frostig, and G.E. Dahl. Measuring the Effects of Data Parallelism on Neural Network Training.Journal of Machine Learning Research, 20:1–49, 2019

  80. [88]

    Stochastic gradient descent for non-smooth optimization: Convergence results and optimal averaging schemes

    Ohad Shamir and Tong Zhang. Stochastic gradient descent for non-smooth optimization: Convergence results and optimal averaging schemes. InInternational conference on machine learning, pages 71–79. PMLR, 2013

  81. [89]

    Scaling Laws from the Data Manifold Dimension.Journal of Machine Learning Research (JMLR), 23(9):1–34, 2022

    Utkarsh Sharma and Jared Kaplan. Scaling Laws from the Data Manifold Dimension.Journal of Machine Learning Research (JMLR), 23(9):1–34, 2022

  82. [90]

    On the empirical distribution of eigenvalues of a class of large dimensional random matrices.Journal of Multivariate analysis, 54(2):175–192, 1995

    Jack W Silverstein and Zhi Dong Bai. On the empirical distribution of eigenvalues of a class of large dimensional random matrices.Journal of Multivariate analysis, 54(2):175–192, 1995

  83. [91]

    Simon, Dhruva Karkada, Nikhil Ghosh, and Mikhail Belkin

    James B. Simon, Dhruva Karkada, Nikhil Ghosh, and Mikhail Belkin. More is better in modern machine learning: when infinite overparameterization is optimal and overfitting is obligatory. InInternational Conference on Learning Representations (ICLR), 2024

  84. [92]

    Weijie Su, Stephen Boyd, and Emmanuel J. Candès. A Differential Equation for Modeling Nesterov’s Accelerated Gradient Method: Theory and Insights.Journal of Machine Learning Research, 17(153):1– 43, 2016

  85. [93]

    On the importance of initialization and momentum in deep learning

    Ilya Sutskever, James Martens, George Dahl, and Geoffrey Hinton. On the importance of initialization and momentum in deep learning. InProceedings of the 30th International Conference on Machine Learning (ICML), volume 28, pages 1139–1147, 2013

  86. [94]

    Accelerated sgd for non-strongly-convex least squares

    Aditya Varre and Nicolas Flammarion. Accelerated sgd for non-strongly-convex least squares. In Proceedings of Thirty Fifth Conference on Learning Theory (COLT), volume 135 ofProceedings of Machine Learning Research, pages 2062–2126, 2022. 16 DIMENSION-ADAPTED MOMENTUM OUTSCALES SGD

  87. [95]

    Last iterate convergence of SGD for Least-Squares in the Interpolation regime

    Aditya Vardhan Varre, Loucas Pillaud-Vivien, and Nicolas Flammarion. Last iterate convergence of SGD for Least-Squares in the Interpolation regime. InAdvances in Neural Information Processing Systems (NeurIPS), volume 34, pages 21581–21591, 2021

  88. [96]

    Fast and Faster Convergence of SGD for Over- Parameterized Models and an Accelerated Perceptron

    Sharan Vaswani, Francis Bach, and Mark Schmidt. Fast and Faster Convergence of SGD for Over- Parameterized Models and an Accelerated Perceptron. InProceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics (ICML), volume 89 ofProceeding...

  89. [97]

    Will we run out of data? Limits of LLM scaling based on human-generated data.arXiv preprint arXiv:2211.04325, 2022

    Pablo Villalobos, Anson Ho, Jaime Sevilla, Tamay Besiroglu, Lennart Heim, and Marius Hobbhahn. Will we run out of data? Limits of LLM scaling based on human-generated data.arXiv preprint arXiv:2211.04325, 2022

  90. [98]

    Re-examining Double Descent and Scal- ing Laws under Norm-based Capacity via Deterministic Equivalence.arXiv preprint arXiv:2502.01585, 2025

    Yichen Wang, Yudong Chen, Lorenzo Rosasco, and Fanghui Liu. Re-examining Double Descent and Scal- ing Laws under Norm-based Capacity via Deterministic Equivalence.arXiv preprint arXiv:2502.01585, 2025

  91. [99]

    A variational perspective on accelerated methods in optimization.proceedings of the National Academy of Sciences, 113(47):E7351–E7358, 2016

    Andre Wibisono, Ashia C Wilson, and Michael I Jordan. A variational perspective on accelerated methods in optimization.proceedings of the National Academy of Sciences, 113(47):E7351–E7358, 2016

  92. [100]

    Small-scale proxies for large-scale Transformer training instabilities

    Mitchell Wortsman, Peter J Liu, Lechao Xiao, Katie E Everett, Alexander A Alemi, Ben Adlam, John D Co-Reyes, Izzeddin Gur, Abhishek Kumar, Roman Novak, et al. Small-scale proxies for large-scale Transformer training instabilities. InThe Twelfth International Conference on Lear...

  93. [101]

    Last Iterate Risk Bounds of SGD with Decaying Stepsize for Overparameterized Linear Regression

    Jingfeng Wu, Difan Zou, Vladimir Braverman, Quanquan Gu, and Sham Kakade. Last Iterate Risk Bounds of SGD with Decaying Stepsize for Overparameterized Linear Regression. InProceedings of the 39th International Conference on Machine Learning (ICML), volume 162 ofProceedings of ...

  94. [102]

    A Unified Analysis of Stochastic Momentum Methods for Deep Learning

    Yan Yan, Tianbao Yang, Zhe Li Li, Qihang Lin, and Yi Yang. A Unified Analysis of Stochastic Momentum Methods for Deep Learning. InProceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 2955–2961. International Joint Confere...

  95. [103]

    Tensor programs iv: Feature learning in infinite-width neural networks

    Greg Yang and Edward J Hu. Tensor programs iv: Feature learning in infinite-width neural networks. InInternational Conference on Machine Learning, pages 11727–11737. PMLR, 2021

  96. [104]

    Tensor programs V: tuning large neural networks via zero-shot hyperparameter transfer

    Greg Yang, Edward J Hu, Igor Babuschkin, Szymon Sidor, Xiaodong Liu, David Farhi, Nick Ryder, Jakub Pachocki, Weizhu Chen, and Jianfeng Gao. Tensor programs V: tuning large neural networks via zero-shot hyperparameter transfer. InProceedings of the 35th International Conferenc...

  97. [105]

    Tensor programs ivb: Adaptive optimization in the infinite-width limit

    Greg Yang and Etai Littwin. Tensor programs ivb: Adaptive optimization in the infinite-width limit. arXiv preprint arXiv:2308.01814, 2023

  98. [106]

    Corner Gradient Descent.arXiv preprint arXiv:2504.12519, 2025

    Dmitry Yarotsky. Corner Gradient Descent.arXiv preprint arXiv:2504.12519, 2025

  99. [107]

    SGD with Memory: Fundamental Properties and Stochastic Acceleration.International Conference on Learning Representations (ICLR), 2025

    Dmitry Yarotsky and Maksim Velikanov. SGD with Memory: Fundamental Properties and Stochastic Acceleration.International Conference on Learning Representations (ICLR), 2025

  100. [108]

    Which Algorithmic Choices Matter at Which Batch Sizes? Insights From a Noisy Quadratic Model

    Guodong Zhang, Lala Li, Zachary Nado, James Martens, Sachdeva Sachdeva, George Dahl, Christopher Shallue, and Roger Grosse. Which Algorithmic Choices Matter at Which Batch Sizes? Insights From a Noisy Quadratic Model. InAdvances in Neural Information Processing Systems (NeurIP...

  101. [109]

    Hanlin Zhang, Depen Morwani, Nikhil Vyas, Jingfeng Wu, Difan Zou, Udaya Ghai, Dean Foster, and Sham M. Kakade. How Does Critical Batch Size Scale in Pre-training? .International Conference on Learning Representations (ICLR), 2025

  102. [110]

    Foster, and Sham Kakade

    Difan Zou, Jingfeng Wu, Vladimir Braverman, Quanguan Gu, Dean P. Foster, and Sham Kakade. The Benefits of Implicit Regularization from SGD in Least Squares Problems (NeurIPS). InAdvances in Neural Information Processing Systems, volume 34, pages 5456–5468, 2021

  103. [111]

    Benign overfitting of constant-stepsize SGD for linear regression.Journal of Machine Learning Research, 24(326):1–58, 2023

    Difan Zou, Jingfeng Wu, Vladimir Braverman, Quanquan Gu, and Sham M Kakade. Benign overfitting of constant-stepsize SGD for linear regression.Journal of Machine Learning Research, 24(326):1–58, 2023. DIMENSION-ADAPTED MOMENTUM OUTSCALES SGD 17 Appendix: Dimension-adapted Momen...

  104. [112]

    The power law random features model (PLRF) 4 18 DIMENSION-ADAPTED MOMENTUM OUTSCALES SGD

  105. [113]

    Scaling laws for SGD on the PLRF

    Continuized analysis of general stochastic momentum algorithms 4 3.1. Scaling laws for SGD on the PLRF. 5

  106. [114]

    hidden-manifold

    Using Momentum to outscale SGD 8 4.1. Compute-optimal regime. 8 References 11 Appendix A. Related Work 20 Appendix B. Additional Algorithm Set-up 23 B.1. Stochastic gradient descent (SGD) 23 B.2. Classic (constant) momentum (SGD-M) 26 B.3. Dimension-Adapted Nesterov Accelerati...

  107. [115]

    high-dimensional

    By hitting both sides byΣVT, we get that ΣV T VΣ T U T D1/2 ˙b= 0⇒ΣΣ T U T D1/2 ˙b= 0. Thus for all nonzero singular valuesσj (or equivalently for all nonzero eigenvaluesλj of ˇK = W T DW), uT j D1/2 ˙b= 0. Additionally, we have that W T D1/2uj =VΣ T U T uj =VΣ T ej =σ jV ej =...

  108. [116]

    We denoteγ def =γ 2 + γ3 δ for classic momentum andγ def =γ 2 for both DANA algorithms and SGD

    For various assumptions specific to the individual algorithms, there exists anM > 0large enough and constants ˜C(α, β, M,alg),˜c(α, β, M,alg)such that if γBt > Mthen: (62)˜c× F(t) + 1 γB K(t) ≤P(t)≤ ˜C× F(t) + 1 γB K(t) . We denoteγ def =γ 2 + γ3 δ for classic momentum andγ de...

  109. [117]

    Proposition D.2(Upper bound for large σ’s).Let α >0

    □ This implies a first bound on the mass ofµF for largeσ. Proposition D.2(Upper bound for large σ’s).Let α >0. Suppose2 α + 2β >1. Then there is a constant C(α, β)such that∀M >0,we have µF([M,+∞])≤C(α, β). Additionally, for anyϵ >0, there exists somed >0such that∀d≥ d,µ F([1 +...

  110. [118]

    There exists M, ˜C depending only onαsuch that for anyC∈[d −2αM, 1 M ], we have: µK([C,2C])≤ ˜C×C 2− 1 2α . Proof. Using Lemma D.2, we only need to estimate1 π Im((x + iη)(1 −m (x + iη))d)for x∈ [d−2αM, 1 M ]and η(x) = (log(1/ϵ)/c) max{x1+1/(2α), π 2α x1−1/(2α) d }. Applying [...

  111. [119]

    On the other hand , ifα > 3 4, we haveFpp(t) = Fac(t)for t≍f 1/2+α 3/2+α but the optimal is to taket⋆ the largest

    = f− (2α+2β−1) 3β+α . On the other hand , ifα > 3 4, we haveFpp(t) = Fac(t)for t≍f 1/2+α 3/2+α but the optimal is to taket⋆ the largest. This brings us to the second case. Fac(t) =F 0(t): We apply Lemma E.1 with γ0 = 2− 1 α , p 0 = 1 2α , γ 1 = 0,andp 1 = 2α. □ If α > 3 4, we ...

  112. [120]

    Ifα < 3 4, the optimal is to chooset⋆ the largest which brings us to the other case

    = f−1+ 1 4α. Ifα < 3 4, the optimal is to chooset⋆ the largest which brings us to the other case. Fac(t, d) =F0(t, d):In that case, we define γ0 = 2− 1 α , p 0 = 1 2α , γ 1 = 0,andp 1 = 2α. For α < 3 4, we have p1 −γ 1 > 0 > p0 −γ 0 and hence applying Lemma E.1, it brings an o...

  113. [121]

    For α >3 4, the optimal is to chooset⋆ the smallest going back to the first case

    = f− 2α α+3/2. For α >3 4, the optimal is to chooset⋆ the smallest going back to the first case. We conclude that forα >3 4, in Phase IIIa,d⋆ =d ⋆ 2 and forα < 3 4 in Phase IIIb,d⋆ =d ⋆ 1.□ E.3. DANA-decaying, compute-optimal curves In all this section, we will use the hyperpa...

  114. [122]

    However, if α > 3+ √ 5 4 then p1 −γ 1 < 0and p0 −γ 0 < 0

    ≍f − (2α+2β−1)(4α−1) 2(2α2 +4αβ−β) . However, if α > 3+ √ 5 4 then p1 −γ 1 < 0and p0 −γ 0 < 0. Hence the optimal is to chooset⋆ the largest which brings us to the second case. Fac( f d , d) =F0( f d , d):In that case we define γ0 = 2− 1 2α 1− 1 2α , p 0 = 1, γ 1 = 0,andp 1 = 2...

  115. [123]

    On the other hand, ifα < 3+ √ 5 4 , then p0 −γ 0 > 0, p1 −γ 1 > 0and the compute optimal is to taket⋆ the smallest, i.e

    ≍f − 2α(4α−1) 4α−1+4α2. On the other hand, ifα < 3+ √ 5 4 , then p0 −γ 0 > 0, p1 −γ 1 > 0and the compute optimal is to taket⋆ the smallest, i.e. going back to the first case. We conclude that forα >3+ √ 5 4 , in Phase IIIa,d ⋆ =d ⋆ 1 and for 3+ √ 5 4 in Phase IIIb,d⋆ =d ⋆ 2.□ ...

  116. [124]

    However, whenα >3+ √ 5 4 , then p0 −γ 0 < 0, p1 −γ 1 < 0and the optimal choice is to taket ⋆ the largest, leading to the second case

    ≍f − (4α−1)2 2α(6α−1). However, whenα >3+ √ 5 4 , then p0 −γ 0 < 0, p1 −γ 1 < 0and the optimal choice is to taket ⋆ the largest, leading to the second case. Fac( f d , d) =F0( f d , d):In that case we define γ0 = 2− 1 2α 1− 1 2α , p 0 = 1, γ 1 = 0,andp 1 = 2α. If α >3+ √ 5 4 ,...

  117. [125]

    On the other hand, ifα <3+ √ 5 4 , then p0 −γ 0 > 0, p1 −γ 1 > 0and the compute optimal is to taket⋆ the smallest, i.e

    ≍f − 2α(4α−1) 2(2α2 +2α−1/2). On the other hand, ifα <3+ √ 5 4 , then p0 −γ 0 > 0, p1 −γ 1 > 0and the compute optimal is to taket⋆ the smallest, i.e. going back to the first case. We conclude that forα >3+ √ 5 4 , in phase IIIa,d ⋆ =d ⋆ 1 and for 3+ √ 5 4 in Phase IIIb,d⋆ =d ⋆...

  118. [126]

    coin-flipping

    by the fact that DANA-c shifts the trade-off (later in training)for smallerαthan DANA-d.□ 66 DIMENSION-ADAPTED MOMENTUM OUTSCALES SGD E.5. Summary on compute-optimality results We provide some specific details about compute-optimality for each algorithm in the different phases...

  119. [127]

    Moreover, we have estimates on the Kesten constant∥K∥

    For anyε >0, there is anMsufficiently large so that forBt∈[M, d 2α/M], Z t 0 K(s)K(t−s)≤(2 +ε)∥ K∥K(t). Moreover, we have estimates on the Kesten constant∥K∥. Lemma F.4(Boundedness of∥ K∥, Corollary G.1 [79]).When2α >1and γ2(B+ 1)<2, ∥K∥= 1 2B ∞X j=1 j−2α 1− 1 2 γ2(B+ 1) (1 +o...

  120. [128]

    There exists some constantsC(α) > 0, ¯d(α) > 1 such that for δ >0, B∈N ∗, γ2 > 0, γ3 > 0, if γ3B≤ δ2 16 , γ2B≤ δ2 16, and d≥ ¯d(α), we have the bounds (note that for2α <1, we usev∼C×d) d(1−2α)+ C (γ2 + γ3 δ )≤ ∥K∥ ≤ ∥K∥ ≤Cd (1−2α)+ (γ2 + γ3 δ ). Proof.We will instead show d(1−...

  121. [129]

    There exists some constant c >0such that if max{γ2B, γ3B δ } ≤δ2 16 and γ2 + γ3 δ ≤cd −(1−2α)+, then for anyt≥ 0with (γ2 + γ3 δ )Bt≤M d2α we have K∗ K (t)≤ϵ K(t). Proof.We apply Propositions G.4 to G.6 to obtain that ift≥δ −1, K(t)≍(γ 2 2 + (γ3 δ )2)Bmin{1,((γ 2 + γ3 δ )Bt)−2+...

  122. [130]

    Hence, applying Lemma C.3 we can bound for someC >0 ∞X k=1 h F∗ K ∗ki (t)≤C× F∗ K (t)

    Using Proposition G.8 we know that forc >0small enough,∀t≥0if(γ 2 + γ3 δ )Bt≤M d2α, K∗ K (t)≤ϵ Kand∥ K∥<1. Hence, applying Lemma C.3 we can bound for someC >0 ∞X k=1 h F∗ K ∗ki (t)≤C× F∗ K (t). We only have left to respectively lower and upper bound[F∗K ] (t)and F∗ K (t). We w...

  123. [131]

    Additionally, for the second integral to converge (and the third integral to be negligible) we requireδ >2−1 α

    It’s not a problem since below we have bounds from [79] (see Proposition H.15). Additionally, for the second integral to converge (and the third integral to be negligible) we requireδ >2−1 α. DIMENSION-ADAPTED MOMENTUM OUTSCALES SGD 109 □ The caset≤ γ2 γ3 . Proposition H.14.Su...

  124. [132]

    The proof is very similar to the one of [79, Proposition H.4]

    There exists a constantC(α, β)such that ∀t≥0: Fac(t)≤C×F 0(t)if2α <1,2β >1 Fac(t) = 0if2β <1 Proof. The proof is very similar to the one of [79, Proposition H.4]. First, if2 β > 1, cβ = 0and hence ∀t≥0,F ac(t) = 0. Now, suppose that2α <1, 2β >1. From Lemma H.1, we know that∀σ ...

  125. [133]

    Then there exists a constantC(α, β, H)such that∀t≥0, 1 C (F0(t) +F ac(t) +F pp(t))≤F(t)≤C(F 0(t) +F ac(t) +F pp(t))

    Use Parametrization H.1 with δ >max{1,2 + 2β−1 α ,2− 1 α }and2δ /∈N. Then there exists a constantC(α, β, H)such that∀t≥0, 1 C (F0(t) +F ac(t) +F pp(t))≤F(t)≤C(F 0(t) +F ac(t) +F pp(t)). Proof. It is clear that for anyt≥ 0and on the range of σ’s where we can apply Proposition H...

  126. [134]

    Under Parametrization H.1, ifκ1 < (1 − 2α)+ or κ2 < κ1 + 1−κ b, then fordlarge enough, lim inf t≥0 Z t t/2 K(t, s) ds >1

    Let δ >max{1, 4 − 1 α }, 2δ /∈N. Under Parametrization H.1, ifκ1 < (1 − 2α)+ or κ2 < κ1 + 1−κ b, then fordlarge enough, lim inf t≥0 Z t t/2 K(t, s) ds >1. Proof.We first consider the case whereκ2 <2κ 1 + 2α−κ b. This ensures that √γ3B γ2B ≳d −α. 112 DIMENSION-ADAPTED MOMENTUM ...

  127. [135]

    Under Parametrization H.1, suppose κ1 < (1 − 2α)+ or κ2 < κ1 + 1 −κ b

    Let δ >max{1, 4− 1 α }, 2δ /∈ N. Under Parametrization H.1, suppose κ1 < (1 − 2α)+ or κ2 < κ1 + 1 −κ b. Then for d large enough, P(t) t→∞ → ∞. Proof.This is a direct consequence of Lemma H.3 and Lemma H.4□ H.9. Sufficient condition for stability: upper-bound on the kernel norm...

  128. [136]

    We will later heuristically extend the results in this section under this assumption to the general(DANA) algorithm in Section I.5

    We also supposeδ large enough (independent ofd). We will later heuristically extend the results in this section under this assumption to the general(DANA) algorithm in Section I.5. Remark I.1.We supposed α >1 2 as it will become clear that forα <1 2, this algorithm is equivale...

  129. [137]

    Suppose that 2ρ >max{2α+2β−1 α ,4− 1 α }

    Consider Parametrization I.1 withB = 1, κ≥ 0, γ2 > 0. Suppose that 2ρ >max{2α+2β−1 α ,4− 1 α }. We have • (Sufficient condition) For any ϵ > 0there exists g(κ, ϵ) > 0and d0 large enough, such that for any d≥d 0 if (κ > 1 2α, γ2 = g and c≤g ) or (κ < 1 2α, γ2 = g and c≤gd 2α(κ−...

  130. [138]

    effective learning rate

    The results are valid up to a constant independent of the dimension Finally, one will note that(√γ3BT )2 ≍τ (t)2 the corresponding time of DANA-decaying around compute- optimality timeT. This implies the following: The previously computedγ3 for DANA-constant induces the same d...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.