Operationally Detectable Causal Loops

• Operationally detectable causal loops are loop structures whose presence is verified through explicit operational procedures rather than mere graph inspection.
• Methodologies include simulation-based metrics, conditional-independence tests, and intervention-sensitive affects relations to quantify loop contributions.
• These approaches form a layered framework where detectability depends on model assumptions, data type, and even spacetime geometry constraints.

Operationally detectable causal loops are loop structures whose presence, absence, contribution, or coherence can be established by a specified operational procedure rather than by graph inspection alone. Across the cited literature, the phrase does not denote a single unified criterion. In system dynamics, it refers to loops in a known formal model whose contributions to simulated behavior can be computed and visualized over time (1909.01138). In constraint-based causal discovery, it refers more narrowly to what conditional-independence information can certify about pairwise membership in a common strongly connected component, with the central result being detectability of the absence of specific cycles rather than general positive cycle identification (Mooij et al., 2020). In intervention-based frameworks for cyclic and fine-tuned causal models, operational detectability is defined through intervention-sensitive affects relations, including higher-order affects, and may or may not coincide with relativistic no-signalling constraints (Vilasini et al., 2021). Other works broaden the topic further by distinguishing mixed causal–constitutive closure from ordinary causal cycles (Ohmura et al., 19 Jun 2026), by treating coherence between alternative causal orders rather than literal loops (Siddiqui et al., 29 Mar 2026), or by analyzing loop-like operators in algebraic quantum field theory as closed paths in a causal poset rather than feedback cycles of influence (Ciolli et al., 2011). The subject is therefore best understood as a family of operational notions for identifying, ruling out, or characterizing loop structure under different assumptions.

1. Conceptual scope and formal meanings

The literature distinguishes several non-equivalent meanings of a “causal loop.” In system dynamics, a loop is a closed causal path in a stock–flow model, and the central question is which loops are actually generating behavior at each point in time (1909.01138). In causal discovery over possibly cyclic structural causal models, loops are represented by directed cycles in a directed mixed graph (DMG), and the key graph-theoretic notion is membership in a strongly connected component (SCC),

$sc_G(i) := an_G(i)\cap de_G(i),$

so that two variables lie in the same SCC exactly when each is an ancestor of the other (Mooij et al., 2020). In interventional operational frameworks, a loop is not identified from graph topology alone, but from patterns of observable intervention-induced influence; the relevant primitive is an affects relation rather than merely an arrow in a graph (Vilasini et al., 2021).

This plurality of meanings matters because operational detectability depends on what data and assumptions are available. A loop may be detectable from a known simulation model and a simulation trace, yet not from observational time series alone. Conversely, a loop may exist in an underlying fine-tuned model while remaining operationally invisible because the same observable behavior is compatible with an acyclic model. The distinction between underlying cyclicity and operational detectability is explicit in the literature on hidden causal loops, higher-order affects, and fine-tuned signalling structures (Vilasini et al., 2021).

A further distinction concerns the type of closure under study. "Closure of Self-Determining System Based on Causal and Constitutive Relations" (Ohmura et al., 19 Jun 2026) rejects pure causal cycles as the basis of self-determining closure and instead defines causal–constitutive loops (CC-loops), in which a causal transmission mechanism $F$ is constituted by variables $c_j$ while itself participating in causal production. The minimal model is

$b := F(a)$

$c_0 := G(b)$

$c_1 := H(b)$

$F :\sim \mathcal{B}(c_0, c_1).$

This is not a temporal cycle of the form $X \to Y \to Z \to X$, but a mixed structure combining causal and constitutive asymmetries.

The quantum-causal literature introduces another separation. "Complementarity Beyond Definite Causal Order" (Siddiqui et al., 29 Mar 2026) studies coherent superpositions of two acyclic orders $A\prec B$ and $B\prec A$, not literal loops. The operationally measurable quantity there is causal coherence, not cycle membership. This suggests that operationally detectable nonclassical causal structure need not be cyclic in the ordinary sense.

2. Detection within known dynamic models

In system dynamics, operational detectability is strongest when the model equations, parameterization, stock–flow structure, and simulation trajectory are all known. "LoopX: Visualizing and understanding the origins of dynamic model behavior" (1909.01138) formulates the problem as linking observed simulation behavior to known model structure. Large models may contain many structurally valid feedback loops, but only some are dynamically important at a given time. LoopX therefore operationalizes dynamic importance through the Loops That Matter (LTM) methodology: simulation of the model, extraction of the equation network, computation of link and loop metrics at each $F$0, and ranking and visualization of those metrics.

The formal basis is a hierarchy of scores. For loop $F$1 containing $F$2 links,

$F$3

The normalized dominance measure is the relative loop score,

$F$4

which the paper states lies between $F$5 and $F$6. Its sign gives polarity and its magnitude gives instantaneous fractional contribution relative to the analyzed loop set. A loop is dominant when its relative contribution is greatest; the paper specifically says loops contributing more than $F$7 of the total are dominant (1909.01138).

LoopX also defines a filtering heuristic,

$F$8

computed over the simulation horizon. This is used to retain variables whose proportional contribution changes substantially over time. The operational criterion for detectability is correspondingly model-implied and simulation-conditioned: a loop is detectable if it is present in the known structure, its contribution can be computed from the simulation-plus-equations using LTM, and it attains a nontrivial relative loop score or passes a display threshold (1909.01138).

"Seamlessly Integrating Loops That Matter into Model Development and Analysis" (Schoenberg et al., 2020) embeds this logic in a modeling environment and makes the same point more directly: loops are detected automatically from a running model. The central primitive is the link score,

$F$9

with $c_j$0 if $c_j$1 or $c_j$2. For flow-to-stock links, the paper gives

$c_j$3

Loop score is then the product of link scores around the loop, and relative loop score is normalized so that absolute values sum to $c_j$4. The paper states that the relative loop score “measures the percentage contribution [of] a loop to the changes of all variables in the model at each point in time” (Schoenberg et al., 2020).

The importance of these model-based approaches is their asymmetry with respect to causal discovery from data. They do not infer unknown structure from uncontrolled observations. They compute time-varying loop importance from a known formal model. This yields strong operational detectability within model analysis, including machine-generated causal loop diagrams, simplified CLDs, loop legends, and animation, but only under the premise that the causal structure is already encoded in the model equations (1909.01138).

3. Observational discovery from data: asymmetric identifiability

When only observational conditional-independence information is available, operational detectability becomes much weaker and more asymmetric. "Constraint-Based Causal Discovery using Partial Ancestral Graphs in the presence of Cycles" (Mooij et al., 2020) considers i.i.d. observational samples from the observational distribution of a simple possibly cyclic SCM, with latent confounding allowed and no selection bias. The graph is a DMG, and separation is generalized from $c_j$5-separation to $c_j$6-separation. The central transfer result is that every cyclic DMG has an acyclic surrogate $c_j$7 such that

$c_j$8

This permits standard FCI to remain sound and complete, but only under the paper’s new cyclic semantics for DPAGs (Mooij et al., 2020).

The strongest positive claim is deliberately one-sided. The consistency corollary says that with consistent CI tests and sufficient data, FCI consistently estimates a DPAG from which one obtains consistent estimates of the absence/presence of causal relations, the absence of confounders, the absence/presence of direct causal relations, and the absence of causal cycles. The loop-specific result is pairwise and negative: if two variables $c_j$9 are in the same SCC, then in the complete DPAG they must display a highly symmetric signature, including a bicircle edge and identical incident mark patterns from every third node. Therefore, failure of that signature rules out common SCC membership (Mooij et al., 2020).

The converse does not hold. Proposition $b := F(a)$0, as summarized in the supplied material, shows that if a pair does satisfy the SCC-like DPAG pattern, this does not identify a loop. There can exist both a cyclic DMG in which $b := F(a)$1 are in the same SCC and an acyclic graph in which they are not, yet both induce the same DPAG. The paper therefore does not support general positive detection of the existence of a cycle from observational data alone (Mooij et al., 2020).

This asymmetry recurs elsewhere. The 2021 thesis "Approaches to causality and multi-agent paradoxes in non-classical theories" (Vilasini, 2021) emphasizes that in unfaithful or fine-tuned models, $b := F(a)$2 can hold even when $b := F(a)$3 is a cause of $b := F(a)$4. This strengthens the general lesson: absence of an operational signature often supports only a restricted negative inference, while positive loop identification requires stronger assumptions or richer interventions.

A different observational route appears in nonlinear dynamical time series. "Causal Feedback Discovery using Convergence Cross Mapping from Sea Ice Data" (Nji et al., 13 May 2025) defines a feedback loop operationally as bidirectional convergent cross mapping. With delay vectors

$b := F(a)$5

cross mapping estimates, for example,

$b := F(a)$6

$b := F(a)$7

A feedback loop $b := F(a)$8 is then inferred when cross-map skill $b := F(a)$9 converges in both directions and the paper’s significance criterion $c_0 := G(b)$0 is satisfied. The real-data claim is that feedback is identified between sea_ice_extent and specific humidity, LW_down, SST, and t2m, each with $c_0 := G(b)$1 (Nji et al., 13 May 2025). This operational criterion is much more affirmative than the DPAG result, but it is tied to state-space reconstruction assumptions and does not address latent confounding in the same way.

4. Interventional criteria, higher-order affects, and hidden loops

The most explicit operational definition of detectability via interventions appears in "A general framework for cyclic and fine-tuned causal models and their compatibility with space-time" (Vilasini et al., 2021). The basic detectable relation is affects: $c_0 := G(b)$2 This is not equivalent to ordinary correlation. In fine-tuned models there can be causation without correlation, and even operationally detectable influence may fail to appear in passive observational statistics (Vilasini et al., 2021).

The paper distinguishes affects causal loops (ACLs) from hidden causal loops (HCLs). An ACL is a set of affects relations that can only arise from a cyclic causal structure. A hidden causal loop is a directed cycle in the underlying model whose affects relations and correlations are also realizable in an acyclic structure. Thus not all loops are operationally detectable. Hidden loops are genuinely present but operationally invisible (Vilasini et al., 2021).

The paper identifies several sufficient operational certificates of cyclicity. ACL1 is mutual pairwise affects,

$c_0 := G(b)$3

ACL2 is a chain of single-variable affects closing back on itself. ACL3–ACL6 involve irreducible set-to-set or higher-order affects relations. The theorem stated in the supplied material is that any set of affects relations containing ACL1–ACL6 can only arise from a cyclic causal structure (Vilasini et al., 2021). This is a strong positive detectability result, but it depends on interventions rather than conditional independence alone.

Higher-order affects generalize the criterion: $c_0 := G(b)$4 iff for some $c_0 := G(b)$5,

$c_0 := G(b)$6

These relations matter because pairwise affects can vanish in fine-tuned models even when causal influence is present. Irreducible higher-order affects can still certify elementwise causation, and patterns of such relations can certify cyclicity (Vilasini et al., 2021). The thesis "Compatibility of Cyclic Causal Structures with Spacetime in General Theories with Free Interventions" (Grothus, 2022) extends this operational program, adds indecreasability, potential cause graphs, and loop graphs, and states a complete and constructive criterion: a set of affects relations implies an ACL if and only if the corresponding presence loop graph is not empty.

The same literature also clarifies why detectability can fail. Hidden loops may be fine-tuned so that their entire correlation and affects profile can be reproduced by an acyclic model. This establishes a three-way distinction: loops that are structurally present, loops that are operationally detectable, and loops that are operationally certifiable as necessarily cyclic. The distinction is central to the modern interventional treatment of causal loops (Vilasini et al., 2021).

5. Spacetime compatibility and geometry-dependent impossibility

The relation between operationally detectable loops and relativistic constraints is not fixed; it depends on how one distinguishes signalling from causation and on the geometry of spacetime. "Impossibility of superluminal signalling in Minkowski space-time does not rule out causal loops" (Vilasini et al., 2022) shows that if one forbids only superluminal signalling, rather than superluminal causation, then one can have a causal model with a loop that is operationally detectable by interventions and yet does not allow signalling outside the future light cone. The paper constructs examples on observed binary variables $c_0 := G(b)$7 with common observational relation

$c_0 := G(b)$8

including a fine-tuned cyclic model in which $c_0 := G(b)$9 affects $c_1 := H(b)$0 and $c_1 := H(b)$1 affects $c_1 := H(b)$2, with no pairwise affects relations among $c_1 := H(b)$3. The supplement argument summarized in the supplied material shows that no acyclic causal model can generate this pattern (Vilasini et al., 2022).

This possibility is dimension-sensitive. The same paper states that the construction works in $c_1 := H(b)$4-dimensional Minkowski spacetime because the joint future of two events has a frame-independent earliest point. It also states that the example is not embeddable in $c_1 := H(b)$5-dimensional Minkowski spacetime, leaving the higher-dimensional case open (Vilasini et al., 2022). That open question is then resolved in "Impossibility of superluminal signalling rules out causal loops in conical spacetimes" (Grothus et al., 18 Jun 2026).

The 2026 result proves a geometry-dependent no-go theorem: in conical spacetimes, including ordinary Minkowski spacetime with more than one spatial dimension, all compatible embeddings of affects causal loops are degenerate. The main theorem is summarized as

$c_1 := H(b)$6

Since compatibility is the paper’s formalization of NSS, the operational conclusion is that NSS rules out all non-degenerate operationally detectable causal loops in conical spacetimes (Grothus et al., 18 Jun 2026). The paper also emphasizes that this result is theory-independent: it applies to classical, quantum, and post-quantum causal models because it is framed at the level of affects relations and graph-theoretic causal inference rather than specific dynamics.

This geometry dependence refines the earlier thesis literature. The 2022 thesis (Grothus, 2022) had already argued that $c_1 := H(b)$7-dimensional Minkowski spacetime is exceptional because its light-cone structure forms a lattice and permits equality of certain joint futures. It also introduced stability conditions that rule out ACL1–ACL6a in any spacetime when stable embeddings are demanded. The later conical-spacetime theorem turns that program into an explicit higher-dimensional impossibility result (Grothus et al., 18 Jun 2026).

6. Boundary cases, adjacent frameworks, and unresolved issues

Several adjacent literatures clarify what operational detectability is not. "Closure of Self-Determining System Based on Causal and Constitutive Relations" (Ohmura et al., 19 Jun 2026) argues that closure defined only in terms of causal relations “necessarily entails circular causality” and that cycles formed exclusively by causal relations are excluded from the relevant notion of self-determining closure. Instead, the operationally suggestive structure is a CC-loop,

$c_1 := H(b)$8

with constitutive relations requiring at least two independently manipulable variables. The paper does not supply a detection algorithm, but it does give intervention-based asymmetry for causation, an independent-manipulability requirement for constitutive variables, and boundary criteria based on participation in CC-loops. This suggests a route toward operationalization, but the paper itself remains conceptual rather than algorithmic (Ohmura et al., 19 Jun 2026).

"The Topology of Causality" (Gogioso et al., 2023) likewise does not define loops explicitly, but it provides a framework in which the closest analogue of a loop witness is failure of global causal gluing. It proves that causality is equivalent to continuity in the lowerset topology: $c_1 := H(b)$9 It also proves that the presheaf of causal functions is not always a sheaf, and characterizes sheaf failure by solipsistic contextuality. Compatible local causal assignments can fail to glue into a global causal function. The paper does not equate this with a closed causal loop, but it does provide an operational obstruction to globally consistent causal structure (Gogioso et al., 2023).

"Complementarity Beyond Definite Causal Order" (Siddiqui et al., 29 Mar 2026) draws another boundary. The paper studies coherent superpositions of two acyclic orders and introduces an operationally measurable quantity,

$F :\sim \mathcal{B}(c_0, c_1).$0

measured via order-qubit interference visibility. This is operational detectability of nonclassical causal order, but not of a causal loop. The paper is explicit that it does not study cyclic causal structures (Siddiqui et al., 29 Mar 2026).

Finally, "Causal posets, loops and the construction of nets of local algebras for QFT" (Ciolli et al., 2011) uses the term “loop” in a wholly different sense: a loop is a closed path in a poset of spacetime regions. Those loops generate groups and local $F :\sim \mathcal{B}(c_0, c_1).$1-algebras, with causality imposed by quotienting causal commutators. This is an algebraic framework for loop-like operators in AQFT, not a theory of detecting feedback cycles of influence.

The unresolved issues differ across these traditions. Observational CI methods identify absence of specific cycles but not generic cycle existence (Mooij et al., 2020). Interventional frameworks identify many positive loop certificates, but hidden causal loops remain possible (Vilasini et al., 2021). Spacetime compatibility depends inherently on geometry, with $F :\sim \mathcal{B}(c_0, c_1).$2-dimensional and higher-dimensional cases behaving differently (Vilasini et al., 2022). System-dynamics methods can detect behavior-generating loops in known models, but they do not discover causal structure from raw data alone (1909.01138). Taken together, these results suggest that “operationally detectable causal loops” is not a unitary doctrine but a layered research program whose conclusions depend on whether the operative inputs are model equations, observational independences, interventions, time-series reconstructions, or spacetime embeddings.