Causally Agnostic Measurement Frameworks

• Causally agnostic measurement is a framework that defines measurement methods without presupposing a fixed causal structure, enabling flexible analysis across quantum, relativistic, and inferential contexts.
• In acausal MBQC, fixed measurement bases combined with process matrices absorb byproduct effects, yielding perfect output probabilities (P₀=1) compared to standard adaptive protocols.
• Relativistic and proxy measurement models show that causal order must be imposed post hoc through localized apparatus couplings or latent variable techniques, highlighting limits of strong agnosticism.

Causally agnostic measurement denotes a family of research programs in which measurement is specified, analyzed, or certified without presupposing a single fixed causal dependence structure for settings, outcomes, or measured variables. The phrase is not univocal. In measurement-based quantum computing, it names a concrete nonadaptive measurement scheme realized only in the process-matrix framework (Morimae, 2014). In relativistic quantum field theory, by contrast, the dominant conclusion is negative: measurement is not causally agnostic in any strong sense, and causality must be imposed a posteriori unless the apparatus and its localized coupling are built into the formalism (Fuksa, 2021). In causal inference and causal representation learning, the same theme reappears as the problem of treating observations or learned representations as proxy measurements of latent causal variables, and of deciding when such measurements are adequate for downstream causal tasks (Yao et al., 23 May 2025).

1. Scope and principal meanings

The expression does not pick out a single formalism. It instead collects several technically distinct ideas that share one structural feature: the measurement description is not fixed by a single assumed causal story from the outset.

In quantum information, the strongest literal use appears in acausal measurement-based quantum computing. There, “acausal” means that local measurement bases are fixed in advance and do not depend on earlier outcomes, even though standard MBQC ordinarily requires adaptive feed-forward along a causal cone because of byproduct operators (Morimae, 2014). In relativistic measurement theory, the phrase is used mainly to deny such a possibility: the operator formalism by itself does not determine which observables are relativistically measurable, and causality enters as an admissibility condition unless one models a localized system–probe interaction (Fuksa, 2021).

A different use arises in device-independent and causal-model analyses of measurement statistics. There, one does not trust a microscopic Hilbert-space realization of the measuring device; instead one asks whether the observed statistics are compatible with a classical causal model, or whether the event structure itself must be treated as channel-relative rather than absolute (Lee, 2018). In monitored quantum dynamics, still another variant appears: effective causal structure is inferred from intervention-to-outcome influence rather than assumed to coincide with the ordinary forward light cone, and the resulting influence pattern can depend on initial and final states and on post-selection (Wang et al., 18 Jan 2026).

In causal inference, measurement error, and causal representation learning, causally agnostic measurement usually means that the variables available to analysis are not taken to be the causal variables themselves. Observed variables, anomaly readouts, or learned representation blocks are treated as measurements or proxies of latent causal variables, and the central question becomes when these measurements preserve enough structure for causal discovery or causal estimation (Dai et al., 2022). This suggests that the phrase is best understood as a cross-disciplinary umbrella for methods that relax, relocate, or explicitly parameterize causal commitments in the measurement description, rather than as the name of a single theory.

2. Acausal measurement-based quantum computing

Morimae’s construction starts from the standard MBQC setting: a resource state $\sigma$, usually a graph state, is divided into a computation subsystem $C$ of $N$ qubits and an output subsystem $O$ of $n$ qubits. In ordinary MBQC, measurements on $C$ are adaptive. The basis chosen for a later qubit depends on earlier outcomes because the randomness of measurement induces byproduct operators that must be compensated. This adaptive dependence is the operational content of the causal cone. Morimae also uses a no-signaling argument from earlier work: if byproduct operators could be completely avoided within ordinary no-signaling physics, then the party holding $C$ could deterministically prepare arbitrary output states on $O$, enabling signaling if $O$ were distant. On that basis, byproducts are not optional nuisances but a structural reason why standard MBQC is causally ordered (Morimae, 2014).

Against that background, acausal MBQC is defined operationally as MBQC in which each local basis is fixed in advance and does not depend on any other outcome. In the distributed version there are $N$ Alices holding the qubits of $C$0 and $C$1 Bobs holding the qubits of $C$2. Each Alice$C$3 measures in a fixed basis $C$4, independently of all other outcomes. The theory enabling this is not ordinary no-signaling quantum mechanics but the process-matrix formalism of Oreshkov, Costa, and Brukner.

The process-matrix framework keeps standard quantum theory inside each local laboratory while dropping global causal and no-signaling constraints. A local event is represented by a CP, trace-non-increasing map, encoded by a Choi–Jamiołkowski operator, and joint probabilities are computed as

$C$5

A valid process matrix must satisfy positivity and normalization conditions. Morimae uses the multipartite extension of this structure to build a resource process matrix for acausal MBQC.

The construction begins with the graph $C$6 underlying a standard causal MBQC. A decorated graph $C$7 is obtained by adding one degree-1 vertex to each vertex of the computation subsystem $C$8. If $C$9 is the associated graph state, then the resource process matrix is

$N$0

Equivalently,

$N$1

where each $N$2 is a controlled-$N$3 gate between an original computation qubit and its added decoration qubit. This is the precise content of the statement that, for projective local measurements, the resource process matrix is equivalent, up to normalization and trivial ancillas, to a decorated graph state.

The computational effect is that the process matrix absorbs the byproduct dependence. Using the CJ operators for the fixed local projective measurements, Morimae shows that the joint probability of outcomes $N$4 is exactly the same as in the positive branch of ordinary causal MBQC, namely the branch where the adaptive corrections behave as if every relevant measurement outcome were $N$5. Thus the local settings are nonadaptive, but the output statistics coincide with those of the desired causally ordered MBQC branch.

This construction is explicitly beyond ordinary no-signaling quantum theory. Morimae emphasizes that the resulting theory is signaling in the usual sense: because the process matrix effectively chooses the correct branch, the Alices can prepare the correct output state in $N$6 without byproduct uncertainty, so information can be encoded into the output and recovered by the Bobs. The same point appears in the causal-inequality discussion. If $N$7 is the probability that all Bobs obtain $N$8, then in causally ordered MBQC

$N$9

whereas in the acausal MBQC constructed from the process matrix

$O$0

The induced correlations therefore violate the relevant causal inequality.

Morimae also notes a restricted simulation inside ordinary quantum theory: acausal MBQC can be simulated by causal MBQC with postselection, specifically by postselecting the red decoration qubits in the decorated-graph construction. This suggests a small-scale simulation, but only at exponentially small success probability. Accordingly, the construction is best understood as a generalized operational resource rather than an ordinary MBQC protocol available inside standard no-signaling quantum mechanics.

3. Relativistic quantum measurement and the rejection of strong causal agnosticism

A directly opposed conclusion is developed in relativistic quantum measurement theory. The central claim there is that quantum field theory is not causally agnostic in any strong operational sense, because the ordinary Hilbert-space and operator formalisms do not by themselves encode relativistic causal admissibility. A self-adjoint operator may exist in the algebra and still fail to be relativistically measurable. The paper formulates Einstein causality in AQFT region language: if $O$1 and $O$2 are causally disjoint, then

$O$3

but emphasizes that this microcausal kinematics is not by itself a complete measurement theory (Fuksa, 2021).

The key obstruction is operational. In Sorkin-type scenarios, Alice and Charlie can be spacelike separated while Bob lies in a region between them, and inserting Bob’s measurement can turn otherwise innocent projection-postulate prescriptions into superluminal signaling schemes. The paper therefore distinguishes the algebraic existence of an observable from its operational relativistic measurability. In the operator language, in decoherent histories, and even in path-integral formulations, causality must be checked or imposed afterward.

The most constructive response comes from the Fewster–Verch apparatus-inclusive framework. There the system and probe are both quantum fields, coupled only in a compact spacetime region $O$4. The interaction action is

$O$5

The coupling locality is the essential input. With uncoupled system–probe algebra $O$6, coupled algebra $O$7, and retarded/advanced response maps, one defines the scattering morphism

$O$8

Given a probe state $O$9, the induced system observable associated with a probe observable $n$0 is

$n$1

This framework yields explicit causal control. If a probe observable $n$2 is localized in $n$3, then

$n$4

so a readout entirely outside the causal influence of the coupling carries no information about the system. More generally, the induced observable can be localized in any connected open causally convex set containing $n$5. The resulting pre-instruments and post-selected state updates replace the naive projection postulate with localized measurement dynamics.

The general conclusion is therefore restrictive. Relativistic measurement is not causally agnostic in the sense of allowing arbitrary candidate measurements first and recovering causality later from the bare formalism. The operator formalism, AQFT, and path integrals all require causality to be imposed a posteriori unless one moves to an apparatus-inclusive theory with localized couplings. A plausible implication is that strong causal agnosticism is hardest to sustain precisely where spacetime localization is physically non-negotiable.

4. Device-independent, relational, and monitored-dynamics variants

A separate quantum-information strand removes trust in device internals rather than in causal structure itself. In device-independent certification of joint measurements, the question is whether the full observed statistics of a joint measurement are compatible with a classical causal model. For preparations $n$6, local outputs $n$7, Charlie’s two-bit measurement outcome $n$8, and independent latent variables $n$9, the classical model factorizes as

$C$0

Non-classical measurement is then defined as incompatibility with this causal decomposition. The corresponding Bell-like witness is

$C$1

and a noisy Bell measurement violates it whenever $C$2. The same paper notes that each fixed post-selected branch can still admit a classical model for $C$3, so the full-measurement criterion is finer-grained than post-selection-based certification (Lee, 2018).

Device-independent quantum cryptography sharpens the limits of this strategy. A general DIQKD security proof with untrusted devices still assumes that raw-key elements are generated by causally independent measurement processes. Formally, the model requires the usual inter-party commutation

$C$4

and, in addition, cross-round commutation inside each lab,

$C$5

The asymptotic key-rate bound is

$C$6

Thus device independence is not full causal agnosticism: internal microscopic implementation is untrusted, but cross-round causal dependence is still excluded (Masanes et al., 2010).

Extended Wigner’s Friend scenarios move the point of flexibility elsewhere. The augmented-circuit framework for EWFSs introduces a setting $C$7 for each putative measurement, where $C$8 means unitary evolution only and $C$9 means a unitary followed by projectors identifying classical records. All conventional EWFS predictions become setting-conditioned probabilities inside one enlarged circuit. The paper’s diagnosis of Frauchiger–Renner contradictions is that they require an extra assumption of setting-independence for predictions that are not actually setting-independent. The framework therefore preserves a single DAG-level causal structure while allowing events to be fundamentally subjective rather than absolute (Vilasini et al., 2022).

Monitored quantum dynamics supplies yet another operational version. There the goal is to measure influence directly from outcome statistics in systems with measurements and post-selection, where effective causal structure can differ from the usual forward light cone. The basic quantity is the cross-entropy quantum causal influence

$C$0

and, in trajectory-resolved form,

$C$1

Its simulations reveal inverted light cones, measurement-reflected influence, and an arrow of time controlled by initial and final entropy. In analytically tractable models the asymmetry is proportional to $C$2, and in hybrid Clifford circuits the resulting correlation time obeys a scaling collapse at the measurement-induced phase transition. Here causal structure is inferred from measured influence rather than imposed geometrically in advance (Wang et al., 18 Jan 2026).

5. Measurement as proxy: latent variables, measurement error, and representation quality

Outside foundational quantum theory, causally agnostic measurement commonly means that the available variables are not taken to be the causal variables themselves. One studies how much causal structure can be recovered when only measurements or proxies of latent causes are observed.

In causal discovery under measurement error, the latent variables $C$3 obey a linear non-Gaussian acyclic model, while the observed variables satisfy

$C$4

with mutually independent measurement errors independent of $C$5. Ordinary conditional-independence structure is not preserved by this observation map, but the paper introduces the Transformed Independent Noise condition. For sets $C$6 and $C$7,

$C$8

For disjoint observed sets $C$9, the key invariance is

$O$0

This enables recovery of the ordered group decomposition of the latent DAG using only independence tests on noisy measurements. The construction is agnostic to the latent values themselves and avoids over-complete ICA, but it is not agnostic to structure: it relies on linearity, non-Gaussianity, acyclicity, and additive independent measurement error (Dai et al., 2022).

A closely related distinction appears in causal anomaly analysis, where each feature has a latent true variable $O$1 and an observed readout $O$2. In the clean regime,

$O$3

Mechanistic anomalies act on the latent structural assignment,

$O$4

whereas measurement anomalies act on the observation channel,

$O$5

The causal difference is explicit: a measurement anomaly cuts the edge $O$6 and does not propagate downstream, while a mechanistic anomaly perturbs descendants. The paper proves structural identifiability only up to observational equivalence and notes a built-in ambiguity for sink nodes, where mechanistic and measurement interventions can induce the same observable conditional-independence structure (Suhr et al., 30 Jan 2026).

Causal representation learning reframes the problem in measurement-model terms. There the latent causal variables are $O$7, learned representation blocks are $O$8, and a measurement model is

$O$9

The paper’s central causal notion is task-relative validity: a measurement model is causally valid for a statistical estimand $O$0 when

$O$1

That criterion is weaker than exact latent recovery and depends on causal role. Invertible reparameterizations can preserve causal validity for confounders and instruments, but not generally for latent treatments or outcomes. To assess measurement structure nonparametrically, the paper tests exclusivity via conditional independence and defines the Test-based Measurement EXclusivity score

$O$2

The score is agnostic to the exact measurement functions $O$3, but not to the hypothesized parentage matrix $O$4 or to the validity of the conditional-independence tests (Yao et al., 23 May 2025).

A broader methodological analogue appears in work on psychosocial constructs. There the claim is that “all measures are formative in that they are formed from observed indicators; all measures are reflective in that they are reflective of a more complex underlying reality.” The proposed multiple-versions-of-treatment interpretation allows associations involving a constructed measure $O$5 to retain a causal interpretation even when strong univariate reflective or formative assumptions fail. This is not full causal agnosticism, but it is an explicit rejection of overcommitting to a single latent-variable measurement ontology (VanderWeele, 2020).

6. Debates, limits, and competing interpretations

The literature is sharply divided on whether causally agnostic measurement is a satisfactory endpoint, a useful operational stance, or an impossibility claim in disguise.

One line rejects causal agnosticism as a solution to the measurement problem. De Ronde argues that the projection postulate “does not provide a solution to the measurement problem by itself,” and that a genuine solution “requires an explanation which relates in a causal model a superposition of multiple states and the one measurement outcome.” Instrumentalism is treated there as the main route that avoids commitment to causal modelling, but is criticized as insufficient if one wants realist understanding. The paper accordingly distinguishes efficient-cause, final-cause, and immanent-cause models, and proposes the last rather than endorsing causal abstention (Ronde, 2013).

Another line permits a stronger empirical agnosticism while keeping an underlying physical principle. Gillis argues that signal causality, not local causality, is the common principle linking relativity and quantum measurement. On that view, nonlocal measurement effects are real, but the sequencing of those nonlocal effects is in principle unobservable because the relevant information cannot be instantiated at the most elementary scales. The paper therefore supports what it effectively describes as an ontologically ordered but empirically causally agnostic picture of measurement: objective sequencing may exist, yet observable physics remains insensitive to it (Gillis, 2010).

Taken together, these works suggest that “causally agnostic measurement” names a spectrum rather than a doctrine. At one end are generalized operational models such as acausal MBQC, where causal dependence is displaced into a process matrix and the price is signaling or postselected simulation. At another are relativistic and apparatus-localized approaches, which deny that measurement can be causally agnostic once spacetime admissibility is taken seriously. Between them lie proxy-based and representation-based frameworks, where the aim is not to eliminate causal structure but to weaken the claim that observed variables or measurement events must coincide with a unique, absolute causal narrative.

The common technical lesson is negative and positive at once. Strong causal agnosticism is rarely free: it typically requires a generalized process theory, an explicit locality constraint, a nontrivial latent-variable model, or a channel-relative semantics for events. Yet these same constraints show why the topic persists. Measurement is often used precisely where causal order, causal role, or causal identity is least transparent, and the most successful formalisms are those that make that ambiguity explicit instead of concealing it.