Causal Fisher-Information Inequalities

• Causal Fisher-Information Inequalities (CFIIs) are constraints derived from classical causal models defined by directed acyclic graphs, conditional independences, and modular parameter dependence.
• They translate the structure of classical causal assumptions into Fisher-information bounds that, when violated, signal incompatibility with the model and a metrological advantage.
• The framework generalizes classical Cramér–Rao bounds by using information resistance in series, thereby providing a falsification criterion that links causal inference with precision estimation.

Searching arXiv for the primary CFII paper and closely related Fisher-information inequality work. Causal Fisher-Information Inequalities (CFIIs) are Fisher-information constraints implied by a classical causal model specified by a directed acyclic graph (DAG), its associated conditional independences, and modular parameter dependence. In this formulation, Fisher information is not treated merely as a local estimation metric; it becomes a necessary compatibility condition for an entire causal model class. If the Fisher informations extracted from operational contexts violate a CFII, the observed statistics are incompatible with every model in that classical causal class. The same violation also certifies a metrological advantage, because it implies a precision unattainable by any member of the corresponding classical class (Bang et al., 18 May 2026).

1. Definition and conceptual scope

The central statement of the CFII framework is that, once an experiment is assumed to admit a classical causal model specified by a DAG, its implied conditional independences, and modular parameter dependence, the Fisher informations of the relevant contexts must satisfy causal Fisher-information inequalities. In the paper’s notation, this takes the abstract form

$$\mathcal G\!\left(\{F_s(\theta)\}_{s\in \mathcal S_0}\right)\ge 0,$$

for every operational model compatible with the classical causal class $\mathcal M$. A violation, $\mathcal G<0$, places the observed statistics outside the feasible Fisher-information region of that class (Bang et al., 18 May 2026).

This formulation is explicitly broader than earlier Fisher-information witnesses tied to segmented dynamics or discrete trajectories. The paper states that the more natural interpretation is causal: the issue is not “trajectory” as such, but the underlying classical causal structure. This suggests viewing CFIIs as a systematic causal-inference framework in which inverse Fisher information functions as an “information resistance” associated with classical mediation (Bang et al., 18 May 2026).

A plausible implication is that CFIIs belong to a wider family of Fisher-information inequalities, but with a distinctive role. Other Fisher-information results constrain estimation under privacy channels, Bayesian priors, generalized divergences, or stochastic dynamics; CFIIs instead translate a classical causal hypothesis directly into a precision bound, and then into a falsification criterion when that bound is violated. Related Fisher-information inequality programs include generalized Cramér–Rao inequalities built from a modified $\chi^\beta$-divergence (Bercher, 2013), data-processing inequalities under local differential privacy (Barnes et al., 2020), mutual-information bounds in terms of Fisher information (Górecki et al., 2024), explicit Cramér–Rao and Van Trees bounds for Wishart-randomized Gaussian models (Letac, 2022), and entropy–Fisher–large-deviation inequalities for Markov jump processes (Hilder et al., 2018).

2. Classical causal assumptions and operational model class

A classical causal model class $\mathcal M$ consists of three ingredients. First, the causal structure is encoded by a DAG. For variables $X_1,\dots,X_n$,

$$p(x_1,\dots,x_n)=\prod_{j=1}^n p(x_j\mid \mathrm{pa}(x_j)).$$

Second, the DAG implies conditional independences such as

$$U \perp V \mid W \quad \Longleftrightarrow \quad p(u,v\mid w)=p(u\mid w)p(v\mid w).$$

Third, the unknown parameter $\theta$ has modular parameter dependence: each local mechanism or kernel has its own allowed $\theta$-dependence, and that dependence is not freely shared across modules (Bang et al., 18 May 2026).

In the operational setting, one observes distributions $\mathcal M$0, where $\mathcal M$1 labels contexts such as measurement choices, intermediate interventions, segmentation strategies, or control settings. The key logical chain is that Fisher information is a functional of the probability model; DAG factorization and conditional independences constrain the score structure; modular parameter dependence forces local score contributions to be orthogonal in the classical causal decomposition; therefore the Fisher information cannot compose arbitrarily and must satisfy model-dependent inequalities (Bang et al., 18 May 2026).

The role of modularity is especially important. In the CFII framework, the local parameter dependences are structurally separated. This separation is what later forces the additivity of inverse Fisher information along classical causal paths and forbids the score-correlation terms responsible for metrological gain. A common misconception is to treat the resulting inequalities as ad hoc metrological benchmarks. The paper’s position is stronger: they are necessary conditions for the full classical causal model class defined by the DAG, the conditional independence structure, the modular parameter split, and the classical coarse-graining logic (Bang et al., 18 May 2026).

3. Causal-path series law and information resistance

The backbone result is the causal-path series law for a classical path

$\mathcal M$2

where $\mathcal M$3 is an intermediate classical mediator and $\mathcal M$4 is the endpoint record. The path model assumes

$\mathcal M$5

with additive total parameter

$\mathcal M$6

Under these assumptions, the paper proves the causal-path CFII

$\mathcal M$7

Here $\mathcal M$8 is the Fisher information of the upstream module, $\mathcal M$9 is the Fisher information of the downstream module, and $\mathcal G<0$0 is the effective Fisher information for estimating the additive total parameter from endpoint data $\mathcal G<0$1 alone (Bang et al., 18 May 2026).

The paper interprets

$\mathcal G<0$2

as an information resistance, so that the theorem becomes

$\mathcal G<0$3

This is the information-theoretic analogue of resistors in series: classical causal bottlenecks add resistance and do not remove it. The generalization to a $\mathcal G<0$4-step chain is

$\mathcal G<0$5

The significance of this formulation is structural rather than merely computational. It makes the constraint depend on causal mediation and modularity, not on a particular physical implementation (Bang et al., 18 May 2026).

The resulting inequality is a necessary condition for compatibility with the entire classical causal-path class. Therefore, if an experiment yields

$\mathcal G<0$6

the data are incompatible with every model in that class. This is stronger than failure of one candidate model: it is a class-level impossibility statement covering the DAG $\mathcal G<0$7, the assumed conditional independence, the modular parameter split, and the classical coarse-graining logic (Bang et al., 18 May 2026).

4. Violation, Fisher-information synergy, and metrological meaning

The paper identifies the gain mechanism behind CFII violation as Fisher-information synergy, namely off-diagonal score correlations that classical modularity forbids. In a two-parameter description,

$\mathcal G<0$8

In a classical modular causal decomposition, $\mathcal G<0$9, because the two module scores are orthogonal by construction. That orthogonality is exactly what enforces the series penalty (Bang et al., 18 May 2026).

If $\chi^\beta$0, the effective Fisher information for the additive parameter $\chi^\beta$1, with $\chi^\beta$2, is

$\chi^\beta$3

The paper shows that this exceeds the classical series benchmark

$\chi^\beta$4

if and only if

$\chi^\beta$5

Equivalently, in inverse form,

$\chi^\beta$6

to be compared with the modular classical resistance

$\chi^\beta$7

The metrological gain therefore comes from positive score correlation between modules, precisely the feature excluded by classical modularity (Bang et al., 18 May 2026).

This dual role is central. A CFII violation is simultaneously a causal-model impossibility statement and a metrological witness. The same datum that falsifies the classical causal class also certifies a precision unattainable within that class. This suggests a direct bridge between causal-model falsification and resource certification: the witness is not external to the estimation problem but derived from the Fisher-information geometry enforced by the assumed causal structure (Bang et al., 18 May 2026).

5. Single-qubit coherent-rotation example and long-chain amplification

The paper’s cleanest example is a single qubit with Hamiltonian

$\chi^\beta$8

prepared in

$\chi^\beta$9

and measured in the $\mathcal M$0 basis. The binary outcome bias is

$\mathcal M$1

so the Fisher information is

$\mathcal M$2

At the coherent point $\mathcal M$3, the paper shows

$\mathcal M$4

Then for any nontrivial split $\mathcal M$5,

$\mathcal M$6

The CFII is therefore violated deterministically. The classical causal-path benchmark is

$\mathcal M$7

whereas the actual Fisher information is $\mathcal M$8, giving an exact factor-of-two improvement (Bang et al., 18 May 2026).

The paper also shows estimator-level achievability. For the binary fringe

$\mathcal M$9

the maximum-likelihood estimator is asymptotically efficient, so the RMSE approaches

$X_1,\dots,X_n$0

At the deterministic point, this beats the classical path frontier by $X_1,\dots,X_n$1 in standard deviation (Bang et al., 18 May 2026).

To exclude the possibility that the classical path fails only because the split was chosen badly, the paper defines the split-optimized classical benchmark

$X_1,\dots,X_n$2

and the ratio

$X_1,\dots,X_n$3

If

$X_1,\dots,X_n$4

the CFII is violated for every admissible split. The paper reports broad regions of the generic qubit landscapes where $X_1,\dots,X_n$5, indicating robustness against this adversarial classical optimization (Bang et al., 18 May 2026).

The analysis then extends to a $X_1,\dots,X_n$6-step chain. If the Fisher information is constant, $X_1,\dots,X_n$7, then for every $X_1,\dots,X_n$8-segment classical causal decomposition,

$X_1,\dots,X_n$9

The standard-deviation advantage therefore scales as $$p(x_1,\dots,x_n)=\prod_{j=1}^n p(x_j\mid \mathrm{pa}(x_j)).$$0. Conceptually, refining the classical causal story into more mediators worsens the classical benchmark, because each extra mediator adds another information resistance in series (Bang et al., 18 May 2026).

6. Finite-data certification, adversarial classical models, and relation to other Fisher-information inequalities

The paper does not restrict itself to noiseless asymptotic reasoning. It introduces an AI-assisted adversarial finite-data stress test using a visibility-reduced fringe

$$p(x_1,\dots,x_n)=\prod_{j=1}^n p(x_j\mid \mathrm{pa}(x_j)).$$1

with readout error $$p(x_1,\dots,x_n)=\prod_{j=1}^n p(x_j\mid \mathrm{pa}(x_j)).$$2. The corresponding Fisher information is

$$p(x_1,\dots,x_n)=\prod_{j=1}^n p(x_j\mid \mathrm{pa}(x_j)).$$3

For the $$p(x_1,\dots,x_n)=\prod_{j=1}^n p(x_j\mid \mathrm{pa}(x_j)).$$4-step test, the witness is

$$p(x_1,\dots,x_n)=\prod_{j=1}^n p(x_j\mid \mathrm{pa}(x_j)).$$5

The Fisher information is estimated from local scores, and the uncertainty is propagated using the delta method. A classifier likelihood-ratio estimator is also used when analytic scores are unavailable (Bang et al., 18 May 2026).

The classical comparator is an AI-optimized modular causal path,

$$p(x_1,\dots,x_n)=\prod_{j=1}^n p(x_j\mid \mathrm{pa}(x_j)).$$6

optimized over latent mediator structure and local kernels while remaining modular and classical. The reported result is that this adversary can saturate the CFII frontier but not cross it: $$p(x_1,\dots,x_n)=\prod_{j=1}^n p(x_j\mid \mathrm{pa}(x_j)).$$7 with the best numerical restarts reaching values like $$p(x_1,\dots,x_n)=\prod_{j=1}^n p(x_j\mid \mathrm{pa}(x_j)).$$8. Noise reduces the advantage and eventually destroys it, but for the chosen readout error the violation remains certifiable up to a finite dephasing threshold. The important structural point is that optimized modular classical causal models can hug the boundary yet do not enter the forbidden region (Bang et al., 18 May 2026).

Within the broader Fisher-information literature, CFIIs occupy a distinct position. Generalized Cramér–Rao inequalities derived from a modified $$p(x_1,\dots,x_n)=\prod_{j=1}^n p(x_j\mid \mathrm{pa}(x_j)).$$9-divergence extend Fisher information to arbitrary norms, arbitrary powers of estimation error, escort distributions, and generalized $$U \perp V \mid W \quad \Longleftrightarrow \quad p(u,v\mid w)=p(u\mid w)p(v\mid w).$$0-Gaussians (Bercher, 2013). Local differential privacy imposes data-processing inequalities in which the surviving Fisher information depends on score tails and the privacy parameter $$U \perp V \mid W \quad \Longleftrightarrow \quad p(u,v\mid w)=p(u\mid w)p(v\mid w).$$1 (Barnes et al., 2020). Mutual-information bounds in terms of Fisher information convert local Fisher-information control into global information and Bayesian quadratic-cost bounds, with classical and quantum forms (Górecki et al., 2024). Wishart-randomized Gaussian covariance models yield explicit Fisher information, inverse Fisher information, and closed-form Cramér–Rao and Van Trees bounds through a two-operator algebra (Letac, 2022). Markov jump processes admit a generalized relative Fisher information linked to entropy distance and a large-deviation rate functional, with applications to coarse-graining (Hilder et al., 2018). These works show that Fisher-information inequalities can encode privacy constraints, Bayesian structure, generalized divergence geometry, or dynamical dissipation. CFIIs add a different principle: classical causal assumptions themselves imply a Fisher-information frontier, and violating that frontier both disproves the classical causal account and certifies a metrological resource (Bang et al., 18 May 2026).