Causal Reasoning with Bipartite Graphical Causal Models

arXiv:2608.198312026-08-21

A bathtub example exposes a blind spot in standard causal reasoning tools

Standard causal frameworks (causal Bayesian networks, structural causal models) struggle with systems that sit at equilibrium with feedback loops, like a bathtub with water flowing in and draining out at a steady state. This paper introduces bipartite graphical causal models (BGCMs), which draw both variables and equations as graph nodes so that an intervention specifies exactly which equation was replaced, resolving cases where fixing the same variable to the same value through different equations gives different results. Using a full analysis of the bathtub system, the authors show this new notation maps cleanly onto distinct physical procedures, and they build a new graph-separation rule and calculus on top of it.

What they did

The problem: in a bathtub at equilibrium (inflow equals outflow), there are several distinct physical ways to fix the water depth to the same value, but they lead to different pressures and outflows. Standard notation like do(X=x) can't tell these apart because it only names the variable, not the mechanism being changed.
The method: the authors draw a bipartite graph with two kinds of nodes, variables and equations, connecting a variable to an equation whenever it appears in it. They extend Simon's causal ordering algorithm to derive a partial ordering over clusters of equations, and write interventions as do(f_j: X_v = value) to specify which equation is being replaced.
A new separation rule: they define B-separation, a generalization of the standard d-separation rule for reading off conditional independence, which also accounts for the deterministic relationships forced by the equations; this extends to cases where some inputs are treated as fixed rather than random.
Results: a complete table of all possible interventions on the bathtub shows which ones yield a uniquely solvable system and correspond to real physical procedures, and which are infeasible. The same framework is also tested on a supply-and-demand market example that has a genuine feedback cycle.
The resulting BGCM framework is shown to strictly generalize causal Bayesian networks and structural causal models while still supporting graph-based causal reasoning, including a do-calculus for reasoning about which relationships stay invariant across different conditions.

Figure 1: The bathtub system (rendered by Google Gemini).

Table 1: Feasibility of hard interventions do(fj:Xv=ξv) for the bathtub. Checkmarks indicate uniquely solvable systems; ∅ indicates the system is not uniquely solvable.
do(fj:Xv=ξv)	f1	f2	f3
XO=ξO	✓	∅	∅
XP=ξP	✓	✓	∅
XD=ξD	✓	✓	✓

Figure 2: Three different interventions on the bathtub that all set the water depth to a fixed value (Example 20).

Table 2: Solution functions for all hard interventions on the bathtub model.
	XO	XP	XD
observational	XI	XI2XK2	XI2XK2Xg
do⁡(XI=ξI)	ξI	ξI2XK2	ξI2XK2Xg
do⁡(XK=ξK)	XI	XI2ξK2	XI2ξK2Xg
do⁡(Xg=ξg)	XI	XI2XK2	XI2XK2ξg
do(f1:XO=ξO)	ξO	ξO2XK2	ξO2XK2Xg
do(f1:XP=ξP)	ξPXK	ξP	ξPXg
do(f1:XD=ξD)	XKXgξD	XgξD	ξD
do(f2:XP=ξP)	XI	ξP	ξPXg
do(f2:XD=ξD)	XI	XgξD	ξD
do(f3:XD=ξD)	XI	XI2XK2	ξD

Why it matters

Many real systems in physics, economics, and biology involve feedback loops that existing causal tools can't properly represent, so this new notation gives a way to reason correctly about them by making explicit which mechanism an intervention actually changes. It clarifies that interventions which look identical on paper (fixing the same variable to the same value) can be physically and causally different, which matters for designing experiments or policies.

Terms in this paper

causal Bayesian network (CBN) · a standard model that represents causal relations using a directed acyclic graph plus conditional probabilities
structural causal model (SCM) · a standard model that defines each variable by an equation involving its causes and noise
bipartite graph · a graph with two types of nodes, here variables and equations, connected only across the two types
B-separation · a new graphical rule for reading off conditional independence in bipartite causal graphs
do-calculus · a set of rules for relating probabilities before and after an intervention

Original abstract (English)

Causal Bayesian networks (CBNs) and structural causal models (SCMs) are the dominant frameworks for graphical causal reasoning, but they cannot adequately represent all real-world causal systems. In particular, systems at equilibrium---where feedback mechanisms create cyclic causal dependencies---can exhibit causal semantics that are fundamentally incompatible with these frameworks: different interventions that enforce the same variable value may have different effects, rendering the standard ``perfect intervention'' do($X = x$) ambiguous. We propose bipartite graphical causal models (BGCMs), in which the structure of a system of equations is encoded by a bipartite graph with variable and equation nodes. In this framework, a hard intervention do($f_j : X_v = \xi_v$) specifies which equation is replaced, which variable is targeted, and at what value---resolving the ambiguity of the standard notion. We demonstrate, through a detailed case study of a physical system, that this representation naturally corresponds to distinct real-world interventions. We formulate a Markov property in terms of a new graphical separation criterion (B-separation) that exploits the functional determinism inherent in the equations, and we extend it to settings with non-random inputs. We show how this gives rise to a do-calculus for reasoning about domain invariances. BGCMs strictly generalize CBNs and SCMs while retaining the ability to perform graphical causal reasoning.

Authors · Joris M. Mooij

Read on arXiv

Latest papers

All papers →

Latest from METAL LAB

Figures: Joris M. Mooij et al., arXiv:2608.19831, arxiv-nonexclusive