WorksMathematics
A Generative Model for Higher-Order Network Dynamics
Abstract
Conventional network representations inferred from multivariate time series describe statistical relationships between pairs of components, even when the underlying dynamics involve group interactions that cannot be represented by pairwise interactions alone. Several multivariate information-theoretic measures have been proposed to characterise dependence beyond pairs. However, due to a lack of systems with known generating mechanisms, these measures have not been carefully validated or compared for empirical use. To address this, we here introduce a generative model of stochastic network dynamics with pairwise and higher-order interactions specified by the user. The higher-order interactions are represented by directed hyperedges, and the model provides a controlled benchmark for evaluating information-theoretic measures. Removing only the higher-order term gives a pairwise-only reference. The experiments test whether a measure detects changes induced by the higher-order term relative to this reference and recovers the participating nodes and their source-target roles. A separate experiment tests discrimination between a product coupling and a specified nonlinear pairwise control, with equal root-mean-square coupling magnitude under the pairwise-only reference. Several measures detected sufficiently strong higher-order interactions across the tested network topologies and coupling functions. In the separate discrimination experiment, the measures also separated the product coupling from the specified nonlinear pairwise control. Recovery, however, became harder as the number of candidate interactions grew. The framework therefore provides a testbed where the link between higher-order signatures and mechanisms can be tested, rather than assumed, before measures are applied to empirical data.
Introduction
The analysis of complex systems from multivariate time series has long relied on graph representations, in which nodes correspond to system components and edges summarise statistical dependence between pairs of observed time series [9, 22]. Higher-order interactions extend beyond pairwise relationships, allowing multiple source nodes to influence a target jointly through a coupling that cannot be expressed as a sum of separate source contributions [4, 6]. At the level of observed data, studies of brain recordings and financial and epidemiological time series have identified statistical structure beyond that captured by pairwise summaries [17, 28, 35].
Multivariate information-theoretic measures are increasingly used to characterise statistical dependence among groups of variables in time series [18, 26, 35]. Yet a higher-order statistical pattern can arise even when the underlying interactions are purely pairwise [27]. Assessing what these measures reveal about higher-order mechanisms requires evaluation on systems with known generating dynamics. We address this gap with a generative model of network dynamics in which pairwise and higher-order interactions are specified in advance and can be varied or switched off independently. We use this model for three tasks. For detection experiments, we test whether a measure detects changes induced by the higher-order term relative to a pairwise-only reference. For recovery experiments, the identity of the planted interaction is withheld, and we test whether a measure can identify its participating nodes and direction among all candidate interactions. For discrimination experiments, we test whether a measure separates the non-decomposable product coupling from a specified nonlinear pairwise control, with equal root-mean-square coupling magnitude under the pairwise-only reference. The central contribution is a generative benchmark for evaluating how statistical measures respond to known higher-order mechanisms and how reliably they recover the associated network structure.
A growing body of methods characterises higher-order structure in empirical data [3, 5]. Topological approaches describe the geometry and organisation of data-derived structures [24, 28], whereas information-theoretic measures characterise the amount and organisation of statistical dependence among groups of variables [26, 35]. These traditions have largely developed in parallel, but Varley et al. [34] recently compared them directly using synthetic point clouds with known topology and multivariate functional magnetic resonance imaging (fMRI) time series. They associated enclosed voids in three-dimensional point clouds with information-theoretic synergy. In the fMRI data, greater synergy dominance was associated with more numerous cavities and greater average persistence. The two approaches can therefore converge on related structure in observed data. As the authors note, however, both describe the statistical behaviour of a system rather than the mechanism that generated it, so their agreement does not establish that the shared structure arose from a higher-order mechanism. A generative model with known interaction mechanisms allows this relationship to be tested under controlled conditions.
This approach is relevant to neuroscience, where network representations are often inferred from statistical dependencies between regional time series instead of observations of interactions among brain regions [7]. Recent work has also compared information-theoretic and topological descriptors of higher-order dependence in fMRI data [29]. For pairwise connectivity, this problem motivated the work of Smith et al. [32], which simulated fMRI data from networks with known connectivity and evaluated whether network modelling methods recovered the planted structure. For higher-order mechanisms, Robiglio et al. [25] examined synergistic signatures and reconstruction of group interactions in simplicial Ising and social contagion models. Antonacci et al. [1] assessed predictive improvements over additive models using analytical examples and stochastic autoregressive simulations. We extend the simulation approach of Smith et al. [32] to stochastic network dynamics with higher-order interactions. Our benchmark evaluates detection and recovery of participating nodes and source-target roles across coupling functions and network topologies, using pairwise-only references for comparison.
The model builds on the neural state equation of dynamic causal modelling (DCM) [10], which Smith et al. [32] used to generate synthetic fMRI data with known connectivity. A separate higher-order term can be added or removed while retaining the other terms and their parameter values. The experiments assess how detection and recovery performance varies across coupling functions and network structures. A separate experiment tests whether the measures can separate the product coupling from a specified nonlinear pairwise control.
The remainder of the paper is organised as follows. Section 2 briefly defines higher-order interactions, their directed representation, and the distinction between interaction mechanisms and statistical dependence, and introduces the measures evaluated. Further background is provided in Appendix A. Section 3 specifies the generative model and the experimental design. Section 4 presents the base-case results for detection, recovery and discrimination, followed by further tests of detection and recovery across network topologies and coupling functions. Section 5 concludes with the implications and limitations of the framework and outlines directions for future work.
Background
We define higher-order interactions by the form of the coupling in the generating dynamics [4, 6]. For two different source nodes and acting on a target , let denote their contribution as a function of their states . We call this coupling higher-order if there are no univariate functions and such that
for all . This criterion identifies an interaction through non-additivity in the source states [14]. The product coupling satisfies this criterion, whereas a sum of separate source contributions remains pairwise even when those contributions are non-linear. We apply the criterion to coupling functions expressed in the original node states.
We represent each such interaction by a directed hyperedge , where is the target and is the unordered pair of source nodes [11]. The hyperedge identifies the participating nodes and their roles, while specifies how the source states jointly affect the target. Exchanging and leaves the hyperedge unchanged, whereas changing the target within the triplet defines a different directed interaction.
The coupling functions describe interaction mechanisms, whereas information-theoretic measures describe statistical dependence in the resulting time series. This dependence reflects the combined effects of the pairwise and higher-order couplings, external inputs, and process noise. Pairwise mechanisms can produce statistical patterns beyond those captured by pairwise summaries [27]. A multivariate statistical signature therefore does not establish that a higher-order mechanism generated it, and a weak response from a measure does not establish that a higher-order interaction is absent.
We use pairwise mutual information as a reference for dependence between nodes. O- and S-information [26] and Streitberg information [15] describe dependence among groups of variables without assigning source and target roles. For recovery of these roles, we use the synergy term of a partial information decomposition under the minimum mutual information convention [2], together with joint-source transfer entropy [23, 30]. The former quantifies information supplied jointly about a target beyond that provided by the more informative source, while the latter quantifies directed predictive dependence. We also use a combined S/O ranking score for candidate node sets, defined in Section 3.3.4. Definitions and interpretations of the measures are provided in Appendix § 8.§ 3, and estimator settings in § 3.3.7. Further background on the DCM neural state equation and higher-order mechanisms is provided in Appendices § 8.§ 1 and § 8.§ 2.
Methods
This section defines the generative model and the experiments used to evaluate information-theoretic measures. In the four-node base case, we test whether the measures detect changes induced by a known higher-order interaction relative to a pairwise-only reference. We then assess recovery of its participating nodes and source-target roles with the interaction’s identity withheld. A separate experiment tests discrimination between the product coupling and a nonlinear additive control. Further experiments assess how detection and recovery vary across alternative coupling functions and more complex topologies. Sections 3.1 and 3.2 define the model and coupling functions. Section 3.3 describes the base case, validation and experimental procedures. Section 3.4 describes the topology experiments and any changes to the base-case settings.
Generative model
We define the generative model by extending the DCM neural state equation (Equation A.1) so that it carries explicit higher-order interactions. Rather than replacing that structure, we retain its decay, pairwise connectivity, and external-input terms, and add a source-pair interaction term together with additive process noise. Most experiments in this work concern directed triplets, in which an unordered pair of source nodes acts jointly on a distinct target node. The state of node evolves as
Here, is the vector of node states and is the self-decay rate, shared by all nodes. The matrix is the weighted pairwise connectivity matrix, where gives the strength and sign of the directed influence of source on target ; self-decay is represented separately, so . The binary tensor encodes the directed-triplet topology: indicates that the unordered source pair jointly acts on target . The restriction ensures that each source pair is counted only once. The coefficient is a scalar that sets the strength and sign of the interaction, and is its coupling function. Finally, is the vector of external inputs, is a matrix that determines how those inputs enter each node; the are independent standard Wiener processes, and is the process-noise intensity. The input process is generated independently of the Wiener processes .
Unless otherwise stated, the experiments use the product coupling function . Substituting this function into Equation (2) gives a joint contribution to the drift of target . The base case contains the single directed interaction , specified under the convention by and , with all other entries of equal to zero. § 3.2 explains the choice of the product coupling and its extension to larger source sets. The dedicated coupling-function sensitivity analysis replaces the product with the alternative functions defined in that section while leaving all other model components unchanged.
The drift in Equation (2) decomposes the dynamics into four interpretable components. The decay term provides self-regulation, pulling each node back toward baseline and opposing unbounded activity in the absence of sustained input. The pairwise term represents pairwise connectivity, with each source contributing additively to the target according to the signed weight . The higher-order term captures joint source-pair effects that cannot be represented as a sum of separate pairwise influences. Finally, the input term models experimentally imposed or exogenous drive, with controlling which nodes receive each input channel. The terms inside the square brackets form the drift, or systematic part of the state dynamics. The Wiener term instead adds random fluctuations directly to the evolution of the node states. It is separate from the external inputs , and no additional observation noise is included in the primary model.
The decay and pairwise terms form the linear backbone of the model, . Setting , for all , and gives , whose solution is the matrix exponential , so stability is governed by the eigenvalues of . The backbone is asymptotically stable when , where is the spectral abscissa of ; intuitively, self-decay must dominate recurrent pairwise amplification. The product higher-order term can produce nonlinear amplification, so large , strong inputs, or large initial conditions can produce unbounded trajectories even when the linear backbone is stable. Linear stability does not guarantee that solutions of the nonlinear stochastic model remain finite for all times. We assess numerical boundedness over the simulation interval and examine timestep convergence in selected experimental conditions. We integrate the stochastic state equation using Euler–Maruyama,
where denotes the drift in Equation (2).
Coupling function
The tensor determines which source-pair interactions are present, whereas the coupling function determines how the states of each source pair are combined before entering the drift of target . We require every to satisfy two local conditions at the origin,
The first condition ensures that the higher-order term contributes nothing at baseline, leaving as an equilibrium of the unforced deterministic drift. The second ensures that the interaction does not enter the Jacobian at the origin, which remains . The higher-order term is therefore absent from the first-order Taylor linearisation of the drift at the origin. These are local conditions and do not imply that the interaction is invisible to linear or pairwise statistics computed away from the origin. We additionally require to be non-decomposable in the sense of Equation (1), so that the joint source-pair contribution cannot be absorbed into separate pairwise terms.
The main experiments use the product coupling . Its contribution to target takes the sign of multiplied by the product of the source states, so for positive it is positive when the sources have the same sign and negative when their signs differ. This choice is motivated on three grounds. First, the two-source product has precedent in the DCM framework, where Stephan et al. [33] introduce a term in which the activity of one region modulates the influence of another. We use the same bilinear product but treat the sources as an unordered pair because . Second, it is the lowest-degree polynomial term that depends jointly on both sources. Such products also form the first additional coordinates beyond the original node states when polynomial dynamics are represented in an enlarged state space [13]. Third, the product satisfies both conditions in Equation (4) and the non-decomposability criterion by construction.
For experiments involving a larger source set , the product coupling is extended as
This extension satisfies the corresponding local conditions and is non-decomposable into additive functions of proper subsets of . Its -fold mixed partial derivative is one, whereas any term omitting one of the sources has a zero derivative of that order.
The coefficient sets the strength and sign of a source-pair interaction. It is not directly comparable with a pairwise coefficient because the realised scale of depends on the amplitudes of both source states. Scaling every source state by the same factor multiplies the product coupling by . We therefore report realised component-wise scales and assess boundedness for each experimental condition.
To assess whether the conclusions depend on the bilinear product, we compare it with three alternative couplings described in Section 3.3.5; their results are reported in § 4.1.3 and § 4.2.3. For any source set , these are defined by the factorised family
where labels the coupling and its source-response function, rescales each source to an order-one input, and rescales the resulting coupling output. The two-source case is obtained by setting .
The source-wise transformations, applied separately to each source before forming their joint product, are
The regularised denominator makes smooth at zero, while its large-amplitude behaviour is approximately . It therefore provides a smooth signed-root response without the nondifferentiability of the exact signed square root. All three source-response functions satisfy and . For , their factorised couplings satisfy the corresponding local conditions in Equation (4). For non-zero and positive , the mixed partial derivative involving all source variables at the origin is . These couplings therefore cannot be expressed as sums of functions of proper subsets of .
The four mechanisms represent different source responses. The product is bilinear and unbounded. Tanh is monotone and saturating, sine is bounded and oscillatory, and the smooth signed root is monotone and sublinear. These alternatives test sensitivity to coupling function.
Experimental setup & Validation
The model of § 3.1 and § 3.2 is defined up to its topology, coupling function, weights, input process, and noise intensities. This section fixes those choices in a single four-node configuration, reports the numerical and dynamical validation of that configuration, and specifies the null-calibration procedures used for detection and recovery experiments. This configuration is also the reference point for the topology experiments in § 3.4. The experiments in § 3.4 retain its dynamics and coupling convention while varying the network topology, reusing initial states and stochastic paths for nodes shared across layouts.
Base case
Figure 1 shows the base case, and Table 1 lists its full configuration. This configuration is used first to assess whether the planted higher-order interaction can be detected and then whether its participating nodes and direction can be recovered. The planted interaction corresponds to the unordered node set and the directed hyperedge , where nodes 1 and 2 are the sources and node 0 is the target. With the convention that is the influence of source on target , in node order , the weighted pairwise template and its spectrally normalized adjacency matrix are

Figure 1. Four-node base system. Grey arrow pairs represent reciprocal directed edges whose weights may differ. The coloured band and arrow represent the planted higher-order interaction . Green arrows denote independent external inputs; node 0 receives no direct input. Alt text: Four-node network with reciprocal pairwise links and a joint interaction from nodes one and two to node zero. External inputs enter nodes one, two and three.
| Component | Chosen value |
| Nodes | |
| Decay | |
| Pairwise topology | |
| Pairwise weights | |
| Higher-order topology | |
| Higher-order weight | |
| Input contribution | for , , |
| Input signals | , independent, ( at ) |
| Process noise | Independent isotropic Wiener noise, ; at , the per-step increment standard deviation is |
| Initial state | |
| Integration | Euler–Maruyama, , (5001 samples); discard , retaining 4501 samples |
| Randomness | Distinct recorded seeds across replicates; initial states, OU paths, and Wiener paths reused across within each replicate; observation noise |
Table 1. Configuration of the four-node base case.
The use of heterogeneous rather than identical connection strengths is motivated by the network simulations of Smith et al. [32], in which non-zero off-diagonal strengths were drawn with mean 0.4 and standard deviation 0.1 and restricted to . Their numerical values are not transferred directly because the decay convention, topology, and overall dynamical scaling differ from ours. Instead, we use a fixed heterogeneous template whose non-zero entries range from 0.05 to 0.30. After spectral normalization, the corresponding non-zero entries of range from 0.0471 to 0.2828. Fixing the relative weights across replicates ensures that comparisons across isolate the higher-order term, while retaining the connection heterogeneity used in realistic network simulations.
The overall pairwise strength is controlled by a single spectral scale rather than by tuning individual edges. Four nodes are the smallest system containing the planted triplet together with an additional node that does not belong to the planted interaction. Node 3 lies on every pairwise path from either source to the target, so source-target statistical dependence can arise even without direct source-target edges. The linear backbone is asymptotically stable while the spectral abscissa remains below the decay rate . We set with , leaving a stability margin of .
External inputs enter as , the product of the input matrix and the signal . Each component of is an independent stationary Ornstein–Uhlenbeck process,
with zero mean, unit variance, and independent Wiener processes . We integrate it exactly at each step via
where . We draw , so that each input process is stationary from the first simulated sample. We set , corresponding to at . Sources 1 and 2 and node 3 each receive a separate channel through a sparse with one common gain , avoiding node-specific input tuning. The target receives no direct input. After warmup, its activity is explained by decay, the node-3-mediated pairwise route, process noise, and, when present, the higher-order term. Process noise represents endogenous fluctuations not captured by the specified inputs or coupling terms.
Initial states, OU paths, and process-noise paths vary across independent replicates but are drawn from fixed, recorded seeds. Whenever conditions are compared across the higher-order coupling coefficient , they reuse the same initial state, OU path, and process-noise path within each replicate. Duration, input correlation time, input gain, warmup, and sampling are also held fixed. This design isolates the effect of changing the higher-order coupling strength. We use OU inputs to represent temporally correlated drive alongside the white process noise. Because neighbouring observations are therefore not independent, Table 1 reports the simulation duration and alongside the sample count. Null calibration uses independent simulations with the same OU parameters and sampling procedure as the evaluation simulations. We do not permute individual time points, because permutation would remove their temporal dependence.
Model validation
The linear stability margin established above applies to the pairwise backbone near baseline, but does not establish numerical accuracy or boundedness of the complete nonlinear model. We therefore assessed timestep convergence at over 12 replicates for each . Within each replicate, the same initial state and underlying OU and Wiener paths were used across timesteps. For each replicate, we express the target-state RMSE as a percentage of the RMS of its reference trajectory at , evaluating both trajectories at common sampling times. Relative to the reference step , the median target-state RMSE at was 1.89% for and 3.03% for . Halving the step to reduced these values to 0.79% and 1.19%, but doubled the number of integration steps. We therefore retained as the computational operating point for the full benchmark. Complete results are reported in Figure A.1 and Table A.2 in the Appendix.
One-factor-at-a-time sensitivity tests used , the null and recovery operating points; the full sweep is evaluated in Section 3.3.3. We varied the pairwise spectral fraction over and the input correlation time, input gain, and process-noise intensity over , , and times their base values, respectively. Initial states and stochastic paths were generated using the same seeds across the tested parameter values. For every setting and each , all 30 simulations remained finite and satisfied . This is a finite-time check for numerical blow-up rather than a proof of global stability. At , the median target-node higher-order-to-decay drift RMS ratio remained at least 1.09, and the higher-order-to-pairwise ratio remained at least 2.41, confirming that the planted term remained dynamically substantial across the tested conditions.
Finally, although the two sources receive independent external inputs, their pairwise connections induce dependence between their states. Their median zero-lag Pearson correlation ranged from 0.219 to 0.517, showing that the product interaction was evaluated against moderate network-induced source dependence rather than synchronisation imposed by a shared input. Condition-wise results are reported in Figure A.2 and Table A.3 in the Appendix.
Higher-order interaction detection
Detection first validates the generative model by confirming that the planted higher-order term leaves a measurable signature before recovery is attempted. It is evaluated over the coupling strengths in Table 1 using an adaptation of the c-sensitivity concept of Smith et al. [32], which measures how often each estimator distinguishes the higher-order condition from the pairwise-only generative null.
Every estimator is applied to the post-warm-up time series of the three participating nodes, . We analyse the node states directly, without first removing the effects predicted by the pairwise dynamics. Before estimation, each series is centred and scaled to unit variance.
O-information, S-information, and Streitberg information are computed jointly on the three variables. To test what an analysis restricted to individual target–source relationships can detect, we use
Taking the maximum gives the pairwise baseline its strongest score, while avoiding the mixture of unique, redundant, and synergistic contributions contained in joint-source mutual information.
The calibration procedure is applied separately to each metric. An independent ensemble of simulations with provides the calibration scores . These simulations are not reused as detection replicates. Let and denote the empirical mean and standard deviation of the calibration scores. We define the absolute standardized departure of a score as
The detection threshold is
the empirical 95th percentile of the standardized departures in the calibration ensemble. Once estimated, , , and are held fixed throughout the detection experiment. The threshold targets a 5% false-positive rate for each measure. The false-positive probability may differ from 0.05 because the threshold is estimated from a finite calibration ensemble.
For each value of , we generate separate detection replicates. Within each replicate, the initial state, OU input path, and Wiener process-noise path are held fixed across , so that the higher-order coefficient is the only manipulation. The detection condition is independent of the calibration ensemble and provides a held-out check of the false-positive rate.
If is the score from detection replicate , the null-calibrated c-sensitivity of metric is
This is the fraction of detection replicates whose absolute standardized deviations from the calibration mean exceed . For each measure and value of , we report pointwise Wilson 95% confidence intervals for the detection probability conditional on , , and . These intervals quantify uncertainty from the detection replicates without including uncertainty from estimating the calibration parameters. The detection rule is two-sided because the included information measures do not share a common signed alternative: evidence may appear as either an increase or a decrease relative to the null distribution.
This definition is inspired by, but differs from, the original c-sensitivity of Smith et al. [32], which compared scores for true network edges with the within-network distribution of scores for absent edges. Our adaptation instead estimates how often a metric detects departure from a pairwise-only generative null under the specified model and sampling design. It therefore evaluates the estimator against known generative conditions, using independent calibration and detection ensembles and a pairwise-only counterfactual. The resulting c-sensitivity is a repeated-simulation detection rate, not the probability that an observed dataset was generated by a higher-order mechanism.
Triplet recovery
Following detection, the recovery benchmark withholds the planted identity and asks whether it can be inferred from the observations, more closely reflecting analysis of empirical data. Unordered-set recovery identifies the three participating nodes, whereas directed recovery also distinguishes the sources from their target. Both use the same simulation ensembles but rank different candidate spaces. The chance probabilities assume uniform random ranking and can differ from selection probabilities under the model.
Recovery is evaluated at , the smallest tested coupling strength at which KSG S-information, our primary detector, and the KSG mutual-information reference both attained a -sensitivity of 1 across all 40 detection replicates (Section 4.1.1). This value is fixed across all recovery measures and evaluated using simulations separate from the detection experiment. We use independent simulations for calibration and 60 test replicates, each using the same seeds for the and conditions. After warmup, estimation time points are spaced by 14 integration steps, or approximately one OU correlation time.
In the unordered-set experiment, the symmetric measures rank the four possible three-node sets using states recorded at the same retained time points. We use KSG S-information, KSG O-information, the combined S/O score, and an entropy-kNN estimate of O-information from HOI. S-information uses the upper null tail because larger values indicate stronger collective dependence, whereas both O-information estimates use the lower tail because more negative values indicate a more synergy-dominated departure. The chance probability of ranking the planted set first is . Because these measures are symmetric, they identify the participating nodes but not their source and target roles.
In the directed experiment, the target-specific measures rank all assignments of one target and two sources. For candidate , the sources are evaluated as and and the target as , with . Conditional variants additionally condition on . The target-specific measures use their upper null tails because larger values indicate stronger target-specific synergy or predictive dependence. The inferred direction represents temporal prediction and does not by itself establish causality. We report the correct target, the correct source pair among candidates with target 0, and the exact candidate , with chance probabilities , , and , respectively.
Each measure and candidate is calibrated against its own distribution from the independent ensemble. Candidate-specific calibration accounts for differences in baseline scores caused by network position and source–target assignment. For an upper-tail score ,
with the inequality reversed for a lower-tail score. The combined score is . For each candidate, is calibrated against leave-one-out values from the independent null ensemble, thereby accounting for dependence between S-information and O-information.
Candidates are ranked by increasing empirical probability of equally or more extreme scores under the pairwise-only model. Ties are broken by decreasing when larger scores are favoured and decreasing when smaller scores are favoured. For each measure and candidate, and are the mean and standard deviation of its calibration scores. These probabilities provide a common ranking scale rather than multiplicity-adjusted significance tests. The planted identity is withheld until the rankings are complete. The condition provides the negative control, and the condition evaluates recovery when the interaction is present. We report pointwise Wilson 95% confidence intervals for the top-rank selection probabilities, conditional on the calibration ensemble.
Coupling-function sensitivity analysis
The coupling-function sensitivity analysis included two complementary experiments. The first compared coupling functions in the two-source base case (§ 4.1.3). At each value of , the coupling function was the only model component varied; the topology, remaining model parameters, initial states, external-input paths, and Wiener process-noise paths were held fixed. The second repeated this controlled comparison across interaction sizes (§ 4.2.3). Within each fixed layout and interaction size, only the coupling function was varied.
In the base case, the product coupling was compared with the sourcewise tanh, sine, and smooth signed-root functions defined in § 3.2. Their normalisation constants were estimated from 200 independent simulations. This scaling ensemble was separate from the detector-null and test ensembles. For each source, we selected such that
This placed the source states on a comparable input scale. The amplitude constant was then chosen so that each alternative function had the same RMS magnitude as the product on the scaling ensemble. All constants were estimated once from the scaling ensemble and then held fixed across all evaluation simulations. After this rescaling, the remaining difference from the product was measured by
Detector performance was evaluated in 40 test replicates at . At , we also compared the RMS of the higher-order state increments at target node 0 across coupling functions.
For the interaction-order analysis, the same coupling family was applied in the network to source sets containing two, three, or four nodes, corresponding to total interaction sizes of three, four, or five nodes after including the target. Each layout used its own set of 200 independent scaling simulations. Its source-normalisation constants were determined by applying the source-scaling rule defined above for the two-source, four-node coupling comparison.
The product coupling used in the main analyses remained unnormalised. For this sensitivity experiment only, we added the normalised-product control
This control tests whether differences across interaction sizes arise from multiplying an increasing number of small source states. The normalised product and the three alternative functions were scaled to the RMS magnitude of the size-three product coupling. Their constants were then fixed and used in 40 test replicates at for each interaction size.
Each layout used its own independent set of 200 detector-null simulations, separate from the scaling and test ensembles. The metric-specific detection thresholds derived from these null simulations were fixed before detector performance was evaluated. Unordered-set recovery used a separate ensemble of 100 calibration simulations and 60 test replicates per layout, with initial states, OU input paths, and Wiener process-noise paths reused across coupling functions. Candidate-specific calibration and ranking followed Section 3.3.4. All conditions were assessed using the finite-time boundedness check in Section 3.3.2. Timestep comparisons for the four coupling functions in the base case at are reported in § 4.1.3.
Discrimination from a decomposable control
Detection asks whether the planted term changes the system, while recovery asks whether its participating nodes and roles can be identified. We additionally tested whether the measures distinguish the non-decomposable product coupling from a specified decomposable alternative. In the four-node base case, we compared
The additive control can be incorporated into separate nonlinear pairwise contributions to the target equation, whereas the product cannot be expressed as a sum of single-source functions. This distinction is algebraic. The experiment tests whether it produces distinguishable statistical signatures in the observed dynamics. Both functions vanish with zero gradient at the origin, satisfying Equation (4). The baseline is the condition, in which neither additional coupling contributes and the original pairwise dynamics and stochastic inputs remain unchanged.
Using 200 independent baseline scaling simulations, we fitted so that
The coefficient was held fixed across . This gives the coupling outputs the same RMS on the baseline ensemble, without equating their means, temporal dependence or realised effects on the target. Both couplings entered the same target equation with coefficient , with all other dynamical parameters unchanged. Initial states, OU inputs and Wiener paths were reused across conditions and strengths within each replicate.
Simulations ran to with and a 10-unit warmup. Retaining every 250th sample yielded 999 observations spaced by 5 time units. This sampling scheme was selected using independent linear-Gaussian baseline checks to reduce the distortion observed in densely sampled KSG estimates (see Appendix § 10.§ 1). Metric definitions and estimator settings were unchanged.
At each , 200 independent additive simulations calibrated the two-sided -sensitivity rule of Equation (7), replacing the baseline reference with the additive reference at the corresponding strength. Discrimination -sensitivity was measured as the fraction of 200 held-out product simulations whose scores fell outside the calibrated additive-reference interval. Applying the same criterion to 200 held-out additive simulations provided the false-positive rate.
An independent validation at used 1000 additive calibration simulations and 400 held-out simulations per coupling form. The score-distribution figure uses this validation ensemble, including 400 baseline simulations; the across- curves use the 200-replicate test ensembles. Scaling, calibration and test ensembles were independent. Wilson 95% intervals are pointwise and conditional on the fitted thresholds. This comparison tests discrimination from the signed-square additive control; it does not establish discrimination from other choices of decomposable coupling.
Estimators and implementation
Before estimation, each node-state series is centred and scaled to unit variance. Mutual information and conditional mutual information are computed with the IDTxl toolbox [38], which wraps the JIDT estimators [16], and O- and S-information are computed directly with JIDT. The Gaussian estimates evaluate the corresponding information measures for a multivariate Gaussian distribution using the estimated covariance matrix. All KSG estimates use , the document-ed default in both toolkits, as a compromise between local resolution and sampling variability. This value was fixed before analysis across all conditions, null ensembles and candidates, and no noise or dynamic-correlation exclusion was added. The HOI entropy-kNN estimate of O-information [20] also uses . Streitberg information uses the reference Tsallis- kNN implementation of Liu et al. [15], with and , and is therefore not labelled as either Gaussian or KSG. Package-level permutation tests are not used; detection and recovery are instead calibrated against independently simulated ensembles, as described in Sections 3.3.3 and 3.3.4, while discrimination is calibrated against independent additive-control simulations at the corresponding coupling strength, as described in § 3.3.6.
Generalisation across topology regimes
Figure 2 shows the topology regimes used to test whether detectability and recovery generalise beyond the four-node base case. Across thirteen fixed deterministic layouts (Table 2), we vary network size over , interaction multiplicity over with shared or separate targets, interaction overlap, and total interaction size over . Each layout embeds the pairwise motif of the base case around every planted interaction and is spectrally normalised to ; all remaining dynamical, stochastic, and estimator settings are inherited from § 3.3. External-input allocation follows the planted target structure. In non-reciprocal layouts, every node except the planted targets receives an independent external input. Reciprocal layouts retain the base-case allocation, in which node 0 is undriven and nodes 1, 2, and 3 receive independent inputs. Comparisons between shared-target and separate-target layouts therefore include the corresponding change in which nodes receive external drive. To approximate an undirected three-node hyperedge, the reciprocal condition includes the three directed terms , , and , each with coefficient . The normalised condition instead assigns to each term, giving the coefficient vector norm .

Figure 2. Topology regimes used in the generalisation experiments. Grey lines denote pairwise links, coloured bands connect higher-order sources, and arrows indicate their targets. The higher-order interactions and their coefficients are listed in Table 2. Alt text: Network diagrams compare the directed base case, reciprocal interactions, a larger network, multiple interactions, shared sources or targets, and an interaction with more than two sources and one target.
| Layout | Planted directed interactions | Coefficient | ||
| Base case and network size | ||||
| Directed base | 4 | 1 | ||
| Network size | 8 | 1 | ||
| Network size | 16 | 1 | ||
| Interaction multiplicity | ||||
| , shared target | 16 | 2 | , | |
| , disjoint | 16 | 2 | , | |
| , shared target | 16 | 4 | , , , | |
| , separate targets | 16 | 4 | , , , | |
| Overlap | ||||
| Shared source | 16 | 2 | , | |
| Shared source and target | 16 | 2 | , | |
| Interaction size | ||||
| Size four | 16 | 1 | ||
| Size five | 16 | 1 | ||
| Reciprocity | ||||
| Reciprocal cell | 4 | 3 | , , | |
| Reciprocal, normalised | 4 | 3 | , , |
Table 2. Topology layouts used in the generalisation experiments. and denote the numbers of nodes and planted interactions, respectively. Interactions are written as an unordered source set followed by its target. Unless varied explicitly, all layouts inherit the base-case model settings.
Detection and recovery are calibrated separately for each layout against its own independently simulated ensemble, using the same calibration and test sizes as the base case. For interactions with more than two sources, mutual information remains the maximum target–source pairwise reference, while O-information, S-information, and Streitberg information are evaluated on the complete planted node set. Because is not rescaled with source-set size, the realised higher-order component is reported separately. For planted interactions, scores are retained separately and recovery is evaluated using recall at rank , mean average precision, and the probability that all planted interactions occupy the first ranks. For , mean average precision reduces to mean reciprocal rank. Directed MMI-PID and transfer-entropy recovery are restricted to three-node interactions, so the size-four and size-five conditions contribute only to detectability and unordered-set recovery. The reciprocal configuration has one true unordered set, , but three true directed assignments; hence for directed recovery.
Within each layout, the initial states, OU input paths, and Wiener paths are reused across . When a node appears in more than one layout, it retains the same node-specific initial condition and stochastic paths, allowing comparisons across topologies.
Reciprocal product interactions can produce positive nonlinear feedback. Trajectories that fail the finite-time numerical-validity criterion are neither clipped nor replaced. For each condition, we report the fraction of requested trajectories that remained valid and calculate detection and recovery scores using only those trajectories. These scores therefore describe performance conditional on numerical validity. For reciprocal layouts, we focus on , where most trajectories remained valid, and report results at stronger coupling together with their validity counts.
Results
This section reports what the information-theoretic measures recover from the generative model. In the four-node base case, we first assess how often each measure detects changes induced by the higher-order term relative to the reference as coupling strength increases. We then test whether the participating nodes and their source-target roles can be identified when the interaction’s identity is withheld. A further experiment tests whether the measures distinguish the product coupling from a specified nonlinear pairwise control scaled to have the same RMS as the product. We then ask how far the detection and recovery results extend, applying the same measures and null-calibration procedure across the topology regimes of § 3.4. All detection and recovery scores reported below are calibrated against an independently simulated ensemble with the higher-order term removed, so these results are always expressed relative to what the same measure produces on the same system without a higher-order interaction. Discrimination scores instead use the nonlinear pairwise control at the corresponding coupling strength as their calibration reference.
Base-case benchmark
The four-node base case separates validation of the generative model from structural inference. Detection evaluates the known planted triplet to confirm that activating the higher-order term produces a measurable effect relative to the pairwise-only condition. Recovery then withholds this structural information and tests whether the measures can identify the participating nodes and their source–target roles, more closely reflecting their use with empirical data. Discrimination then tests whether the statistical signatures of the product coupling can be distinguished from those of a specified decomposable nonlinear pairwise alternative.
Higher-order interaction detectability
We varied the higher-order coupling strength over . For each metric, the detection threshold was estimated from 200 independent calibration simulations. C-sensitivity was then evaluated across 40 test replicates; within each replicate, the initial state, external Ornstein–Uhlenbeck input paths, and Wiener process noise paths were held fixed across .
The main result was that the planted higher-order interaction was detectable and became more readily distinguishable from the pairwise-only null as increased. This shows that the model produces a controllable higher-order effect, detected most clearly by the non-parametric estimators. Detection remained close to its independent level for and increased consistently from onwards. At , KSG mutual information and KSG S-information detected the planted interaction in 39/40 and 33/40 replicates, respectively. At , both detected all 40 replicates (c-sensitivity 1.000; 95% Wilson CI [0.912, 1.000]). KSG O-information and Streitberg information showed more moderate responses, reaching c-sensitivities of 0.575 and 0.625 at , whereas the Gaussian estimators remained at or below 0.275 throughout the sweep (Figure 3). Exact metric-by- values are provided in Appendix Figure A.4.

Figure 3. Detection of the higher-order interaction as increases. C-sensitivity is the fraction of 40 test replicates exceeding metric-specific thresholds calibrated using 200 independent simulations. The dashed line marks the 5% false-positive rate targeted by the calibration threshold.
Alt text: Detection rates versus coupling strength. KSG mutual information and S-information rise to full detection; KSG O-information and Streitberg information increase more moderately, while Gaussian estimates remain low.
Because the test replicates contain no higher-order term, any detection among them is a false positive, and the calibration threshold is designed to keep this rate at 0.05. The observed rate was higher, at 0.125 (5/40; 95% Wilson CI [0.055,0.261]) for the KSG and Streitberg estimators and 0.075–0.100 for the Gaussian estimators. This elevation did not recur in the independently simulated topology ensemble, where the corresponding rates were 0.000–0.075. Both the calibration and test ensembles differed between these experiments, so this comparison does not isolate the source of the difference in false-positive rates. Detection rates at larger should therefore be read relative to this observed baseline rather than the nominal 0.05.
The pairwise KSG mutual-information baseline responded most strongly, showing that changes induced by the higher-order term were detectable using a score computed from target-source pairs. This result provides a sensitivity benchmark, but does not establish irreducibly multivariate dependence. Among the measures evaluated on the full triplet, KSG S-information was the most sensitive: it separated from the null earlier than O-information and Streitberg information and detected the interaction in all 40 replicates at . Because pairwise dynamics can also produce non-zero S-information, the evidence for the planted mechanism comes from this null-calibrated increase rather than from the S-information value alone.
KSG O-information had lower detection rates than KSG S-information. Mean KSG O-information decreased from −0.631 at to −0.753 at and became more negative in 38/40 corresponding replicates (Appendix Table A.4). These results describe changes in estimated O-information relative to the pairwise-only reference. Interpreting their signs and magnitudes as population properties requires accounting for estimation bias, including the sampling effects examined in Appendix § 10.§ 1. Streitberg information also departed from the null, but neither this response nor the O-information change rules out a lower-order statistical explanation.
The KSG estimators had higher detection rates than their Gaussian counterparts under the sampling scheme used. For jointly zero-mean Gaussian sources, , including when the sources are correlated. Differences in estimation bias and variability may also contribute to the observed differences in detection rates.
Alt text: Drift magnitudes fluctuate without sustained growth. External input dominates across the network. At the target, higher-order and decay contributions increase with coupling strength, while the pairwise contribution remains comparatively small.
Finally, the detection pattern occurred in bounded, sustained dynamics. All 40 replicates remained numerically valid at each tested coupling strength, with no trajectories approaching the numerical-validity threshold. At and , the ensemble-mean higher-order contribution remained sustained throughout the post-warm-up interval. Its magnitude increased from to , as expected for stronger coupling, but did not continue to grow over time within either condition (Figure 4). This temporal stability indicates a persistent coupling effect rather than runaway dynamics. At , its whole-network norm was 0.36 times the pairwise norm, whereas at the affected target its mean RMS magnitude was 4.61 times the pairwise contribution. This difference arises because the higher-order term acts only on node 0, while pairwise contributions act throughout the network. Although process noise is excluded from the drift-component figure, its scale remains relevant when interpreting the operating regime. Across the detection replicates at , the median target-node process-noise increment RMS was per integration step, comparable to the median higher-order increment RMS of . The planted interaction is therefore detectable despite operating in the presence of stochastic fluctuations of comparable instantaneous scale.

Figure 4. Drift-component magnitudes across the network (a) and at the planted target (b) for , , and . Lines and shading show the mean standard deviation across 20 replicates, smoothed over two seconds. The target receives no direct input, so panel (b) omits the external-input curve. The panels use different vertical scales; process noise is excluded because it is not a drift term.
Appendix Figure A.3 shows the node trajectories for one replicate. Because both conditions share the same initial state, inputs, and noise, the non-target trajectories largely overlap, and the clearest difference is the larger excursions of the target, node 0, where the higher-order term enters directly.
Recovery of the planted triplet
The detectability results showed that, when the planted triplet was supplied, the higher-order term produced a detectable departure from the reference. Having validated that the model produces a detectable higher-order effect, we next considered a task closer to empirical analysis. We withheld the interaction’s identity and asked whether the measures could recover its participating nodes and source–target orientation from all possible candidates.
Recovery was evaluated at , the fixed operating point specified in Section 3.3.4. Candidate-specific null distributions were estimated from 100 independent calibration simulations, followed by 60 test replicates, each using the same seeds for the and conditions. Symmetric measures ranked the four admissible unordered three-node sets, whereas target-specific measures ranked all 12 directed source-pair–target assignments. Each candidate was calibrated against its own null distribution to account for differences in baseline dependence across the network.
The principal result was that the participating nodes could be recovered more reliably than their source–target orientation. In the control, selection of the planted unordered set remained close to the chance probability of 1/4. At , KSG S-information ranked the planted set first in 0.933 of replicates (95% CI [0.841,0.974]). The combined high-S/low-O score reached 0.750, while neither O-information estimator exceeded 0.383 (Figure 5). KSG S-information therefore provided the clearest identification of the participating nodes. Adding O-information to the combined score reduced performance because O-information was less selective among the candidate sets.

Figure 5. Unordered-set recovery across 60 replicates. Cells show top-1 selection frequencies at (a), (b), and their difference (c). The black outline marks the planted set . Alt text: Three heatmaps compare triplet recovery without and with higher-order coupling and the resulting changes. KSG S-information recovers the planted triplet most strongly, followed by the combined high-S, low-O score.
Recovering the directed interaction was more difficult. Unconditional KSG MMI-PID ranked the exact assignment first in 0.550 of replicates (95% CI [0.425,0.669]), compared with the chance-level frequency of 0.083 under the control (Figure 6). It recovered the correct source pair in 0.767 of replicates when target 0 was supplied, but selected the correct target in 0.667. Target identification was therefore the main limitation on exact recovery.

Figure 6. Top-ranked directed assignments at (a), (b), and their difference (c) across 60 replicates. Cells show selection frequencies; the black outline marks the planted interaction, with nodes 1 and 2 as sources and node 0 as the target. Alt text: Three heatmaps compare directed-assignment recovery without and with higher-order coupling and the resulting changes. Unconditional KSG MMI-PID recovers the planted interaction most strongly; joint-source transfer entropy recovers an incorrect target instead.
Conditioning MMI-PID on the target’s current state reduced exact recovery to 0.183. One possible explanation is that the short prediction horizon leaves little additional source information after conditioning on the target’s present state. Conditioning also increases the dimension of the estimation problem, which may affect recovery.
Joint-source transfer entropy showed a sharper contrast. It recovered the correct source pair in 0.833 of replicates when the target was supplied, but its exact recovery rate in the blind search was only 0.067. It frequently favoured , whose proposed sources both connect directly to node 3 in the pairwise network. Joint-source transfer entropy measures the total predictive information supplied by the source pair and therefore does not isolate information that requires both sources jointly. Unconditional MMI-PID was consequently more selective for the planted directed interaction.
The detection and recovery experiments serve different purposes in the benchmark. Detection evaluates the known planted triplet to confirm that activating the higher-order term produces an observable statistical signature and to compare estimator sensitivity. Recovery withholds the planted identity, more closely reflecting applications in which the interacting nodes and their roles are unknown. In the base case, KSG S-information recovered the unordered participating set reliably, suggesting that it can help identify candidate groups. Recovering the exact source–target assignment was far less reliable. This distinction illustrates the importance of generative benchmarks that test both whether higher-order effects are detectable and how reliably their structure can be recovered.
Sensitivity to the coupling function
We next assessed to what extent the base-case results depended on the bilinear product coupling. The product was retained as the reference and compared with the sourcewise tanh, sine, and smooth signed-root couplings defined in § 3.2. Their calibration and comparison are described in Section 3.3.5. The product remains the coupling used in all other experiments unless otherwise stated.
The purpose of this experiment was to test whether detection generalised beyond the product coupling while controlling for interaction strength. We replaced the product with sourcewise tanh, sine, and smooth-root couplings. Here, “sourcewise” means that each source was transformed separately before the transformed values were multiplied: . This preserves an explicit dependence on both sources while changing how their individual values contribute to the interaction. Each alternative was calibrated on source states from independent simulations to have the same output RMS as the product.
The calibrated outputs were related but not equivalent (Figure 7a). All three transformations are approximately linear near the origin, so their outputs initially follow the product. At larger source magnitudes, tanh saturates, sine moves beyond its near-linear region, and the smooth root compresses extreme values. These effects produce the visible curvature away from the identity line. The vertical dispersion occurs because the same product can arise from different individual values of and , which the sourcewise transformations treat differently. Correlations with the product ranged from 0.922 to 0.941, with NRMSE values of 0.323–0.374. The alternatives therefore retained broad similarity to the product while implementing meaningfully different coupling functions.

Figure 7. Calibration and realised coupling scales. (a) Alternative outputs versus the product for the same source states from 200 independent simulations, scaled to a common RMS. Darker hexagons indicate greater density, and the black line marks equal outputs. Titles report correlation and normalised error (NRMSE) relative to the product, with higher and lower NRMSE indicating closer agreement. (b) Target-node component RMS per step at . Points show medians and bars show the 10th–90th percentiles across 40 replicates. The axis is logarithmic. No input component is shown because the target receives no direct input.
Alt text: (a) Alternative coupling outputs closely track the product but diverge at larger magnitudes. (b) Target-component magnitudes are similar across coupling functions, with higher-order and process-noise increments exceeding pairwise increments.
Figure 7b confirms that the alternative couplings affected the target at nearly the same strength once embedded in the dynamics. At , their median higher-order increment RMS ranged from to per step, a difference of less than 1%. Thus, no coupling function produced a stronger planted signal than the others. These higher-order increments were also similar in magnitude to the process-noise increment RMS of , so they did not simply overwhelm the stochastic fluctuations. Differences in detector performance can therefore be interpreted primarily as sensitivity to the coupling function rather than to unequal realised signal strengths.
The main result was the consistency of KSG S-information across coupling functions (Figure 8a). All four mechanisms showed a strong increase in c-sensitivity as increased. At , c-sensitivity reached 0.825 for the product and 0.950 for each sourcewise alternative. At and , all 40 replicates exceeded the KSG S-information null threshold for the product, tanh, sine, and smooth-root couplings. At , all functions shared the same c-sensitivity of 0.125 because the higher-order term was absent and the trajectories were identical. KSG S-information therefore showed the same overall transition from the pairwise-only reference under every coupling function.

Figure 8. Detection across coupling functions. (a) KSG S-information c-sensitivity across in 40 replicates, with 95% Wilson intervals and a 5% false-positive rate targeted by the calibration threshold. (b) Detector c-sensitivity scores at , using thresholds from 200 independent simulations.
Alt text: (a) KSG S-information reaches full detection as coupling increases for all four functions. (b) KSG mutual information and S-information detect all replicates; other detectors show lower, sometimes function-dependent sensitivity.
Several similarities also appeared across the full detector suite at (Figure 8b). The pairwise KSG mutual-information baseline and KSG S-information both reached a c-sensitivity of 1.000 for all four functions. The Gaussian estimators were also broadly consistent across functions, but remained weak. Gaussian mutual information ranged from 0.200 to 0.250, Gaussian O-information from 0.100 to 0.175, and Gaussian S-information from 0.100 to 0.150. Within these narrow ranges, the product generally gave the highest values and sine the lowest.
The clearest differences occurred for KSG O-information and Streitberg information. KSG O-information increased from 0.575 for the product to 0.850 for each sourcewise alternative. Streitberg information showed the opposite pattern. Its c-sensitivity was 0.625 for the product, 0.425 for both tanh and sine, and 0.475 for the smooth root. Thus, tanh and sine produced identical results for these two measures, while the smooth root achieved the same KSG O-information performance but gave slightly stronger Streitberg detection.
Because Figure 7 established that the four mechanisms had comparable realised strengths, these differences are best interpreted as sensitivity to coupling form rather than signal magnitude. The main group-level result was robust, since KSG S-information detected all four mechanisms once the coupling was sufficiently strong. However, the relative performance of KSG O-information and Streitberg information remained mechanism-dependent. We retain the product in the remaining experiments because it provides the simplest interpretable coupling, while this analysis shows that the principal detection result extends to all three sourcewise alternatives.
All trajectories remained bounded through . At , refining the timestep from to 0.005 produced median post-warm-up target-state errors of 2.84–2.94% across the four mechanisms, providing no indication that numerical instability drove the observed similarities or differences.
Discriminating non-decomposable and decomposable couplings
To assess discrimination, we tested whether the measures could distinguish the product coupling from a specified additive alternative after scaling the additive output to have the same RMS as the product on the scaling ensemble. Figure 9 compares the score distributions for the baseline (), additive control and product coupling (both at ). At , most measures showed little overlap between the product and additive distributions, whereas KSG S-information showed greater overlap. Because calibration is two-sided, either higher or lower product scores can provide discrimination.

Figure 9. Score distributions for the baseline (), additive control and product coupling (), shown separately for each measure. Each distribution contains 400 held-out simulations from the independent validation ensemble. Histograms use shared bins within each panel and are individually normalised to unit area.
Alt text: Seven histograms compare baseline, additive-control and product-coupling scores. Product and additive distributions are clearly separated for most measures, with greater overlap for KSG S-information.
In the independent validation at , using 1000 calibration simulations, KSG S-information rejected the additive reference for 356 of 400 product simulations (89%), with a 4% false-positive rate on additive controls. The other six measures rejected the reference for all 400 product simulations, with false-positive rates between 3.25% and 6%.
The separate across- experiment also showed substantial separation (Figure 10a). At every tested nonzero strength, all measures except KSG S-information rejected the additive reference for all 200 product simulations. KSG S-information achieved rejection rates between 79.5% and 95%. However, additive-control rejection rates ranged from 4% to 15.5% at these strengths (Figure 10b), exceeding the nominal 5% level for some measures. Some measures incorrectly flagged additive simulations more often than others, so their high discrimination rates must be considered alongside their false-positive rates. In the larger independent validation at , discrimination remained strong while false-positive rates were close to the nominal 5% level. This larger validation was not repeated at the other coupling strengths.
Figure 10. Discrimination across coupling strengths. (a) Discrimination c-sensitivity: the fraction of product simulations whose scores fall outside the calibrated additive-reference interval, using the nonlinear pairwise control rather than the baseline as the reference. (b) Fraction of additive-control simulations falling outside the same interval (false-positive rate). At each , thresholds were calibrated using 200 independent additive simulations and evaluated on 200 held-out simulations per coupling form. Shading shows pointwise Wilson 95% intervals conditional on the fitted thresholds; dotted lines mark the nominal 5% level. At , the coupling conditions reduce to the same baseline. Alt text: (a) Discrimination is high across all measures at nonzero coupling strengths, ranging from about eighty to one hundred percent. (b) Additive-control false-positive rates vary across measures, often exceeding the nominal five percent.
Taken together, the results show substantial discrimination between the product and signed-square additive couplings across all seven measures, already at the smallest tested nonzero strength, . These results must be considered alongside the elevated false-positive rates for some measures in the across- experiment. The larger independent validation provides stronger evidence at .
These findings complement detection and recovery by addressing whether higher-order mechanisms can be distinguished from what pairwise dynamics could also produce. Here, the measures separated the planted product from a specified nonlinear pairwise alternative with the same output RMS on the scaling ensemble. This provides evidence that their discrimination extends beyond detecting a change relative to . The comparison demonstrates distinguishable statistical signatures of the two tested couplings, but does not establish that a measure uniquely identifies non-decomposability or distinguishes it from every possible pairwise alternative.
Generalisation across topology regimes
We next tested whether the base-case findings generalised as network and interaction structure became more complex. We varied network size, interaction multiplicity and arrangement, interaction size, and reciprocity according to Table 2. Each layout was calibrated against its own independently simulated null ensemble. Shared-target and separate-target layouts also differ in which nodes receive input, as described in § 3.4. Their comparison therefore assesses the combined effect of these design changes.
Detection was evaluated across the full coupling sweep, while recovery was evaluated at for the non-reciprocal layouts. These layouts retained every trajectory. In the reciprocal layouts, numerical validity decreased as increased. Their high- scores therefore apply only to the trajectories that remained valid and may not represent the full set of runs. We consequently treat , where most reciprocal trajectories remained valid, as the primary reciprocal result and report higher- scores alongside their validity counts.
Higher-order interaction detectability across topologies
We first examined how network size and interaction structure affected the detectability of the planted interactions. Figure 11 compares null-calibrated c-sensitivity across the tested layouts. Detailed metric-by- values are provided in Appendix Figure A.8 for network size, Figure A.9 for interaction multiplicity, and Figure A.10 for interaction overlap and size. Each interaction was evaluated on its own target and source set using the threshold from that layout’s independent calibration ensemble. Values for the directed layout differ slightly from those in Section 4.1.1 because the topology experiments use independently simulated calibration and test ensembles. For layouts with multiple interactions, the curves average the separate interaction-level scores. Simultaneous detection of all planted interactions is examined separately below. Numerical validity is also reported separately for the reciprocal conditions.

Figure 11. Null-calibrated detection across topology regimes. Curves show interaction-level c-sensitivity across ; for layouts with multiple planted interactions, scores were calculated separately and then averaged. The dashed line marks the 5% false-positive rate targeted by the calibration threshold. All trajectories were valid except in the reciprocal layouts, whose detection scores are conditional on numerical validity. The final panel reports the corresponding valid/40 counts at = 10, 20, and 30.
Alt text: Detection curves remain similar across network sizes but vary with interaction multiplicity and overlap. Four- and five-node interactions show little detection. Reciprocal interactions are detectable, but fewer trajectories remain numerically valid at stronger coupling.
The topology experiments produced a consistent overall pattern. Detection of size-three interactions was robust to increasing network size, but shared targets made simultaneous detection of multiple interactions harder. Source overlap strengthened the response of KSG -information, whereas interactions with more sources became undetectable because the unscaled product coupling became extremely small. Reciprocal interactions were detectable at , but their stronger-coupling results were limited by numerical validity.
Increasing network size did not reduce detection of the planted size-three interaction. Because the measures were evaluated on the known planted triplet, adding nodes did not enlarge the candidate search. Instead, it added indirect pairwise pathways that could reshape dependence within the triplet and obscure the higher-order signature. At , KSG S-information reached a c-sensitivity of 0.925 and the pairwise KSG mutual-information baseline reached 0.900 for , 8, and 16; both reached 1.000 at . The broader pairwise network context therefore did not appreciably weaken detection within the planted triplet.
Interaction multiplicity had a larger effect when several interactions converged on the same target. Two interactions remained strongly detectable whether their targets were shared or separate. With four interactions at , mean c-sensitivity was 0.931 for KSG mutual information and 0.831 for KSG S-information when the target was shared, compared with 1.000 for both measures when the targets were separate. We also counted a replicate as successful only when all four planted interactions exceeded their respective thresholds. Under this stricter criterion, the success fractions were 0.850 for KSG mutual information and 0.550 for KSG S-information in the shared-target layout, but 1.000 for both measures when the targets were separate. Shared targets therefore affected simultaneous detection more strongly than the mean interaction-level scores suggest.
Interaction overlap affected KSG O-information differently. At , its mean c-sensitivity was 0.938 when two interactions shared a source, compared with 0.738 when their sources and targets were disjoint. When both a source and the target were shared, c-sensitivity fell to 0.638. The stronger response was therefore associated specifically with source overlap, rather than overlap in general. This may occur because the two shared-source products, and , both contain fluctuations from . However, O-information is symmetric and does not identify which variable is the shared source. Streitberg information also varied with interaction arrangement, whereas the Gaussian estimators remained weak.
Increasing interaction size produced a different failure. For both the size-four and size-five conditions at , c-sensitivity was 0.050 for KSG mutual information and 0.025 for KSG S-information, with no consistent increase across the coupling sweep. This resulted from a collapse in the realised higher-order contribution as more small-amplitude source states were multiplied together (Appendix Figure A.7). At , the whole-network higher-order increment RMS decreased by approximately two orders of magnitude with each additional source, from for the size-three interaction to for the size-five interaction, placing the size-four and size-five contributions approximately 400 and 41000 times below the process-noise increment RMS of . Their failure therefore does not show that the estimators cannot detect interactions involving larger sets, but that the fixed-coefficient product coupling does not preserve realised interaction strength as the number of sources increases. Section 4.2.3 addresses this directly by comparing the raw product with scale-normalised and alternative coupling functions.
The reciprocal conditions require more cautious interpretation. At , 33/40 trajectories in the equal-strength condition and 38/40 in the normalised condition remained valid. KSG S-information detected a collective dependence change on the reciprocal triplet in 31/33 and 22/38 of these trajectories, giving c-sensitivities of 0.939 and 0.579, respectively (Appendix Figure A.11). These are the primary reciprocal results because most trajectories were represented.
The loss of validity at stronger coupling is consistent with nonlinear positive feedback. The reciprocal terms form a closed loop in which each node influences the other two. When , this feedback can increase the triplet’s magnitude and overcome the stabilising linear decay. Scaling each reciprocal coefficient to weakens this feedback, explaining the greater numerical validity of the normalised condition.
At , -sensitivity reached 1.000 and 0.969, but only 10/40 trajectories in the equal-strength condition and 32/40 in the normalised condition remained valid; at , validity fell to 1/40 and 14/40. Because valid and diverging trajectories were not compared, these high-coupling scores may be selection biased and should be interpreted only as detection within the bounded subset. The numerical diagnostics are summarised in Appendix Table A.5, while the complete reciprocal -sensitivity sweep and numerical-validity counts are reported in Appendix Figure A.11.
Triplet recovery across topologies
Recovery was more difficult than detection because the planted interactions had to be identified among all admissible candidates. As network size increased from to , the unordered search expanded from 4 to 560 candidate sets and the directed search from 12 to 1680 assignments. The main result was that recovery declined as the candidate space grew, even though the planted interaction remained detectable in § 4.2.1. Recovery was also lower when multiple interactions converged on the same target.
Figure 12 shows that, across network sizes, KSG S-information unordered mean average precision decreased from 0.992 at to 0.333 at . The combined high-S/low-O score declined less sharply, from 0.908 to 0.544, and at ranked the planted set first in 0.450 of replicates, compared with 0.200 for S-information alone.

Figure 12. Unordered-set recovery across topology regimes. Heatmaps show mean average precision at and and the change between them. The right column reports validity at ; scores are conditional on validity. Alt text: Heatmaps compare unordered-set recovery across topologies. Recovery improves with coupling for most three-node interactions, but declines in larger networks and with shared targets. Larger interactions show little improvement; reciprocal results use fewer valid trajectories.
As interaction multiplicity increased, recovery depended strongly on whether the targets were separate. For two interactions, high-S/low-O unordered mean average precision was 0.325 with separate targets but 0.078 with a shared target. For four interactions, the corresponding values were 0.334 and 0.028. Recovery with a shared source was similar to recovery with disjoint sources and targets. High-S/low-O unordered mean average precision was 0.358 when the interactions shared a source and 0.325 when they were disjoint, but decreased to 0.165 when both a source and the target were shared. Directed MMI-PID similarly reached 0.378 with a shared source, compared with 0.363 for disjoint interactions, but fell to 0.143 when both a source and the target were shared. The stronger O-information detection response under source overlap did not improve its recovery, which remained similar in the shared-source and disjoint layouts (0.205 and 0.200).
For the size-four and size-five interactions in Figure 12, unordered recovery remained near its matched-null level, consistent with their weak realised coupling and absent detectability. Directed recovery was not evaluated because the two-source MMI-PID decomposition does not extend directly to these larger source sets.
Finally, unordered recovery in the reciprocal configurations was assessed primarily at , where 49/60 trajectories in the equal-strength condition and all 60 in the normalised condition remained valid. KSG S-information achieved mean average precision values of 0.939 and 0.833, respectively. This strong performance reflects the model structure, as the three reciprocal terms act within the same triplet, , reinforcing its multivariate dependence without requiring the individual directions to be distinguished. Topology-specific null calibration further separates this higher-order signature from dependence generated by the pairwise backbone. The stronger equal-strength result is consistent with its larger combined higher-order contribution relative to the normalised condition.
Figure 13 shows that unconditional MMI-PID was the only directed measure to achieve meaningful recovery beyond the base-case and reciprocal layouts. Its mean average precision nevertheless declined from 0.780 at to 0.469 at . At , conditional MMI-PID and joint-source transfer entropy reached only 0.006 and 0.002, respectively, and remained at or near their null levels across the other topology regimes. Unconditional MMI-PID was also consistently more successful when interactions had separate rather than shared targets.

Figure 13. Directed recovery across topology regimes. Heatmaps show mean average precision at and and the change between them. The right column reports validity at ; grey cells indicate layouts for which directed recovery was not evaluated. Alt text: Heatmaps compare directed recovery across topologies. Unconditional MMI-PID shows the largest improvements with coupling, though recovery declines in larger networks and with shared targets. Reciprocal results use fewer valid trajectories; larger interactions were not evaluated.
The results provide the primary evidence for reciprocal recovery. At this coupling strength, 49/60 trajectories in the equal-strength condition and all 60 in the normalised condition remained valid. For directed recovery, MMI-PID reached 0.691 and 0.542, but all three planted directions occupied the first three ranks in only 0.265 and 0.050 of valid replicates. These results support reliable identification of the participating set and partial recovery of direction, but not consistent reconstruction of the complete reciprocal structure. The unordered and directed scores are not directly comparable because they evaluate different ranking tasks. Symmetric measures rank participating node sets, whereas target-specific measures rank directed source–target assignments over a larger candidate space.
Although the reciprocal scores increased at , they were based on only 17/60 trajectories in the equal-strength condition and 43/60 in the normalised condition. We did not establish that these valid trajectories were representative of those that diverged, so the apparent improvement may reflect selection bias. The values are therefore reported in the figures as conditional sensitivity results and are not interpreted as evidence that stronger coupling improved recovery. The result for the normalised condition, for which all trajectories remained valid, provides the strongest reciprocal evidence; the equal-strength estimate remains conditional on the 49/60 valid trajectories.
Overall, detection remained robust across network sizes, whereas both unordered and directed recovery became less reliable as the candidate space grew and when multiple interactions shared the same target.
Sensitivity to coupling function across interaction sizes
The product coupling used in the main experiments multiplies the states of all sources participating in an interaction. When this coupling is applied without an amplitude adjustment, its realised magnitude decreases as interaction size increases because it multiplies an increasing number of small-amplitude source states. We refer to this original implementation as the unscaled product. We repeated the detection and recovery experiments in the network for interactions containing three, four, or five nodes. The unscaled product was compared with an RMS-normalised product control and three other couplings: sourcewise tanh, sine, and smooth root, as defined in Section 3.3.5. These alternatives were also scaled to the RMS magnitude of the size-three unscaled product for each interaction size.
Here, RMS normalisation refers only to an offline calibration used to place the coupling-function outputs on a common amplitude scale. For each interaction size, source scales were first estimated from an independent ensemble to place the source states on comparable input scales. A separate output-amplitude constant was then estimated for each coupling function and interaction size so that its coupling-output RMS was the same as that of the size-three unscaled product. These constants were fixed before the evaluation simulations and were not adjusted across trajectories or time. The higher-order term was therefore not dynamically normalised in the state equation: , the planted interaction structure, and all other model terms remained unchanged. The unscaled product retained its original form and natural interaction-size-dependent amplitude. This design separates sensitivity to coupling form from the attenuation produced by multiplying additional small-amplitude sources.
At size three, KSG mutual information and KSG S-information detected every coupling condition in all replicates (Figure 14(a)). The unscaled and RMS-normalised products were identical at this size because the size-three unscaled product defined the common amplitude reference. KSG O-information was more sensitive to coupling form, increasing from 0.68 for the two product conditions to 0.88–0.93 for the other couplings.

Figure 14. Detection across coupling functions in the network at . Panels (a)–(c) correspond to interactions containing three, four, and five nodes. Rows identify the information-theoretic measure and columns identify the coupling function. Cells show c-sensitivity across 40 test replicates using thresholds estimated from 200 independent simulations. The unscaled product retains its natural interaction-size-dependent amplitude. For each interaction size, the RMS-normalised product and the three other couplings use source scales and fixed output-amplitude constants calibrated to the RMS magnitude of the size-three unscaled product, as described in Section 3.3.5.
Alt text: Three heatmaps compare detection across coupling functions and interaction sizes. KSG mutual information and S-information detect three-node interactions reliably. For four- and five-node interactions, amplitude rescaling restores strong detection lost with the unscaled product.
The effect of RMS normalisation became clear at larger interaction sizes in Figure 14. For the unscaled product, KSG S-information c-sensitivity fell to 0.05 at size four and 0.03 at size five. By contrast, it reached 1.00 at size four and 0.97–1.00 at size five for the RMS-normalised product and the three other couplings. KSG mutual information also reached 1.00 for each of these four conditions at both sizes. These results show that the apparent loss of detectability for larger product interactions arose primarily from the interaction-size-dependent attenuation of the unscaled product rather than from an intrinsic inability of the measures to detect interactions involving larger sets.
Recovery remained more difficult than detection (Figure 15). At size three, the combined high-S/low-O score achieved a mean reciprocal rank of 0.544 for both product conditions, whereas the other couplings increased it to 0.662–0.664. Additional summaries of the same candidate rankings, which are not displayed in the figure, gave top-1 recovery rates of 0.450 for the product conditions and 0.583 for each of the other couplings.

Figure 15. Recovery of the planted unordered interaction in the network at . Panels (a)–(c) correspond to interaction sizes three, four, and five; is the number of candidate sets. Rows identify the ranking measure and columns identify the coupling function. Cells show mean reciprocal rank across 60 replicates using candidate-specific calibrations. Higher values indicate better localisation, and the common colour scale is logarithmic. Because each condition contains one planted interaction, mean reciprocal rank equals mean average precision. Alt text: Three heatmaps compare recovery across coupling functions. The combined score performs best for three-node interactions. Recovery scores are much lower for four- and five-node interactions, although alternative couplings improve many rankings.
At size four, RMS normalisation increased the mean reciprocal rank of the combined score from 0.0019 for the unscaled product to 0.0065 for the RMS-normalised product. The other couplings produced larger improvements, reaching 0.017–0.023. Exact top-1 recovery was nevertheless absent apart from one unscaled-product replicate under HOI kNN O-information. With the smooth-root coupling, the combined score placed the planted set at a median rank of 104.5 among 1,820 candidates. At size five, no method achieved top-1 recovery; the highest mean reciprocal rank was 0.0079, obtained by KSG S-information with the smooth-root coupling.
Scaling the coupling outputs to a common RMS on the ensemble restored high detection rates for KSG mutual information and KSG S-information at interaction sizes four and five. Alternative coupling functions also improved some recovery rankings. First-rank recovery remained rare or absent for these larger interactions. The common RMS was imposed on the scaling ensemble, and realised contributions at non-zero coupling may differ. Interaction size and the number of candidate sets increased together, so these comparisons reflect their combined effects on recovery. Across the benchmark, detecting statistical changes induced by specified higher-order network interactions was more successful than recovering their participating nodes and source-target roles.
Conclusion
We introduced a generative model of stochastic network dynamics and a framework for evaluating information-theoretic measures against higher-order interactions. Across coupling functions, network topologies and interaction sizes, the experiments assess detection of changes relative to a pairwise-only reference and recovery of the participating nodes. Directed recovery additionally evaluates source-target roles for three-node interactions. The framework also tests whether the planted higher-order coupling can be distinguished from a specified decomposable nonlinear pairwise alternative. Used independently, the model can generate continuous multivariate time series with controlled pairwise and higher-order structure. The framework allows new measures, mechanisms and network configurations to be tested using the same experimental setup.
Detection and recovery experiments produced different results. At sufficiently strong coupling, KSG mutual information and KSG S-information achieved high detection rates across the tested coupling functions and network sizes. Scaling coupling outputs to a common RMS on the ensemble also supported detection of interactions with more sources. Recovery was less successful in larger candidate spaces and in shared-target layouts. Identifying source–target assignments was harder than identifying the participating nodes. In the base case, KSG O-information had lower detection rates than KSG S-information. In the discrimination experiments, the measures separated the product coupling from the specified nonlinear additive control. False-positive rates were close to the 5% level in the independent validation at , and exceeded this level for some measures in the across- experiment. Together, these results show that statistical signatures can respond to a planted higher-order mechanism and distinguish it from a specified nonlinear pairwise alternative, while recovering its participating nodes and direction remains more challenging, particularly as the candidate space grows or multiple interactions share a target.
These conclusions are limited to the conditions tested. All model states were observed, with process noise and no observation model. Detection and recovery used a known pairwise-only reference, and discrimination used the specified nonlinear additive control. Performance summaries for reciprocal layouts were conditional on numerical validity. The higher-order mechanisms examined were built by multiplying separate functions of individual source states, which leaves out interactions defined through a non-factorised joint function or a collective threshold rule, and the estimators were used with fixed, prespecified settings. The results should therefore be read as outcomes for the mechanisms and settings examined here rather than as a general ranking of higher-order detection methods. The modular construction of the framework supports extensions in each of these directions, including partial observation, measurement noise, learned rather than known pairwise and higher-order references, further coupling functions, and larger networks with structured candidate searches. Because the planted structure and the corresponding pairwise-only reference are retained, such extensions can test which aspects of a higher-order mechanism remain detectable and recoverable as the setting becomes more realistic.
For empirical time-series analysis, our results highlight the distinction between detecting statistical signatures and identifying the interactions that generated them. Validation frameworks with known generating mechanisms are therefore essential for assessing whether inferred interactions correspond to those actually present in the simulated system. Our results suggest that KSG S-information may help identify candidate groups of interacting nodes, while the combined S/O score may improve recovery in larger candidate spaces. To guide inference for a particular type of data, the model should be adapted to reflect the relevant properties of the system and its recordings, allowing measures to be validated under conditions appropriate to that application. Future work could examine whether, and under what conditions, other statistical measures or inference models improve recovery across different types of synthetic data. Adding measurement models and developing model-inversion methods could provide a route to estimating interaction structures and their parameters from observed signals. Together, these developments could help establish when statistical signatures support reliable conclusions about the mechanisms underlying different types of empirical time series data.
Funding
This work was supported by the Digital Sciences for Society of Tilburg University [DSFS 2023IC0N006]. KVH gratefully acknowledges funding from the Dutch Research Council [NWO; VIDI grant no. 09150171910043].
Acknowledgements
OpenAI Codex, ChatGPT, and Claude were used as supporting tools for manuscript revision, code development, numerical consistency checks, and submission preparation. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the article.
Data and code availability
This study uses synthetic time series data generated by the model described in this article. The code for the model and experimental pipeline is maintained at https://github.com/ValentinaSanchezMelchor/higher-order-network-dynamics.
Conflict of interest
The authors declare no conflicts of interest.
Background on the model and information-theoretic measures
DCM neural state equation
Dynamic causal modelling (DCM) was introduced by Friston et al. [10] as a Bayesian framework for inferring effective connectivity between brain regions from neuroimaging data. It combines a model of neural dynamics with an observation model linking neural activity to measured signals. Smith et al. [32] used a DCM neural state equation coupled to a haemodynamic model to generate synthetic fMRI time series with known connectivity. We use their neural state formulation as the starting point for our stochastic network model.
In the formulation used by Smith et al. [32], the neural state equation is
where contains the neural states of regions and is its time derivative. The matrix has diagonal entries , representing decay within each node, and off-diagonal entries that specify the coupling from source to target . The coefficient scales the decay and pairwise coupling rates. Without pairwise coupling or input, a node decays with time constant . The vector contains inputs, and determines how these inputs affect each node. In their simulations, the inputs switched between two levels according to a Poisson process, with added neural noise.
We analyse the simulated node states , without a haemodynamic observation model. We specify the generating parameters so that we can focus on evaluating the information-theoretic measures under known interaction mechanisms, without introducing the additional statistical inference problem of estimating those parameters using DCM. In our model, self-decay is represented by and pairwise coupling by a matrix with . The symbol therefore has a different meaning from the matrix in Equation A.1, which includes self-decay. The linear state-dependent terms agree when and the off-diagonal entries of our coupling matrix equal times the corresponding off-diagonal entries of the DCM matrix. We allow and the pairwise coupling strengths to be specified separately, add higher-order couplings and additive process noise, and use Ornstein-Uhlenbeck inputs in the simulations. The stochastic state equation is given in § 3.1, and the input processes are described in § 3.3.1.
Higher-order mechanisms and directed representations
Higher-order descriptions of a system can refer to its network structure, its coupling functions, or statistical dependence in its outputs [6, 27]. A hyperedge identifies a group of participating nodes, but even when it joins three or more nodes, the associated coupling can be a sum of pairwise contributions [21]. The number of participating nodes therefore does not determine whether the coupling requires a higher-order description. In this paper, we use the non-additivity criterion in Equation (1) to identify higher-order mechanisms in the original node states. Directed hyperedges represent the participating sources and target, while information-theoretic measures characterise the statistical dependence in the resulting time series.
The functions and in Equation (1) may be non-linear, so the criterion concerns additivity [14]. When that decomposition exists, and describe separate pairwise contributions from sources and to target . For example, the coupling is non-linear but additive in the source states and can be represented by two non-linear pairwise terms. For the product coupling , a non-zero increment in changes the coupling by . The effect of changing one source therefore depends on the state of the other, which prevents an additive decomposition into separate source contributions.
To specify where a joint source contribution acts, we use the 1-directed hypergraph convention of Gallo et al. [11], in which each hyperedge has a single target. Let denote the node set. A directed triplet is represented by
where is the target and is the unordered source pair. Exchanging the source labels leaves the hyperedge unchanged, while choosing a different target defines a different directed interaction. This extends the convention for , which represents a contribution from source to target . The direction identifies the state equation receiving the joint contribution. The same interaction can also be represented as a focal interaction with support and focal node in the framework of Moutuou [19].
Information-theoretic measures describe properties of the joint distribution of the variables being analysed. This distribution reflects the combined effects of the network topology, coupling functions, interaction strengths, inputs, and process noise. Pairwise mechanisms can produce multivariate dependence beyond that captured by pairwise summaries [27]. A statistical signature therefore does not uniquely identify a higher-order mechanism, and a weak response from a measure does not establish that such a mechanism is absent. In our approach, comparisons with the pairwise-only reference assess changes induced by the added higher-order term. The separate comparison with a non-linear additive control tests whether a measure can discriminate between those two coupling mechanisms.
Information-theoretic measures
We characterise statistical dependence in the simulated node states using the information-theoretic measures summarised in Table A.1. The table gives their definitions, interpretations, labels used in the figures, and roles in the experiments. It also includes the combined S/O score, a ranking statistic for candidate node sets. Estimator settings and implementation details are given in § 3.3.7.
Let denote a random vector of node states, and let be the vector obtained by removing component . For continuous random variables or vectors , , and , and denote differential entropy and conditional differential entropy, respectively [8]. The entropy expressions below assume that the relevant densities exist and that all entropy terms are finite. Mutual information is , and conditional mutual information is . Total correlation is [36], and dual total correlation is [12].
O-information is defined as [26]. Positive values indicate redundancy-dominated dependence, and negative values indicate synergy-dominated dependence. Dependent variables can also have , so a zero value does not imply mutual independence. S-information is defined as and summarises overall statistical dependence, including pairwise contributions. It is non-negative and equals zero if and only if the components of are mutually independent.
Streiberg information is constructed using all partitions of the analysed variables [15]. Let denote the set of partitions of , where each partition consists of non-empty, disjoint blocks whose union is . For a block , let denote the joint marginal distribution of the states indexed by , and write and . Streiberg information is defined as , where and is the number of blocks. The Tsallis divergence is , where and are the densities of and ; we use . If the variables can be partitioned into at least two mutually independent groups, their joint distribution factorises as and Streiberg information is zero at the population level.
| Quantity / figure labels | Definition | Interpretation | Use |
| Mutual information [8, 31]<br>Gaussian MI; KSG MI | Larger values indicate greater pairwise dependence | Detection;<br>discrimination | |
| O-information [26]<br>Gaussian O-info; KSG O-info; HOI kNN O-info | : redundancy-dominated dependence; : synergy-dominated dependence | Detection;<br>discrimination;<br>set recovery | |
| S-information [26]<br>Gaussian S-info; KSG S-info | Larger values indicate greater overall dependence, including pairwise contributions | Detection;<br>discrimination;<br>set recovery | |
| Streitberg information [15]<br>Streitberg (Tsallis-kNN) | Zero if the variables can be partitioned into two or more mutually independent groups | Detection;<br>discrimination | |
| MMI-PID synergy [2, 37]<br>KSG MMI-PID; Conditional KSG MMI-PID | <br>Conditional form: replace each by | Additional target information from the joint sources relative to the more informative source | Directed<br>recovery |
| Joint-source transfer entropy [23, 30]<br>KSG joint-source TE | Information about the target’s next state from the source pair, conditional on the target’s current state | Directed<br>recovery | |
| Combined S/O score<br>KSG high-S / low-O | Combines upper-tail S-information and lower-tail O-information departures from the pairwise-only reference | Set recovery |
Table A.1. Definitions, interpretations, figure labels, and experimental roles of the information-theoretic measures and combined S/O ranking score. Interpretations of the measures refer to population quantities. Italic text gives the figure labels. Gaussian labels refer to covariance-based estimates, KSG to Kraskov-Stögbauer-Grassberger estimation, and kNN to -nearest neighbours. HOI identifies the software package. In the “Use” column, and indicate the upper and lower reference tails used for recovery. Detection and discrimination use two-sided comparisons.
Partial information decomposition (PID) separates the mutual information between a joint source vector and a target into redundant information, information unique to each source, and synergy [37]. We use the minimum mutual information (MMI) definition of redundancy, [2]. The resulting synergy is . It quantifies the additional information about the target obtained by considering both sources together, relative to the more informative source. For directed recovery, we use , , and , where is the integration step. The conditional version uses the current target state and is defined by .
Using the same source and target variables, the one-step joint-source transfer entropy (TE) is defined as [23, 30]. It quantifies the information about the target’s next state available from the current source pair, conditional on the target’s current state. This quantity includes predictive information available from single sources as well as conditional synergy. Its direction refers to temporal prediction, and a non-zero value does not establish a causal interaction.
For unordered-set recovery, we combine estimates of S- and O-information using the ranking score
The quantities and are empirical upper-tail probabilities for S-information and lower-tail probabilities for O-information, respectively, obtained from pairwise-only reference simulations for each node set. The score increases as either tail probability decreases, combining departures towards higher S-information and lower O-information. For each node set, we calculate an empirical null probability for using its reference distribution, which incorporates dependence between the S- and O-information estimates. We rank the sets by these probabilities, following the procedure in Section 3.3.4.
Base-model validation
This appendix provides the numerical results supporting the model-validation checks reported in Section 3.3.2. The convergence analysis evaluates the selected integration timestep, while the sensitivity analysis tests whether the base model remains numerically finite and retains a dynamically substantial higher-order contribution under perturbations of its pairwise and input parameters and of its process-noise intensity.
Timestep convergence

Figure A.1. Timestep convergence of the base model. Panel (a) shows the target-state RMSE at and , relative to the reference solution at . Small points show the 12 individual replicates, large points show the medians, and error bars span the 10th to 90th percentiles. The reference step is omitted because its error is zero by construction. Panel (b) shows the target-state difference from the reference solution for one representative replicate after warm-up. Initial states, OU paths, and Wiener paths were reused across timesteps, with fine-scale Wiener increments aggregated to construct the corresponding coarse increments.
Alt text: (a) Target-state errors decrease with a smaller integration timestep, both with and without higher-order coupling. (b) The smaller timestep generally produces smaller deviations from the reference trajectory over time.
| Finite runs | All-state RMSE (%) | Target-state RMSE (%) | ||
| 0 | 0.02 | 12/12 | 2.631 [2.532, 2.689] | 1.892 [1.519, 2.215] |
| 0 | 0.01 | 12/12 | 1.087 [1.053, 1.147] | 0.788 [0.711, 1.000] |
| 20 | 0.02 | 12/12 | 2.636 [2.538, 2.697] | 3.034 [2.445, 3.501] |
| 20 | 0.01 | 12/12 | 1.092 [1.057, 1.149] | 1.189 [1.059, 1.532] |
Table A.2. Timestep-convergence results. RMSE values are expressed as percentages of the corresponding reference RMS at and are reported as the median with the 10th-90th percentile interval across 12 replicates.
Parameter sensitivity

Figure A.2. Sensitivity of the base model at . The columns vary the pairwise spectral fraction , the input correlation time relative to its base value , the common input gain relative to its base value , and the process-noise intensity relative to its base value . Panels (a)–(d) show the median target-node higher-order-to-decay and higher-order-to-pairwise drift RMS ratios. The dashed line marks a ratio of one. Panels (e)–(h) show the median zero-lag Pearson correlation between the two source states. Shaded regions span the 10th to 90th percentiles across 30 replicates.
Alt text: Top panels show higher-order drift exceeding pairwise and decay contributions across parameter variations. Bottom panels show source correlation increasing with pairwise strength and input timescale, while remaining nearly unchanged with input gain and process noise.
| Condition | Setting | Linear margin | Finite runs | Finite runs |
| (a) Numerical validity | ||||
| Baseline | , , , | 0.65 | 30/30 | 30/30 |
| Pairwise strength | 0.80 | 30/30 | 30/30 | |
| Pairwise strength | 0.50 | 30/30 | 30/30 | |
| Input correlation time | 0.65 | 30/30 | 30/30 | |
| Input correlation time | 0.65 | 30/30 | 30/30 | |
| Input gain | 0.65 | 30/30 | 30/30 | |
| Input gain | 0.65 | 30/30 | 30/30 | |
| Process noise | 0.65 | 30/30 | 30/30 | |
| Process noise | 0.65 | 30/30 | 30/30 | |
| Process noise | 0.65 | 30/30 | 30/30 | |
| Condition | Setting | HO/decay | HO/pairwise | |
| (b) Dynamics at | ||||
| Baseline | , , , | 0.366 [0.277, 0.473] | 1.287 [1.215, 1.378] | 4.810 [3.683, 7.099] |
| Pairwise strength | 0.219 [0.117, 0.320] | 1.391 [1.279, 1.475] | 7.516 [5.706, 10.275] | |
| Pairwise strength | 0.517 [0.439, 0.620] | 1.199 [1.141, 1.303] | 4.352 [3.167, 6.270] | |
| Input correlation time | 0.338 [0.258, 0.446] | 1.303 [1.214, 1.413] | 3.548 [2.759, 5.070] | |
| Input correlation time | 0.392 [0.305, 0.511] | 1.226 [1.170, 1.284] | 6.413 [4.663, 9.883] | |
| Input gain | 0.367 [0.276, 0.473] | 1.089 [0.977, 1.221] | 2.405 [1.846, 3.576] | |
| Input gain | 0.367 [0.277, 0.472] | 1.311 [1.244, 1.386] | 7.218 [5.521, 10.624] | |
| Process noise | 0.367 [0.277, 0.472] | 1.303 [1.215, 1.367] | 4.814 [3.671, 7.046] | |
| Process noise | 0.367 [0.277, 0.472] | 1.300 [1.222, 1.376] | 4.812 [3.680, 7.072] | |
| Process noise | 0.367 [0.276, 0.473] | 1.253 [1.166, 1.345] | 4.812 [3.693, 7.154] |
Table A.3. Condition-wise parameter-sensitivity results. Finite runs are reported separately for and . A run was considered finite when all states were finite and . The linear margin is . In panel (b), is the zero-lag Pearson correlation between the source states. HO/decay and HO/pairwise are the ratios of the higher-order drift RMS to the decay and pairwise drift RMS, respectively, at target node 0. These three quantities are evaluated at and reported as the median with the 10th-90th percentile interval across 30 replicates. Here, , , and .
Supplementary figures and tables

Figure A.3. Post-warm-up trajectories at (a) and (b) for one replicate. Initial states, OU inputs, and Wiener paths are identical across conditions. The higher-order term enters only the target equation directly, producing larger target excursions; indirect changes in the other nodes are negligible at the plotted scale. Alt text: Two time-series panels compare the same replicate without and with higher-order coupling. The target shows larger excursions with coupling.
Figure A.4. C-sensitivity for each detection metric and higher-order coupling strength in the four-node base case. The heatmap presents the same values as Figure 3 while making the metric-by- comparisons explicit. Alt text: Detection heatmap showing the strongest increases for KSG mutual information and S-information as coupling grows.
Table A.4. KSG O-information of the planted triplet across the coupling sweep in the four-node base case. Values are taken over the 40 test replicates at each . For replicate , the paired difference is , where denotes the KSG estimate of O-information. Each pair shares the same initial state, input paths and Wiener paths. The More negative column reports the number of replicates with out of the 40 test replicates. The final column reproduces the null-calibrated -sensitivity of Figure 3. Every value is negative at every coupling strength, as are all 200 simulations of the independent calibration ensemble.
| Mean | Range | Median | More negative | -sensitivity | |
| 0 | 0/40 | 0.125 | |||
| 0.2 | 20/40 | 0.100 | |||
| 0.5 | 18/40 | 0.100 | |||
| 1 | 23/40 | 0.100 | |||
| 2 | 25/40 | 0.125 | |||
| 5 | 30/40 | 0.175 | |||
| 10 | 37/40 | 0.400 | |||
| 20 | 38/40 | 0.575 | |||
| 30 | 38/40 | 0.625 |
Table A.4.
Table A.5. Search-space dimensions and numerical diagnostics for the topology-generalisation regimes at . is the number of directed interactions, is the total number of nodes in each interaction including its target, is the number of unordered candidate node sets, and is the number of directed source-set–target candidates. In each reciprocal configuration, the three directed interactions correspond to one unordered node set. Detection and recovery validity are reported as valid/requested trajectories. The final column gives the mean, across valid detection trajectories, of the higher-order-to-process-noise increment RMS ratio, with both RMS quantities evaluated over the planted target nodes. Directed recovery was not evaluated for the size-four and size-five interactions. The planted interactions and coefficient normalisations are defined in Table 2.
| Regime | Detection valid | Recovery valid | HO/noise RMS | |||||
| Network size | ||||||||
| Directed base case | 4 | 1 | 3 | 4 | 12 | 40/40 | 60/60 | 0.996 |
| Eight-node network | 8 | 1 | 3 | 56 | 168 | 40/40 | 60/60 | 0.994 |
| Sixteen-node network | 16 | 1 | 3 | 560 | 1680 | 40/40 | 60/60 | 0.994 |
| Interaction multiplicity | ||||||||
| Two interactions, shared target | 16 | 2 | 3 | 560 | 1680 | 40/40 | 60/60 | 1.45 |
| Two interactions, disjoint targets | 16 | 2 | 3 | 560 | 1680 | 40/40 | 60/60 | 0.969 |
| Four interactions, shared target | 16 | 4 | 3 | 560 | 1680 | 40/40 | 60/60 | 2.17 |
| Four interactions, separate targets | 16 | 4 | 3 | 560 | 1680 | 40/40 | 60/60 | 0.941 |
| Overlap and interaction size | ||||||||
| Shared source | 16 | 2 | 3 | 560 | 1680 | 40/40 | 60/60 | 0.934 |
| Shared source and target | 16 | 2 | 3 | 560 | 1680 | 40/40 | 60/60 | 1.45 |
| Size-four interaction | 16 | 1 | 4 | 1820 | — | 40/40 | 60/60 | 0.0104 |
| Size-five interaction | 16 | 1 | 5 | 4368 | — | 40/40 | 60/60 | |
| Reciprocal configurations | ||||||||
| Reciprocal, equal strength | 4 | 3 | 3 | 4 | 12 | 10/40 | 17/60 | 0.839 |
| Reciprocal, normalised | 4 | 3 | 3 | 4 | 12 | 32/40 | 43/60 | 0.365 |
Table A.5.
Sampling and calibration checks for discrimination
Figures A.5 and A.6 provide the sampling and calibration checks supporting Section 3.3.6. The sampling comparison changes both recording duration and observation spacing; the calibration comparison evaluates two reference ensembles on the same held-out simulations.

Figure A.5. Baseline S-information estimates across six replicates per condition. (a) Independent multivariate Gaussian observations (4501 observations), dense trajectories (, 4501 observations), and widely spaced trajectories (, 999 observations, spacing 5 time units). (b) Expanded view of the independent and widely spaced conditions. Points show individual estimates and short horizontal lines their means. The dashed line marks the analytic stationary S-information of the discretised model, 0.1657 nats (natural units of information). Agreement under wider spacing supports the discrimination protocol, without establishing unbiased estimation under nonzero coupling. Alt text: Dense sampling inflates KSG S-information; independent and widely spaced observations give estimates near the analytic baseline.

Figure A.6. Calibration comparison at using (a) the original 200-run additive reference and (b) an independent 1000-run additive reference. Both are evaluated on the same 400 held-out simulations per coupling form. Orange diamonds show the fraction of product simulations outside the reference interval; blue circles show false-positive rates on additive controls. Bars show Wilson 95% intervals conditional on the fitted thresholds; dotted lines mark the nominal 5% level. The comparison concerns these two reference ensembles, rather than a systematic assessment of calibration size. Alt text: Both reference ensembles yield high discrimination; false-positive rates are closer to five percent with the independent reference.

Figure A.7. Whole-network component RMS increments and numerical validity at , evaluated using 10 diagnostic simulations per layout, separate from the detection and recovery ensembles. Markers show means across numerically valid diagnostic trajectories, the dashed line marks the process-noise RMS, and the right panel reports valid/requested diagnostic trajectory counts.
Alt text: Higher-order increments shrink sharply for larger interactions; only reciprocal layouts have numerically invalid diagnostic trajectories.
Figure A.8. Null-calibrated c-sensitivity across the network-size regimes. Panels show the four-node directed base case and the corresponding eight- and sixteen-node configurations. Rows denote the seven detection metrics, columns denote the higher-order coupling strength , and each cell reports the fraction of valid test trajectories that exceeded the corresponding topology-specific null threshold. All 40 requested trajectories were numerically valid at every coupling strength in these regimes.
Alt text: Three heatmaps show similar detection patterns across network sizes, with KSG mutual information and S-information reaching full detection at strong coupling.
Figure A.9. Null-calibrated c-sensitivity across the interaction-multiplicity regimes. The panels compare two and four planted interactions under shared-target and separate-target configurations. Rows denote detection metrics and columns denote . For each metric and coupling strength, c-sensitivity was calculated separately for each planted interaction and then averaged across the interactions. All 40 requested trajectories were numerically valid in each condition.
Alt text: Four heatmaps compare detection of multiple interactions, showing weaker detection when four interactions share a target.
Figure A.10. Null-calibrated c-sensitivity across the interaction-overlap and interaction-size regimes. Panels (a) and (b) compare interactions sharing a source with interactions sharing both a source and their target. Panels (c) and (d) show single raw-product interactions containing four and five nodes, respectively, including the target. Rows denote detection metrics and columns denote . Values in the multi-interaction conditions are means across the separately retained interaction-level results. All 40 requested trajectories were numerically valid in each condition.
Alt text: Heatmaps show detection increasing with coupling for overlapping interactions, but remaining low for four- and five-node unscaled-product interactions.

Figure A.11. Null-calibrated c-sensitivity and numerical validity in the reciprocal configurations. Panels (a) and (b) show mean interaction-level c-sensitivity under equal local interaction strength and a total coefficient norm of , respectively. Rows denote detection metrics and columns denote . Panel (c) reports the number of numerically valid trajectories out of the 40 requested trajectories. Because c-sensitivity is calculated conditional on numerical validity, values at the largest coupling strengths are based on progressively smaller retained subsets, particularly in the equal-strength condition. Alt text: Heatmaps show stronger detection among valid simulations as coupling increases, but fewer simulations remain numerically valid, especially in the equal-strength configuration.
References
- [1]Antonacci, Y., Barà, C., Sparacino, L., Marinazzo, D., Faes, L. & Stramaglia, S. (2026) From Statistical to Structural Synergy: A Predictability Framework to Quantify the Effects due to High-Order Mechanisms. arXiv preprint arXiv:2607.07286.arxiv.org/abs/2607.07286
- [2]Barrett, A. B. (2015) Exploration of synergistic and redundant information sharing in static and dynamical Gaussian systems. Physical Review E, 91(5), 052802.pubmed.ncbi.nlm.nih.gov/26066207
- [3]Battiston, F., Amico, E., Barrat, A., Bianconi, G., Ferraz de Arruda, G., Franceschiello, B., Iacopini, I., Kéfi, S., Latora, V., Moreno, Y., Murray, M. M., Peixoto, T. P., Vaccarino, F. & Petri, G. (2021) The Physics of Higher-Order Interactions in Complex Systems. Nature Physics, 17(10), 1093–1098.DOI
- [4]Battiston, F., Centcetti, G., Iacopini, I., Latora, V., Lucas, M., Patania, A., Young, J.-G. & Petri, G. (2020) Networks beyond Pairwise Interactions: Structure and Dynamics. Physics Reports, 874, 1–92.DOI
- [5]Bianconi, G. (2021) Higher-Order Networks: An Introduction to Simplicial Complexes. Cambridge University Press, Cambridge.
- [6]Bick, C., Gross, E., Harrington, H. A. & Schaub, M. T. (2023) What Are Higher-Order Networks?. SIAM Review, 65(3), 686–731.DOI
- [7]Bullmore, E. E. & Sporns, O. (2009) Complex Brain Networks: Graph Theoretical Analysis of Structural and Functional Systems. Nature Reviews Neuroscience, 10(3), 186–198.DOI
- [8]Cover, T. M. & Thomas, J. A. (2006) Elements of Information Theory. Wiley-Interscience, Hoboken, NJ, 2nd edition.
- [9]Estrada, E. (2011) The Structure of Complex Networks: Theory and Applications. Oxford University Press, Oxford.DOI
- [10]Friston, K. J., Harrison, L. & Penny, W. (2003) Dynamic Causal Modelling. NeuroImage, 19(4), 1273–1302.pubmed.ncbi.nlm.nih.gov/12948688
- [11]Gallo, L., Muolo, R., Gambuzza, L. V., Latora, V., Frasca, M. & Carletti, T. (2022) Synchronization Induced by Directed Higher-Order Interactions. Communications Physics, 5, 263.DOI
- [12]Han, T. S. (1978) Nonnegative Entropy Measures of Multivariate Symmetric Correlations. Information and Control, 36(2), 133–156.DOI
- [13]Lacasa, L. (2026) On the Equivalence between Nonlinear Graph-Based Dynamics and Linear Dynamics on Higher-Order Networks. arXiv preprint arXiv:2602.21727.arxiv.org/abs/2602.21727
- [14]Lim, M. & Hastie, T. (2015) Learning Interactions via Hierarchical Group-Lasso Regularization. Journal of Computational and Graphical Statistics, 24(3), 627–654.pubmed.ncbi.nlm.nih.gov/26759522
- [15]Liu, Z., Barahona, M. & Peach, R. (2025) Information-Theoretic Measures on Lattices for Higher-Order Interactions. In Proceedings of the 28th International Conference on Artificial Intelligence and Statistics, volume 258 of Proceedings of Machine Learning Research, pages 2206–2214. PMLR.DOI
- [16]Lizier, J. T. (2014) JIDT: An Information-Theoretic Toolkit for Studying the Dynamics of Complex Systems. Frontiers in Robotics and AI, 1, 11.DOI
- [17]Luppi, A. I., Mediano, P. A. M., Rosas, F. E., Holland, N., Fryer, T. D., O’Brien, J. T., Rowe, J. B., Menon, D. K., Bor, D. & Stamatakis, E. A. (2022) A Synergistic Core for Human Brain Evolution and Cognition. Nature Neuroscience, 25(6), 771–782.
- [18]Luppi, A. I., Rosas, F. E., Mediano, P. A. M., Menon, D. K. & Stamatakis, E. A. (2024) Information Decomposition and the Informational Architecture of the Brain. Trends in Cognitive Sciences, 28(4), 352–368.DOI
- [19]Moutouou, E. M. (2026) Graphs as Focal Interaction Systems in Higher-Order Dynamics. arXiv preprint arXiv:2603.04215. Accepted for publication in Journal of Physics: Complexity.
- [20]Neri, M., Vinchi, D., Ferreyra, C., Robiglio, T., Ates, O., Ontivero-Ortega, M., Brovelli, A., Marinazzo, D. & Combrisson, E. (2024) HOI: A Python Toolbox for High-Performance Estimation of Higher-Order Interactions from Multivariate Data. Journal of Open Source Software, 9(103), 7360.DOI
- [21]Neuhäuser, L., Scholkemper, M., Tudisco, F. & Schaub, M. T. (2024) Learning the Effective Order of a Hypergraph Dynamical System. Science Advances, 10(19), eadh4053.
- [22]Newman, M. E. J. (2010) Networks: An Introduction. Oxford University Press, Oxford.DOI
- [23]Novelli, L., Wollstadt, P., Mediano, P., Wibral, M. & Lizier, J. T. (2019) Large-Scale Directed Network Inference with Multivariate Transfer Entropy and Hierarchical Statistical Testing. Network Neuroscience, 3(3), 827–847.
- [24]Otter, N., Porter, M. A., Tillmann, U., Grindrod, P. & Harrington, H. A. (2017) A Roadmap for the Computation of Persistent Homology. EPJ Data Science, 6(1), 17.
- [25]Robiglio, T., Neri, M., Cognolles, D., Agostinelli, C., Battiston, F., Lucas, M. & Petri, G. (2025) Synergistic Signatures of Group Mechanisms in Higher-Order Systems. Physical Review Letters, 134, 137401.DOI
- [26]Rosas, F. E., Mediano, P. A. M., Gastpar, M. & Jensen, H. J. (2019) Quantifying High-Order Interdependencies via Multivariate Extensions of the Mutual Information. Physical Review E, 100(3), 032305.DOI
- [27]Rosas, F. E., Mediano, P. A. M., Luppi, A. I., Varley, T. F., Lizier, J. T., Stramaglia, S., Jensen, H. J. & Marinazzo, D. (2022) Disentangling High-Order Mechanisms and High-Order Behaviours in Complex Systems. Nature Physics, 18(5), 476–477.DOI
- [28]Santoro, A., Battiston, F., Petri, G. & Amico, E. (2023) Higher-Order Organization of Multivariate Time Series. Nature Physics, 19(2), 221–229.DOI
- [29]Santoro, A., Neri, M., Poetto, S., Orsenigo, D., Diano, M., Gatica, M. & Petri, G. (2026) Charting higher-order models of brain function beyond pairwise interactions. Nature Communications, 17, 9207.
- [30]Schreiber, T. (2000) Measuring Information Transfer. Physical Review Letters, 85(2), 461–464.DOI
- [31]Shannon, C. E. (1948) A Mathematical Theory of Communication. Bell System Technical Journal, 27(3), 379–423.
- [32]Smith, S. M., Miller, K. L., Salimi-Khorshidi, G., Webster, M., Beckmann, C. F., Nichols, T. E., Ramsey, J. D. & Woolrich, M. W. (2011) Network Modelling Methods for fMRI. NeuroImage, 54(2), 875–891.DOI
- [33]Stephan, K. E., Kasper, L., Harrison, L. M., Daunizeau, J., den Ouden, H. E. M., Breakspear, M. & Friston, K. J. (2008) Nonlinear Dynamic Causal Models for fMRI. NeuroImage, 42(2), 649–662.DOI
- [34]Varley, T. F., Mediano, P. A. M., Patania, A. & Bongard, J. (2025) The Topology of Synergy: Linking Topological and Information-Theoretic Approaches to Higher-Order Interactions in Complex Systems. PLOS Computational Biology, 21(11), e1013649.arxiv.org/abs/2504.10140
- [35]Varley, T. F., Pope, M., Faskowitz, J. & Sporns, O. (2023) Multivariate Information Theory Uncovers Synergistic Subsystems of the Human Cerebral Cortex. Communications Biology, 6(1), 451.DOI
- [36]Watanabe, S. (1960) Information Theoretical Analysis of Multivariate Correlation. IBM Journal of Research and Development, 4(1), 66–82.DOI
- [37]Williams, P. L. & Beer, R. D. (2010) Nonnegative Decomposition of Multivariate Information. arXiv preprint arXiv:1004.2515.arxiv.org/abs/1004.2515
- [38]Wollstadt, P., Lizier, J. T., Vicente, R., Finn, C., Martínez-Zarzuela, M., Mediano, P., Novelli, L. & Wibral, M. (2019) IDTxl: The Information Dynamics Toolkit xl: A Python Package for the Efficient Analysis of Multivariate Information Dynamics in Networks. Journal of Open Source Software, 4(34), 1081.




