Causal Action Principle Overview

• Causal action principle is a framework that defines interventions using expected utility maximization in decision theory and nonlinear variational methods in physics.
• It applies causal graphical models for evaluating interventional consequences and extends to settings with unknown causal models and strategic interactions.
• In causal fermion systems, the principle employs spectral methods on operator measures to derive emergent spacetime structure, geometry, and field dynamics.

The term causal action principle is used in at least two technically distinct literatures. In causal decision theory and causal graphical modeling, it denotes the rule that a rational agent should evaluate actions by their interventional consequences, choosing actions that maximize expected utility with respect to distributions of the form $P(\cdot \mid \mathrm{do}(A=a))$ rather than observational conditionals. In causal fermion systems and causal variational principles, it denotes a nonlinear variational principle on measures over operator spaces, where spacetime is identified with the support of the minimizing measure and causal relations are encoded spectrally in operator products (Gonzalez-Soto et al., 2019, Finster et al., 2011).

1. Decision-theoretic formulation

In the decision-theoretic usage, a causal action principle is formulated on a Causal Decision Problem under Uncertainty (CDPU), a tuple

$$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \mathcal{G}, \succeq),$$

where $(\mathcal{A}, \mathcal{E}, \mathcal{C}, \succeq)$ is a classical decision problem under uncertainty and $\mathcal{G}$ is a causal graphical model relating actions, uncertain events, and consequences (Gonzalez-Soto et al., 2019). The environment is represented by a DAG whose joint distribution factorizes as

$$P(X_1,\dots,X_n)=\prod_{i=1}^n P(X_i \mid \mathrm{Pa}(X_i)),$$

with the usual assumptions of acyclicity, the Markov condition, and invariance under interventions.

The defining move is to interpret each admissible action $a \in \mathcal{A}$ as an intervention $\mathrm{do}(A=a)$ on $\mathcal{G}$. The action-value criterion is then causal rather than associational: $$a^\ast \in \arg\max_{a \in \mathcal{A}} \mathbb{E}[u(Y)\mid \mathrm{do}(A=a),\mathcal{G}] = \arg\max_{a \in \mathcal{A}} \sum_y u(y)\,P(Y=y \mid \mathrm{do}(A=a),\mathcal{G}).$$ In the binary target case emphasized in the paper, where $Y\in\{0,1\}$ and $$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \mathcal{G}, \succeq),$$0 is the desired outcome, this reduces to

$$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \mathcal{G}, \succeq),$$1

This rule is justified by combining expected-utility rationality with Pearl’s intervention semantics: if actions are interventions, then the relevant probabilities are $$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \mathcal{G}, \succeq),$$2, not $$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \mathcal{G}, \succeq),$$3. The distinction matters precisely when confounding or more complex causal structure makes observational and interventional conditionals differ (Gonzalez-Soto et al., 2019).

A closely related representation theorem extends the same logic to a setting in which the agent’s available probabilities are explicitly grounded in causal models. For a known causal model $$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \mathcal{G}, \succeq),$$4, the criterion is

$$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \mathcal{G}, \succeq),$$5

For a finite family of possible causal models $$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \mathcal{G}, \succeq),$$6, the same paper shows that rational choice can be represented using a model-averaged causal distribution rather than a single known graph (Soto et al., 2019).

2. Unknown causal models, sequential learning, and strategic interaction

When the causal mechanism is unknown, the decision-theoretic causal action principle is generalized by introducing a probability distribution $$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \mathcal{G}, \succeq),$$7 over a family $$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \mathcal{G}, \succeq),$$8 of causal models. The resulting causal expected utility is

$$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \mathcal{G}, \succeq),$$9

and rational choice is represented by

$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \succeq)$0

This replaces the classical Savage-style use of states of the world by a distribution over causal models, while preserving expected-utility form (Soto et al., 2019).

The online decision-making procedure discussed in the CDPU framework assumes that the graph structure is known while the conditional probability tables are unknown. The agent maintains beliefs over parameters, computes $(\mathcal{A}, \mathcal{E}, \mathcal{C}, \succeq)$1 under the current beliefs, chooses the maximizing action, observes the outcome, and updates beliefs by Bayesian updating. The reported empirical result is that such a causal agent achieves similar average reward to a Q-learning agent while also learning a causal model of the environment; both outperform a random action-selection agent in the cited medical treatment scenario (Gonzalez-Soto et al., 2019).

The same causal decision-theoretic logic has also been extended to games. In the proposed causal analogue of Bayesian strategic games, the states of nature are interpreted as admissible causal models, and each player evaluates action profiles by causal consequence distributions. A Causal Nash Equilibrium is then an action profile $(\mathcal{A}, \mathcal{E}, \mathcal{C}, \succeq)$2 such that, for every player $(\mathcal{A}, \mathcal{E}, \mathcal{C}, \succeq)$3,

$(\mathcal{A}, \mathcal{E}, \mathcal{C}, \succeq)$4

where $(\mathcal{A}, \mathcal{E}, \mathcal{C}, \succeq)$5 is defined from causal probabilities of the form $(\mathcal{A}, \mathcal{E}, \mathcal{C}, \succeq)$6 weighted by the player’s belief over causal models (Soto et al., 2019).

The main limitations stated in this line of work are also explicit. The CDPU paper does not provide formal theorems with regret bounds or convergence rates, and its asymptotic equivalence to known maximum-expected-utility procedures is stated as a conjecture rather than a theorem (Gonzalez-Soto et al., 2019). The unknown-model representation theorem is normative rather than algorithmic: it does not supply a concrete causal discovery procedure for learning $(\mathcal{A}, \mathcal{E}, \mathcal{C}, \succeq)$7 from data (Soto et al., 2019).

3. Variational principle in causal fermion systems

In causal fermion systems, the causal action principle is a variational principle on a triple $(\mathcal{A}, \mathcal{E}, \mathcal{C}, \succeq)$8, where $(\mathcal{A}, \mathcal{E}, \mathcal{C}, \succeq)$9 is a complex Hilbert space, $\mathcal{G}$0 is the set of self-adjoint finite-rank operators with at most $\mathcal{G}$1 positive and $\mathcal{G}$2 negative eigenvalues, and $\mathcal{G}$3 is a positive measure on $\mathcal{G}$4. Spacetime is defined as

$\mathcal{G}$5

For $\mathcal{G}$6, letting $\mathcal{G}$7 denote the non-trivial eigenvalues of $\mathcal{G}$8, one defines the spectral weights

$\mathcal{G}$9

and the Lagrangian

$$P(X_1,\dots,X_n)=\prod_{i=1}^n P(X_i \mid \mathrm{Pa}(X_i)),$$0

The causal action is

$$P(X_1,\dots,X_n)=\prod_{i=1}^n P(X_i \mid \mathrm{Pa}(X_i)),$$1

to be minimized subject to a fixed volume constraint, a trace constraint, and a boundedness constraint involving $$P(X_1,\dots,X_n)=\prod_{i=1}^n P(X_i \mid \mathrm{Pa}(X_i)),$$2 (Finster et al., 2011).

This variational structure is called “causal” because the spectral classification of the operator product $$P(X_1,\dots,X_n)=\prod_{i=1}^n P(X_i \mid \mathrm{Pa}(X_i)),$$3 simultaneously defines causal separation and suppresses spacelike contributions. If all eigenvalues $$P(X_1,\dots,X_n)=\prod_{i=1}^n P(X_i \mid \mathrm{Pa}(X_i)),$$4 have the same absolute value, then $$P(X_1,\dots,X_n)=\prod_{i=1}^n P(X_i \mid \mathrm{Pa}(X_i)),$$5 and $$P(X_1,\dots,X_n)=\prod_{i=1}^n P(X_i \mid \mathrm{Pa}(X_i)),$$6 are spacelike separated and $$P(X_1,\dots,X_n)=\prod_{i=1}^n P(X_i \mid \mathrm{Pa}(X_i)),$$7. Timelike and lightlike separation are likewise defined spectrally. In this way, the support of a minimizing measure carries an emergent topology and causal structure without assuming a background manifold (Finster et al., 2011).

The same framework also defines spin spaces $$P(X_1,\dots,X_n)=\prod_{i=1}^n P(X_i \mid \mathrm{Pa}(X_i)),$$8, a spin scalar product $$P(X_1,\dots,X_n)=\prod_{i=1}^n P(X_i \mid \mathrm{Pa}(X_i)),$$9, wave functions $a \in \mathcal{A}$0, the fermionic operator $a \in \mathcal{A}$1, and the closed chain $a \in \mathcal{A}$2. On this basis, the theory develops notions of Clifford subspaces, a spin connection, a metric connection, and curvature. In appropriate continuum limits, the resulting metric connection converges to Levi-Civita parallel transport in Minkowski space and agrees with the Lorentzian manifold connection up to higher-order curvature corrections in globally hyperbolic settings (Finster et al., 2011).

A major feature of this literature is that the causal action principle is meant to supply a common origin for geometry, matter, and interaction. In the continuum limit around the Dirac sea, the Euler–Lagrange equations reduce to a nonlinear system of Dirac and classical gauge field equations, including Maxwell-type equations sourced by Dirac currents. The same framework also suggests a discreteness or “quantization effect” in simplified variational models, where numerical minimizers become purely atomic and highly symmetric (Finster et al., 2011).

4. Homogeneous momentum-space form and continuum dynamics

A translation-invariant or homogeneous version of the causal action principle is formulated directly in momentum space by replacing the universal measure on operator space with a positive definite operator-valued measure $a \in \mathcal{A}$3 on a compact momentum domain $a \in \mathcal{A}$4. The translation-invariant kernel of the fermionic projector is

$a \in \mathcal{A}$5

the closed chain is

$a \in \mathcal{A}$6

and the homogeneous causal Lagrangian is

$a \in \mathcal{A}$7

where $a \in \mathcal{A}$8 are the eigenvalues of $a \in \mathcal{A}$9. The action is

$\mathrm{do}(A=a)$0

subject to the trace constraint

$\mathrm{do}(A=a)$1

and the dimension constraint

$\mathrm{do}(A=a)$2

with $\mathrm{do}(A=a)$3 the eigenvalues of $\mathrm{do}(A=a)$4 (Finster et al., 2022).

On a compact momentum domain, one obtains weak-* compactness of minimizing sequences up to unitary conjugation, existence of minimizers, and explicit Euler–Lagrange equations. Under suitable regularity assumptions on the first variation, there exist real parameters $\mathrm{do}(A=a)$5 such that for every $\mathrm{do}(A=a)$6,

$\mathrm{do}(A=a)$7

and

$\mathrm{do}(A=a)$8

with $\mathrm{do}(A=a)$9. Equivalently, the support of $\mathcal{G}$0 lies where the positivity inequality is saturated (Finster et al., 2022).

In Minkowski space, the causal action has also been analyzed directly at the level of the Lagrangian and of surface layer integrals. In that setting, with Dirac wave functions interacting with classical electromagnetism and linearized gravity, the continuum-limit analysis shows that the same variational structure encodes Maxwell equations, Einstein equations, and conserved surface layer integrals that reproduce the standard symplectic form and inner products of bosonic and fermionic field theory (Finster, 2017). The linearized field equations have been formulated as

$\mathcal{G}$1

and solved systematically in Minkowski space. The resulting theory exhibits retarded solutions, a multitude of homogeneous solutions associated with direction-dependent local phase freedom, and non-propagating perturbations (Finster, 2023).

The thesis literature develops the same dynamics at a more abstract level. Linearized solutions define conserved surface layer integrals, yielding a symplectic form and a Hamiltonian time evolution when the causal fermion system admits a notion of time. The same formalism also produces stochastic and non-linear correction terms; their appearance is described as reminiscent of dynamical collapse models in quantum theory (Kleiner, 2020).

5. Gradient-like evolution and numerical minimizers

A different development treats the causal action principle and related causal variational principles as non-convex energies on spaces of measures and studies their relaxation by action-driven flows. In the compact setting, with a non-negative continuous Lagrangian $\mathcal{G}$2 on a compact space $\mathcal{G}$3, the action

$\mathcal{G}$4

is combined with a De Giorgi minimizing-movements step and an additional linear penalization: $\mathcal{G}$5 The resulting discrete solutions define Hölder-continuous curves of measures, and the linear penalization guarantees finite length and existence of limit points for the action-reparametrized curves. The limiting measures satisfy approximate Euler–Lagrange equations with controlled $\mathcal{G}$6 error. The same construction is adapted to the finite-dimensional causal action principle using moment measures and a metric on pairs $\mathcal{G}$7 representing operator-valued configurations (Finster et al., 1 Mar 2025).

Numerical work in low-dimensional causal fermion systems studies the causal action principle for weighted counting measures

$\mathcal{G}$8

with action

$\mathcal{G}$9

For $$a^\ast \in \arg\max_{a \in \mathcal{A}} \mathbb{E}[u(Y)\mid \mathrm{do}(A=a),\mathcal{G}] = \arg\max_{a \in \mathcal{A}} \sum_y u(y)\,P(Y=y \mid \mathrm{do}(A=a),\mathcal{G}).$$0, the causal relations admit a geometric description in terms of causal cones, and the Lagrangian can be written as

$$a^\ast \in \arg\max_{a \in \mathcal{A}} \mathbb{E}[u(Y)\mid \mathrm{do}(A=a),\mathcal{G}] = \arg\max_{a \in \mathcal{A}} \sum_y u(y)\,P(Y=y \mid \mathrm{do}(A=a),\mathcal{G}).$$1

For large $$a^\ast \in \arg\max_{a \in \mathcal{A}} \mathbb{E}[u(Y)\mid \mathrm{do}(A=a),\mathcal{G}] = \arg\max_{a \in \mathcal{A}} \sum_y u(y)\,P(Y=y \mid \mathrm{do}(A=a),\mathcal{G}).$$2, the numerically observed minimizers in the cases $$a^\ast \in \arg\max_{a \in \mathcal{A}} \mathbb{E}[u(Y)\mid \mathrm{do}(A=a),\mathcal{G}] = \arg\max_{a \in \mathcal{A}} \sum_y u(y)\,P(Y=y \mid \mathrm{do}(A=a),\mathcal{G}).$$3 and $$a^\ast \in \arg\max_{a \in \mathcal{A}} \mathbb{E}[u(Y)\mid \mathrm{do}(A=a),\mathcal{G}] = \arg\max_{a \in \mathcal{A}} \sum_y u(y)\,P(Y=y \mid \mathrm{do}(A=a),\mathcal{G}).$$4 are well approximated by discrete Dirac spheres, i.e. configurations in which distinct support points are spacelike separated and only self-interaction terms remain. In the case $$a^\ast \in \arg\max_{a \in \mathcal{A}} \mathbb{E}[u(Y)\mid \mathrm{do}(A=a),\mathcal{G}] = \arg\max_{a \in \mathcal{A}} \sum_y u(y)\,P(Y=y \mid \mathrm{do}(A=a),\mathcal{G}).$$5, the asymptotic action is

$$a^\ast \in \arg\max_{a \in \mathcal{A}} \mathbb{E}[u(Y)\mid \mathrm{do}(A=a),\mathcal{G}] = \arg\max_{a \in \mathcal{A}} \sum_y u(y)\,P(Y=y \mid \mathrm{do}(A=a),\mathcal{G}).$$6

while in the case $$a^\ast \in \arg\max_{a \in \mathcal{A}} \mathbb{E}[u(Y)\mid \mathrm{do}(A=a),\mathcal{G}] = \arg\max_{a \in \mathcal{A}} \sum_y u(y)\,P(Y=y \mid \mathrm{do}(A=a),\mathcal{G}).$$7, the action behaves as

$$a^\ast \in \arg\max_{a \in \mathcal{A}} \mathbb{E}[u(Y)\mid \mathrm{do}(A=a),\mathcal{G}] = \arg\max_{a \in \mathcal{A}} \sum_y u(y)\,P(Y=y \mid \mathrm{do}(A=a),\mathcal{G}).$$8

These computations were carried out using differentiable programming and automatic differentiation, with the parameterization chosen so that positivity, trace, and normalization constraints are built in from the outset (Finster et al., 2022).

6. Scope, adjacent formulations, and open problems

Across these literatures, the phrase causal action principle does not designate a single formalism but a family of constructions in which causality and action are linked at the level of the primitive variational or decision rule. In decision theory, the operative distinction is between observation and intervention. In causal fermion systems, the operative distinction is between background structures and structures emergent from minimizing a measure-theoretic action.

Several open problems are explicit. In the decision-theoretic line, the core rule is conceptually grounded but lacks regret bounds and formal convergence theorems; the asymptotic comparison with standard reinforcement-learning procedures is conjectural rather than proven (Gonzalez-Soto et al., 2019). In the causal-fermion-systems line, the compact momentum-domain theory provides existence and Euler–Lagrange equations, but extending from compact $$a^\ast \in \arg\max_{a \in \mathcal{A}} \mathbb{E}[u(Y)\mid \mathrm{do}(A=a),\mathcal{G}] = \arg\max_{a \in \mathcal{A}} \sum_y u(y)\,P(Y=y \mid \mathrm{do}(A=a),\mathcal{G}).$$9 to $Y\in\{0,1\}$0, controlling ultraviolet divergences, and recovering the unregularized Dirac sea remain open (Finster et al., 2022). The continuum-limit field equations and conserved structures are known in several regimes, but quantitative dependence on the unknown ultraviolet regularization remains substantial (Finster, 2017).

The same theme appears in more recent dynamical work. Action-driven flows show that non-convexity can obstruct convergence of naive gradient-flow analogues; a linear penalization cures this at the price of producing approximate rather than exact Euler–Lagrange solutions (Finster et al., 1 Mar 2025). Numerical work confirms candidate minimizing geometries in low-dimensional settings but leaves higher-dimensional minimizers and boundedness-constrained limits unresolved (Finster et al., 2022).

A broader conceptual motif also appears in adjacent work that is not itself formulated as a causal action principle. “Phenomenological causality” defines causal structure from elementary actions by requiring that each elementary action change only one causal mechanism in a Markov factorization (Janzing et al., 2022). Rule-based causal modeling with abductive logic programs likewise imports Pearl-style interventions by modifying rules and facts so that stable models represent post-intervention worlds (Rückschloß et al., 7 Jul 2025). These developments do not use the same variational machinery, but they reinforce a common pattern: causal structure is characterized operationally by how admissible actions alter mechanisms.

In this sense, the term marks two major technical traditions. One treats rational action as intervention-aware expected-utility maximization. The other treats spacetime, geometry, and fields as consequences of minimizing a causal spectral action. Both are explicitly causal, both assign a privileged role to actions or interventions, and both remain active sites of arXiv-level research (Soto et al., 2019, Finster et al., 2011).