Kan-Do-Calculus (KDC): A Concise Overview

Updated 4 July 2026

Kan-Do-Calculus is a diverse family of formalisms that unifies causal intervention, identifiability, and transportability across different mathematical frameworks.
It employs universal constructions like Kan extensions and adjunctions to reinterpret do-calculus rules within categorical, Bayesian, and geometric settings.
KDC enables the translation of interventional queries into estimable observational expressions, facilitating meta-synthesis and error minimization in causal models.

Kan-Do-Calculus (KDC) is not a standardized term in the arXiv literature. In the causal-inference lineage associated with Pearl’s survey of do-calculus, the expression does not appear as a formal name; later works use or interpret it as a label for several related calculi centered on intervention, Kan extension, localization, and universal constructions. In one usage, KDC denotes the modern deployment of do-calculus in nonparametric structural causal models (SCMs), including identifiability, mediation, transportability, and meta-synthesis. In other usages, it denotes categorical semantics for intervention and conditioning, Bayesian reformulations of causal estimation, a Kan-extension view of learning, geometric algorithms for causal discovery under latent confounding, and even homotopical localization calculi (Pearl, 2012, Yin et al., 2022, Mahadevan, 17 Jun 2026, Riehl et al., 2015).

1. Terminological scope and major usages

The literature attaches the KDC label to several mathematically distinct programs. Some are directly causal; others are categorical or homotopical and use “Kan” and “calculus” in a different sense. The common thread is a calculus of universal transformations: rewriting causal queries, computing Kan extensions, or representing localized morphisms by short normal forms.

Source	Sense of KDC	Core formal objects
(Pearl, 2012)	Umbrella for modern do-calculus	SCMs, DAGs, interventions, identifiability
(Yin et al., 2022)	Categorical do-calculus	Free Markov categories, causal effects, t-separation
(Mahadevan, 17 Jun 2026)	Kan/do bi-adjunction and geometric discovery	Kan extensions, RN derivatives, Lie brackets, BRIDGE, SKFM
(Riehl et al., 2015)	Calculus of modules for Kan extensions	$\infty$ -cosmos, modules, virtual equipment, exact squares
(Lattimore et al., 2019)	Bayesian replacement of do-calculus	Twin-network PGMs, mechanism invariance, posterior predictive
(Pugh et al., 19 Feb 2025)	Learning as Kan extension	Left Kan extensions, adjunctions, error minimization
(Thomas, 2010)	Informal Kan–Dwyer or 3-arrow calculus	Localization, denominators, double fractions
(Gu et al., 2024)	Kan/Goodwillie comparison in retractive spaces	Bousfield–Kan completions, Taylor towers

This multiplicity matters because the same acronym can refer either to Pearlian intervention logic or to category-theoretic and homotopical calculi with different primitives. A plausible implication is that KDC functions less as a single formalism than as a family-resemblance label spanning several research communities.

2. Structural-causal foundations and the three rules

In the classical causal setting, the core object is a nonparametric SCM. An SCM $M$ consists of exogenous variables $U$ with distribution $P(u)$ , endogenous variables $V=\{V_1,\dots,V_n\}$ , structural functions $V_i=f_i(PA_i,U_i)$ , and a causal diagram $G$ encoding parent-child relations. An intervention $X:=x$ replaces the structural functions for nodes in $X$ with constants and yields the intervened model $M_x$ , with post-intervention distribution

$M$ 0

Graphically, interventions are represented by mutilations such as $M$ 1 for incoming-edge deletion and $M$ 2 for outgoing-edge deletion; d-separation in these mutilated graphs is the formal mechanism by which do-calculus relates interventional and observational quantities (Pearl, 2012).

The three rules of do-calculus are rewrite rules for expressions containing $M$ 3. For disjoint node sets $M$ 4 in a causal DAG $M$ 5, they are:

Insertion/deletion of observations

$M$ 6

if $M$ 7 in $M$ 8.

Action/observation exchange

$M$ 9

if $U$ 0 in $U$ 1.

Insertion/deletion of actions

$U$ 2

if $U$ 3 in $U$ 4, where $U$ 5 is the subset of $U$ 6 that are not ancestors of any node in $U$ 7 in $U$ 8 (Pearl, 2012).

These rules formalize when conditioning can be inserted or removed, when intervention can be exchanged for observation, and when an intervention can be dropped altogether. Their importance lies not only in symbolic manipulation but in identifiability: a causal query is identifiable when it depends only on the observational distribution and can therefore be expressed in observed quantities.

3. Completeness, identifiability, and canonical adjustment formulas

A central result of the modern theory is that do-calculus is complete for nonparametric identification of causal effects. Huang and Valtorta and Shpitser and Pearl showed that whenever a causal effect is identifiable from observational data under a causal diagram, there exists a derivation using the three rules that removes all $U$ 9 operators; conversely, if repeated application cannot eliminate all $P(u)$ 0 symbols, the effect is not identifiable. Parallel work by Tian and Pearl, and later Tian and Shpitser, provided graphical criteria and polynomial-time algorithms for deciding identifiability and synthesizing estimands, which Pearl described as having “closed the chapter” of nonparametric identification in DAGs (Pearl, 2012).

Two canonical identification formulas illustrate the calculus. If $P(u)$ 1 satisfies the back-door criterion relative to $P(u)$ 2, then

$P(u)$ 3

If $P(u)$ 4 satisfies the front-door criterion relative to $P(u)$ 5, then

$P(u)$ 6

Both formulas are derivable from Rules 1–3, but their significance is broader: they exemplify how do-calculus moves from a causal query to an estimand expressible in observational terms (Pearl, 2012).

The same completeness results sharply delimit non-identifiability. Typical failure modes include unobserved confounding between $P(u)$ 7 and $P(u)$ 8 with no admissible back-door set and no front-door mediator meeting the front-door conditions, or mediation settings in which the required $P(u)$ 9-specific effects are not identifiable. In these cases, graphical criteria and associated algorithms detect failure and block unjustified estimation attempts (Pearl, 2012).

4. Mediation, transportability, and meta-synthesis

Pearl’s 2012 survey emphasized that do-calculus extends beyond plain effect identification into mediation analysis, transportability, and meta-synthesis (Pearl, 2012). In mediation, the controlled direct effect is

$V=\{V_1,\dots,V_n\}$ 0

For natural direct and indirect effects,

$V=\{V_1,\dots,V_n\}$ 1

and under Assumption Set A these admit the identification formulas

$V=\{V_1,\dots,V_n\}$ 2

$V=\{V_1,\dots,V_n\}$ 3

Assumption Set A requires a measured covariate set $V=\{V_1,\dots,V_n\}$ 4 such that no member of $V=\{V_1,\dots,V_n\}$ 5 is a descendant of $V=\{V_1,\dots,V_n\}$ 6, $V=\{V_1,\dots,V_n\}$ 7 blocks all back-door paths from $V=\{V_1,\dots,V_n\}$ 8 to $V=\{V_1,\dots,V_n\}$ 9 ignoring the one through $V_i=f_i(PA_i,U_i)$ 0, the $V_i=f_i(PA_i,U_i)$ 1-specific effect of $V_i=f_i(PA_i,U_i)$ 2 on $V_i=f_i(PA_i,U_i)$ 3 is identifiable using do-calculus, and the $V_i=f_i(PA_i,U_i)$ 4-specific joint effect of $V_i=f_i(PA_i,U_i)$ 5 on $V_i=f_i(PA_i,U_i)$ 6 is identifiable using do-calculus. Pearl contrasts this with stronger “sequential ignorability” assumptions requiring a single covariate set to deconfound both $V_i=f_i(PA_i,U_i)$ 7 and $V_i=f_i(PA_i,U_i)$ 8 (Pearl, 2012).

Transportability generalizes identification across domains. Pearl and Bareinboim introduced selection diagrams, which augment a shared DAG with selection variables $V_i=f_i(PA_i,U_i)$ 9 wherever mechanisms differ across domains. A target relation such as $G$ 0 is transportable if it is uniquely computable from source observational and interventional distributions together with target observational data. Pearl’s survey states a do-calculus criterion: transportability holds if $G$ 1 is reducible to an expression in which selection variables appear only as conditioning variables in do-free terms (Pearl, 2012). Representative formulas include

$G$ 2

when only the marginal of $G$ 3 differs across domains, and

$G$ 4

when $G$ 5 lies on the $G$ 6 path and differs structurally given $G$ 7 (Pearl, 2012).

Meta-synthesis, a term coined in the same survey, is the principled fusion of heterogeneous observational and experimental results to estimate a target causal relation in a target environment. A target relation is meta-identifiable when it can be decomposed into subrelations transportable from individual studies, after which do-calculus-guided transport formulas synthesize an unbiased estimator. The workflow is explicitly graphical: encode each study by a selection diagram, decompose the target estimand into transportable pieces, and use do-calculus to separate selection variables from interventions (Pearl, 2012). This generalizes ordinary meta-analysis from effect aggregation to structure-sensitive causal synthesis.

5. Category-theoretic formulations

Several later works place KDC in explicitly categorical settings. One line, developed in free Markov categories, associates to a DAG $G$ 8 a causal syntax category $G$ 9 generated by comonoid maps and local mechanisms $X:=x$ 0. In this setting an intervention $X:=x$ 1 is a syntactic rewrite that removes the incoming mechanism for $X:=x$ 2 and substitutes a chosen state $X:=x$ 3. The paper gives syntactic analogues of conditional independence, decomposition, and irrelevance via t-separation, with rules such as

$X:=x$ 4

under t-separation. In causal Bayesian networks with positivity and acyclicity, these specialized categorical rules are as strong as Pearl’s full do-calculus (Yin et al., 2022).

A second line formulates the observation–intervention boundary as a bi-adjunction

$X:=x$ 5

Here $X:=x$ 6 embeds observational contexts into interventional states, $X:=x$ 7 models universal pushforward into intervention, and $X:=x$ 8 models universal pullback or conditioning. In the probabilistic semantics described for a Markov category, the Radon–Nikodym calibration identity

$X:=x$ 9

connects observational and interventional laws, and do-rules are interpreted as instances of push–pull compatibility, projection-formula behavior, and Frobenius reciprocity for Kan extensions (Mahadevan, 17 Jun 2026). This does not replace Pearl’s graphical rules; it reinterprets them as universal constructions.

A third categorical lineage appears in the calculus of modules for an $X$ 0-cosmos. There, modules $X$ 1 form a virtual equipment $X$ 2 with cartesian restriction cells and cocartesian unit cells. Pointwise Kan extensions are defined as right extensions in this virtual equipment and are equivalent to 2-categorical Kan extensions stable under pasting with exact squares, especially comma squares. In cartesian closed $X$ 3-cosmoi, limits and colimits arise as pointwise Kan extensions along $X$ 4 (Riehl et al., 2015). Although this paper does not use the term KDC, later descriptions apply the label to precisely this calculus of modules, companions, conjoints, exact squares, and Beck–Chevalley base change.

6. Bayesian and learning-theoretic reformulations

A different KDC usage argues that causal estimation can be carried out entirely within the standard Bayesian paradigm, provided causal invariance assumptions are encoded explicitly. The key construction is a twin-network probabilistic graphical model containing both the observed world and the post-intervention world, with shared parameters for all non-intervened mechanisms. The joint factorization is

$X$ 5

and the interventional query is computed by the posterior predictive

$X$ 6

Under the modularity assumption that only the intervened mechanism changes, this reproduces Pearl’s truncated product, and standard Bayesian probability manipulations recover back-door and front-door formulas without invoking symbolic do-calculus rules (Lattimore et al., 2019). The same framework also clarifies non-identifiability: when observational data identify only a mixture such as $X$ 7 but not the post-intervention quantity $X$ 8, the posterior remains prior-sensitive even in the large-sample limit (Lattimore et al., 2019).

An even broader generalization appears in the claim that every error minimization algorithm can be presented as a left Kan extension. In the 2-categorical formulation of learning, a left Kan extension along $X$ 9 is a global error minimizer, and Theorem 5.1 states that any set-theoretic error minimization problem can be represented in such a way. Adjunctions yield global minimizers independently of the choice of error functional, while left Kan extensions provide pointwise minimizers when they exist. The enriched formulas

$M_x$ 0

make explicit how learning is recast as weighted aggregation or weighted constraint satisfaction (Pugh et al., 19 Feb 2025). In this usage, KDC is no longer confined to causal identification; it becomes a general calculus for optimization, regularization, and representation of learning algorithms.

7. Geometric discovery, localization, and homotopical extensions

The most recent causal-discovery usage takes KDC into smooth statistical geometry. When interventional and observational measures are mutually absolutely continuous on a smooth statistical manifold, Radon–Nikodym derivatives induce local intervention vector fields $M_x$ 1, and their Lie brackets

$M_x$ 2

measure non-commutativity of infinitesimal interventions. Frobenius residuals such as

$M_x$ 3

or the experimental bracket norm

$M_x$ 4

detect failures of visible integrability and are interpreted as witnesses of latent or unmodeled structure. Two algorithms implement this program: BRIDGE uses RN-ratio estimation, geometric screening, latent-obstruction detection, and downstream score-based search; SKFM learns amortized intervention fields and factors latent curvature spectrally (Mahadevan, 17 Jun 2026). The reported experiments include an endpoint-confounded 7-node chain in which BRIDGE retains 21 of 42 arrows, reducing the acyclic candidate family to 107,121 graphs, an exact compression of 10,630.8×, and a 9-gene S9 setting in which 72 arrows are compressed to 18, reducing an approximately $M_x$ 5 labeled-DAG family to 119,216 acyclic candidates (Mahadevan, 17 Jun 2026).

Outside causal inference proper, one informal use of KDC refers to a Kan–Dwyer-style localization calculus. In a uni-fractionable category, every morphism in the localization is represented by a 3-arrow diagram

$M_x$ 6

with $M_x$ 7 denominators, and equality is characterized by embeddability into a commutative $M_x$ 8 diagram. The resulting fraction category $M_x$ 9 has the universal property of localization, applies to model categories and derived categories, and does not require functorial factorizations (Thomas, 2010). Here KDC denotes a practical short-zigzag calculus rather than intervention.

A further homotopical usage appears in the interaction between Bousfield–Kan completions and Goodwillie calculus in retractive spaces. For a 0-connected base $M$ 00 and the endofunctors $M$ 01, the Bousfield–Kan completion is

$M$ 02

If $M$ 03 is 1-connected relative to $M$ 04, then the coaugmentation $M$ 05 is a weak equivalence for all $M$ 06; if $M$ 07 is only 0-connected relative to $M$ 08, then

$M$ 09

for all $M$ 10 (Gu et al., 2024). In this setting, “Kan” refers to Bousfield–Kan completion and “Do” to Goodwillie’s functor calculus, producing yet another mathematically precise but discipline-specific meaning of KDC.

Across these strands, KDC names a collection of calculi rather than a single universally accepted formalism. In the Pearlian sense it is the graph-based logic of intervention, identifiability, and transport across environments; in categorical versions it becomes a semantics of intervention and conditioning by Kan-like universal properties; in Bayesian and learning-theoretic versions it becomes a calculus of posterior prediction and minimization; and in geometric and homotopical versions it denotes algorithmic or localization frameworks that preserve the themes of universal rewriting, compositionality, and structural constraints (Pearl, 2012, Yin et al., 2022, Mahadevan, 17 Jun 2026, Lattimore et al., 2019, Pugh et al., 19 Feb 2025, Thomas, 2010, Gu et al., 2024).