acceptodds
Under review as a conference paper at ICLR 2027

Cost-Weighted Cross-Entropy for Diffusion-Based Decision-Focused Learning

Abstract

*Decision-Focused Learning* (*DFL*) trains predictive models for downstream decision quality rather than predictive accuracy alone. For stochastic discrete optimization, however, decision-focused learning faces two fundamental challenges: downstream decisions are non-differentiable, while nonlinear objectives may depend on the full conditional outcome distribution rather than point predictions. To address these challenges, we propose *Cost-Weighted Cross-Entropy* (*CWCE*), a differentiable decision-alignment loss that uses realized downstream costs to align the expected-cost profile induced by the learned distribution without differentiating through the discrete optimizer. We establish two complementary theoretical results. First, arbitrarily small distributional error can induce arbitrarily large decision regret, demonstrating that distributional fidelity alone provides no general control of downstream decision quality. Second, when CWCE is incorporated into diffusion training, the excess risk of the resulting joint objective quantitatively controls downstream decision regret. These results motivate combining diffusion training, which promotes fidelity in modeling conditional outcome uncertainty, with CWCE, which provides explicit decision-relevant supervision. Experiments on three stochastic discrete optimization tasks demonstrate consistent improvements over deterministic DFL, ranking-based, and diffusion-based baselines in downstream decision quality and regret.

Primary area:
optimization
Keywords:
Decision-Focused Learning, Surrogate Loss Functions, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_SMC7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a diffusion-based decision-focused learning framework equipped with a novel cost-weighted cross-entropy loss. In addition to the standard diffusion model loss, this new loss aligns the normalized expected cost profile with the learned distribution. While the method achieves substantial experimental gains, comparisons against several foundational DFL baselines are absent.

Critical strengths
  1. The proposed method is technically sound: it introduces an auxiliary loss, alongside the diffusion model loss, to enforce alignment for decision-making tasks.
  2. Empirical results demonstrate considerable performance improvements over existing diffusion-based DFL methods.
Critical weaknesses
  1. Notation requires clarification. For instance, in lines 48–49 of the introduction, symbols �0��0� appear without prior definition, which may confuse readers. The meaning of ntn_t is also left unspecified in line 395.
  2. Baseline comparisons are limited primarily to diffusion-based DFL approaches, omitting other relevant baselines. Although the paper’s core contribution targets improved alignment for diffusion-based solvers, comprehensive comparisons against classic DFL methods (e.g., [1,2]) remain essential.
  3. The related work discussion appears insufficient.

[1] S. Shah, K. Wang, B. Wilder, A. Perrault, and M. Tambe. Decision-focused learning without decision-making: Learning locally optimized decision losses. (NeurIPS), 2022.

[2] J. Mandi, P. J. Stuckey, T. Guns, et al. Smart predict-and-optimize for hard combinatorial optimization problems. (AAAI), 2020.

Questions for authors
  1. Scope of applicability. Section 2 formulates the optimization problem in the discrete setting. Can the proposed method extend to continuous settings? Are the challenges outlined in the paper exclusive to discrete problems? If yes, this distinction should be explicitly stated in the introduction (and also title).
3 comments

W1. Notation requires clarification. For instance, in lines 48–49 of the introduction, symbols appear without prior definition, which may confuse readers. The meaning of is also left unspecified in line 395.

AW1. We thank the reviewer for pointing out the unclear notation. We agree that introducing mathematical symbols without prior definitions in Introduction Section may affect readability. To address this concern, we have revised Lines 48–49 to explain the distinction between deterministic and distribution-aware decision-making in words, without introducing unnecessary mathematical notation. Specifically, we clarify that, for certain structured objectives, such as those that are linear in the decision variables, conditional expected costs depend only on specific conditional expectations, whereas more general nonlinear stochastic objectives may require additional distributional information. This revision improves readability while preserving the original motivation. We clarify that ntn_t denotes the number of test samples, and Dt=(xi,yi)i=1ntD_t=\\{(x_i,y_i)\\}_{i=1}^{n_t} denotes the test set.

W2. Baseline comparisons are limited primarily to diffusion-based DFL approaches, omitting other relevant baselines. Although the paper’s core contribution targets improved alignment for diffusion-based solvers, comprehensive comparisons against classic DFL methods (e.g., [1,2]) remain essential.

AW2. We thank the reviewer for emphasizing the importance of comparisons with classical DFL methods. We would first like to clarify that our original experiments already include several representative deterministic DFL baselines, namely Pointwise Ranking, Listwise Ranking, and the Quadratic surrogate proposed by Shah et al. [1], in addition to PtO and the generative DFL methods Diffusion-DFL and Gen-DFL. Therefore, our evaluation is not limited to diffusion-based DFL approaches.

To further strengthen the experimental comparison, we have additionally included SPO+, proposed by Mandi et al. [2], as suggested by the reviewer. The new results are reported below.

MethodWeb nDQ ↑Web Regret ↓Portfolio nDQ ↑Portfolio Regret ↓Recommendation nDQ ↑Recommendation Regret ↓
Quadratic [1]3.22 ± 1.9980.03 ± 1.7626.11 ± 4.91129.72 ± 7.208.48 ± 1.3259.31 ± 0.39
SPO+ [2]29.50 ± 1.6857.42 ± 1.4422.30 ± 4.21135.16 ± 3.5013.69 ± 8.4355.99 ± 5.39
CWCE (Ours)51.63 ± 5.8339.66 ± 4.6856.77 ± 0.6783.28 ± 1.3636.38 ± 2.7242.31 ± 1.85

These additional results demonstrate that CWCE consistently outperforms both the Quadratic surrogate [1] and SPO+ [2] across all three tasks, achieving higher normalized decision quality (nDQ) and lower regret. Together with our existing comparisons, these results provide a more comprehensive evaluation against classical deterministic and recent generative DFL methods.

W3. The related work discussion appears insufficient.

AW3. We thank the reviewer for this valuable suggestion. We agree that a more systematic discussion of related work would help better contextualize our contributions. To address this concern, we have added a dedicated Related Work section, organized into two main categories: deterministic DFL and generative DFL.

For deterministic DFL, we provide a more comprehensive discussion of differentiable optimization approaches and decision-focused surrogate losses, including SPO+, locally optimized surrogates, and pointwise/listwise ranking-based methods. For generative DFL, we expand the discussion of recent approaches, particularly Gen-DFL and Diffusion-DFL, and clarify their methodological differences from CWCE in terms of decision-aware supervision and the treatment of downstream discrete optimization.

We have also streamlined the Introduction to avoid redundant discussions of individual methods. These revisions provide a more comprehensive overview of existing research and more clearly position CWCE within the DFL literature.

Q1. Scope of applicability. Section 2 formulates the optimization problem in the discrete setting. Can the proposed method extend to continuous settings? Are the challenges outlined in the paper exclusive to discrete problems? If yes, this distinction should be explicitly stated in the introduction (and also title).

AW1. We thank the reviewer for raising this important question. We clarify that CWCE is not inherently restricted to discrete decision spaces and can also be extended to continuous optimization problems. Although our current formulation and experiments focus on discrete optimization, the underlying principle of aligning decision-relevant expected-cost profiles is more broadly applicable.

Extension to continuous decision spaces. The main difference is that a continuous decision space Z\mathcal{Z} generally contains infinitely many feasible decisions, making it impractical to compute CWCE over the entire decision space. However, a similar computational challenge arises in large-scale discrete combinatorial optimization problems, where enumerating all feasible decisions is also prohibitively expensive.

To address this challenge, we can adopt the same finite candidate-set approximation already introduced in our paper (Lines 287–294), following the candidate-based strategies used in Mandi et al. (2022) and Wang et al. (2026). Specifically, we construct a finite candidate set S⊂Z\mathcal{S}\subset\mathcal{Z} and evaluate CWCE over S\mathcal{S} instead of the entire decision space.

We initialize the candidate set using the optimal decisions associated with observed training outcomes:

z^{\star}(y_i)\in\arg\min_{z\in\mathcal{Z}}f(y_i,z).$$ During training, we additionally draw $M$ samples $\\{y_i^{(m)}\\}_{m=1}^{M}\sim p_{\theta}(\cdot\mid x_i)$ and, with probability $p_{\mathrm{solve}}=0.1$, compute $$\widehat z(x_i)\in\arg\min_{z\in\mathcal{Z}} \frac{1}{M}\sum_{m=1}^{M}f(y_i^{(m)},z).$$ We then update $\mathcal{S}\leftarrow\mathcal{S}\cup\{\widehat z(x_i)\}$, allowing the candidate set to adapt to decisions that become relevant under the evolving learned distribution. This procedure avoids exhaustive enumeration of $\mathcal{Z}$ and is applicable to both discrete and continuous optimization problems, provided that the corresponding downstream optimization problems can be solved. Thus, the finite candidate-set formulation offers a practical way to extend CWCE to continuous decision spaces, although its approximation quality depends on the coverage of decision-relevant candidates. **Challenges in discrete and continuous settings.** The fundamental challenges motivating our work are not exclusive to discrete optimization. First, distributional fidelity does not necessarily guarantee downstream decision quality, as even small distributional errors can alter the expected-cost structure and induce suboptimal decisions. Second, for general nonlinear stochastic objectives, deterministic point predictions may fail to capture the distributional information required for optimal decision-making. Both challenges arise regardless of whether the decision space is discrete or continuous. However, discrete optimization presents an additional difficulty: the downstream argmin mapping is generally non-differentiable, making direct gradient propagation through the optimizer particularly challenging. Although differentiability issues may also arise in continuous optimization, continuous problems can admit implicit differentiation or other gradient-based approaches under suitable regularity conditions. CWCE avoids differentiating through the downstream optimizer by directly aligning the expected-cost profiles induced by the learned distribution. **Clarification of scope.** Therefore, our focus on discrete optimization is motivated by the particular difficulty of incorporating decision-aware supervision through non-differentiable downstream optimization, rather than by the assumption that distributional misalignment is exclusive to discrete problems. We have clarified this distinction in the Introduction. The current title, *Cost-Weighted Cross-Entropy for Diffusion-Based Decision-Focused Learning*, does not restrict the proposed framework to discrete optimization and therefore remains appropriate.

Dear Reviewer,

Thank you again for your constructive feedback. We have addressed your concerns with additional clarifications and experimental results in our rebuttal. As the discussion period is approaching its end, we would greatly appreciate it if you could review our responses and let us know whether they adequately address your concerns.

Thank you for your time and consideration.

Reviewer_8hsLOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed a new loss function for Decision Focused Learning in stochastic context, where the predictor is assumed to have distributional output -- i.e. settings where, given a sample from a distribution that represents a set of scenarios, optimal decisions should be taken so as to minimize the expected cost over those scenarios. The authors mainly assume that distribution estimates are obtained by sampling from a diffusion model.

The authors suggest using a loss function that relies on two components: 1) an normalized estimation of the expected cost profile of a finite number of candidates; and 2) weights associated to the cost of each of the candidates under a given realized sample. The resulting functions has a structure that closely resembles a cross-entropy (though it is not one), which motivates the paper title.

They formally prove two statement that suggest how the proposed loss might be effective, and proceed to quantify its efficacy in empirical experiments. The new approach considerably outperforms two existing baselines designed for the same setting and a few deterministic DFL approaches (though it should be said that the considered benchmark have been selected exactly because they are challenging for deterministic DFL).

Critical strengths

The proposed approach is simple and, at least in the considered benchmarks, quite effective against baselines designed for the same settings, but focusing on accurate distribution estimates.

This also suggest that the (formally proved) statement about the significance of the cost values, besides distribution accuracy, is practically relevant and not just a theoretical oddity.

The method reliance on a generic pool of candidates, rather than on the solution corresponding to the current estimated distribution, also allows to collect samples sporadically, and as a result can greatly reduce the computational cost. It should be said that I could not find details on the sampling schedule in the paper and that the impact of sampling frequency is not explicitly investigated.

Critical weaknesses

Probably the main concern I have about this paper is its degree of novelty: the core ideas behind the proposed loss mirror some of those proposed in [1] and the result from Section 4.1 essentially posits that the cost surface may not be aligned with the likelihood surface, even when their optima match. Overall, both the key insights appears to have been established in deterministic DFL and the work essentially seems to generalize them.

In fact, the proposed loss could work out of the box for deterministic DFL (unless I misunderstood something, but that does not seem very likely).

To be fair, generalizing or adapting existing results is still a welcome contribution, but the presentation does not seem to adequately acknowledge the dependence on prior work, and a generalization is usually a weaker contribution compared to a novel insight.

As a second, less critical issue, the theoretical properties of the considered loss are not well established. The paper contains two formal proofs, but both result only suggest that the proposed loss should work well: they do not bound its estimation error or quantify a possible skew in the location of the true local minima. That said, the empirical results are good enough to make the need for a formal characterization less critical.

My last concern is about the chosen DFL context: DFL tends to be computationally expensive, and stochastic problems are among the computationally challenging ones in optimization. This fact tends to be downplayed a bit in the paper, through an attentive choice of the empirical benchmarks. In fact, the structure of all chosen problems is such that: 1) deterministic DFL methods -- which rely on a single scenario -- are structurally limited in their ability to output the optimal solutions; and 2) no second-stage decisions are present -- also called recourse actions -- so as to keep the solution time reasonable.

Again, there is nothing wrong with choosing benchmarks where a method has a good chance of succeeding (it would be unreasonable to do the opposite), but more transparency on the limitations of the chosen settings would be welcome.

[1] Mandi, Jayanta, et al. "Decision-focused learning: Through the lens of learning to rank." International conference on machine learning. PMLR, 2022.

Questions for authors
  1. What are the substantial differences of the proposed loss w.r.t (for example) the learning-to-rank approaches?

  2. Do you have an intuition about the degree by which the smoothing implicit inthe use of CE may skew the location of the true local optima?

After the response

The authors have provided substantial new information and clarifications in their response, and addressed most of my concerns. I am changing my score accordingly.

6 comments

AW1 & AQ1. Novelty of CWCE and its differences from learning-to-rank approaches.

We thank the reviewer for raising this important concern. We agree that CWCE and existing learning-to-rank methods share the motivation of incorporating downstream costs into model training. In particular, we acknowledge that the importance of cost ordering for downstream decision-making has been established in deterministic DFL. However, CWCE is not simply an application of an existing ranking loss to a generative model. Its main distinction lies in directly learning the conditional expected-cost profile under outcome uncertainty, together with a population-level characterization and a quantitative decision-regret guarantee. We clarify these differences below.

1. Existing ranking losses learn a deterministic outcome prediction for each input, whereas CWCE explicitly models outcome uncertainty.

The pointwise and listwise ranking approaches considered in our experiments learn a deterministic outcome prediction y^θ(x)\widehat y_\theta(x) for each input xx. Their predicted decision costs are consequently determined by f(y^θ(x),z)f(\widehat y_\theta(x),z). Pointwise ranking minimizes the squared discrepancy between predicted and realized decision costs, while listwise ranking aligns soft ranking distributions constructed from these costs.

Although both methods use realized costs as supervision, their predictive representation is a single outcome estimate rather than a conditional outcome distribution. In particular, they do not explicitly model the uncertainty of yy conditional on xx.

In contrast, CWCE starts from a conditional generative model pθ(y∣x)p_\theta(y\mid x), which represents uncertainty over possible outcomes. It computes the expected cost of each candidate decision under the learned distribution,

Eθ(z;x)=Ey^∼pθ(⋅∣x)[f(y^,z)],\mathcal E_\theta(z;x)=\mathbb E_{\widehat y\sim p_\theta(\cdot\mid x)}[f(\widehat y,z)],

and constructs the normalized expected-cost profile,

πθ(z∣x)=Eθ(z;x)∑z′∈ZEθ(z′;x).\pi_\theta(z\mid x)=\frac{\mathcal E_\theta(z;x)}{\sum_{z'\in\mathcal Z}\mathcal E_\theta(z';x)}.

CWCE then uses the realized downstream costs to supervise this distribution-induced profile:

LCW(θ;x,y)=−∑z∈Zf(y,z)log⁡πθ(z∣x).\mathcal L_{\mathrm{CW}}(\theta;x,y)=-\sum_{z\in\mathcal Z}f(y,z)\log\pi_\theta(z\mid x).

Therefore, CWCE does not merely seek to align the rankings induced by a deterministic prediction and a realized outcome. Instead, it aims to align the expected-cost structure induced by the learned conditional distribution with that of the true conditional distribution. This distinction is important in stochastic DFL, where the optimal decision is determined by conditional expected costs rather than the costs associated with a single predicted outcome.

2. Deterministic ranking losses can fail to recover the Bayes-optimal decision under outcome uncertainty.

The distinction above has important consequences for decision consistency. We discuss the limitations of two representative ranking losses.

Pointwise ranking. The pointwise squared-cost loss is defined as

LPR(y,y^)=1∣Z∣∑z∈Z(f(y^,z)−f(y,z))2.L_{\mathrm{PR}}(y,\widehat y)=\frac{1}{|\mathcal Z|}\sum_{z\in\mathcal Z}\bigl(f(\widehat y,z)-f(y,z)\bigr)^2.

For a fixed input xx, let Eq(z;x)=Ey∼q(⋅∣x)[f(y,z)]\mathcal E_q(z;x)=\mathbb E_{y\sim q(\cdot\mid x)}[f(y,z)]. Its conditional population risk admits the decomposition

Eq[LPR(y,y^)∣x]=1∣Z∣∑z∈Z[(f(y^,z)−Eq(z;x))2+Var⁡q(f(y,z)∣x)].\mathbb E_q[L_{\mathrm{PR}}(y,\widehat y)\mid x]=\frac{1}{|\mathcal Z|}\sum_{z\in\mathcal Z}\left[\bigl(f(\widehat y,z)-\mathcal E_q(z;x)\bigr)^2+\operatorname{Var}_q(f(y,z)\mid x)\right].

Thus, minimizing pointwise ranking amounts to approximating the true conditional expected-cost profile using a cost profile induced by a single deterministic prediction. If the true expected-cost profile is representable, i.e., for every xx, there exists a prediction y^\hat{y} such that f(y^,z)=Eq(z;x)f(\hat{y},z) = \mathcal E_q(z;x) holds for all z∈Zz\in \mathcal{Z}, then the population pointwise ranking minimizer recovers the Bayes-optimal decision. However, when it is not representable by any deterministic prediction, the resulting approximation may alter the optimal decision.

For example, consider two possible outcomes y1,y2y_1,y_2 with probabilities q(y1∣x)=0.2q(y_1\mid x)=0.2 and q(y2∣x)=0.8q(y_2\mid x)=0.8, and two decisions z1,z2z_1,z_2. Let their costs be

f(y1,z1)=0,f(y1,z2)=10,f(y2,z1)=2,f(y2,z2)=0.f(y_1,z_1)=0,\quad f(y_1,z_2)=10,\quad f(y_2,z_1)=2,\quad f(y_2,z_2)=0.

The true conditional expected costs are

Eq(z1;x)=1.6,Eq(z2;x)=2,\mathcal E_q(z_1;x)=1.6,\qquad \mathcal E_q(z_2;x)=2,

so the Bayes-optimal decision is z1z_1. Suppose the deterministic prediction is restricted to the outcome space {y1,y2}\{y_1,y_2\}. The population pointwise loss is minimized by predicting y2y_2, since its squared cost-profile approximation error is smaller than that of y1y_1. However, this prediction induces decision z2z_2, which is not Bayes-optimal.

This example illustrates that the deterministic point-prediction parameterization can introduce a representation limitation under outcome uncertainty.

Listwise ranking. The listwise ranking loss constructs a soft decision distribution from the realized costs,

pτ(z∣y)=exp⁡(−f(y,z)/τ)∑z′∈Zexp⁡(−f(y,z′)/τ),p_\tau(z\mid y)=\frac{\exp(-f(y,z)/\tau)}{\sum_{z'\in\mathcal Z}\exp(-f(y,z')/\tau)},

and minimizes cross-entropy between the target and predicted soft distributions.

At the population level, when the target soft distribution is representable, the resulting decision satisfies

zLR(x)∈arg⁡max⁡z∈ZEy∼q(⋅∣x)[pτ(z∣y)].z_{\mathrm{LR}}(x)\in\arg\max_{z\in\mathcal Z}\mathbb E_{y\sim q(\cdot\mid x)}[p_\tau(z\mid y)].

However, the Bayes-optimal decision satisfies

z∗(x)∈arg⁡min⁡z∈ZEy∼q(⋅∣x)[f(y,z)].z^*(x)\in\arg\min_{z\in\mathcal Z}\mathbb E_{y\sim q(\cdot\mid x)}[f(y,z)].

These decisions need not coincide because softmax is nonlinear and does not generally commute with expectation.

To illustrate this, consider two outcomes with probabilities 0.10.1 and 0.90.9, respectively, and costs f(y1,z1)=0f(y_1,z_1)=0, f(y1,z2)=100f(y_1,z_2)=100, f(y2,z1)=10f(y_2,z_1)=10, and f(y2,z2)=9f(y_2,z_2)=9. The true expected costs are 99 and 18.118.1, respectively, making z1z_1 Bayes-optimal. However, with τ=1\tau=1, the expected softmax probabilities are approximately 0.34210.3421 for z1z_1 and 0.65790.6579 for z2z_2, incorrectly favoring z2z_2.

3. CWCE establishes an explicit connection between its population objective and downstream decision regret.

Beyond the difference in learning formulation, our theoretical contribution establishes a quantitative relationship between CWCE and downstream decision quality.

First, we show that arbitrarily small distributional error can induce arbitrarily large downstream decision regret, demonstrating that distributional fidelity alone does not generally control decision quality.

Second, we characterize the population CWCE objective through the KL divergence between normalized expected-cost profiles. Specifically, defining πq(z∣x)=Eq(z;x)/∑z′Eq(z′;x)\pi_q(z\mid x)=\mathcal E_q(z;x)/\sum_{z'}\mathcal E_q(z';x), we obtain

Ey∼q(⋅∣x)[LCW(θ;x,y)]=Sq(x)[H(πq(⋅∣x))+DKL(πq(⋅∣x)∥πθ(⋅∣x))],\mathbb E_{y\sim q(\cdot\mid x)}[\mathcal L_{\mathrm{CW}}(\theta;x,y)]=S_q(x)\left[H(\pi_q(\cdot\mid x))+D_{\mathrm{KL}}(\pi_q(\cdot\mid x)\|\pi_\theta(\cdot\mid x))\right],

where Sq(x)=∑z∈ZEq(z;x)S_q(x)=\sum_{z\in\mathcal Z}\mathcal E_q(z;x).

Under the stated assumptions, minimizing population CWCE therefore aligns the normalized expected-cost profiles and preserves their decision ordering.

Furthermore, incorporating CWCE into diffusion training yields the following regret guarantee:

Ljoint(p)−Ljoint(q)≤ϵ⟹Regret⁡q(p)≤2Bϵλ.\mathcal L_{\mathrm{joint}}(p)-\mathcal L_{\mathrm{joint}}(q)\leq\epsilon\quad\Longrightarrow\quad\operatorname{Regret}_q(p)\leq\sqrt{\frac{2B\epsilon}{\lambda}}.

This result establishes a quantitative connection between the excess risk of the joint diffusion and CWCE objective and downstream decision regret. Specifically, minimizing the population excess risk of the joint objective tightens the theoretical upper bound on decision regret, which approaches zero as the excess risk vanishes. This provides theoretical justification for our joint training objective, demonstrating that optimizing distributional fidelity together with decision-aware cost alignment can yield near-optimal downstream decisions under the stated assumptions.

AW2. Theoretical properties of CWCE, including estimation error and local minima.

We thank the reviewer for highlighting the importance of a more comprehensive theoretical characterization of CWCE. We agree that finite-sample estimation error and the behavior of local minima are important considerations. Our existing analysis establishes population-level decision calibration and a quantitative excess-risk-to-regret guarantee. To further address the reviewer's concern, we provide an additional finite-sample analysis that explicitly characterizes the statistical estimation error of the joint diffusion–CWCE objective and its effect on downstream decision regret.

1. Population-level characterization

We first recall that the population CWCE objective admits an exact characterization in terms of the KL divergence between normalized expected-cost profiles:

LCW(p)−LCW(q)=Ex[Sq(x)DKL(πq(⋅∣x)∥πp(⋅∣x))],\mathcal L_{\mathrm{CW}}(p)-\mathcal L_{\mathrm{CW}}(q)=\mathbb E_x\left[S_q(x)D_{\mathrm{KL}}(\pi_q(\cdot\mid x)\|\pi_p(\cdot\mid x))\right],

where Sq(x)=∑z∈ZEq(z;x)S_q(x)=\sum_{z\in\mathcal Z}\mathcal E_q(z;x), and πq\pi_q and πp\pi_p denote the normalized expected-cost profiles induced by the true and learned conditional distributions, respectively.

This identity shows that, under the stated realizability and positivity assumptions, minimizing population CWCE aligns the learned expected-cost profile with the true profile up to a positive normalization factor, thereby preserving the Bayes-optimal decision.

Furthermore, Theorem 4.2 establishes that the population excess risk of the joint objective quantitatively controls downstream decision regret:

Ljoint(p)−Ljoint(q)≤ϵ⟹Regret⁡q(p)≤2Bϵλ,\mathcal L_{\mathrm{joint}}(p)-\mathcal L_{\mathrm{joint}}(q)\le\epsilon\quad\Longrightarrow\quad\operatorname{Regret}_q(p)\le\sqrt{\frac{2B\epsilon}{\lambda}},

where λ>0\lambda>0 is the fixed coefficient of CWCE and BB bounds the expected total true cost.

We next extend this population-level guarantee to finite-sample learning.

2. Finite-sample estimation and decision-regret guarantee

Consider a hypothesis class P\mathcal P of conditional generative distributions and an i.i.d. training dataset Dn={(xi,yi)}i=1nD_n=\{(x_i,y_i)\}_{i=1}^n sampled from q(x,y)q(x,y).

For any p∈Pp\in\mathcal P, define the per-sample joint loss as

gp(x,y)=ℓDM(p;x,y)+λℓCW(p;x,y),g_p(x,y)=\ell_{\mathrm{DM}}(p;x,y)+\lambda\ell_{\mathrm{CW}}(p;x,y),

where ℓDM\ell_{\mathrm{DM}} denotes the diffusion loss after taking the expectation over the diffusion timestep and noise, and

ℓCW(p;x,y)=−∑z∈Zf(y,z)log⁡πp(z∣x).\ell_{\mathrm{CW}}(p;x,y)=-\sum_{z\in\mathcal Z}f(y,z)\log\pi_p(z\mid x).

The corresponding population and empirical objectives are

Ljoint(p)=E(x,y)∼q[gp(x,y)],L^joint(p)=1n∑i=1ngp(xi,yi).\mathcal L_{\mathrm{joint}}(p)=\mathbb E_{(x,y)\sim q}[g_p(x,y)],\qquad \widehat{\mathcal L}_{\mathrm{joint}}(p)=\frac1n\sum_{i=1}^n g_p(x_i,y_i).

We introduce the following assumptions for the finite-sample analysis:

  • A1 (Bounded costs). The decision set Z\mathcal Z is finite, with K=∣Z∣K=|\mathcal Z|, and 0≤f(y,z)≤F0\le f(y,z)\le F for all y,zy,z.
  • A2 (Positive cost profiles). There exists ρ>0\rho>0 such that πp(z∣x)≥ρ\pi_p(z\mid x)\ge\rho for all p∈Pp\in\mathcal P, xx, and zz.
  • A3 (Bounded diffusion loss). There exists MDM>0M_{\mathrm{DM}}>0 such that 0≤ℓDM(p;x,y)≤MDM0\le\ell_{\mathrm{DM}}(p;x,y)\le M_{\mathrm{DM}} for all p,x,yp,x,y.
  • A4 (Realizability). The true conditional distribution qq belongs to P\mathcal P and minimizes the population diffusion loss over P\mathcal P.
  • A5 (Approximate empirical minimization). The learned model p^n\widehat p_n satisfies L^joint(p^n)≤inf⁡p∈PL^joint(p)+ηopt\widehat{\mathcal L}_{\mathrm{joint}}(\widehat p_n)\le\inf_{p\in\mathcal P}\widehat{\mathcal L}_{\mathrm{joint}}(p)+\eta_{\mathrm{opt}}, where ηopt≥0\eta_{\mathrm{opt}}\ge0.

Define the induced joint loss class G={gp:p∈P}\mathcal G=\{g_p:p\in\mathcal P\}, and let Rn(G)\mathfrak R_n(\mathcal G) denote its expected Rademacher complexity.

Theorem (Finite-Sample Estimation and Decision-Regret Guarantee). Under Assumptions A1–A5 and the conditions of Theorem 4.2, define M=MDM+λKFlog⁡(1/ρ)M=M_{\mathrm{DM}}+\lambda KF\log(1/\rho). Then, for any δ∈(0,1)\delta\in(0,1), with probability at least 1−δ1-\delta, the population excess risk of the learned model satisfies

Ljoint(p^n)−Ljoint(q)≤4Rn(G)+2Mlog⁡(2/δ)2n+ηopt.\mathcal L_{\mathrm{joint}}(\widehat p_n)-\mathcal L_{\mathrm{joint}}(q)\le4\mathfrak R_n(\mathcal G)+2M\sqrt{\frac{\log(2/\delta)}{2n}}+\eta_{\mathrm{opt}}.

Consequently, its downstream decision regret satisfies

Regret⁡q(p^n)≤2Bλ[4Rn(G)+2Mlog⁡(2/δ)2n+ηopt].\operatorname{Regret}_q(\widehat p_n)\le\sqrt{\frac{2B}{\lambda}\left[4\mathfrak R_n(\mathcal G)+2M\sqrt{\frac{\log(2/\delta)}{2n}}+\eta_{\mathrm{opt}}\right]}.

The above results provide a finite-sample characterization of the proposed joint objective. In particular, the decision-regret bound explicitly accounts for the statistical complexity of the joint loss class, the number of training samples, and the empirical optimization error. Moreover, when Rn(G)→0\mathfrak R_n(\mathcal G)\to 0 and ηopt→0\eta_{\mathrm{opt}}\to 0 as n→∞n\to\infty, the decision regret converges to zero, providing a finite-sample extension of our population-level calibration guarantee.

Regarding local minima, the optimization error ηopt\eta_{\mathrm{opt}} captures the effect of empirical suboptimality on decision regret, while characterizing the locations of local minima in the nonconvex parameter space would require additional assumptions on the optimization landscape.

Proof. We establish the result in three steps.

Step 1: Boundedness of the joint loss.

By Assumption A1, 0≤f(y,z)≤F0\le f(y,z)\le F. Moreover, Assumption A2 guarantees that ρ≤πp(z∣x)≤1\rho\le\pi_p(z\mid x)\le1. Therefore,

0≤−log⁡πp(z∣x)≤log⁡(1/ρ).0\le-\log\pi_p(z\mid x)\le\log(1/\rho).

It follows that the per-sample CWCE loss is uniformly bounded:

0≤ℓCW(p;x,y)=−∑z∈Zf(y,z)log⁡πp(z∣x)≤KFlog⁡(1/ρ).0\le\ell_{\mathrm{CW}}(p;x,y)=-\sum_{z\in\mathcal Z}f(y,z)\log\pi_p(z\mid x)\le KF\log(1/\rho).

Combining this inequality with Assumption A3 gives

0≤gp(x,y)≤MDM+λKFlog⁡(1/ρ)=M.0\le g_p(x,y)\le M_{\mathrm{DM}}+\lambda KF\log(1/\rho)=M.

Thus, every function in G\mathcal G takes values in [0,M][0,M].

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.