Structured O-Information

• Structured O-information is a refined measure that decomposes classical O-information into contributions from explicit structures like groups, local patterns, and gradients.
• It quantifies the balance of redundancy versus synergy within multivariate systems, indicating dominant interaction types by comparing total and dual total correlations.
• The framework supports localized and dynamical analyses, aiding applications from neural and economic data to complex system modeling by isolating high-order interactions.

Searching arXiv for the cited papers and closely related O-information work to ground the article in the literature. Structured O-information denotes a set of closely related refinements of O-information in which the redundancy–synergy balance is resolved relative to an explicit structure rather than reported only as a single system-level scalar. The reference quantity is the O-information, defined for a multivariate system as the difference between total correlation and dual total correlation, with positive sign indicating redundancy-dominated organization and negative sign indicating synergy-dominated organization. In the literature, this structural refinement appears in several forms: a grouped measure that isolates between-group dependence, a pointwise measure on individual patterns, finite-difference gradients attached to variables or pairs, conditioned lagged versions for dynamical circuits, and a more abstract information-geometric decomposition on partially ordered sets that isolates admissible higher-order interactions (Rosas et al., 2019, Pascual-Marqui et al., 11 Jul 2025).

1. O-information as the reference quantity

For a system $\mathbf{X}^n=(X_1,\dots,X_n)$, the classical O-information is

$$\Omega(\mathbf{X}^n)=TC(\mathbf{X}^n)-DTC(\mathbf{X}^n),$$

with

$$TC(\mathbf{X}^n)=\sum_{i=1}^n H(X_i)-H(\mathbf{X}^n), \qquad DTC(\mathbf{X}^n)=H(\mathbf{X}^n)-\sum_{i=1}^n H(X_i\mid \mathbf{X}^n_{-i}).$$

Equivalently,

$$\Omega(\mathbf{X}^n)=(n-2)H(\mathbf{X}^n)+\sum_{i=1}^{n}\big[H(X_i)-H(\mathbf{X}^n_{-i})\big].$$

Its sign gives the standard interpretation: $\Omega>0$ indicates a redundancy-dominated system, $\Omega<0$ indicates a synergy-dominated system, and $\Omega=0$ indicates either a balance or a case in which opposing structures cancel (Rosas et al., 2019).

Several analytical properties explain why O-information became the basis for later structured variants. It vanishes identically for two variables, $\Omega(X_1,X_2)=0$, and for three variables it coincides with classical co-information or interaction information. It is symmetric in the variables, additive over independent subsystems, and attains extremal values on canonical redundant and synergistic distributions: for binary vectors, the copy system yields $\Omega=n-2$, whereas the XOR system yields $\Omega=2-n$. More generally, if all variables have alphabet size $$\Omega(\mathbf{X}^n)=TC(\mathbf{X}^n)-DTC(\mathbf{X}^n),$$0, then

$$\Omega(\mathbf{X}^n)=TC(\mathbf{X}^n)-DTC(\mathbf{X}^n),$$1

These properties make $$\Omega(\mathbf{X}^n)=TC(\mathbf{X}^n)-DTC(\mathbf{X}^n),$$2 a sign-sensitive high-order statistic rather than a mere measure of overall dependence strength (Rosas et al., 2019).

The original formulation already contains a structured reading. The partition-lattice analysis decomposes total correlation, binding entropy, and hence O-information along paths in the lattice of partitions. In particular, along an assembly path,

$$\Omega(\mathbf{X}^n)=TC(\mathbf{X}^n)-DTC(\mathbf{X}^n),$$3

so the global scalar can be written as a sum of triple interaction-information terms across scales. This is the conceptual bridge from unstructured O-information to later localization schemes (Rosas et al., 2019).

2. Grouped structured O-information

A specific measure named structured O-information was introduced for systems partitioned into non-overlapping groups. Let

$$\Omega(\mathbf{X}^n)=TC(\mathbf{X}^n)-DTC(\mathbf{X}^n),$$4

with covariance matrix partitioned accordingly as $$\Omega(\mathbf{X}^n)=TC(\mathbf{X}^n)-DTC(\mathbf{X}^n),$$5, and let $$\Omega(\mathbf{X}^n)=TC(\mathbf{X}^n)-DTC(\mathbf{X}^n),$$6 denote the block-diagonal matrix retaining the within-group blocks and setting all between-group blocks to zero. The same partition applies to the precision matrix $$\Omega(\mathbf{X}^n)=TC(\mathbf{X}^n)-DTC(\mathbf{X}^n),$$7 and to $$\Omega(\mathbf{X}^n)=TC(\mathbf{X}^n)-DTC(\mathbf{X}^n),$$8 (Pascual-Marqui et al., 11 Jul 2025).

For complex Gaussian data, the grouped quantities are

$$\Omega(\mathbf{X}^n)=TC(\mathbf{X}^n)-DTC(\mathbf{X}^n),$$9

$$TC(\mathbf{X}^n)=\sum_{i=1}^n H(X_i)-H(\mathbf{X}^n), \qquad DTC(\mathbf{X}^n)=H(\mathbf{X}^n)-\sum_{i=1}^n H(X_i\mid \mathbf{X}^n_{-i}).$$0

and the structured O-information is

$$TC(\mathbf{X}^n)=\sum_{i=1}^n H(X_i)-H(\mathbf{X}^n), \qquad DTC(\mathbf{X}^n)=H(\mathbf{X}^n)-\sum_{i=1}^n H(X_i\mid \mathbf{X}^n_{-i}).$$1

Hence

$$TC(\mathbf{X}^n)=\sum_{i=1}^n H(X_i)-H(\mathbf{X}^n), \qquad DTC(\mathbf{X}^n)=H(\mathbf{X}^n)-\sum_{i=1}^n H(X_i\mid \mathbf{X}^n_{-i}).$$2

For real-valued Gaussian data, the same expression is used with a factor of $$TC(\mathbf{X}^n)=\sum_{i=1}^n H(X_i)-H(\mathbf{X}^n), \qquad DTC(\mathbf{X}^n)=H(\mathbf{X}^n)-\sum_{i=1}^n H(X_i\mid \mathbf{X}^n_{-i}).$$3 on the right-hand side (Pascual-Marqui et al., 11 Jul 2025).

The interpretation parallels the ordinary O-information, but only at the group level. If $$TC(\mathbf{X}^n)=\sum_{i=1}^n H(X_i)-H(\mathbf{X}^n), \qquad DTC(\mathbf{X}^n)=H(\mathbf{X}^n)-\sum_{i=1}^n H(X_i\mid \mathbf{X}^n_{-i}).$$4, between-group interactions are redundancy-dominated; if $$TC(\mathbf{X}^n)=\sum_{i=1}^n H(X_i)-H(\mathbf{X}^n), \qquad DTC(\mathbf{X}^n)=H(\mathbf{X}^n)-\sum_{i=1}^n H(X_i\mid \mathbf{X}^n_{-i}).$$5, between-group interactions are synergy-dominated. The distinction from ordinary $$TC(\mathbf{X}^n)=\sum_{i=1}^n H(X_i)-H(\mathbf{X}^n), \qquad DTC(\mathbf{X}^n)=H(\mathbf{X}^n)-\sum_{i=1}^n H(X_i\mid \mathbf{X}^n_{-i}).$$6 is essential: the grouped measure removes within-group effects by block-diagonalization, thereby isolating only the balance among groups. The motivating example is explicit: when three or more groups each contain strong within-group redundancy but interact synergistically across groups, ordinary O-information can remain positive because redundancy occurs more often in the full system, whereas structured O-information correctly reports predominant between-group synergy (Pascual-Marqui et al., 11 Jul 2025).

The same paper gives a statistical interpretation in terms of covariance and precision hypotheses. Structured total coherence is associated with the null

$$TC(\mathbf{X}^n)=\sum_{i=1}^n H(X_i)-H(\mathbf{X}^n), \qquad DTC(\mathbf{X}^n)=H(\mathbf{X}^n)-\sum_{i=1}^n H(X_i\mid \mathbf{X}^n_{-i}).$$7

and structured dual total coherence with

$$TC(\mathbf{X}^n)=\sum_{i=1}^n H(X_i)-H(\mathbf{X}^n), \qquad DTC(\mathbf{X}^n)=H(\mathbf{X}^n)-\sum_{i=1}^n H(X_i\mid \mathbf{X}^n_{-i}).$$8

The formulas extend beyond Gaussianity to elliptical distributions because the dispersion matrix plays the role of covariance and the relevant KL-based differences are invariant to additive entropy constants. A further localization is defined at the group level through connection contributions: $$TC(\mathbf{X}^n)=\sum_{i=1}^n H(X_i)-H(\mathbf{X}^n), \qquad DTC(\mathbf{X}^n)=H(\mathbf{X}^n)-\sum_{i=1}^n H(X_i\mid \mathbf{X}^n_{-i}).$$9

$$\Omega(\mathbf{X}^n)=(n-2)H(\mathbf{X}^n)+\sum_{i=1}^{n}\big[H(X_i)-H(\mathbf{X}^n_{-i})\big].$$0

$$\Omega(\mathbf{X}^n)=(n-2)H(\mathbf{X}^n)+\sum_{i=1}^{n}\big[H(X_i)-H(\mathbf{X}^n_{-i})\big].$$1

so negative $$\Omega(\mathbf{X}^n)=(n-2)H(\mathbf{X}^n)+\sum_{i=1}^{n}\big[H(X_i)-H(\mathbf{X}^n_{-i})\big].$$2 indicates that group $$\Omega(\mathbf{X}^n)=(n-2)H(\mathbf{X}^n)+\sum_{i=1}^{n}\big[H(X_i)-H(\mathbf{X}^n_{-i})\big].$$3 contributes synergy to the between-group network, while positive $$\Omega(\mathbf{X}^n)=(n-2)H(\mathbf{X}^n)+\sum_{i=1}^{n}\big[H(X_i)-H(\mathbf{X}^n_{-i})\big].$$4 indicates redundancy (Pascual-Marqui et al., 11 Jul 2025).

3. Local O-information and pattern-level structure

A different use of structure arises in the local O-information, which assigns a signed value to each individual realization $$\Omega(\mathbf{X}^n)=(n-2)H(\mathbf{X}^n)+\sum_{i=1}^{n}\big[H(X_i)-H(\mathbf{X}^n_{-i})\big].$$5 rather than only to the distribution as a whole. The local total correlation and local dual total correlation are

$$\Omega(\mathbf{X}^n)=(n-2)H(\mathbf{X}^n)+\sum_{i=1}^{n}\big[H(X_i)-H(\mathbf{X}^n_{-i})\big].$$6

where

$$\Omega(\mathbf{X}^n)=(n-2)H(\mathbf{X}^n)+\sum_{i=1}^{n}\big[H(X_i)-H(\mathbf{X}^n_{-i})\big].$$7

The local O-information is then

$$\Omega(\mathbf{X}^n)=(n-2)H(\mathbf{X}^n)+\sum_{i=1}^{n}\big[H(X_i)-H(\mathbf{X}^n_{-i})\big].$$8

Positive $$\Omega(\mathbf{X}^n)=(n-2)H(\mathbf{X}^n)+\sum_{i=1}^{n}\big[H(X_i)-H(\mathbf{X}^n_{-i})\big].$$9 indicates that the specific pattern is redundancy-dominated, negative values indicate synergy-dominated structure, and values near zero correspond to more balanced or weakly structured patterns (Scagliarini et al., 2021).

A central property is that the pointwise quantity averages back to the global one: $\Omega>0$0 Accordingly,

$\Omega>0$1

The finite-alphabet bounds

$\Omega>0$2

apply to the average $\Omega>0$3, not necessarily to each individual $\Omega>0$4. This separates the pattern-level manifestation of synergy and redundancy from the system-level average and makes explicit that a single global statistic can conceal heterogeneous local structure (Scagliarini et al., 2021).

The three-spin Ising example clarifies the distinction. For ferromagnetic coupling $\Omega>0$5, the global O-information is positive; for frustrated negative $\Omega>0$6, it becomes negative. Yet the local analysis shows that configurations with all spins aligned are redundancy-dominated, whereas the six mixed configurations are synergy-dominated regardless of the sign of $\Omega>0$7. What changes with $\Omega>0$8 is the frequency with which these states are visited. The pattern-level sign is therefore not reducible to the global average (Scagliarini et al., 2021).

The four-voice Bach chorale application gives a concrete high-dimensional use case. The analysis preprocesses 172 chorales in major mode, transposes them to C major, encodes each voice as a 13-symbol alphabet, estimates the empirical joint distribution over four-note chords, and computes $\Omega>0$9 for each observed chord. Frequency correlates weakly but positively with local O-information; dissonance significantly lowers $\Omega<0$0; root-position triads tend to be more redundant than first- or second-inversion forms; harmonically closer chords have higher $\Omega<0$1 than harmonically distant ones; and the outer voices display relatively weak redundancy but substantial synergy. The same framework is also applied to words associated with chords, where many common theological words are redundancy-associated while some, such as “Jesu,” are synergy-associated (Scagliarini et al., 2021).

4. Gradients and dynamical structured O-information

Localization can also be expressed by finite differences of the global O-information. The first-order gradient with respect to variable $\Omega<0$2 is

$\Omega<0$3

and the second-order gradient for a pair $\Omega<0$4 is

$\Omega<0$5

The latter is symmetric and can be written as a discrete second difference: $\Omega<0$6 The first-order quantity can also be expressed as

$\Omega<0$7

or as the difference between nonnegative gradients of total correlation and dual total correlation: $\Omega<0$8 Positive gradients indicate localized redundancy-dominated contributions; negative gradients indicate localized synergy-dominated contributions (Scagliarini et al., 2022).

The gradient framework is designed to map high-order effects onto low-order objects. In the seven-spin frustrated Ising model, first-order gradients identify the central spin as the main source of synergy, while peripheral spins are mostly redundant; second-order gradients show redundant pairwise descriptors at low temperature and synergistic pairs at higher temperature, especially the central spin with a peripheral spin and neighboring peripheral-spin pairs. In US macroeconomic data, seven indicators have significant positive first-order gradients, GPDI is significantly synergistic, and GDP has significantly negative second-order gradients with four other variables. The paper argues that these pairwise gradients yield a sparser and more parsimonious network than local O-information alone (Scagliarini et al., 2022).

For time series, the dynamical O-information extends the same logic to directed prediction. Given lagged state vectors

$\Omega<0$9

and future target $\Omega=0$0, the conditioned dynamical quantity is

$\Omega=0$1

Conditioning on the target’s past removes effects due to autoregressive memory, common history, and shared inputs. For two drivers,

$\Omega=0$2

which is the second-order term in the transfer-entropy expansion. The framework uses maximization of $\Omega=0$3 to identify the most redundant multiplet and minimization to identify the most synergistic multiplet; it is insensitive to adding an unrelated variable, and $\Omega=0$4, so it is not a measure of pure pairwise effects (Stramaglia et al., 2020).

The neural-spiking application illustrates the scope of the dynamical formulation. Using 169 channels and 1778 trials from a macaque performing a random dot motion discrimination task, with neurons categorized as H, M, and L, the method shows that both redundant and synergistic circuits influencing H targets peak around 300 ms after the go cue, synergy decays more slowly, redundant circuits are mostly composed of H and M neurons, and L neurons become important in synergistic circuits for larger multiplets. A representative H target supports a redundant circuit of 7 drivers and a synergistic circuit of 5 drivers under the surrogate-based stopping rule (Stramaglia et al., 2020).

5. Information-geometric decomposition on structured spaces

A more abstract precursor of structured O-information appears in the information geometry of partially ordered sets. Instead of using the full Boolean lattice of all subsets, the framework begins with an arbitrary finite poset $\Omega=0$5 with bottom element $\Omega=0$6 and defines the probability simplex

$\Omega=0$7

For any $\Omega=0$8, the lower and upper sets are

$\Omega=0$9

Principal ideals $\Omega(X_1,X_2)=0$0 and principal filters $\Omega(X_1,X_2)=0$1 become the structural primitives of the decomposition (Sugiyama et al., 2016).

Two dual coordinate systems are assigned to distributions on $\Omega(X_1,X_2)=0$2. The exponential-family form uses feature functions

$\Omega(X_1,X_2)=0$3

which gives

$\Omega(X_1,X_2)=0$4

with recursion

$\Omega(X_1,X_2)=0$5

The dual expectation coordinates are

$\Omega(X_1,X_2)=0$6

These coordinates are dually orthogonal, which enables mixed-coordinate projections and KL/Pythagorean identities (Sugiyama et al., 2016).

Given a subset $\Omega(X_1,X_2)=0$7, the mixed coordinates are

$\Omega(X_1,X_2)=0$8

For distributions $\Omega(X_1,X_2)=0$9 and $\Omega=n-2$0, the mixed distribution $\Omega=n-2$1 with respect to $\Omega=n-2$2 is the unique distribution such that $\Omega=n-2$3 outside $\Omega=n-2$4 and $\Omega=n-2$5 on $\Omega=n-2$6. The KL divergence then satisfies the Pythagorean theorem

$\Omega=n-2$7

and along a chain of subsets $\Omega=n-2$8,

$\Omega=n-2$9

Entropy decomposition follows by using the uniform distribution $\Omega=2-n$0, and the same machinery defines refined mutual-information increments along the poset (Sugiyama et al., 2016).

This framework does not define O-information, redundancy, or synergy directly. Its significance is methodological: it provides a structured, orthogonal dependence decomposition on an arbitrary poset, so dependence is organized by the admissible order relations rather than by all subsets of variables. This suggests a geometric substrate for structured redundancy–synergy analyses on incomplete hierarchies, sparse event spaces, or constrained interaction families (Sugiyama et al., 2016).

6. Estimation, applications, and interpretive boundaries

The practical use of structured O-information depends on estimation. A recent development is the score-based O-information estimator SSI, which removes the usual discrete or Gaussian restrictions by estimating O-information for general continuous variables through score matching and diffusion-style denoising. In this framework,

$\Omega=2-n$1

and the O-information gradient is

$\Omega=2-n$2

The key identity expresses KL divergence in terms of score functions of Gaussian-noised variables,

$\Omega=2-n$3

Since the score is linked to a denoiser by

$\Omega=2-n$4

the method estimates total correlation and dual total correlation from joint, marginal, and conditional denoising scores, obtained within a single amortized architecture that takes noised variables, clean context variables, and a masking/time vector as input (Bounoua et al., 2024).

The synthetic validation uses 100k training samples, 10k test samples, a VP-SDE denoising network, an MLP with skip connections in the main experiments, 10-sample Monte Carlo integration for the time integral, and averaging over 5 random seeds. The method is compared with MINE, NWJ, InfoNCE, and CLUB. On redundant systems it tracks the true positive O-information and remains stable as dimensionality grows; on synergistic systems it correctly produces negative O-information while pairwise-MI baselines fail badly; on mixed systems it is the most reliable estimator and captures sign changes; and the gradient estimates identify which variables contribute to redundancy or synergy. The real-data application uses the Allen Brain Observatory Visual Behavior dataset with 80 mice, two conditions (“change” and “no change”), and analyses over 3 and then 6 visual regions. Across time bins, O-information is higher for change trials and lower for no-change trials, indicating stronger redundancy-dominated coordination under change stimuli (Bounoua et al., 2024).

Across these variants, several interpretive boundaries recur. Ordinary O-information is not a full PID decomposition but a scalar summary of whether redundancy or synergy dominates. The local and gradient frameworks refine the global value without replacing it, since the local quantity averages back to $\Omega=2-n$5 and the gradients are finite-difference descriptors. The dynamical formulation is a conditioned selection criterion for informational circuits rather than an exact decomposition of transfer entropy. The grouped measure $\Omega=2-n$6 answers a different question from local or gradient O-information, because it isolates between-group interactions after factoring out within-group structure. A zero value therefore does not admit a unique interpretation: it can reflect weak high-order structure, a balance between redundancy and synergy, or cancellation between different organized substructures (Rosas et al., 2019, Scagliarini et al., 2021, Stramaglia et al., 2020, Pascual-Marqui et al., 11 Jul 2025).

The cumulative picture is that structured O-information is not a single universally fixed construction but a technical program for resolving multivariate redundancy–synergy balance relative to an explicit organization: groups, patterns, variables, pairs, dynamical driver sets, or posets. The common objective is to preserve the sign-based interpretability of O-information while exposing where in the state space, network, hierarchy, or modular architecture the relevant high-order structure resides.