Unitary Causal Decompositions

• Unitary causal decompositions are the quantum counterpart of classical complete common cause explanations, achieved through Choi operator factorization and vanishing quantum conditional mutual information.
• The structural theorem equivalently characterizes factorization, no-influence in unitary dilations, and a direct-sum tensor product decomposition of the input Hilbert space.
• These decompositions underpin quantum causal models by clarifying local causal relationships and informing applications from semicausal circuits to quantum gravity.

Unitary causal decompositions are the quantum counterpart of the classical idea that correlations explained by a complete common cause should factorize. In the formulation introduced for quantum channels, the central question is when an input system $A$ can be regarded as the complete common cause of outputs $B$ and $C$. Under the assumption that quantum dynamics is fundamentally unitary, the answer is expressed as a structural theorem: compatibility with $A$ as a complete common cause is equivalent to a factorization of the channel’s Choi operator, to vanishing quantum conditional mutual information on a normalized Choi state, and to a direct-sum-of-tensor-products decomposition of $H_A$ (Allen et al., 2016).

1. Classical antecedents and the quantum motivation

The starting point is Reichenbach’s principle. In its qualitative form, if two variables $Y$ and $Z$ are statistically dependent, then there should be a causal explanation: $Y$ causes $Z$, $Z$ causes $B$0, or a common cause $B$1 causes both. In contrapositive form, ancestral independence implies statistical independence,

$B$2

Its quantitative form says that if the correlation between $B$3 and $B$4 is explained purely by a common cause, and $B$5 is a complete common cause, then

$B$6

Unitary causal decompositions arise from asking for the quantum analogue of this factorization requirement (Allen et al., 2016).

The classical principle is challenged by Bell correlations. In a Bell experiment, one typically seeks to explain correlations between distant measurement outcomes by a common cause in their joint past together with local influences of settings. Bell’s theorem shows that, within classical causal models, observed violations of Bell inequalities cannot be explained without fine-tuning. This undermines the classical Reichenbach-style common-cause explanation for quantum correlations and motivates a quantum generalization (Allen et al., 2016).

Later process-theoretic work sharpened the surrounding landscape. Oreshkov and Giarmatzi distinguished causality from causal separability, exhibited tripartite quantum processes that are causal but not causally separable, and introduced extensibly causal and extensibly causally separable processes to address activation of non-causality by entangled ancillas (Oreshkov et al., 2015). That broader context clarifies why unitary causal decompositions are not merely a translation of classical conditional independence into operator language, but part of a reworking of causal explanation for intrinsically quantum processes.

2. Channels, Choi operators, and complete common causes

For a quantum channel from $B$7 to $B$8, assuming $B$9 is initially uncorrelated with its environment, the CPTP map is represented by a basis-independent Choi operator

$C$0

acting on $C$1, with

$C$2

Composition with an input state is implemented by the linking operator

$C$3

so that

$C$4

This is the quantum analogue of summing over a hidden common cause in the classical formula $C$5 (Allen et al., 2016).

The notion of causal influence is defined first at the level of unitary dilations. For a unitary channel from inputs $C$6 to outputs $C$7, represented by $C$8, one says that $C$9 has no causal influence on $A$0 iff

$A$1

Equivalently, the marginal at $A$2 is independent of any operation on $A$3 prior to the unitary (Allen et al., 2016).

A channel $A$4 is compatible with “$A$5 is a complete common cause of $A$6 and $A$7” if there exist ancillas $A$8 prepared in product state $A$9, together with a unitary dilation $H_A$0 on $H_A$1 to outputs $H_A$2, such that $H_A$3 has no causal influence on $H_A$4, $H_A$5 has no causal influence on $H_A$6, and the original channel is recovered by tracing out $H_A$7 and linking in the ancilla states: $H_A$8 The role of the ancillas is strictly local: they may influence one output each, but not both (Allen et al., 2016).

3. Equivalent characterizations of the unitary quantum Reichenbach principle

The main theorem states that, for a channel $H_A$9, four conditions are equivalent. The equivalence is the core content of unitary causal decomposition in the original sense (Allen et al., 2016).

Characterization	Content
Quantum compatibility	A unitary dilation exists with ancillas $Y$0 that can only locally influence $Y$1 and $Y$2
Factorization	$Y$3 and $Y$4
QCMI condition	$Y$5 on $Y$6
Hilbert-space structure	$Y$7 and $Y$8

The factorization condition,

$Y$9

is the channel-level quantum analogue of $Z$0. Because the factors commute, the product is an ordinary operator product. The paper explicitly notes that no separate star-product operation is required in this convention (Allen et al., 2016).

The QCMI formulation uses the trace-one operator

$Z$1

Then

$Z$2

is equivalent to the logarithmic identity

$Z$3

and therefore to equality in strong subadditivity. The structural form of equality is supplied by the Hayden–Jozsa theorem, yielding

$Z$4

This is the precise sense in which the input can be split into sectors on which the causal influence toward $Z$5 and $Z$6 separates (Allen et al., 2016).

The theorem extends to any number of outputs. For a channel $Z$7, compatibility with $Z$8 as complete common cause, product factorization with pairwise commutation,

$Z$9

vanishing $Y$0 for each $Y$1, and a direct-sum-of-tensor-products decomposition

$Y$2

are again equivalent (Allen et al., 2016).

A frequent misconception is that such a factorization amounts to unrestricted broadcasting of quantum information. The theorem says something more restrictive: the allowed channels are exactly those for which the input Hilbert space decomposes into sectors that further tensor-factorize, so that local CPTP maps act separately on the relevant factors. This is consistent with the no-broadcasting perspective emphasized in the original analysis (Allen et al., 2016).

4. Quantum causal models and the quantum Markov condition

The factorization theorem is embedded in a formalism of quantum causal models. For a DAG with nodes $Y$3, each node is assigned an input Hilbert space $Y$4 and its dual output $Y$5. Each node carries a local channel $Y$6, acting on $Y$7, and these local channels are required to commute pairwise: $Y$8 The global operator is then

$Y$9

This operator is the quantum Markov state for the DAG (Allen et al., 2016).

Operational predictions are extracted by interventions. A quantum instrument $Z$0 at node $Z$1 is represented by

$Z$2

and the joint outcome probabilities are

$Z$3

At a node with a complete-common-cause role, unitary causal decomposition becomes the local causal Markov condition: if $Z$4 is the complete common cause of $Z$5 and $Z$6, then

$Z$7

Thus the quantum Reichenbach theorem is a local structural axiom inside the DAG semantics (Allen et al., 2016).

Barrett, Lorenz, and Oreshkov later showed that any unitary quantum circuit has a causal structure corresponding to a directed acyclic graph, that marginalizing over local noise sources yields a process satisfying a Markov condition with respect to that graph, and that there is a converse. They also introduced an intrinsically quantum notion analogous to conditional independence, proved a quantum d-separation theorem, and formulated quantum analogues of the three rules of do-calculus (Barrett et al., 2019). This places unitary causal decompositions within a broader program in which causal structure is defined directly at the level of quantum processes rather than only at the level of classical outcomes.

5. Examples, nonexamples, and operational interpretation

Three examples in the original work delineate the scope of the theorem. A bipartite unitary $Z$8 has a trivial dilation, so $Z$9 is the complete common cause of $Z$0 and $Z$1. Accordingly,

$Z$2

on the normalized Choi operator (Allen et al., 2016).

The qubit copying examples are more diagnostic.

Example	Choi structure	Result
Incoherent copy	$Z$3 on $Z$4	Factorizes, $Z$5
Coherent copy	Proportional to $Z$6	Does not factorize, $Z$7
Bipartite unitary $Z$8	Trivial dilation	$Z$9 is complete common cause

The incoherent copy measures $B$00 in the computational basis and prepares $B$01 or $B$02. Its Choi operator factorizes, so $B$03 is a complete common cause of $B$04 and $B$05. By contrast, the coherent copy

$B$06

has a Choi operator proportional to a GHZ projector; on the normalized Choi state,

$B$07

Hence it does not factorize. The paper interprets this through the dilation: a CNOT-style implementation lets the target back-act on the control, so an ancilla can function as an additional common cause. The failure is therefore not a mere technicality of representation, but a statement about causal completeness (Allen et al., 2016).

The Bell case gives the broader causal significance. Classical causal models require

$B$08

for a complete common cause $B$09. Quantum causal models do not demand classical joint distributions across descendant variables. Instead, common-cause structure is encoded by Choi-operator factorization and vanishing QCMI. Channels producing Bell-type correlations across measurement outcomes generally do not satisfy

$B$10

with a single $B$11 as complete common cause of the underlying quantum systems. This is the intended quantum analogue of the failure of classical common-cause explanation in Bell scenarios (Allen et al., 2016).

A constructive reading of the theorem is also available. Given a unitary dilation $B$12 and the hypothesis that $B$13 is the complete common cause of $B$14 and $B$15, one can test no-influence constraints on the dilation, compute $B$16 on the normalized Choi operator, extract the Hayden–Jozsa decomposition of $B$17, and, if needed, build a corresponding dilation from block-diagonal unitaries. This turns the theorem into a practical structural criterion rather than a purely existential one (Allen et al., 2016).

6. Compositional, cyclic, and lattice-theoretic generalizations

Subsequent work broadened the notion of causal decomposition from common-cause factorization of channels to the relation between causal structure and compositional structure of unitary transformations. In cyclic quantum causal models, a unitary process still factorizes into commuting local channels for its causal structure. Within that framework, all unitarily extendible bipartite processes are causally separable, and for unitary processes causal nonseparability is equivalent to cyclicity of the causal structure (Barrett et al., 2020).

Lorenz and Barrett showed that ordinary circuit decompositions need not make all no-influence constraints simultaneously evident. To remedy this, they introduced extended circuit diagrams, whose key additional compositional operation is direct sum alongside sequential and tensor-product composition. For large classes of finite-dimensional unitaries they derived causally faithful extended circuit decompositions, and they formulated the hypothesis that every finite-dimensional unitary transformation has such a decomposition (Lorenz et al., 2020).

A later lattice-theoretic result gave a sharp criterion for when traditional unitary circuits suffice. For a set $B$18 of no-influence constraints, every unitary transformation satisfying $B$19 has a unitary causal decomposition in the traditional circuit formalism iff $B$20 satisfies the $B$21-exclusion property; equivalently, in the canonical concept lattice $B$22, there is no more than one path between each input and output (Lugt et al., 15 Aug 2025). In one dimension, an analogous constructive characterization was obtained for locality-preserving dynamics: for $B$23, a unitary channel is a 1D QCA of radius $B$24 iff it can be decomposed into a unitary routed circuit of nearest-neighbour interactions, and this decomposition can be chosen translation-invariant in the translation-invariant case (Vanrietvelde et al., 27 Jun 2025).

Subsystem-decomposition methods supplied a different generalization. A framework based on changes of tensor-product structure showed that some cyclic circuits can be mapped to temporal circuits on time-delocalised subsystems while preserving composed probabilities. However, in the quantum switch there is no single isomorphism

$B$25

that relates Alice’s and Bob’s temporal decompositions for all local operations. The proof uses invariants of unitary similarity and, for a particular choice of operations, the mismatch

$B$26

to rule out a universal subsystem change $B$27 (Wechs et al., 2024). A plausible implication is that “unitary causal decomposition” now names a family of closely related structural programs rather than a single theorem.

7. Related uses in semicausal circuits and graph-based dynamics

The decomposition idea also appears in explicitly semicausal settings. In a circuit model of black hole evaporation, semicausality means $B$28: degrees of freedom inside the horizon cannot signal to degrees of freedom outside. The structural theorem there states that a tripartite unitary on $B$29 is semicausal iff it admits

$B$30

or, with local pre- and post-unitaries,

$B$31

Iterating this layered semicausal decomposition yields entropy bounds

$B$32

so that any entropy curve lies within a Page-like triangular envelope (Broda, 2023).

In quantum causal graph dynamics, the object is a vertex-preserving causal unitary on a Hilbert space spanned by finite labelled graphs. The main structure theorem states that any such unitary admits a finite-depth decomposition into local unitaries: $B$33 where each $B$34 is 1-local, each $B$35 is localized on $B$36, and

$B$37

for all $B$38. The depth is $B$39, and the result extends block-representation ideas from QCA to time-varying graphs and superpositions of graphs (Arrighi et al., 2016).

A more background-independent variant appears in discrete unitary causal theories of quantum gravity. There, global evolution is decomposed into local unitary replacement rules acting on non-overlapping initial and final evolution regions, and the physical inner product is a sum over histories with amplitudes given by products of local transition amplitudes. The resulting inner product is Hermitian and fully gauge-degenerate under spacetime diffeomorphisms generated by local evolution moves (Wall, 2012). These constructions are not the same theorem as the unitary quantum Reichenbach principle, but they preserve its central theme: causal constraints become explicit through a decomposition of unitary dynamics into structurally local pieces.