Surrogate Gap-Closing Score

• Surrogate Gap-Closing Score is a metric that quantifies the reduction in discrepancy between a tractable surrogate loss and the true performance metric.
• It leverages concepts like direct metric difference, regret transfer, and gradient alignment to rigorously assess gap closure across various domains.
• Practical applications include enhancing fairness, simulation-based likelihood estimation, and guiding sharpness-aware minimization with provable performance bounds.

A Surrogate Gap-Closing Score quantitatively characterizes the degree to which a surrogate loss, regularizer, or model modification succeeds in eliminating the mismatch (“surrogate gap”) between a tractable optimization objective and the true metric, constraint, or dynamical target of interest. Across modern machine learning, statistics, optimization, and physical modeling, the surrogate gap-closing paradigm enables practitioners to (1) define a rigorously measurable gap, (2) construct and analyze surrogate mechanisms aiming to close that gap, and (3) monitor, bound, or guarantee the fraction of the gap eliminated by the surrogate—thereby yielding provable alignment of practical optimization with ultimate performance or fairness goals.

1. Conceptual Foundation: Surrogate Gap and Gap-Closing Scores

The surrogate gap is the quantifiable discrepancy between the objective used during learning (usually a differentiable, tractable surrogate loss $\mathcal{L}$) and the true performance metric or constraint $\mathcal{M}$ one ultimately cares about—such as $F_\beta$-score, fair demographic parity, likelihood fit, or generalization sharpness. The Surrogate Gap-Closing Score (SGCS) is any metric that measures how much of this gap is closed by a particular surrogate modeling or optimization scheme.

Typically, the gap is defined in one of several ways:

• Direct metric difference: $|\mathcal{M}(f) - \widetilde{\mathcal{M}}_\phi(f)|$ for surrogate $\phi$ (e.g., fairness surrogate vs. true demographic parity).
• Regret transfer: The maximum possible regret under the target metric for a given amount of surrogate regret, yielding a ratio $S(\mathcal{L} \to \mathcal{M})$.
• Alignment in gradients or statistics: Discrepancy between model parameter gradients under surrogate and target (e.g., Hessian alignment, likelihood score matching).

The SGCS is formalized either as an absolute reduction in this gap or as the fraction of the initial gap that has been eliminated (e.g., $(\Delta_0 - \Delta_t)/\Delta_0$ at iteration $t$) (Yao et al., 2023, Xu et al., 29 Nov 2025, Shen et al., 12 May 2026).

2. Mathematical Formalizations and Canonical Examples

The surrogate gap and corresponding closing score admit rigorous mathematical forms, often tailored to specific domains:

Domain	Gap Definition	Gap-Closing Score Example
Classification	$\operatorname{Regret}_\mathcal{M}$	$S(\mathcal{L}\to\mathcal{M})$, the finite-sample slope
Fairness	$\mathcal{M}$0	$\mathcal{M}$1 fractional gap closure minus penalty
SBI / Likelihood	$\mathcal{M}$2	Fraction of KL/grad-alignment gap eliminated
Sharpness/generalization	$\mathcal{M}$3 (SAM/SAM+)	Surrogate gap $\mathcal{M}$4; GSAM decrease over SAM
Physical models	$\mathcal{M}$5 PDE residual (e.g., Fokker-Planck)	Decrease in Wasserstein-2 error/ODE-SDE gap
Optimization	Rank-correlation $\mathcal{M}$6	Threshold $\mathcal{M}$7, $\mathcal{M}$8 for monotonic descent
Policy learning	$\mathcal{M}$9	Proportion of regret/gain gap closed

Canonical examples:

• Classification metrics: The surrogate gap-closing score $F_\beta$0 is defined as $F_\beta$1, where $F_\beta$2 is the regret transfer function. A finite $F_\beta$3 gives strong reliability; infinite $F_\beta$4 warns of irreducible mismatch (Pu et al., 8 Mar 2026).
• Score-based methods: Augmenting the cross-entropy loss with simulation score matching in simulation-based inference can halve the KL gap between surrogate and target, yielding a gap-closing score (proportion of original KL gap eliminated) (Shen et al., 12 May 2026).
• Diffusion models: The $F_\beta$5 Fokker–Planck residual is the gap; decreasing it directly controls the Wasserstein-2 gap between ODE and SDE sample distributions (Deveney et al., 2023).
• Fairness: Fractional gap-closure is $F_\beta$6, with variance and instability penalties yielding a composite $F_\beta$7 (Yao et al., 2023).

3. Surrogate Gap-Closing Scores in Loss and Metric Alignment

A prominent application is the construction of surrogate losses that explicitly align their gradient fields with those of the desired non-differentiable metric. For example, in class-imbalanced binary classification, the non-differentiable $F_\beta$8 score induces a stationary condition involving the gradients of confusion matrix rates; standard losses (BCE, MAE) do not match this, creating a surrogate gap that may stall optimization at suboptimal $F_\beta$9. By deriving a surrogate loss whose stationary points yield the exact $|\mathcal{M}(f) - \widetilde{\mathcal{M}}_\phi(f)|$0 gradient condition, the surrogate gap can be made arbitrarily small in the limit $|\mathcal{M}(f) - \widetilde{\mathcal{M}}_\phi(f)|$1, effectively closing the gap (Lee et al., 2021).

Similarly, in information-geometric optimization (IGO) with model-based surrogates, monotonic improvement in the expected objective is only guaranteed if the surrogate quality—quantified via Kendall's $|\mathcal{M}(f) - \widetilde{\mathcal{M}}_\phi(f)|$2 or Pearson's $|\mathcal{M}(f) - \widetilde{\mathcal{M}}_\phi(f)|$3 correlation between surrogate and true objective ranking—exceeds an explicit threshold, e.g., $|\mathcal{M}(f) - \widetilde{\mathcal{M}}_\phi(f)|$4. This threshold is the surrogate gap–closing score: only surrogates whose gap (in rank-correlation) is smaller than $|\mathcal{M}(f) - \widetilde{\mathcal{M}}_\phi(f)|$5 guarantee progress (Akimoto, 2022).

4. Algorithmic Strategies and Practical Computation

Many works operationalize the surrogate gap-closing score within iterative algorithms:

• Fairness optimization: The Balanced Surrogate algorithm adaptively reweights group surrogates to minimize the current surrogate fairness gap. At each iteration $|\mathcal{M}(f) - \widetilde{\mathcal{M}}_\phi(f)|$6, the SGCS is calculated and monitored, tracking both the direct gap but also variance and large margin point instability (Yao et al., 2023).
• Simulation-based inference: Score-augmented neural surrogate likelihoods monitor KL and gradient-alignment gap closure relative to standard baseline surrogates (Shen et al., 12 May 2026).
• Sharpness-aware minimization: GSAM (Surrogate Gap Guided Sharpness-Aware Minimization) performs an explicit secondary update step in the orthogonal direction to the SAM loss, minimizing the surrogate gap in sharpness, with provable (per-iteration) gap reduction (Zhuang et al., 2022).
• Physical generative models: Fokker–Planck residual penalties are added to the loss to minimize the ODE–SDE distributional gap, which is precisely controlled by integrated $|\mathcal{M}(f) - \widetilde{\mathcal{M}}_\phi(f)|$7 norm of the residual (Deveney et al., 2023).

The gap-closing score is thus either recorded as a metric (fractional reduction, absolute difference, or per-iteration improvement) or used directly in regularization/acceptance criteria to ensure or monitor progress toward metric alignment.

5. Regret-Transfer, Bounds, and Theoretical Guarantees

A major use of surrogate gap-closing scores is to provide provable bounds on the worst-case misalignment between surrogate and target metrics. The regret-transfer framework yields finite-sample and asymptotic upper bounds of the form

$|\mathcal{M}(f) - \widetilde{\mathcal{M}}_\phi(f)|$8

where $|\mathcal{M}(f) - \widetilde{\mathcal{M}}_\phi(f)|$9 is the surrogate gap-closing score for loss $\phi$0 and metric $\phi$1. If $\phi$2 is finite and small, improvements in $\phi$3 guarantee proportional improvements in $\phi$4 (Pu et al., 8 Mar 2026). If $\phi$5 is infinite (e.g., when $\phi$6 is logistic loss and $\phi$7 is NDCG), arbitrarily small surrogate loss regret may yield no improvement in the primary metric, reflecting a fundamentally uncloseable surrogate gap.

In decision and treatment policy learning, ratio-type gap-closing scores measure the proportion of potential gain realized by a surrogate-based policy relative to optimal and baseline policies, with doubly robust estimators achieving $\phi$8-convergence (Xu et al., 29 Nov 2025).

In IGO, explicit per-iteration improvements in the expected objective are only guaranteed when the surrogate’s rank-correlation surpasses a threshold; otherwise, no theoretical improvement can be ensured (Akimoto, 2022).

6. Applications Across Domains

Surrogate gap-closing scores have been employed in:

• Metric-aligned loss design: $\phi$9-sensitive surrogates, ranking losses, and evaluation systems for alignment of offline and online metrics (Lee et al., 2021, Pu et al., 8 Mar 2026).
• Algorithmic fairness: Sigmoid and hinge surrogates for demographic parity, with explicit composite gap-closure + stability measurements (Yao et al., 2023).
• Simulation-based inference: Efficient neural approximations of likelihood/posterior surfaces, dramatically narrowing finite-sample errors (Shen et al., 12 May 2026).
• Generative modeling: Fokker–Planck residual minimization for ODE/SDE sample consistency in diffusion frameworks (Deveney et al., 2023).
• Evolutionary strategies: Monotone progress guarantees in IGO/CMA-ES under surrogate-based fitness evaluation (Akimoto, 2022).
• Treatment regime evaluation: Surrogate regret and efficiency metrics for evaluating surrogacy under budget constraints in individualized policies (Xu et al., 29 Nov 2025).
• Quantum systems: Curvature-based surrogate scores for gap closing in Majorana nanowire parameter extraction (Pan et al., 2018).

7. Structural Asymmetries, Limitations, and Trade-offs

Surrogate gap-closing scores make explicit the structural limitations of surrogate-based optimization. As shown in inter-metric transfer analyses, the gap may not be closeable in all directions: for example, pointwise classification surrogates cannot, even at zero regret, guarantee zero ranking metric regret, yielding $S(\mathcal{L} \to \mathcal{M})$0 (Pu et al., 8 Mar 2026). Trade-offs also arise when gap closure in one domain sacrifices performance in another (e.g., regularizing Fokker–Planck residual improves ODE–SDE alignment but may degrade SDE sample quality) (Deveney et al., 2023). In practical deployment, variance and large-margin instability penalization often factors into composite gap-closing metrics (Yao et al., 2023).

These scores, when properly constructed and interpreted, enable both principled loss/algorithm selection and transparent, measurable monitoring of metric-alignment progress in real-world and theoretical optimization scenarios.