Counterfactual Criterion: Frameworks & Applications

• Counterfactual criterion is a formal rule that specifies which elements remain fixed and which may change when constructing alternative, 'what if' scenarios.
• It is applied in quantum protocols, structural causal models, and fairness to assess trace presence, identifiability, invariance, and risk differences.
• Practitioners use counterfactual criteria to enhance model robustness, guide decision-making, and ensure fairness by rigorously evaluating alternative outcomes.

A counterfactual criterion is a formal rule for evaluating a contrary-to-fact claim, or for deciding whether a process, model, predictor, explanation, or protocol legitimately deserves a counterfactual interpretation. The term is not used uniformly across fields. In quantum communication, the criterion is often trace-based and asks whether a pre- and post-selected particle leaves any weak trace in a relevant region (Vaidman, 2016). In structural causal modeling, counterfactuals are evaluated by abduction, intervention, and prediction, while alternative formalisms ask whether a counterfactual distribution is graphically identifiable, canonically representable, or physically realizable [(Balke et al., 2013); (Shpitser et al., 2020); (Lara, 22 Jul 2025); (Raghavan et al., 14 Mar 2025)]. In statistical decision theory, the criterion may be a loss that depends on the full vector of potential outcomes, with additivity as the necessary-and-sufficient condition for identifiable risk differences under strong ignorability (Koch et al., 13 May 2025). In fairness and explainability, the criterion becomes invariance under interventions on protected attributes or the requirement that editing the explained features produces a nearby counterfactual that changes the model output (Kusner et al., 2017, Ge et al., 2021).

1. Conceptual scope

Across the literature, a counterfactual criterion specifies what must be held fixed, what may be altered, and what counts as legitimate evidence about the altered world. In Pearl-style structural models, a counterfactual is evaluated by first using observations to update beliefs about disturbances, then replacing the structural equations for antecedent variables, and finally propagating forward in the modified model (Balke et al., 2013). In quantum protocol analysis, by contrast, the issue is not hypothetical policy intervention but whether a particle was present in a region “in the sense relevant to interaction,” which is operationalized by weak trace (Vaidman, 2016).

This variation is substantive rather than terminological. Some criteria are semantic, specifying how counterfactual worlds are constructed; some are epistemic, specifying when counterfactual quantities are identifiable from data; and some are normative, specifying which counterfactual dependencies are acceptable for decision making, fairness, or explanation. The literature therefore treats “counterfactual criterion” not as a single theorem, but as a family of formal tests attached to different inferential tasks (Koch et al., 13 May 2025, Kusner et al., 2017, Ge et al., 2021).

A recurrent technical distinction is between actual observations and alternative worlds. In SCM-based work, this separation is enforced by intervention semantics or by alternative cross-world couplings [(Balke et al., 2013); (Lara, 22 Jul 2025)]. In explainability and robustness work, it appears as the distinction between the factual input and a counterfactual counterpart or a set of approximate counterfactuals (Ge et al., 2021, Leofante et al., 2023). In quantum and Bell-type analyses, it appears as the distinction between performed and unperformed measurements, or between detected successful events and the full physical protocol (Hess et al., 2016, Wander et al., 2021).

2. Trace-based criteria in quantum protocols

A prominent quantum criterion is due to Vaidman, who defines counterfactuality in terms of trace rather than path rhetoric. The proposed definition is that a process is interaction-free or counterfactual if the pre- and post-selected particle leaves no trace near the object, with the trace identified using the two-state vector formalism. In that language, the trace is present only in the overlap of the forward- and backward-evolving wave functions, so a practical condition is: counterfactual iff there is no overlap near the object or channel (Vaidman, 2016).

This criterion sharply separates detecting the presence of an opaque object from inferring its absence. In the Elitzur–Vaidman interaction-free measurement, successful detection of the object can be counterfactual because the detector click occurs while the photon leaves no trace near the object. Vaidman therefore classifies Noh-style key distribution protocols, which rely only on interaction-free finding of an object, as fully counterfactual. By contrast, protocols that infer an empty region are not counterfactual, because the successful post-selection still leaves a trace in the supposedly empty location. On this basis, “counterfactual” quantum direct communication and state transfer are rejected: for one logical value or in superposition, there is no single transmission region that is trace-free in both cases (Vaidman, 2016).

The later comparative analysis in "Three approaches for analyzing the counterfactuality of counterfactual protocols" formalizes this further by contrasting three criteria: a classical argument, the weak trace criterion, and the Fisher information criterion. The classical argument is judged inconsistent for quantum protocols and should be abandoned. For postselected protocols, however, the weak trace and Fisher information criteria are reported to agree about the degree of counterfactuality, and postselection is argued to be essential for genuine counterfactual communication (Wander et al., 2021).

The experimental paper "Counterfactual communication without a trace in the transmission channel" implements an Aharonov–Vaidman modification intended to eliminate the dominant environmental trace in the transmission channel. The operational test uses arm-dependent frequency tagging with EOMs, and the decisive signature is the absence of the EOM-A sideband, which labels the transmission channel. In successful bit-1 and bit-0 events, no EOM-A peak is observed above the noise floor, which the authors interpret as absence of first-order trace in the channel for the detected photons. The QR-code transfer is explicitly presented only as a proof of principle, not as a loophole-free certification, because losses and undetected photons remain outside the idealized successful-event claim (Pan et al., 2023).

3. Bell-type reasoning, measurement settings, and quantum counterfactuals

In Bell-related literature, the counterfactual criterion concerns whether outcomes of unperformed measurements may be treated as well-defined objects in the derivation of inequalities. "Counterfactual Definiteness and Bell's Inequality" characterizes counterfactual definiteness as requiring functions that map tests onto numbers, with arguments in one-to-one correspondence with physical entities and chosen as independent variables. On this view, Bell’s use of expressions such as $A(\mathbf{j},\lambda)$ already presupposes the simultaneous numerical availability of outcomes for alternative settings, and therefore builds counterfactual definiteness directly into the formalism (Hess et al., 2016).

"Counterfactual restrictions and Bell's theorem" retains the claim that counterfactual reasoning is necessary for Bell inequalities, but weakens the target from full counterfactual definiteness to a matching condition over settings and hidden variables. Its central distinction is between counterfactual indefiniteness and counterfactual restriction. The latter means that some triples $(a,b,\lambda)$ are not physically realizable, even if they are logically conceivable. The paper then argues that Bell-inequality violation can be understood not only through retrocausality or superdeterminism, but through restriction of the physically available counterfactual space (Hance, 2019).

A different quantum-foundational development appears in "Counterfactual quantum measurements" (Banerjee et al., 2 Oct 2025). There the antecedent is restricted to an alternative measurement setting, the consequent is an outcome in the counterfactual world, and the result is not a truth value but a probability termed a supposability. The formal rule is to keep fixed all classical events that are not causally influenced by the antecedent, infer their distribution from the actual evidence, and then average the counterfactual outcome probabilities over those fixed events. The paper calls these retained classical variables fixtures and uses the rule both in a Bell-CHSH example and in a continuous-measurement fluorescence example, where the object of interest is an expected counterfactual homodyne current rather than a binary outcome (Banerjee et al., 2 Oct 2025).

Taken together, these works show that in quantum foundations the counterfactual criterion is often a criterion about admissible cross-context comparison: whether alternative settings may be combined algebraically, whether some counterfactual branches are forbidden, or how actual classical evidence constrains probabilities in an alternative measurement context (Hess et al., 2016, Hance, 2019, Banerjee et al., 2 Oct 2025).

4. Structural causal models, SWIGs, and realizability

In structural-model causality, the classical criterion for evaluating counterfactuals is the abduction–action–prediction procedure. Observations are first used to update beliefs about exogenous disturbances; the antecedent is then implemented by replacing the corresponding structural equations; and the consequent is computed in the modified model. The paper "Counterfactuals and Policy Analysis in Structural Models" emphasizes that this is fundamentally different from conditioning on an observed value of an endogenous policy variable, because intervention severs the ordinary causes of that variable rather than merely selecting cases in which it happened to take a given value (Balke et al., 2013).

For Gaussian linear structural equation models, the criterion becomes analytically explicit. "Counterfactual Reasoning in Linear Structural Equation Models" rewrites Balke–Pearl counterfactual formulas in terms of total effects and observed covariance structure. Under an unconditional plan $X=x_0$, the counterfactual mean and variance of $Y$ are

$\mu_{y^*} = \mu_y + \tau_{yx}(x_0 - \mu_x),$

$$\sigma_{y^*y^*} = \sigma_{yy.x} + (\tau_{yx} - B_{yx})^2 \sigma_{xx},$$

and with observed evidence $R=r$, the variance acquires an additional adjustment term involving $B_{yr}$, $B_{xr}$, and $\Sigma_{rr}$. The same paper extends the framework to interval observations and conditional plans $(a,b,\lambda)$0, and derives an optimality condition on $(a,b,\lambda)$1 that minimizes the counterfactual variance (Cai et al., 2012).

A graphical counterfactual criterion is supplied by SWIGs. "Multivariate Counterfactual Systems And Causal Graphical Models" treats counterfactuals as primitive objects and uses d-separation in a SWIG to read off counterfactual independences. The SWIG global Markov property, the po-calculus, and the extended ID algorithm together provide the paper’s criterion for identification in models with hidden variables: a counterfactual is identified exactly when the required district terms can be obtained through valid splitting and marginalization steps, equivalently when no hedge blocks the construction (Shpitser et al., 2020).

Two more recent developments separate semantic representation from physical accessibility. "Canonical Representations of Markovian Structural Causal Models" argues that the causal graphical model fixes observational and interventional structure but not the counterfactual layer. It replaces explicit structural equations by one-step-ahead process distributions $(a,b,\lambda)$2, or equivalently by normalizations $(a,b,\lambda)$3 in latent Gaussian space, so that different counterfactual conceptions—comonotonic, countermonotonic, or more general stochastic couplings—can be chosen without altering observational or interventional constraints (Lara, 22 Jul 2025). "Counterfactual Realizability" addresses a different question: which Layer-3 distributions can be physically sampled from. Under the Fundamental Constraint of Experimentation, the complete algorithm CTF-REALIZE decides realizability, and under a maximal action set the graphical criterion is that the counterfactual ancestor set must not contain two potential responses of the same variable under different regimes. As a corollary, $(a,b,\lambda)$4 is not realizable when $(a,b,\lambda)$5 and $(a,b,\lambda)$6 are causally related (Raghavan et al., 14 Mar 2025).

The interventionist SCM criterion is not the only semantic option. "Backtracking Counterfactuals" formalizes an alternative in which the structural equations $(a,b,\lambda)$7 remain fixed while exogenous variables change across worlds via a backtracking conditional $(a,b,\lambda)$8. The paper proposes closeness, symmetry, and decomposability as desiderata on that conditional, and shows that backtracking counterfactuals are computed by cross-world conditioning on $(a,b,\lambda)$9 rather than by equation replacement (Kügelgen et al., 2022).

5. Counterfactual loss and policy choice

In statistical decision theory, the counterfactual criterion is a loss functional that depends on the full vector of potential outcomes rather than only the realized outcome under the chosen action. "Statistical Decision Theory with Counterfactual Loss" defines a counterfactual loss as

$X=x_0$0

with associated risk

$X=x_0$1

The motivating claim is that classical decision theory cannot assess whether a chosen treatment was right relative to feasible alternatives, because it ignores what would have happened under unchosen actions (Koch et al., 13 May 2025).

The central technical issue is identification. Under IID sampling, consistency, and strong ignorability, distributions involving one potential outcome at a time are identifiable, but the joint distribution of $X=x_0$2 generally is not. The paper proves that differences in counterfactual risk between any two decision systems are identifiable if and only if the loss is additive in the potential outcomes: $X=x_0$3 If $X=x_0$4, the risk itself is exactly identified; otherwise the risk is identified only up to a decision-independent constant (Koch et al., 13 May 2025).

This additivity criterion is not merely technical. For binary decisions, additive counterfactual losses can be matched by standard losses yielding the same optimal policy, but for $X=x_0$5 the equivalence generally fails. The paper’s overtreatment example,

$X=x_0$6

penalizes choosing a more invasive treatment when a less invasive one would also have succeeded. The result is that counterfactual and standard criteria can produce different treatment rankings precisely in multi-action problems (Koch et al., 13 May 2025).

The broader implication is that a counterfactual criterion in decision theory is a criterion about action quality relative to the whole potential-outcome landscape, not merely about predictive accuracy on factual data. This is why the paper treats additivity as the pivotal necessary-and-sufficient condition.

6. Fairness, explanation, and robustness

In fairness research, the canonical counterfactual criterion is counterfactual fairness. A predictor $X=x_0$7 is counterfactually fair if, conditional on the observed features and protected attribute, its distribution is invariant under interventions that change the protected attribute: $X=x_0$8 The central sufficient condition is equally important: if $X=x_0$9 is a function only of variables that are not descendants of the protected attribute $Y$0, then $Y$1 is counterfactually fair (Kusner et al., 2017).

This descendant criterion is also the basis of causal-model aggregation under fairness constraints. "Pooling of Causal Models under Counterfactual Fairness via Causal Judgement Aggregation" proposes two graph-level procedures—Removal-Pooling and Pooling-Removal—that guarantee fairness by removing protected variables and all their descendants from the final pooled graph. The two algorithms differ only in whether that removal occurs before or after graph aggregation, with the former being more conservative (Zennaro et al., 2018).

In explainable AI, "Counterfactual Evaluation for Explainable AI" turns the criterion into a test of faithfulness. An explanation is deemed faithful if editing the features selected by the explanation yields a counterfactual input that substantially changes the model output. The paper operationalizes this using Validity, Proximity, and the Counterfactual Evaluation score

$Y$2

together with soft probabilistic variants when model probabilities are available. The stated advantage over erasure-based criteria is that counterfactual replacement preserves more realistic inputs and avoids artifacts caused by deletion (Ge et al., 2021).

Robustness work adds another layer. "Promoting Counterfactual Robustness through Diversity" argues that a single nearest counterfactual is fundamentally fragile, because nearby inputs can have very different nearest counterfactuals for geometric reasons. The theoretical ideal is an exhaustive set of $Y$3-approximate counterfactuals, which the paper proves is weakly $Y$4-robust under symmetry and triangle inequality. Because exhaustive reporting is often infeasible, the proposed algorithm uses distance filtering, a diversity filter, and binary search to return a compact representative set, with diversity defined either by cosine angle or by pairwise distance. The empirical comparison reports smaller set distances under perturbations than DiCE while maintaining competitive proximity and runtime (Leofante et al., 2023).

These works share a common structure. The criterion is no longer “what would happen under $Y$5?” alone, but “what property of a model or explanation remains stable, fair, or faithful when an appropriate counterfactual comparison is made?” (Kusner et al., 2017, Ge et al., 2021, Leofante et al., 2023).

7. Strategic reasoning, LLMs, and persistent disputes

Game theory uses a distinct but related counterfactual criterion. In "Common Counterfactual Belief of Rationality Subsumes Superrationality On Symmetric Games," Common Counterfactual Belief of Rationality requires not only that players believe all are rational, but that this remains true under counterfactual deviations of their own strategies. In symmetric normal-form games, every Hofstadter equilibrium is shown to satisfy both minimax-rationalizability and individual rationality under this criterion, though the converses fail (Fourny, 2017).

Large-language-model evaluation repurposes the idea again. "Better Think Thrice: Learning to Reason Causally with Double Counterfactual Consistency" defines

$Y$6

where the model is queried on an original question, a counterfactual version produced by an intervention, and a double-counterfactual version that restores the original state. The criterion is meant to test both causal intervention and counterfactual prediction without requiring labeled counterfactual data. It is also used as a hard inference-time rejection rule, with the reported mean number of attempts to achieve consistency equal to $Y$7 (Lin et al., 18 Feb 2026).

Several disputes recur across domains. In quantum communication, classical path-based reasoning is criticized as inconsistent, and the debate shifts to weak trace, Fisher information, and measurement-setting-based probabilistic formalisms (Wander et al., 2021, Banerjee et al., 2 Oct 2025). In Bell theory, the necessity of counterfactual reasoning is widely affirmed in the cited papers, but its formal locus is disputed: some accounts emphasize counterfactual definiteness, while others emphasize restriction of the counterfactual domain rather than outright indefiniteness (Hess et al., 2016, Hance, 2019). In structural causality, the counterfactual layer is explicitly treated as extra structure not identified by observational or interventional constraints alone, which is why canonical representations and normalization procedures are introduced (Lara, 22 Jul 2025).

A plausible implication is that “counterfactual criterion” names a structural role rather than a single mathematical object. In every domain covered here, the criterion determines how an alternative world is constructed, what remains fixed across worlds, and what inferential or normative conclusion follows from that construction. The differences among weak trace, d-separation, additivity, invariance, diversity, and intervention-reversal consistency are therefore differences in counterfactual ontology and task definition, not merely differences in notation.