Causal-Path Series Law Overview

• Causal-Path Series Law is an umbrella term for stepwise causal formulations that include iterative averaging (via the g-formula), entropy monotonicity, and additive log-scale decomposition.
• It employs the law of iterated expectation to handle time-varying confounding, ensuring robust estimation through recursive averaging or backward recursion in longitudinal settings.
• The framework also integrates spectral causal discovery and rigorous falsification criteria to validate claims in rare-event analysis and structural causal modeling.

Searching arXiv for the cited papers and adjacent terminology to ground the article in current literature. {"queries":[{"query":"arXiv (Naimi et al., 18 Jun 2026) A Law of Iterated Expectation Primer for Causal Inference"},{"query":"arXiv (Gyenis, 19 Feb 2026) The Causal Second Law"},{"query":"arXiv (Haghighat et al., 29 May 2026) Formalizing and falsifying causal pathways of rare events"},{"query":"arXiv (Tusoni et al., 16 Jul 2025) Robust Causal Discovery in Real-World Time Series with Power-Laws"}]} Found the relevant papers and metadata on arXiv, including the three directly pertinent works on the g-formula, the causal second law, and causal pathways of rare events, plus the time-series paper on power-law spectral causal discovery. In the current arXiv literature, “Causal-Path Series Law” is not a standard standalone theorem. The phrase most closely matches a family of law-like results about how causal structure is traversed, aggregated, or constrained along a pathway: successive averaging over confounders in causal identification, nondecreasing causal entropy from robust cause to effect, and additive log-scale pathway scores for rare-event explanations. These formalisms arise in distinct settings—causal inference, philosophy of special-science causation, structural causal models for rare events, and robust time-series causal discovery—and they are mathematically related more by the idea of stepwise propagation through causal structure than by a single unified law (Naimi et al., 18 Jun 2026, Gyenis, 19 Feb 2026, Haghighat et al., 29 May 2026, Tusoni et al., 16 Jul 2025).

1. Terminological status and scope

The term is best understood as an umbrella label for several nearby constructions rather than a canonical name in the literature. One line of work treats causal pathways as successive conditional averaging governed by the law of iterated expectation and realized as the g-formula. Another treats causal pathways as entropy-constrained regular transitions in measure-preserving dynamics. A third formalizes binary event pathways in SCMs and derives additive decomposition identities and falsifiable score bounds. A fourth, adjacent rather than equivalent, tracks the evolution of power-law spectral features in time series and performs causal discovery in that transformed domain (Naimi et al., 18 Jun 2026, Gyenis, 19 Feb 2026, Haghighat et al., 29 May 2026, Tusoni et al., 16 Jul 2025).

Literature locus	Closest “law-like” content	Core mathematical object
Causal identification	g-formula as successive averaging	Conditional expectations and integrals
Causal entropy	Causal second law	Phase volume $m(\cdot)$
Rare-event pathways	Additivity and score bounds	Log-probability explanation scores
Time-series discovery	Spectral trend propagation	Windowed power-law parameters $a,\lambda$

This distribution of meanings matters because the same informal expression—following a causal path step by step—supports different technical claims in different subfields. In causal inference, the relevant claim is identifiability under assumptions. In the causal second law, it is entropy monotonicity from robust cause to effect. In rare-event pathway analysis, it is additive decomposition on the log scale plus formal rejection criteria. In spectral time-series discovery, it is robustness gained by moving from raw observations to slower-evolving summaries.

2. Sequential averaging and the g-formula

The most direct mathematical precursor to a “series law” for causal paths is the law of iterated expectation:

$E(Y)=E\!\left[E(Y\mid X)\right]$

and equivalently

$E(Y)=\int E(Y\mid X=x)\,dP(x).$

This identity is purely statistical: the overall mean is recovered by averaging subgroup means over the distribution of the conditioning variable. The causal role of the identity appears when the target is a counterfactual mean such as $E(Y^a)$, which is generally unobserved. Under causal consistency, conditional exchangeability, and positivity, the law of iterated expectation becomes the g-formula, so that a causal estimand is rewritten as an observed-data functional (Naimi et al., 18 Jun 2026).

For a time-fixed confounding structure with baseline confounder $W$, the derivation proceeds as

$E(Y^a)=\sum_w E(Y^a\mid W=w)P(W=w),$

then

$=\sum_w E(Y^a\mid A=a,W=w)P(W=w),$

and then

$=\sum_w E(Y\mid A=a,W=w)P(W=w).$

The three assumptions have distinct roles. Causal consistency states that if $A=a$, then $a,\lambda$0. Conditional exchangeability states

$a,\lambda$1

Positivity requires

$a,\lambda$2

Taken together, they permit a counterfactual mean to be expressed as a standardization over observed strata.

The paper distinguishes two nonparametrically equivalent forms of this standardization. The non-iterative conditional expectation (NICE) form is

$a,\lambda$3

or, for discrete $a,\lambda$4,

$a,\lambda$5

The iterative conditional expectation (ICE) form is

$a,\lambda$6

These are mathematically identical:

$a,\lambda$7

The distinction is therefore computational rather than causal. NICE emphasizes a single weighted average; ICE emphasizes nested prediction and averaging. This suggests that, if “Causal-Path Series Law” is intended to denote a rule for moving through a causal structure by repeated averaging, the most exact counterpart is the g-formula’s use of the law of iterated expectation (Naimi et al., 18 Jun 2026).

3. Longitudinal pathway recursion

The series-like character becomes most explicit in longitudinal settings, where causal identification requires repeated averaging over time-varying confounders. In the two-timepoint example with treatment variables $a,\lambda$8, confounders $a,\lambda$9, and final outcome $E(Y)=E\!\left[E(Y\mid X)\right]$0, the paper gives the NICE form under treatment regime $E(Y)=E\!\left[E(Y\mid X)\right]$1 as

$E(Y)=E\!\left[E(Y\mid X)\right]$2

The corresponding ICE form is

$E(Y)=E\!\left[E(Y\mid X)\right]$3

Because $E(Y)=E\!\left[E(Y\mid X)\right]$4 is affected by earlier treatment and also affects later treatment and outcome, ordinary regression adjustment fails; the identifying functional must instead average sequentially over the evolving confounder distribution (Naimi et al., 18 Jun 2026).

The backward recursion is written as

$E(Y)=E\!\left[E(Y\mid X)\right]$5

$E(Y)=E\!\left[E(Y\mid X)\right]$6

$E(Y)=E\!\left[E(Y\mid X)\right]$7

This recursion gives the most literal version of a causal-path series computation in the material: one regresses and averages backward through the pathway of time-varying confounding. The numerical examples reinforce the equivalence of the two forms. In a time-fixed binary-confounder example, the contrast is reported as

$E(Y)=E\!\left[E(Y\mid X)\right]$8

With mixed discrete and continuous confounders, the estimates are

$E(Y)=E\!\left[E(Y\mid X)\right]$9

In the two-timepoint longitudinal example, the paper reports

$E(Y)=\int E(Y\mid X=x)\,dP(x).$0

These examples support a common interpretation: causal effects can be computed by successively averaging over the distributions of confounders along a causal pathway, provided the identification assumptions hold (Naimi et al., 18 Jun 2026).

A common misconception is that the law of iterated expectation is itself causal. The paper explicitly rejects that reading: the identity is statistical for any probability distribution; it becomes a causal standardization formula only under causal consistency, positivity, and conditional exchangeability.

4. Entropy monotonicity from robust cause to effect

A different candidate for a “Causal-Path Series Law” is the causal second law, which is explicitly formulated as an entropy monotonicity principle for special-science causal regularities. In that framework, special-science descriptions correspond to measurable regions of phase space, and causal regularities are regular transitions between those regions under an underlying dynamical system. A robust causal regularity is formalized as

$E(Y)=\int E(Y\mid X=x)\,dP(x).$1

equivalently,

$E(Y)=\int E(Y\mid X=x)\,dP(x).$2

or

$E(Y)=\int E(Y\mid X=x)\,dP(x).$3

A proportional relaxation is also defined:

$E(Y)=\int E(Y\mid X=x)\,dP(x).$4

Here “robust” means that almost every physical realization of the cause evolves into the effect at the relevant characteristic time (Gyenis, 19 Feb 2026).

The key assumptions are state-supervenience, measure-preservation, a deterministic dynamical-system framework for the main proof, and robustness. Measure preservation is expressed as

$E(Y)=\int E(Y\mid X=x)\,dP(x).$5

for all measurable $E(Y)=\int E(Y\mid X=x)\,dP(x).$6. Causal entropy is defined as the phase volume $E(Y)=\int E(Y\mid X=x)\,dP(x).$7 or $E(Y)=\int E(Y\mid X=x)\,dP(x).$8, and can also be represented as a monotone function of phase volume, in particular by a Boltzmann-style form

$E(Y)=\int E(Y\mid X=x)\,dP(x).$9

The central theorem states that if $E(Y^a)$0 is measure-preserving and $E(Y^a)$1, then

$E(Y^a)$2

The proof uses

$E(Y^a)$3

and measure preservation to infer

$E(Y^a)$4

Thus causal entropy cannot decrease from robust cause to effect (Gyenis, 19 Feb 2026).

The paper also gives sufficient conditions for strict increase. One is the presence of multiple possible causes for the same effect: if distinct $E(Y^a)$5 robustly lead to $E(Y^a)$6 and their evolved images are not almost everywhere identical, then

$E(Y^a)$7

Another is a mismatch between the descriptive resources of the special science and the full fine-grained physical dynamics, so that the smallest describable cause region leading to the effect must be broader than a fine-grained pull-back.

Several caveats are integral to the result. The law is compatible with multiple realizability, theory-supervenience, and open systems embedded in a larger dynamically isolated composite. It is also compatible with a more non-metaphysical reading of supervenience. Most importantly, it does not imply a causal time arrow: entropy nondecrease from cause to effect is distinct from entropy nondecrease in time. The paper further argues that the standard reversibility objection does not threaten the law, because the time-reversed regions need not be special-science descriptions, and even if they are, measure preservation blocks a reverse robust-cause relation when the effect has larger phase volume (Gyenis, 19 Feb 2026).

5. Pathway additivity and falsification for rare events

A third formalization closest to the phrase “Causal-Path Series Law” appears in the analysis of causal pathways of rare events in SCMs. A causal pathway is defined as a tuple $E(Y^a)$8 built from binary event variables $E(Y^a)$9, a DAG describing causal relations among them, a unique sink node $W$0 as the target event, a set of root causes $W$1, and a pathway subgraph $W$2 that coincides with the full DAG except possibly for some edges into the root-cause set. The framework is designed to formalize explanations such as “the rare target happened via this pathway,” while retaining testable implications (Haghighat et al., 29 May 2026).

The scores are logarithmic and interventional. The general explanation score is

$W$3

For any subset $W$4 of nodes, the target explanation score is

$W$5

The pathway explanation score is

$W$6

The cluster explanation score is

$W$7

These definitions already imply a series-like structure: contributions multiply at the probability level and add on the log scale (Haghighat et al., 29 May 2026).

The paper proves several decomposition results. For disjoint $W$8,

$W$9

and

$E(Y^a)=\sum_w E(Y^a\mid W=w)P(W=w),$0

For pathway scores, the key modified additivity identity is

$E(Y^a)=\sum_w E(Y^a\mid W=w)P(W=w),$1

This is the strongest candidate in the paper for a genuine pathway decomposition law. The paper also shows that the pathway score is always no larger than the corresponding target score:

$E(Y^a)=\sum_w E(Y^a\mid W=w)P(W=w),$2

where $E(Y^a)=\sum_w E(Y^a\mid W=w)P(W=w),$3 is the complement of $E(Y^a)=\sum_w E(Y^a\mid W=w)P(W=w),$4. The gap measures how much of the non-root part of the pathway still has to be completed after intervening on the root causes.

The framework is explicitly falsification-oriented. A necessary condition for a high pathway score is the absence of large log-likelihood gaps. Defining

$E(Y^a)=\sum_w E(Y^a\mid W=w)P(W=w),$5

the paper proves

$E(Y^a)=\sum_w E(Y^a\mid W=w)P(W=w),$6

Thus a pathway cannot score highly if a non-root node is much less likely than its parent configuration. The paper also provides rejection theorems. One states that the mechanisms of non-root causes working as expected can be rejected with p-value

$E(Y^a)=\sum_w E(Y^a\mid W=w)P(W=w),$7

where $E(Y^a)=\sum_w E(Y^a\mid W=w)P(W=w),$8 is pathway abstraction accuracy. Another gives, under feature monotonicity, the bound

$E(Y^a)=\sum_w E(Y^a\mid W=w)P(W=w),$9

with

$=\sum_w E(Y^a\mid A=a,W=w)P(W=w),$0

These results connect verbal pathway claims to observational and interventional rejection criteria (Haghighat et al., 29 May 2026).

A central warning follows from the paper’s chain-of-events example:

$=\sum_w E(Y^a\mid A=a,W=w)P(W=w),$1

Strong local explanations therefore do not guarantee a strong global pathway explanation. That point is often obscured in informal “series” reasoning.

6. Adjacent developments, distinctions, and recurring misconceptions

An adjacent but distinct use of pathwise causal reasoning appears in robust causal discovery for noisy time series with power-law spectra. The method PLaCy begins from the observation that many real-world stochastic time series exhibit

$=\sum_w E(Y^a\mid A=a,W=w)P(W=w),$2

or, for amplitude,

$=\sum_w E(Y^a\mid A=a,W=w)P(W=w),$3

equivalently

$=\sum_w E(Y^a\mid A=a,W=w)P(W=w),$4

Each series is segmented into overlapping windows; for window $=\sum_w E(Y^a\mid A=a,W=w)P(W=w),$5, the discrete Fourier transform is

$=\sum_w E(Y^a\mid A=a,W=w)P(W=w),$6

Ordinary least squares in log-log space yields windowwise intercepts $=\sum_w E(Y^a\mid A=a,W=w)P(W=w),$7 and slopes $=\sum_w E(Y^a\mid A=a,W=w)P(W=w),$8, after which multivariate Granger causality is applied to the transformed data with significance threshold $=\sum_w E(Y^a\mid A=a,W=w)P(W=w),$9. The appendix states that $=\sum_w E(Y\mid A=a,W=w)P(W=w).$0 is the main carrier of causal information, while $=\sum_w E(Y\mid A=a,W=w)P(W=w).$1 may help as a covariate in some stationary cases but is less useful or even harmful as a target in nonstationary regimes (Tusoni et al., 16 Jul 2025).

This framework is not a “causal-path series law” in the theorem-like sense. It is a preprocessing-and-testing method that shifts causal discovery from raw time-domain observations to spectral trend features. Its relevance here is conceptual: rather than infer causality directly from noisy trajectories, it tracks the slower evolution of a structural summary, thereby suppressing spurious lagged correlations. The paper states that the causal graph is invariant under the spectral transformation used in PLaCy, and reports strong performance in difficult regimes, including F1 $=\sum_w E(Y\mid A=a,W=w)P(W=w).$2 and TNR $=\sum_w E(Y\mid A=a,W=w)P(W=w).$3 on the Rivers dataset, and F1 $=\sum_w E(Y\mid A=a,W=w)P(W=w).$4 and TNR $=\sum_w E(Y\mid A=a,W=w)P(W=w).$5 on AirQuality. It also notes limitations: the method is not well suited to very short series, may struggle when spectra vary only slowly over time, does not address latent confounders, and is not positioned as a universal replacement for other causal discovery methods (Tusoni et al., 16 Jul 2025).

Across these literatures, several misconceptions recur. First, there is no single accepted theorem called “Causal-Path Series Law.” Second, the law of iterated expectation is not inherently causal; it becomes the g-formula only under causal assumptions. Third, the causal second law does not establish a time arrow. Fourth, additive or high local pathway scores do not entail a good global rare-event explanation. Fifth, spectral causal discovery on power-law features is not a new causal theory but a robust representation-and-testing framework.

Taken together, the closest encyclopedic characterization is therefore this: the phrase denotes a cluster of results in which causal structure is handled stepwise—by repeated averaging, monotone phase-volume inclusion, or additive log-score decomposition—rather than a single theorem with a fixed formal statement.