Surrogate Information Gain (SIG)

• Surrogate Information Gain (SIG) is a family of proxy objectives that replace intractable direct information measures with computationally feasible alternatives while preserving decision relevance.
• In offline quantum experimental design, SIG quantifies the prior variance of measurement signals, achieving notable efficiency gains such as up to an 85% reduction in measurement time.
• Extensions of SIG span optimal experimental design, reinforcement learning, and decision-tree corrections, addressing diverse computational, statistical, and practical challenges.

Surrogate Information Gain (SIG) denotes a class of information-gain-style objectives in which a target notion of informativeness is replaced by a tractable, better calibrated, or better behaved proxy. The term is not standardized across the literature. In the most explicit usage in the cited corpus, SIG is a formal variance-based figure of merit for offline quantum experimental design (Varona-Uriarte et al., 29 Aug 2025). In other papers, closely related constructions appear as adjoint-compatible surrogates of expected information gain, surrogate-loss-induced statistical information, or corrected information-gain criteria, but without adopting the same name (Montella et al., 27 Mar 2026, Duchi et al., 2016, Leroux et al., 2018). The acronym is also ambiguous: in a medical-LLM reinforcement-learning paper, SIG means Shapley Information Gain, not Surrogate Information Gain (Ding et al., 19 Aug 2025).

1. Terminological scope and conceptual boundaries

The literature does not present a single canonical SIG formalism. What is shared is a recurring pattern: a direct information objective is unavailable, too expensive, too biased, or poorly aligned with the operational task, and is therefore replaced by a surrogate that preserves some of its semantics while improving tractability or practical behavior. This suggests a broader “SIG family” (Editor’s term) of objectives that stand in for raw information gain under explicit computational, statistical, or decision-theoretic constraints.

Usage in the literature	Core quantity	Status
Surrogate Information Gain	$$\mathbb{E}\{\mathrm{Var}_{\vec A\sim p(\vec A)}[P_{\pm x}(\tau\mid \vec A)]\}$$	Formal SIG (Varona-Uriarte et al., 29 Aug 2025)
Shapley Information Gain	Shapley-weighted fact-acquisition reward	Different expansion of SIG (Ding et al., 19 Aug 2025)
Surrogates of EIG	Instantaneous and Gaussian-tilting approximations of EIG	Conceptually aligned, different naming (Montella et al., 27 Mar 2026)

A technical consequence is that SIG should not be treated as a single universally accepted object analogous to Shannon mutual information. In some settings it is a variance proxy; in others it is a posterior approximation, a generalized entropy gap, a corrected split criterion, or a policy-value analogue of information. The most stable cross-domain interpretation is therefore functional rather than terminological: SIG is an information-gain surrogate designed to preserve the ranking or decision content of a harder target quantity.

2. Formal SIG in offline quantum experimental design

The clearest formal definition appears in “Computationally Tractable Offline Quantum Experimental Design for Nuclear Spin Detection” (Varona-Uriarte et al., 29 Aug 2025). There the target Bayesian quantity is the expected information gain

$$\mathrm{EIG}(\tau)\equiv \mathbb{E}\Big\{S[p(\vec A)]-S[p(\vec A\mid \pm x)]\Big\},$$

where $\tau$ is the experimental control variable, $\vec A$ is the vector of unknown hyperfine parameters, and $P_{\pm x}(\tau\mid \vec A)$ is the model-predicted binary measurement probability. Because full Bayesian estimation is expensive and the number of spins is not known a priori, the paper introduces

$$\mathrm{SIG}(\tau)\equiv \mathbb{E}\Big\{\mathrm{Var}_{\vec A\sim p(\vec A)}[P_{\pm x}(\tau\mid \vec A)]\Big\}.$$

For the binary-outcome setting used in the paper, this simplifies to

$$\mathrm{SIG}(\tau)=\mathrm{Var}_{\vec A\sim p(\vec A)}[P_x(\tau\mid \vec A)].$$

This SIG is therefore the prior variance of the measurable signal at a candidate control setting. Its interpretation is explicit: a setting $\tau$ is informative when plausible parameter values produce strongly varying signals, and uninformative when they produce nearly identical signals. The paper also states an important limitation: because SIG uses $P_x$ rather than $\log P_x$, and because variance is nonlinear, SIG is not an extensive quantity.

The computational workflow is fully offline. One samples $$\mathrm{EIG}(\tau)\equiv \mathbb{E}\Big\{S[p(\vec A)]-S[p(\vec A\mid \pm x)]\Big\},$$0 from the prior, simulates $$\mathrm{EIG}(\tau)\equiv \mathbb{E}\Big\{S[p(\vec A)]-S[p(\vec A\mid \pm x)]\Big\},$$1 on a candidate $$\mathrm{EIG}(\tau)\equiv \mathbb{E}\Big\{S[p(\vec A)]-S[p(\vec A\mid \pm x)]\Big\},$$2-grid, computes $$\mathrm{EIG}(\tau)\equiv \mathbb{E}\Big\{S[p(\vec A)]-S[p(\vec A\mid \pm x)]\Big\},$$3, ranks the grid points by SIG, and retains the top $$\mathrm{EIG}(\tau)\equiv \mathbb{E}\Big\{S[p(\vec A)]-S[p(\vec A\mid \pm x)]\Big\},$$4 points for data acquisition and SALI training. In the nitrogen-vacancy-center application, the high-field regime yielded an 85% reduction in measurement time for a modest reduction in performance, while the low-field simulation study predicted a 60% reduction by combining improved temporal resolution with SIG-based point selection (Varona-Uriarte et al., 29 Aug 2025).

3. Adjoint-compatible surrogates of expected information gain

In controlled dynamical systems, the central obstacle is different. “Adjoint-Compatible Surrogates of the Expected Information Gain for Optimal Experimental Design in Controlled Dynamical Systems” starts from the exact Bayesian objective

$$\mathrm{EIG}(\tau)\equiv \mathbb{E}\Big\{S[p(\vec A)]-S[p(\vec A\mid \pm x)]\Big\},$$5

for unknown parameter $$\mathrm{EIG}(\tau)\equiv \mathbb{E}\Big\{S[p(\vec A)]-S[p(\vec A\mid \pm x)]\Big\},$$6, control $$\mathrm{EIG}(\tau)\equiv \mathbb{E}\Big\{S[p(\vec A)]-S[p(\vec A\mid \pm x)]\Big\},$$7, and sensor-activation policy $$\mathrm{EIG}(\tau)\equiv \mathbb{E}\Big\{S[p(\vec A)]-S[p(\vec A\mid \pm x)]\Big\},$$8, and uses the chain rule

$$\mathrm{EIG}(\tau)\equiv \mathbb{E}\Big\{S[p(\vec A)]-S[p(\vec A\mid \pm x)]\Big\},$$9

The paper identifies the design-dependent posterior $\tau$0 as the sole obstruction to time-additivity and adjoint compatibility, and replaces it by tractable approximations (Montella et al., 27 Mar 2026).

The first surrogate is the instantaneous surrogate, obtained by replacing the posterior with the prior. It is myopic, because every incremental term is scored as if no previous data had been observed. The second is the Gaussian tilting surrogate, which replaces the posterior by a deterministically tilted prior

$\tau$1

where $\tau$2 is an accumulated Fisher-information-like matrix driven by the design. A third, multi-center tilting surrogate, replaces the single reference point by a mixture of local tilts to improve robustness under multimodal priors.

The theoretical picture is unusually explicit. The instantaneous surrogate overestimates each true incremental information term by exactly the redundancy $\tau$3, so its bias is structurally interpretable. The Gaussian tilting surrogate is exact in the linear-Gaussian setting. The paper’s main claim is therefore not merely approximation quality, but structural compatibility: the surrogate objective becomes time-additive, differentiable, and suitable for adjoint-based optimal control, while preserving substantially more Bayesian character than standard Fisher criteria (Montella et al., 27 Mar 2026).

A related kernel-theoretic construction appears in “Relative Information Gain and Gaussian Process Regression,” which defines

$\tau$4

as the additional information gained when reducing observation noise variance from $\tau$5 to $\tau$6. The paper shows that a scaled relative information gain interpolates between ordinary information gain and effective dimension, and that it has the same growth rate as effective dimension (Flynn, 5 Oct 2025). This suggests a recurrent SIG design principle: retain the information-theoretic interpretation of a log-determinant objective while modifying it to match sharper statistical complexity or optimization requirements.

4. Surrogate losses, generalized entropy, and posterior scoring

A different formalization arises in multiclass surrogate-risk theory. “Multiclass Classification, Information, Divergence, and Surrogate Risk” shows that every multiclass surrogate loss $\tau$7 induces a generalized entropy

$\tau$8

and therefore a statistical information measure

$\tau$9

Conversely, every closed concave entropy $\vec A$0 can be realized by a convex loss through

$\vec A$1

This yields a precise surrogate-information-gain interpretation: surrogate losses do not merely approximate classification risk; they induce their own generalized information gains, and these gains determine whether optimizing the surrogate preserves optimal quantizer or representation choice (Duchi et al., 2016).

The same surrogate logic reappears in evaluation rather than training in “A Proper Scoring Rule for Virtual Staining.” There the true cell-wise posterior $\vec A$2 is unavailable, so the paper evaluates model-implied posteriors by the average log score

$\vec A$3

and defines information gain relative to the marginal feature distribution $\vec A$4 as

$\vec A$5

Because the logarithmic score is strictly proper, the relative score remains strictly proper. In practice, $\vec A$6 is estimated from many generated samples using KDE or GMM fitting, so the operational metric is a plug-in surrogate for ideal posterior information gain. The paper’s empirical comparison is revealing: for feature F7, cDDPM and Pix2pixHD have nearly indistinguishable marginal KLD values $\vec A$7 and $\vec A$8, and rank distances $\vec A$9 and $P_{\pm x}(\tau\mid \vec A)$0, but information gains $P_{\pm x}(\tau\mid \vec A)$1 and $P_{\pm x}(\tau\mid \vec A)$2, respectively, exposing posterior-quality differences that marginal metrics hide (Tonks et al., 26 Feb 2026).

5. Corrected and proxy information-gain criteria in classical machine learning

In classical machine learning, SIG-like constructions often appear as corrections to raw information gain or as proxy objectives that guide search. “Information gain ratio correction: Improving prediction with more balanced decision tree splits” diagnoses a residual pathology of C4.5 gain ratio: because split information becomes small for highly unbalanced splits, the ratio can overvalue branches that isolate small pure subsets of low predictive interest. The proposed Balanced Gain Ratio

$P_{\pm x}(\tau\mid \vec A)$3

preserves Quinlan’s anti-high-arity correction while attenuating denominator inflation when split information is small. The paper reports, for the Letter dataset, tree depth around 20 under the corrected criterion versus around 100 under standard gain ratio (Leroux et al., 2018). This is not formal SIG terminology, but it is a direct example of a corrected information-gain objective standing in for a pathological original.

“Improved Information Gain Estimates for Decision Tree Induction” addresses a different defect: biased entropy estimation. For classification, it replaces the naive plug-in entropy estimator by Grassberger’s estimator; for regression, it considers both a Gaussian UMVUE correction and a nonparametric 1-NN differential entropy estimator. The aim is to make empirical information gain a better finite-sample surrogate for the population mutual information used in split selection. The classification experiments report Grassberger winning on 18 datasets, naive entropy on 8, with 4 ties and a Wilcoxon signed-rank $P_{\pm x}(\tau\mid \vec A)$4 (Nowozin, 2012).

Two additional lines of work make the proxy role of information gain explicit. “Multivalued Subsets Under Information Theory” maximizes standard information gain over grouped-value binary partitions rather than the original attribute partition, effectively replacing ordinary IG by a subset-optimized binary-partition score (Dabhade, 2011). “A new algorithm for Subgroup Set Discovery based on Information Gain” uses entropy-based information gain as a dynamic thresholding criterion for selector filtering during subgroup search, and only later refines patterns with ORR and $P_{\pm x}(\tau\mid \vec A)$5-value. The paper is explicit that IG acts as the single optimization index during search rather than as the final sole quality measure (Gómez-Bravo et al., 2023). In both cases, information gain functions as a surrogate objective: it shapes search toward promising structures even when the ultimate desideratum is different.

6. Application-specific extensions, acronym ambiguity, and decision-oriented uses

The acronym SIG is not stable across application areas. In “ProMed: Shapley Information Gain Guided Reinforcement Learning for Proactive Medical LLMs,” SIG means Shapley Information Gain, not Surrogate Information Gain. The reward is

$P_{\pm x}(\tau\mid \vec A)$6

where $P_{\pm x}(\tau\mid \vec A)$7 are softmax-normalized Shapley values of atomic clinical facts and $P_{\pm x}(\tau\mid \vec A)$8 is a generated proxy for the model’s current understanding. The reward is then used both in MCTS-guided trajectory construction and in token-level RL credit assignment. The paper reports average gains of 6.29% over state-of-the-art methods and 54.45% over the reactive paradigm (Ding et al., 19 Aug 2025). Although the expansion differs, the construction still exemplifies a common pattern: raw information gain is replaced by a context-aware, interaction-sensitive proxy.

A closely related reinforcement-learning example is “CIG: Exploration via Conditional Information Gain,” which derives a tractable surrogate of trajectory-level Bayesian information gain for deep model-based exploration. Starting from $P_{\pm x}(\tau\mid \vec A)$9, the paper approximates posterior uncertainty with an ensemble, replaces the trajectory predictive mixture by a moment-matched Gaussian, and then introduces a trace-reduced log-determinant objective over an ensemble-disagreement kernel. The resulting per-step reward has the conditional form

$$\mathrm{SIG}(\tau)\equiv \mathbb{E}\Big\{\mathrm{Var}_{\vec A\sim p(\vec A)}[P_{\pm x}(\tau\mid \vec A)]\Big\}.$$0

which preserves conditioning on both lifetime experience and within-rollout redundancy while remaining tractable (Joseph et al., 20 May 2026).

Outside explicit information-theoretic objectives, several papers quantify the value of surrogate information through decision or inferential gain. In censored survival analysis, the proportion of treatment effect explained by surrogate information up to time $$\mathrm{SIG}(\tau)\equiv \mathbb{E}\Big\{\mathrm{Var}_{\vec A\sim p(\vec A)}[P_{\pm x}(\tau\mid \vec A)]\Big\}.$$1 is

$$\mathrm{SIG}(\tau)\equiv \mathbb{E}\Big\{\mathrm{Var}_{\vec A\sim p(\vec A)}[P_{\pm x}(\tau\mid \vec A)]\Big\}.$$2

which measures how much residual treatment effect remains after accounting for surrogate information (Parast et al., 2016). In individualized treatment regimes, the most SIG-like quantity is the surrogate efficiency

$$\mathrm{SIG}(\tau)\equiv \mathbb{E}\Big\{\mathrm{Var}_{\vec A\sim p(\vec A)}[P_{\pm x}(\tau\mid \vec A)]\Big\}.$$3

the gain from surrogate-guided treatment relative to random treatment assignment at the same budget $$\mathrm{SIG}(\tau)\equiv \mathbb{E}\Big\{\mathrm{Var}_{\vec A\sim p(\vec A)}[P_{\pm x}(\tau\mid \vec A)]\Big\}.$$4 (Xu et al., 29 Nov 2025). In surrogate-powered inference, the contribution of surrogate labels appears as variance reduction in the asymptotic covariance of an augmented estimator, with regularization and adaptive labeling used to make that gain reliable under many noisy surrogates (Chen et al., 26 Dec 2025). In efficient testing using surrogate information, the gain is framed as retaining valid power for a primary-outcome treatment effect while measuring the expensive primary outcome only in the subgroup where the surrogate is weak (Knowlton et al., 21 Apr 2025).

Taken together, these works show that SIG has developed less as a single formal invariant than as a design pattern. The pattern is consistent: replace an inaccessible, expensive, or behaviorally inadequate information criterion by a surrogate that preserves the task-relevant ordering of designs, questions, splits, or policies. The exact surrogate varies—signal variance, tilted-prior predictive information, generalized entropy reduction, Shapley-weighted fact acquisition, corrected split scoring, or decision-value gain—but the underlying purpose is the same.