Integrated Information Decomposition (ΦID)

• Integrated Information Decomposition (ΦID) is a framework that decomposes time-lagged mutual information into distinct, non-negative atoms, clarifying how past variables inform future states.
• It extends Partial Information Decomposition by using a product redundancy lattice to fully classify information dynamics such as storage, transfer, and synergistic interactions.
• ΦID has practical applications in analyzing neural networks, coupled oscillators, and complex systems, offering new measures for integration and dynamic causation while addressing limitations of scalar metrics.

Integrated Information Decomposition (ΦID) is a mathematical framework that generalizes Partial Information Decomposition (PID) to decompose the information flow from sets of variables in the past to sets of variables in the future. ΦID systematically resolves the total time-lagged mutual information—or "excess entropy"—in a complex dynamical system into non-overlapping, non-negative atoms that capture distinct modes of information dynamics, such as redundancy, unique transfer, and synergy. Unlike scalar metrics of "integrated information" (e.g., the traditional Φ-measures of Integrated Information Theory), ΦID provides a full taxonomy of past-to-future information conversions, unifies forward and backward PIDs, and enables the construction of new, interpretable measures of dynamical and informational complexity in multivariate systems (Mediano et al., 2019, Mediano et al., 2021).

1. Theoretical Foundations: PID to ΦID

ΦID builds on the Williams–Beer PID, which decomposes the mutual information $I(X_1, X_2; Y)$ into four non-negative atoms (redundancy, unique$_1$, unique$_2$, synergy). For two sets of variables, e.g., past $(X_1, X_2)$ and future $(Y_1, Y_2)$, ΦID generalizes PID by introducing a product redundancy lattice $\mathcal{A} \times \mathcal{A}$, where $\mathcal{A}$ is the set of antichains over the elements $\{1,2\}$. Each lattice node $\bm{\alpha} \to \bm{\beta}$ specifies which past combinations $\bm{\alpha}$ carry information about which future combinations $_1$0.

A double-redundancy function $_1$1 is defined that is compatible (recovering original PID for singleton targets) and monotonic on the lattice. Under Möbius inversion, this yields 16 non-negative "ΦID atoms" $_1$2 that sum to the total excess entropy $_1$3. Classical PID terms are sums over subsets of these atoms (Mediano et al., 2019, Mediano et al., 2021, Jansma et al., 2024).

2. ΦID Lattice and Taxonomy of Information Dynamics

Each ΦID atom corresponds to a qualitatively distinct dynamical phenomenon. Table 1 lists the standard classification for two components:

Mode	Representative Atoms	Information Flow Class
Storage	$_1$4, $_1$5, $_1$6	Within-component persistence
Copy	$_1$7, $_1$8	Duplication to future sets
Transfer	$_1$9, $_2$0	Unique transfer
Erasure	$_2$1, $_2$2	Loss of redundancy
Downward causation	$_2$3, $_2$4, $_2$5	Joint past to element
Upward causation	$_2$6, $_2$7, $_2$8	Element(s) to joint future

Every classic information-dynamics metric (transfer entropy, active information storage, whole-minus-sum $_2$9) decomposes as specific aggregates of ΦID atoms. Scalar measures can thus conflate fundamentally different functional roles that ΦID separates (Mediano et al., 2019, Mediano et al., 2021).

3. Mathematical Formalism and Computation

A central operation underlying ΦID is Möbius inversion on finite partially-ordered sets (lattices):

• The redundancy lattice for PID is the set of antichains $(X_1, X_2)$0 over $(X_1, X_2)$1 variables.
• The ΦID lattice is the product $(X_1, X_2)$2 for $(X_1, X_2)$3 sources and $(X_1, X_2)$4 targets.
• A redundancy function $(X_1, X_2)$5 is posited for each node; atoms are recovered via

$(X_1, X_2)$6

• For $(X_1, X_2)$7, there are $(X_1, X_2)$8 atoms.

The fast Möbius transform (Jansma et al., 2024) enables practical PID and ΦID intractable at $(X_1, X_2)$9 using prior systems of equations by providing a closed-form for the Möbius function:

$(Y_1, Y_2)$0

where $(Y_1, Y_2)$1 is the complement, $(Y_1, Y_2)$2 the downward ideal. This reduces decomposition to a single sparse matrix-vector product, with significant speed-up for $(Y_1, Y_2)$3 (Jansma et al., 2024).

4. Redundancy Functions and Interpretive Constraints

The choice of redundancy function is pivotal. The Minimum Mutual Information (MMI) measure is the common estimator for redundancy and, by duality, synergy. However, when applied to temporal processes, MMI-based synergy cannot always distinguish true integration from spurious autocorrelation. For example, systems with only independent autocorrelated components can yield identical MMI-synergy to parity-preserving systems, despite only the latter exhibiting genuine higher-order integration. More paradoxically, some systems may show increased synergy under functional disintegration, demonstrating that MMI-based synergy is not a reliable indicator of integrated information (Varley, 2024).

Alternative redundancy measures, particularly those based on probability-mass exclusions ($(Y_1, Y_2)$4), demonstrate improved capacity to distinguish autocorrelation from synergistic integration and allow local, multi-target decompositions practical for real-world, noisy dynamics (e.g., neural avalanches) (Varley, 2022).

5. Empirical Applications in Complex Systems

ΦID has been deployed across a range of continuous and discrete systems:

• Neural and cortical cultures: ΦID-based decompositions on both in silico (simulated) and in vitro neuronal networks have revealed that the unique-information transfer atom most reliably tracks underlying structural connectivity, whereas redundant atoms contribute residual information and synergistic atoms reflect emergent or collective network dynamics (Menesse et al., 2024).
• Coupled oscillators and cellular automata: Integrated information metrics from ΦID peak at critical bands associated with dynamical complexity, such as metastability in Kuramoto networks or emergent glider structures in cellular automata, surpassing classic entropy and metastability indices (Mediano et al., 2021).
• Complex networks and time-resolved analysis: Local (pointwise) ΦID analysis applied to neural avalanches shows that different information-processing modes (storage, transfer, synergy) exhibit distinct profiles across the temporal evolution of an event, with nontrivial integration predominantly occurring early in the cascade (Varley, 2022).

6. Limitations, Cautions, and Future Directions

Several challenges and limitations persist:

• The MMI redundancy function can overestimate redundancy and underestimate synergy; use of richer measures such as the shared-probability-mass-exclusion framework is advised, particularly in high-autocorrelation or ambiguous dynamic environments (Varley, 2024, Varley, 2022).
• Computation remains combinatorially intensive for $(Y_1, Y_2)$5 variables, though the fast Möbius transform provides a viable approach for moderate system sizes (Jansma et al., 2024).
• Empirically, ΦID is highly sensitive to uncorrelated noise and finite-sample estimation biases; denoising and robust statistical validation are necessary (Mediano et al., 2021).
• Extending ΦID to continuous variables, larger networks, and integrating pointwise local decompositions remains an active area of research.

Best practice is to validate redundancy/synergy measures on interpretable ground-truth models and corroborate synergy findings with independent causal inference or perturbation techniques.

7. Broader Impact and Theoretical Integration

ΦID resolves the total information flow in a system into a multidimensional, interpretative taxonomy that unifies classical information-dynamics notions and enables new targeted measures of integration, transfer, and emergence specific to context. It subsumes traditional scalar complexity measures such as $(Y_1, Y_2)$6, reveals the internal structure of integrated information, and supports rigorous inquiry into phenomena such as upward/downward causation. By providing a principled, additive, and localizable decomposition, ΦID offers a universal framework for dissecting the computational and informational substrate of complex systems, from neuronal cultures to climate models and social dynamics (Mediano et al., 2019, Mediano et al., 2021, Varley, 2022, Menesse et al., 2024, Jansma et al., 2024).