acceptodds
Under review as a conference paper at ICLR 2027

Statistical guarantees for multimodal contrastive learning via pointwise mutual information estimation

Abstract

Multimodal contrastive learning has emerged as a powerful approach for learning representations from paired multimodal data. In this work, we develop a theoretical framework for analyzing the finite-sample behavior of contrastive learning with the InfoNCE objective. Our approach is based on viewing the learned similarity score as an estimator of the pointwise mutual information (PMI) function, which characterizes the population-optimal similarity score up to an additive constant. By exploiting the curvature of the InfoNCE objective together with a localized analysis around the PMI score, we derive fast convergence rates for the PMI estimation error despite the dependence among negative pairs in InfoNCE. When PMI is Hölder smooth, we further show that InfoNCE achieves the minimax-optimal rate up to logarithmic factors for both joint-input neural networks and CLIP-style dual encoders. For downstream learning, we show that accurate recovery of PMI implies approximate sufficiency of the learned representations and yields a bound on the excess risk of zero-shot classification. We provide numerical experiments to support our theoretical results.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
contrastive learning, pointwise mutual information, multimodal learning, zero-shot learning, minimax estimation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_D7aLOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposed a finite-sample statistical analysis of multimodal contrastive learning by interpreting InfoNCE as an estimator of pointwise mutual information (PMI). Via localized complexity analysis, it leads to fast convergence rates and establishes minimax-optimal PMI estimation rates for both joint-input networks and CLIP-style dual encoders. The results are further extended to representation sufficiency and zero-shot classification.

Critical strengths
  • The paper provides a clear theoretical analysis of InfoNCE, establishing fast convergence rates while accounting for the dependence among negative pairs within a batch. The localized analysis offers a stronger statistical guarantee than conventional global generalization bounds.

  • The matching minimax upper and lower bounds establish the optimality of PMI estimation under Holder smoothness assumptions. Extending these results to CLIP-style dual encoders, along with theoretical guarantees for downstream representation learning, further strengthens the contribution.

Critical weaknesses
  • The theoretical setting differs from practical CLIP architectures and training procedures. The minimax-optimal rate for dual encoders allows the embedding dimension and network capacity to increase with the sample size, while the analysis assumes unnormalized similarity scores and empirical risk minimization over fixed batches. These assumptions limit the direct applicability of the theoretical results to practical CLIP models.

  • The connection between PMI estimation and downstream performance remains somewhat limited. Proposition 5.2 depends on the excess risk of the ideal PMI-based classifier and an additional prompt coverage condition, which may lead to loose guarantees in realistic settings. Moreover, the real-data experiments show relatively low zero-shot classification accuracy without directly evaluating PMI estimation error or representation sufficiency. Thus, the practical significance of the downstream guarantees remains unclear.

Questions for authors
  • Theorem 4.5 establishes minimax-optimal rates for CLIP-style dual encoders, but the embedding dimension and network capacity are allowed to increase with the sample size. How should these results be interpreted for practical CLIP models with fixed embedding dimensions and normalized similarity scores? Clarification of which guarantees remain applicable under these settings would help assess the scope of the theoretical contribution.

  • The theoretical analysis assumes empirical risk minimization over fixed disjoint batches, whereas practical contrastive learning typically uses stochastic optimization with reshuffled batches. How essential are these assumptions to the fast convergence rates? A clarification of their role in the analysis would help establish the applicability of the results to practical training procedures.

  • Proposition 5.2 includes the excess risk of the ideal PMI-based classifier and an additional prompt coverage constant. Under what conditions does this bound provide a meaningful guarantee for downstream classification? Clarifying the effect of these terms would help assess the significance of the theoretical results.

5 comments

We thank the reviewer for the helpful comments and thoughtful feedback. We respond to each point below. All major changes in the revised paper have been highlighted in blue, and the list of references can be found in the global response.


Response to W1 and Q1 and Q2

(i) W1 and Q1 (embedding dimension). Our Theorem 4.5 is interpreted as follows: for a given sample size nn, we can find network sizes, including the embedding dimension pp, such that the InfoNCE minimizer attains the minimax rate. Choosing the network size as a function of the sample size is very common in statistical learning theory for neural networks (e.g. Schmidt-Hieber '20, Nakada & Imaizumi '20, Han et al. '25, Ma et al. '26). This is also known as the sieve method in nonparametric statistics (Shen & Wong '94), where the class has to grow with nn so that its approximation error vanishes as nn grows. It is a standard technique for deriving convergence rates in terms of the sample size nn.

For a fixed embedding dimension pp, a minor modification of Theorem 4.5 also gives the result in terms of the approximation and estimation errors. Specifically, we obtain

ErrQ(s^)≲p−2α/min⁡dX,dY+p(dX+dY)/min⁡dX,dYn,\mathrm{Err}_Q(\widehat s)\lesssim p^{-2\alpha/\min\\{d_X,d_Y\\}} +\frac{p^{(d_X+d_Y)/\min\\{d_X,d_Y\\}}}{n},

up to logarithmic factors, where the first term is the approximation error and the second is the estimation error. We now state this as Corollary E.3 in the revised paper. For a fixed pp the approximation error does not go to zero as nn grows, so it remains in the bound. If we instead choose p≍nmin⁡dX,dY/(2α+dX+dY)p\asymp n^{\min\\{d_X,d_Y\\}/(2\alpha+d_X+d_Y)} in the above, the two terms balance and we can recover the minimax rate n−2α/(2α+dX+dY)n^{-2\alpha/(2\alpha+d_X+d_Y)} for the Hölder class.

(ii) W1 and Q1 (normalized scores). We have now added results for normalized scores (see Lemma E.4 of the revised paper). We consider the class of scores ⟨f/∣f∣2,g/∣g∣2⟩/τ\langle f/\\|f\\|_2,g/\\|g\\|_2\rangle/\tau, where ff and gg are ReLU networks whose norms are bounded away from zero. These scores are bounded by 1/τ1/\tau, so Proposition 4.1 and Theorem 4.2 of the original paper apply without change. The only change occurs when we compute the covering number of this class, where we use that the normalization u↦u/∣u∣2u\mapsto u/\\|u\\|_2 is Lipschitz away from zero. This gives the same bound as for unnormalized scores, i.e.,

ErrQ(s^)≲ϵnor2+min⁡S,W2Ln\mathrm{Err}_Q(\widehat s)\lesssim\epsilon_{\rm nor}^2+\frac{\min\\{S,W^2L\\}}{n}

up to logarithmic factors, where ϵnor\epsilon_{\rm nor} is the approximation error of the class and min⁡S,W2L\min\\{S,W^2L\\} is the size of the networks as in Corollary 4.4.

(iii) W1 and Q2 (fixed batches). We first note that many existing finite-sample theories of contrastive learning analyze the empirical loss or its minimizer with fixed batches, or under an even more restrictive assumption on the negative pairs (e.g. Saunshi et al. '19, HaoChen et al. '21, Lei et al. '23, Chen et al. '24, Oko et al. '25, Lin & Mei '25, Hieu et al. '25, Li et al. '26). Following this literature, we also analyzed the empirical risk minimizer with fixed batches. Having said this, the fixed-batch assumption is not essential for our analysis, and the same results extend to reshuffled batches. Specifically, for each epoch e=1,…,Ee=1,\ldots,E, let Be,1,…,Be,NB_{e,1},\ldots,B_{e,N} be the reshuffled batches and consider the objective that averages the batch losses over all epochs, i.e.,

L^K,E(s)=1E∑e=1E1N∑b=1NℓK(s;Be,b),\widehat L_{K,E}(s) =\frac1E\sum_{e=1}^E\frac1N\sum_{b=1}^N\ell_K(s;B_{e,b}),

where ℓK(s;B)\ell_K(s;B) is the InfoNCE loss of a batch BB as in the paper. Similar objectives that average a contrastive loss over a collection of mini-batches have been also considered in Cho et al. '24. Then the same concentration bound of Proposition B.7 also holds for the loss averaged over the mini-batches, which allows us to obtain the same convergence rates as in the fixed-batch setting (see Theorem E.1 in Appendix E.1 of the revised paper, and also our response to Reviewer TU7F for further details).

Regarding stochastic optimization, our paper focuses on the finite-sample statistical analysis, as is typical in statistical learning theory. Whether SGD converges to the empirical risk minimizer is a question of optimization. The InfoNCE loss is highly nonconvex in the network parameters, so a convergence theory for SGD on the contrastive loss should be treated as a separate work, and it is beyond the scope of this paper. As we pointed out in the original paper, we view this as an interesting direction for future work. We also note that our results still apply to the output of any training procedure, including SGD, in two ways. First, we have now added Theorem E.1 to the revised paper, which holds for every approximate minimizer, and the optimization error ϵopt\epsilon_{\rm opt} appears in addition to the approximation and estimation errors, so any guarantee for the convergence of the optimization algorithm can be combined with it. Second, the lower bound of Theorem 4.6 holds for every estimator based on the nn pairs, including the output of SGD with reshuffled batches, so no training procedure can improve on the rate in terms of nn when the PMI is Hölder smooth.


Response to W2 and Q3

(i) W2 and Q3 (downstream guarantee). We thank the reviewer for this question. We discussed the conditions under which Proposition 5.2 gives a meaningful guarantee in Appendix D.2 of the original paper. Since most of this discussion appeared only in the appendix, we have also added a short summary of it to the main text of the revised paper after Proposition 5.2. We summarize it here.

The first term, 2R(z(⋅;s⋆))−R⋆2\\{R(z(\cdot;s^\star))-R^\star\\}, is the excess risk of the classifier that uses the true PMI. It does not depend on the learned score s^\widehat s or on the sample size nn. It is small when the logits of the true PMI approximately recover the class probabilities. For example, let AcA_c be the set of texts that describe class cc and take the prompt distribution νc=PY(⋅∣Ac)\nu_c=P_Y(\cdot\mid A_c) with the class bias bc=log⁡PY(Ac)b_c=\log P_Y(A_c). If the image affects the text only through its class, then the logits of the true PMI are exactly the log class probabilities log⁡P(C=c∣X=x)\log P(C=c\mid X=x), and the first term is zero (see Appendix D.2 of the original paper for further details). This corresponds to the conditions discussed in Uesaka et al. '25, who also give a bound when the conditions hold only approximately. In the revised paper we also state this bound in Appendix D.2.

The second term, 34C~,ErrQ(s^)\frac34\widetilde C\\,\mathrm{Err}_Q(\widehat s), reflects how accurately the learned score s^\widehat s estimates the PMI s⋆s^\star. For the prompt distributions considered above we showed C~=∑c1/PY(Ac)\widetilde C=\sum_c1/P_Y(A_c) in Appendix D.2 of the original paper. In this case the second term equals 34∑c=1m1PY(Ac),ErrQ(s^)\frac34\sum_{c=1}^m\frac1{P_Y(A_c)}\\, \mathrm{Err}_Q(\widehat s), so it is of the same order as the PMI error ErrQ(s^)\mathrm{Err}_Q(\widehat s). When the first term is small, our results for the PMI error then directly translate into guarantees for zero-shot classification.

(ii) W2 (real data experiments). We would like to first clarify that the PMI error and the representation sufficiency both require the true joint distribution of images and texts, so they cannot be evaluated on real data. In this experiment we instead report zero-shot accuracy, which is a standard evaluation metric for contrastive pretraining (e.g. Radford et al. '21) and can be seen as a proxy for the quality of the learned score.

Regarding the low accuracy, we did not tune the training for accuracy, and both encoders are trained from scratch on at most 2202^{20} image-text pairs (far fewer than the 400M pairs used for CLIP in Radford et al. '21). Even so, the top-1 accuracy on ImageNet-1K, which has 10001000 classes, is about 3.83.8\\% at n=220n=2^{20}, which is still far above the random chance (0.10.1\\%). On Caltech-101, which has 102102 classes, the top-1 and top-5 accuracies are also up to 20.720.7\\% and 47.547.5\\%, respectively. More importantly, the goal of this experiment is not to obtain high accuracy but to examine whether the predictions of our theory hold on real data. The results show that they do, i.e., the performance is mainly governed by the total number of pairs nn, and the batch size KK is less important.

To further address the reviewer's comment about evaluating the PMI error, we note that our theory also bounds the excess InfoNCE risk (see Theorem 4.2 of the original paper), and this risk can be estimated on real data. We therefore additionally evaluated the held-out InfoNCE loss on the WIT dataset, which is a good proxy for the excess InfoNCE risk. It is computed on pairs of the WIT dataset that are not used for training, for the models of the original paper (fixed batches) and for models trained with batches reshuffled at every epoch. As the table below shows, the loss decreases clearly with nn, and the effect of the batch size KK is smaller than that of nn.

Held-out InfoNCE loss on the WIT dataset

nnbatchesK=64K=6425625610241024
2172^{17}fixed6.8596.7697.046
reshuffled6.5196.8307.051
2182^{18}fixed5.7845.7926.024
reshuffled5.3915.8766.107
2192^{19}fixed4.7224.9095.156
reshuffled4.2334.8015.197
2202^{20}fixed3.6273.8264.136
reshuffled3.0573.5774.086
Reviewer_p5GVOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper develops a finite-sample theory for multimodal contrastive learning by treating the InfoNCE score as an estimator of pointwise mutual information, up to an additive constant. A self-concordance argument links the InfoNCE excess risk to the L²(Q) error of PMI, with constants that do not depend on the batch size. Maurer's inequality for weakly interacting functions handles the reuse of samples as negatives within a batch. Combined with localization, this gives an oracle inequality whose rate depends on the total number of pairs n = NK. For Hölder-smooth PMI, both joint-input ReLU networks and CLIP-style dual encoders reach the minimax rate up to log factors, and a matching lower bound is given. The paper also bounds approximate sufficiency and the zero-shot excess risk in terms of the PMI error. Synthetic experiments show the predicted 1/n scaling and little dependence on K, and the appendix adds a small image-text experiment on WIT.

Critical strengths
  1. The paper handles the dependence created by reusing samples as negatives within a batch, which is a real technical obstacle, and gets a fast O(1/n) rate in the total number of pairs instead of the n^{-1/2} that global analyses give. The upper bound is matched by a minimax lower bound and connected to sufficiency and zero-shot classification, so the story is complete from estimation to downstream use. This is the main reason for a positive assessment: it answers a clear question left open by Oko et al. and Lin & Mei.

  2. The writing is clear and the technical work is careful. Section 4.3 explains well why localization gives the faster rate. I checked Prop. 4.1, Lemma A.4, the proof of Theorem 4.2 and the lower bound, and found no problems. The synthetic experiments also line up with the predicted scaling.

Critical weaknesses
  1. The guarantees are far from what practical CLIP training looks like.
    • The constants are exponential in M₀, C₁ ∝ e^{2M₀}, C₂ ∝ e^{7M₀}, and Prop. 5.1 has a factor e^{2M₀}. With τ = 0.01, CLIP logits span roughly ±100, so the bounds say very little in practice.
    • The CLIP-style rate in Theorem 4.5 needs the embedding dimension to grow polynomially with n (Lemma C.4). The main text hides this behind "network sizes can be chosen", although fixed embedding dimension is exactly what makes dual encoders interesting.
    • The rate is in the ambient dimension d_X + d_Y.
    • For single prompts, the coverage constant C̃ = Σ 1/P_Y({y_c}) in Prop. 5.2 is essentially infinite in a continuous text space.

Some of this is inherent to the setting, but much of it can be addressed. The paper should state the growth of p in the main theorem, discuss the M₀ and τ dependence openly, and consider adapting to intrinsic dimension, as Han et al. (JASA 2025) and Gui et al. (2025) do.

  1. The novelty over existing work is moderate.
    • Oko et al. (2025) and Lin & Mei (2025) already provide the sufficiency framework and sample-complexity bounds.
    • Han et al. (JASA 2025) already give minimax-optimal deep estimation of the mutual density ratio (i.e., PMI) via Bregman objectives, including manifold adaptation, and should be discussed more directly.
    • The genuinely new piece is localization under batch dependence. The tools used (self-concordance, Maurer's inequality, chaining, Schmidt-Hieber approximation, Assouad) are standard, and the downstream results are short corollaries.

The experiments don't add much beyond this. The synthetic setting is low-dimensional and well-specified, and it checks the 1/n scaling but not the dependence on α or d. In the real-data experiment, ImageNet zero-shot top-1 is between 0.7% and 3.8%, which is too low to support conclusions about the role of K. Positioning is fixable; the experiments could be strengthened by varying α or d and by using a setting where zero-shot accuracy is meaningful.

Questions for authors
  1. For the CLIP-style class, what can be said when the embedding dimension p is fixed? For example, does the rate change if the PMI has low rank, and can the bound be stated in terms of that rank?

  2. How do the constants scale with M₀ and τ, and is there a realistic regime (e.g., τ = 0.01) where the bounds are non-vacuous? Can the exponential dependence be improved?

  3. For single-prompt zero-shot classification, C̃ seems to be infinite when the text distribution is continuous. Is there a version of Prop. 5.2 that covers realistic prompts, e.g., by smoothing over neighbourhoods of the prompt?

  4. In the WIT experiment, a larger K with a fixed number of epochs means fewer optimization steps. How is this separated from the statistical effect of K?

5 comments

We thank the reviewer for the helpful comments and thoughtful feedback. We respond to each point below. All major changes in the revised paper have been highlighted in blue, and the list of references can be found in the global response.


Response to W1 and Q1 and Q2 and Q3

W1(a) and Q2. We first note that the exponential dependence of the constants in M0M_0 is common in the theory of contrastive learning (e.g. Oko et al. '25, Lin & Mei '25, Li et al. '26) and also in the related literature on density ratio estimation (e.g. Han et al. '25). And indeed, we can also show that this dependence is rather unavoidable. To see this, consider C1C_1 in Proposition 4.1 as the reviewer says, and let X,Y∈0,1X,Y\in\\{0,1\\} have uniform marginals and P(X≠Y)=e−M0/2P(X\ne Y)=e^{-M_0}/2, so that s⋆=log⁡(2−e−M0)s^\star=\log(2-e^{-M_0}) if x=yx=y and s⋆=−M0s^\star=-M_0 if x≠yx\ne y. Consider the score s(x,y)=M0s(x,y)=M_0 if x=yx=y and s(x,y)=−M0s(x,y)=-M_0 if x≠yx\ne y, so that s−s⋆=c,1x=ys-s^\star=c\\,\mathbf 1\\{x=y\\} with c=M0−log⁡(2−e−M0)≥M0−log⁡2c=M_0-\log(2-e^{-M_0})\ge M_0-\log 2. Then we can show that

ErrQ(s)=c24,;;EK(s)≤c,P(X≠Y)=c2,e−M0,\mathrm{Err}_Q(s)=\frac{c^2}{4}, \\;\\; \mathcal E_K(s)\le c\\,P(X\ne Y)=\frac c2\\,e^{-M_0},

and so ErrQ(s)≥c2,eM0,EK(s)\mathrm{Err}_Q(s)\ge\frac c2\\,e^{M_0}\\,\mathcal E_K(s), i.e., C1C_1 must be at least of order M0eM0M_0e^{M_0}. We have now added this result with the details in Proposition A.5 of the revised paper, and we also improved C1C_1 in Proposition 4.1 from O(M0e2M0)O(M_0e^{2M_0}) to O(M0eM0)O(M_0e^{M_0}) to match this lower bound.

Regarding the dependence on the temperature τ\tau, the proof of Proposition 4.1 of the revised paper shows that we can take C1=4(2B+2∣s⋆∣∞+2),ebC_1=4\bigl(2B+2\\|s^\star\\|_\infty+2\bigr)\\,e^{b} when ∣s∣≤B|s|\le B for all s∈Ss\in\mathcal S and s⋆≥−bs^\star\ge-b. In the revised paper we bounded BB, ∣s⋆∣∞\\|s^\star\\|_\infty, and bb by M0M_0 of Assumption 3.2, which gives C1=8(2M0+1)eM0C_1=8(2M_0+1)e^{M_0}. For normalized CLIP-style scores (i.e., s(x,y)=⟨f(x),g(y)⟩/τs(x,y)=\langle f(x),g(y)\rangle/\tau with f,gf,g on the unit sphere) we have B=1/τB=1/\tau. If ∣s⋆∣∞≤1/τ\\|s^\star\\|_\infty\le1/\tau and b=O(1)b=O(1), then C1C_1 is actually O(1/τ)O(1/\tau), whereas bounding ∣s⋆∣∞\\|s^\star\\|_\infty and bb by M0=1/τM_0=1/\tau would give e1/τ/τe^{1/\tau}/\tau, which is very conservative. For example, with τ=0.01\tau=0.01 these are of order 10310^3 and 103e10010^3e^{100}. This shows that the constant C1C_1 can be much smaller in practice than its bound in terms of M0M_0. We stated our results with the single constant M0M_0 only for simplicity, because our work focuses on the statistical rate in terms of nn rather than optimizing the constants. We have added this discussion to Appendix A.3 of the revised paper.

W1(b) and Q1. Our Theorem 4.5 is interpreted as follows: for a given sample size nn, we can find network sizes, including the embedding dimension pp, such that the InfoNCE minimizer attains the convergence rate stated in the theorem. The dependence of the architecture on the sample size is very common in statistical learning theory for neural networks, where the size of the network is chosen as a function of nn (e.g. Schmidt-Hieber '20, Nakada & Imaizumi '20, Han et al. '25, Ma et al. '26). This is also known as the sieve method in nonparametric statistics (Shen & Wong '94), where the class has to grow with nn so that its approximation error vanishes as nn grows. In the CLIP-style score, increasing the embedding dimension pp with nn allows the score to approximate a Hölder-smooth PMI more accurately. We now state this more explicitly in the revised paper after Theorem 4.5.

For a fixed embedding dimension pp, a minor modification of Theorem 4.5 gives ErrQ(s^)≲p−2α/min⁡dX,dY+p(dX+dY)/min⁡dX,dY/n\mathrm{Err}_Q(\widehat s)\lesssim p^{-2\alpha/\min\\{d_X,d_Y\\}} +p^{(d_X+d_Y)/\min\\{d_X,d_Y\\}}/n up to logarithmic factors (see Corollary E.3 of the revised paper). The first term is the approximation error, and for a fixed pp it does not decrease as nn grows, so it remains in the bound. This is unavoidable in general, since a score with embedding dimension pp has rank at most pp (see Lemma E.2 of the revised paper). However, as the reviewer suggests, a fixed pp suffices when the PMI has low rank. If the PMI has rank rr, we take p=rp=r so that there is no error due to the rank, and the encoders only need to approximate the rr factors of the PMI, which are functions of xx alone or of yy alone. By the existing approximation results for ReLU networks, this approximation error vanishes as nn grows, and we can show the rate n−2α/(2α+max⁡dX,dY)n^{-2\alpha/(2\alpha+\max\\{d_X,d_Y\\})} up to logarithmic factors. We have now added this low-rank case as Theorem E.5 of the revised paper.

W1(c). Our paper makes no assumption on a low-dimensional structure of the data, so the rate in Theorem 4.5 depends on the ambient dimension dX+dYd_X+d_Y, and Theorem 4.6 shows that this rate is minimax optimal up to logarithmic factors, that is, it cannot be improved without further assumptions. If the data have a low-dimensional structure as in Han et al. '25, the same argument gives a rate in terms of the intrinsic dimension. For simplicity consider the joint-input class Snet\mathcal S_{\rm net}. Corollary 4.3 of the original paper gives ErrQ(s^)≲ϵnet2+min⁡S,W2L/n\mathrm{Err}_Q(\widehat s)\lesssim\epsilon_{\rm net}^2+\min\\{S,W^2L\\}/n up to logarithmic factors. Approximating a Hölder function within ϵnet=m−α\epsilon_{\rm net}=m^{-\alpha} on the whole cube [0,1]dX+dY[0,1]^{d_X+d_Y} requires S≍mdX+dYS\asymp m^{d_X+d_Y}, which gives the rate of Theorem 4.5. However, the approximation is needed only on the support of the data, which is much smaller than the whole cube when the data are low-dimensional. Specifically, if the support has intrinsic dimension less than qq for some q<dX+dYq<d_X+d_Y (in the sense of the Minkowski dimension), then Theorem 5 of Nakada & Imaizumi '20 shows that S≍mqS\asymp m^{q} suffices for the same accuracy ϵnet=m−α\epsilon_{\rm net}=m^{-\alpha}, and we obtain

ErrQ(s^)≲m−2α+mqn\mathrm{Err}_Q(\widehat s)\lesssim m^{-2\alpha}+\frac{m^{q}}{n}

up to logarithmic factors. Choosing m≍n1/(2α+q)m\asymp n^{1/(2\alpha+q)} then gives the rate n−2α/(2α+q)n^{-2\alpha/(2\alpha+q)} instead of the rate n−2α/(2α+dX+dY)n^{-2\alpha/(2\alpha+d_X+d_Y)} of Theorem 4.5.

W1(d) and Q3. As the reviewer mentioned, C~=∞\widetilde C=\infty for a single prompt when the text space is continuous. Since text is usually discrete, we thought that this would not be a problem in Proposition 5.2. When the text space is continuous, as the reviewer suggests, we can smooth over a neighbourhood AcA_c of each prompt, that is, take νc=PY(⋅∣Ac)\nu_c=P_Y(\cdot\mid A_c). This choice was also discussed in Appendix D.2 of the original submission. Proposition 5.2 then holds with C~=∑c1/PY(Ac)\widetilde C=\sum_c1/P_Y(A_c), which is finite as long as each AcA_c has positive probability.


Response to W2

We thank the reviewer for raising this point. We would like to clarify the novelty of our work over the existing work below.

W2(a). One of our main contributions is to show that estimating the PMI is enough to control the approximate sufficiency of Oko et al. '25 and Lin & Mei '25, as the title of our paper "via pointwise mutual information estimation" says. We do not claim the sufficiency framework itself as our contribution. Proposition 5.1 bounds the sufficiency by the PMI error, so a rate for the PMI error translates to a rate for the sufficiency. Additionally, their sample-complexity bounds control the difference between the empirical and the population loss uniformly over the class, which gives a slow rate. In the setting of Corollary 4.4 such a bound would be of order min⁡S,W2L/n\sqrt{\min\\{S,W^2L\\}/n}, whereas our bound is of order min⁡S,W2L/n\min\\{S,W^2L\\}/n, up to logarithmic factors. We thus improve the statistical rate compared to Oko et al. '25 and Lin & Mei '25.

W2(b). Our work and Han et al. '25 have substantially different goals. Han et al. design a new objective based on Bregman divergence for estimating the PMI directly. On the other hand, we analyze InfoNCE, the objective that is used in contrastive pretraining, and the PMI appears because it is the population minimizer of this objective. Our question is whether the objective used in practice for contrastive learning is statistically efficient, and we show that it attains the minimax rate uniformly in the batch size. The dependence structure is also different. The Bregman objective of Han et al. is a U-statistic of order 2, which has been well studied in the classical theory of U-processes. In contrast, the InfoNCE loss of a batch depends on all KK pairs in the batch through the K2K^2 many scores s(Xi,Yj)s(X_i,Y_j), i,j=1,…,Ki,j=1,\ldots,K, and KK may grow with nn, so it requires a more delicate analysis. Finally, unlike Han et al., we also consider CLIP-style dual encoders and establish guarantees for sufficiency and zero-shot classification. We have added a remark on this comparison to Section 4.3 of the revised paper.

W2(c). Yes, localization under batch dependence is the main technical contribution of our paper, and we agree that tools such as chaining, the approximation results for ReLU networks, and Assouad's lemma are standard. However, although self-concordance and the inequality by Maurer '19 are known, to the best of our knowledge neither of them has been used before in the analysis of contrastive learning, and Maurer's inequality has also not been used within the localization framework in the literature to obtain fast rates. A direct approach is to treat each batch (or tuple) as one independent observation (e.g. Saunshi et al. '19, Lei et al. '23). Then with this approach, standard localization (e.g. Wainwright '19) can give a fast rate only in terms of the number of batches N=n/KN=n/K, so the bound becomes worse as the batch size grows. We overcome this difficulty with a novel analysis that combines localization with a control of the dependence among the pairs within a batch, which gives a fast rate in nn uniformly in KK.

Further, as the reviewer points out, the downstream results are short. We believe this is indeed an advantage of our paper, because the sufficiency and zero-shot guarantees follow easily from the bound on the PMI error and do not require a separate involved analysis.


Response to W3 and Q4

(i) We thank the reviewer for this suggestion. We conducted an additional synthetic experiment where we vary the smoothness α∈0.5,1.5,2.5\alpha\in\\{0.5,1.5,2.5\\} and the dimension dX=dY=d∈1,2d_X=d_Y=d\in\\{1,2\\}. For each nn we choose the network width from 8,16,32,64,128,256\\{8,16,32,64,128,256\\} that gives the smallest InfoNCE loss on a held-out dataset. The table below shows the relative PMI error ErrQ(s^)/VarQ(s⋆)\mathrm{Err}_Q(\widehat s)/\mathrm{Var}_Q(s^\star) of the score function given by a joint-input ReLU network. As the table shows, the error decreases with nn in every case, and it decreases faster for a larger α\alpha and more slowly for a larger dd, which is in line with our theoretical result (see Theorem 4.5 of the original paper).

Relative PMI error

ddα\alphan=211n=2^{11}2132^{13}2152^{15}2172^{17}
10.50.950.2940.150.0811
11.50.360.10.04350.0166
12.50.2750.07270.02580.00874
20.52.91.180.770.543
21.51.790.7930.480.241
22.52.060.6890.390.19

(ii) Regarding the low accuracy on ImageNet-1K, we believe it arises mainly because both encoders are trained from scratch on at most 2202^{20} image-text pairs (far fewer than the 400M pairs used for CLIP in Radford et al. '21) and the training is not tuned for accuracy. The purpose of our experiment was to examine how well real data align with our theory, and not to achieve high accuracy. Even so, the accuracy on ImageNet-1K, which has 10001000 classes, is still above the random chance (0.10.1\\%). We also note that the accuracy is not low on every dataset. On Caltech-101, which has 102102 classes, the top-1 and top-5 accuracies are up to 20.720.7\\% and 47.547.5\\% at n=220n=2^{20}, respectively. On this dataset as well, the accuracy increases by a large factor as nn grows, whereas it changes much less when KK varies.

(iii) Q4. As the reviewer points out, in the original paper we train for 3030 epochs for every KK, so a larger KK has fewer updates. To reduce the effect of the optimization in training, we compute the accuracy of each run at the epoch with the smallest held-out InfoNCE loss, instead of at a fixed number of epochs. We also train the larger batch sizes for more epochs (120120 and 480480 epochs for K=256K=256 and 10241024, respectively), so that they are not limited by the number of updates. As the table below shows, the accuracy is still similar across KK, and the differences across KK are much smaller than the effect of nn.

Top-5 accuracy (IN-1K: ImageNet-1K, C-100: CIFAR-100, Cal-101: Caltech-101)

nnIN-1K, 64IN-1K, 256IN-1K, 1024C-100, 64C-100, 256C-100, 1024Cal-101, 64Cal-101, 256Cal-101, 1024
2172^{17}3.123.112.4710.0210.8410.3613.9316.2712.03
2182^{18}4.724.854.8411.9410.8610.3922.3620.0519.78
2192^{19}7.618.068.0212.4913.3813.3935.5635.9838.82
2202^{20}11.2811.5212.4316.4417.9218.3347.5546.2648.19
Reviewer_TU7FOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper provides a theoretical analysis of the finite-sample statistical properties of multimodal contrastive learning, in particular the InfoNCE objective, from the perspective of viewing the learned similarity score as an estimator of pointwise mutual information (PMI). The main contributions are as follows. First, the authors establish a general finite-sample oracle inequality based on a localized analysis, which directly controls the PMI estimation error and yields fast convergence rates. Second, they specialize this to joint-input and CLIP-style encoders and show that, for Hölder-smooth PMI, the resulting estimators achieve the minimax-optimal rate up to logarithmic factors, together with a matching lower bound. Third, they show that the PMI estimation error controls the approximate sufficiency of the learned representations and the excess risk of zero-shot classification. The analysis is validated through experiments on synthetic data and on real image-text data (WIT).

Critical strengths
  • While the fact that the population optimum of InfoNCE is PMI is already known, this paper directly quantifies it as a finite-sample estimation error and establishes minimax optimality by providing both a matching upper bound and lower bound. Proving not only the upper bound but also the lower bound gives the theoretical contribution a degree of completeness. (I note, however, that I have not verified all of the proofs in full detail.)
Critical weaknesses
  • My major concern is that the fixed-batch assumption on which the analysis relies is too far removed from realistic training settings, which substantially limits the practical relevance of the main results. In actual contrastive learning, batches are reshuffled at every epoch, and the analysis assumes a fixed partition into disjoint batches instead. The paper does not make clear whether fixing the batches is essential to the derivation or whether it is merely an analytical convenience that could be relaxed, and this makes it difficult to judge how far the results extend to the settings used in practice. Because the key theoretical message, namely that performance is governed by the total number of pairs nn rather than the number of batches NN, is stated under this assumption, the gap from real training is not something I can overlook. In addition, the boundedness assumption (Assumption 3.2) is not shown to hold for realistic distributions, which further weakens the case for the applicability of the theory.

  • The empirical validation does not adequately compensate for this gap. The absolute zero-shot classification accuracy on real data is low, making it difficult to draw conclusions beyond confirming the qualitative trend predicted by the theory, and it is unclear whether this low performance is itself a consequence of the fixed batch construction. Furthermore, the batch presentation order is fixed for the synthetic data but reshuffled every epoch for the real data, and the paper does not explain which of these settings is intended to be consistent with the theory. The dependence on the batch size KK observed on Caltech-101 is also left unexplained, which raises doubts about the central claim of independence from KK.

Questions for authors

All of the following questions are relevant to my assessment of the paper.

  1. How essential is the fixed-batch assumption for obtaining the derived convergence rates and minimax optimality? Or is it an analytical convenience, so that an extension to batches that are reshuffled every epoch is in principle feasible? If such an extension is feasible, please indicate the degree of technical difficulty that you anticipate.

  2. To what extent do you expect the boundedness of PMI and of the score (Assumption 3.2) to hold realistically for actual data distributions? In addition, how does the constant M0M_0 affect the convergence rate and the bounds in the downstream-task guarantees?

5 comments

We thank the reviewer for the helpful comments and thoughtful feedback. We respond to each point below. All major changes in the revised paper have been highlighted in blue, and the list of references can be found in the global response.


Response to W1 and Q1

(i) First, regarding the fixed-batch assumption, we agree that our problem setting has some gaps from real training (we also pointed this out in the original paper as a direction for future work), but we would like to emphasize that much of the existing literature on finite-sample results for contrastive learning assumes the fixed-batch setting and our formulation follows this standard. For example, Chen et al. '24, Oko et al. '25, Lin & Mei '25, and van Elst & Ghoshdastidar '25 all assume that the batches are fixed as in our paper. Other existing works such as Saunshi et al. '19, HaoChen et al. '21, Lei et al. '23, Nakada et al. '23, Hieu et al. '25, and Li et al. '26 consider settings that are even more restrictive than our fixed-batch setting, in which the negative pairs are drawn independently for each positive pair or each positive pair is contrasted with all the other samples as negatives so that only one batch is used. In this sense our setting is closer to how negative samples are used in practice, where they are reused within the same batch, and we allow the batch size KK to be fixed or to grow with nn.

(ii) We can also show that the fixed-batch assumption is not essential to our theoretical results. We would first like to clarify that reshuffling the batches at every epoch, as the reviewer mentions, involves stochastic updates of the model parameters. Our paper focuses on the statistical analysis and does not study the optimization. Since the InfoNCE loss is nonconvex with respect to the network parameters, a convergence analysis of stochastic updates would be a separate work in itself and is beyond the scope of this paper.

In the statistical setting, however, our results extend easily to reshuffled batches, and we have added this extension to the revised paper (see Theorem E.1 in Appendix E.1). Specifically, for each epoch e=1,…,Ee=1,\ldots,E, let Be,1,…,Be,NB_{e,1},\ldots,B_{e,N} be the reshuffled batches and consider the objective that averages the batch losses over all epochs, i.e.,

L^K,E(s)=1E∑e=1E1N∑b=1NℓK(s;Be,b),\widehat L_{K,E}(s) =\frac1E\sum_{e=1}^E\frac1N\sum_{b=1}^N\ell_K(s;B_{e,b}),

where ℓK(s;B)\ell_K(s;B) is the InfoNCE loss of a batch BB. For E=1E=1 this is the fixed-batch loss. Training with batches reshuffled at every epoch takes one SGD step on each ℓK(s;Be,b)\ell_K(s;B_{e,b}), so it is the natural empirical objective for SGD with reshuffled batches. Similar objectives that average a contrastive loss over a collection of mini-batches have also been considered in the literature (e.g. Cho et al. '24).

Since the batches within each epoch are disjoint and thus independent, for each epoch ee the moment generating function (MGF) of Ws(e)−Wt(e)W_s^{(e)}-W_t^{(e)} can be bounded as in the proof of Proposition B.7 of the original paper, where Ws(e)−Wt(e)W_s^{(e)}-W_t^{(e)} is the centered loss difference Ws−WtW_s-W_t of Appendix B.1 computed on the batches of epoch ee. Jensen's inequality then gives

Eexp⁡(βE∑e=1E(Ws(e)−Wt(e)))≤1E∑e=1EE eβ(Ws(e)−Wt(e)),\mathbb E\exp\Big(\frac{\beta}{E}\sum_{e=1}^E\big(W_s^{(e)}-W_t^{(e)}\big)\Big) \le\frac1E\sum_{e=1}^E\mathbb E\,e^{\beta(W_s^{(e)}-W_t^{(e)})},

so the same MGF bound holds for the centered loss difference of L^K,E\widehat L_{K,E}, and hence the same concentration bound and the theoretical results follow directly.

Response to W2

(i) We thank the reviewer for raising this point. We believe the low accuracy arises mainly because both encoders are trained from scratch on at most 2202^{20} image-text pairs (far fewer than the 400M pairs used for CLIP in Radford et al. '21) and the training is not tuned for accuracy, since the purpose of this experiment is only to examine how well real data align with our theory. Following the reviewer's comment, we also repeated the experiment of the original paper with the batches reshuffled at every epoch (see the table below), and the accuracies with fixed and with reshuffled batches are not much different. We also point out that while the accuracy is low on ImageNet-1K, even there it is above the random chance (0.1%0.1\%), so the agreement with our theory observed in this experiment still supports our theoretical results.

Top-1 accuracy (IN-1K: ImageNet-1K, C-100: CIFAR-100, Cal-101: Caltech-101)

nnbatchesIN-1K, 64IN-1K, 256IN-1K, 1024C-100, 64C-100, 256C-100, 1024Cal-101, 64Cal-101, 256Cal-101, 1024
2172^{17}fixed0.860.790.652.132.111.822.001.982.45
reshuffled0.890.770.652.232.321.752.382.561.86
2182^{18}fixed1.381.291.172.422.422.364.995.243.70
reshuffled1.501.251.082.732.911.916.877.374.19
2192^{19}fixed2.292.282.013.223.083.2913.8312.7812.72
reshuffled2.762.421.923.643.563.4517.4613.819.65
2202^{20}fixed3.773.923.754.944.594.4020.7019.2020.40
reshuffled4.764.493.895.904.553.9425.1519.9816.12

Top-5 accuracy (IN-1K: ImageNet-1K, C-100: CIFAR-100, Cal-101: Caltech-101)

nnbatchesIN-1K, 64IN-1K, 256IN-1K, 1024C-100, 64C-100, 256C-100, 1024Cal-101, 64Cal-101, 256Cal-101, 1024
2172^{17}fixed2.912.742.309.659.457.4711.077.418.64
reshuffled3.012.602.269.339.077.8910.329.427.16
2182^{18}fixed4.424.283.8510.6110.029.2316.8718.3612.89
reshuffled4.854.003.6310.0610.357.9823.1021.4415.03
2192^{19}fixed6.776.765.9212.5611.9911.9335.5831.5827.94
reshuffled8.007.025.8313.6013.3312.2337.7332.1826.85
2202^{20}fixed10.5910.9710.1716.6314.6114.7847.5338.6139.01
reshuffled12.8612.0010.5218.8614.9913.9149.8742.9934.57

(ii) Regarding the batch presentation order, we would like to clarify that the batches are fixed in both experiments of the original paper, as in our theory, and the two experiments are different only in how the loss is optimized. In the synthetic experiments, every step uses the gradient of the full loss, so there is no presentation order, and therefore the synthetic experiments are exactly in the setting of our theory. On real data, using the gradient of the full loss is computationally infeasible, so we take one step per batch and visit the same fixed batches in a random order at each epoch. While this is slightly different from our theoretical setting, the objective is the same and the results are still qualitatively aligned with our theory.

(iii) Regarding the claim of independence from KK, we note that our paper does not claim that the downstream accuracy is identical for every KK. Our theoretical results state that the bounds on the PMI error hold with constants that do not depend on the batch size KK (see Theorem 4.5 of the original paper), so the convergence rate in terms of nn is the same for every KK and matches the minimax lower bound of Theorem 4.6. For the real data experiment, the original paper states that the downstream performance improves primarily with nn and is comparatively less sensitive to the batch size, and it already mentions that the differences across batch sizes are somewhat larger on Caltech-101.

For the dependence on KK on Caltech-101 that the reviewer mentioned, we believe that it is not a statistical effect of the batch size. First, the effect of nn on the accuracy is much larger than that of KK. As the table in (i) shows, the accuracy on Caltech-101 increases by a large factor as nn grows, whereas it changes much less with KK at a fixed nn. Second, in actual training several other factors such as the optimization are also involved. To reduce the effect of the optimization, we then conducted an additional experiment in which we find the epoch with the smallest held-out InfoNCE loss for each training run and evaluate the downstream accuracy at this epoch (best-val in the table below). As the table shows, at this epoch the differences across KK mostly become smaller, and with reshuffled batches they nearly disappear when the sample size is large (i.e., n=220n=2^{20}).

Caltech-101 top-5 accuracy

nnbatches30 epochs, 6430 epochs, 25630 epochs, 1024best-val, 64best-val, 256best-val, 1024
2172^{17}fixed11.077.418.6414.8914.9216.67
reshuffled10.329.427.1612.5314.5416.17
2182^{18}fixed16.8718.3612.8921.6717.7115.75
reshuffled23.1021.4415.0317.5425.7226.27
2192^{19}fixed35.5831.5827.9432.8139.9035.22
reshuffled37.7332.1826.8542.3240.0636.54
2202^{20}fixed47.5338.6139.0141.6947.2950.38
reshuffled49.8742.9934.5745.4344.7744.08

Response to Q2

(i) The boundedness assumption is indeed common in the theory of contrastive learning (e.g. Tosh et al. '21, Chen et al. '24, Oko et al. '25, Lin & Mei '25, Li et al. '26) and also in the related literature on density ratio estimation (e.g. Nguyen et al. '10, Han et al. '25). Furthermore, the boundedness assumption on the PMI is reasonable in practice. It says e−M0≤p(x,y)/{pX(x)pY(y)}≤eM0e^{-M_0}\le p(x,y)/\{p_X(x)p_Y(y)\}\le e^{M_0}, or equivalently e−M0≤p(y∣x)/pY(y)≤eM0e^{-M_0}\le p(y\mid x)/p_Y(y)\le e^{M_0}. The upper bound p(y∣x)≤eM0pY(y)p(y\mid x)\le e^{M_0}p_Y(y) means that seeing an image xx does not make any caption yy much more likely than before. This is reasonable because in practice one image can be described by many different captions. The lower bound holds when the data contain some mismatched pairs. For example, if a fraction λ\lambda of the captions is unrelated to their images, then the density of the data is p(x,y)=(1−λ) p0(x,y)+λ pX(x)pY(y)≥λ pX(x)pY(y)p(x,y)=(1-\lambda)\,p_0(x,y)+\lambda\,p_X(x)p_Y(y) \ge \lambda\,p_X(x)p_Y(y), where p0p_0 is the density of the correctly matched pairs. Hence the PMI is lower bounded by log⁡λ\log\lambda. Image-text datasets collected from the web often contain such mismatched pairs (e.g. Jia et al. '21, Li et al. '22).

(ii) M0M_0 appears in our bounds on the PMI error and in the downstream guarantees only through the constants, so it does not affect the convergence rates or the minimax optimality. For example, the constant in Proposition 4.1 of the revised paper is O(M0eM0)O(M_0e^{M_0}) (see also our response to Reviewer p5GV). This dependence is unavoidable, as shown in Proposition A.5 of the revised paper. A similar exponential dependence also appears in prior work such as Oko et al. '25, Lin & Mei '25, Han et al. '25, and Li et al. '26.

Comments on the paper

2 comments

Global response

We sincerely thank all reviewers for their careful reading and valuable feedback. We have responded to each reviewer's comments individually. We have also uploaded a revised paper in which all major changes are highlighted in blue. We hope that our responses and these changes help address the reviewers' comments. If you have further questions, we would be very happy to clarify.

The main changes in the revised paper are as follows.

(1) Reshuffled batches (Reviewers TU7F and D7aL): In the revised paper, we have added Theorem E.1, which extends Theorem 4.2 to the loss averaged over batches that are reshuffled at every epoch. It also covers approximate minimizers through an additional optimization error.

(2) Embedding dimension (Reviewers p5GV and D7aL): In the revised paper, we have explicitly stated after Theorem 4.5 how the embedding dimension pp grows with nn. We have also added results for a fixed pp (Lemma E.2 and Corollary E.3) and Theorem E.5, which shows that a fixed pp suffices when the PMI has low rank. In addition, we have added results for normalized scores (Lemma E.4).

(3) Constants (Reviewers TU7F and p5GV): In the revised paper, we have improved the constant of Proposition 4.1 from O(M0e2M0)O(M_0e^{2M_0}) to O(M0eM0)O(M_0e^{M_0}). We have also added Proposition A.5, which shows that the factor M0eM0M_0e^{M_0} cannot be improved.

Other changes in the revised paper are also highlighted in blue.

Below we collect the references cited in our responses.

  • Chen et al. '24. Understanding transferable representation learning and zero-shot transfer in CLIP. ICLR 2024.
  • Cho et al. '24. Mini-batch optimization of contrastive loss. Transactions on Machine Learning Research, 2024.
  • Han et al. '25. Deep mutual density ratio estimation with Bregman divergence and its applications. Journal of the American Statistical Association, 2025.
  • HaoChen et al. '21. Provable guarantees for self-supervised deep learning with spectral contrastive loss. NeurIPS 2021.
  • Hieu et al. '25. Generalization analysis for deep contrastive representation learning. AAAI 2025.
  • Jia et al. '21. Scaling up visual and vision-language representation learning with noisy text supervision. ICML 2021.
  • Lei et al. '23. Generalization analysis for contrastive representation learning. ICML 2023.
  • Li et al. '22. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. ICML 2022.
  • Li et al. '26. Statistical consistency and generalization of contrastive representation learning. ICML 2026.
  • Lin & Mei '25. A statistical theory of contrastive learning via approximate sufficient statistics. NeurIPS 2025.
  • Ma et al. '26. Optimal in-context adaptivity and distributional robustness of Transformers. NeurIPS 2026.
  • Maurer '19. A Bernstein-type inequality for functions of bounded interaction. Bernoulli, 2019.
  • Nakada & Imaizumi '20. Adaptive approximation and generalization of deep neural network with intrinsic dimensionality. Journal of Machine Learning Research, 2020.
  • Nakada et al. '23. Understanding multimodal contrastive learning and incorporating unpaired data. AISTATS 2023.
  • Nguyen et al. '10. Estimating divergence functionals and the likelihood ratio by convex risk minimization. IEEE Transactions on Information Theory, 2010.
  • Oko et al. '25. A statistical theory of contrastive pre-training and multimodal generative AI. arXiv:2501.04641, 2025.
  • Radford et al. '21. Learning transferable visual models from natural language supervision. ICML 2021.
  • Saunshi et al. '19. A theoretical analysis of contrastive unsupervised representation learning. ICML 2019.
  • Schmidt-Hieber '20. Nonparametric regression using deep neural networks with ReLU activation function. The Annals of Statistics, 2020.
  • Shen & Wong '94. Convergence rate of sieve estimates. The Annals of Statistics, 1994.
  • Tosh et al. '21. Contrastive learning, multi-view redundancy, and linear models. ALT 2021.
  • Uesaka et al. '25. Weighted point set embedding for multimodal contrastive learning toward optimal similarity metric. ICLR 2025.
  • van Elst & Ghoshdastidar '25. Tight PAC-Bayesian risk certificates for contrastive learning. SIAM Journal on Mathematics of Data Science, 2025.
  • Wainwright '19. High-Dimensional Statistics: A Non-Asymptotic Viewpoint. Cambridge University Press, 2019.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.