Science typically studies systems by decomposing them into independent parts — factorizing the joint into a product of marginals. This strategy works spectacularly well, until the connections between parts carry more information than the parts themselves.
Circulatory Fidelity starts from a simple question: when does breaking a system apart destroy something essential? The answer turns out to be precise, quantifiable, and surprisingly universal.
Any system of interacting variables has a precision matrix — a mathematical object that encodes how much each variable constrains the others. This matrix decomposes cleanly into two parts:
When we treat variables as independent — whether in statistical modeling, scientific explanation, or everyday reasoning — we are implicitly setting R to zero. CF measures the cost of that choice.
Note on scope: This page develops the framework through the bivariate Gaussian — the clearest pedagogical case, where every quantity has a closed-form expression. The framework itself is not restricted to Gaussians or to two variables. IC is defined for any joint distribution via Fisher information geometry; the decomposition Λ = D + R applies wherever a precision structure exists. The Gaussian case is where the identities are provably exact (Class 10); extensions carry appropriate epistemic markers.
The precision matrix isn't arbitrary — it is the Fisher Information Matrix for a Gaussian model. This identity (Λ = I(μ)) means that the decomposition Λ = D + R is simultaneously a decomposition of statistical information: how much information comes from individual variables (D) versus their relationships (R).
This is why IC has information-theoretic meaning: it directly quantifies what fraction of the system's statistical information is carried by relationships. The decomposition is not a modeling choice — it is a fact about the geometry of the model's parameter space.
Scope: This identity is exact for Gaussian models. For non-Gaussian distributions, the precision matrix generalizes to the Fisher Information Matrix, and the decomposition principle carries over with appropriate modifications to the off-diagonal structure.
How much information lives in the connections? IC answers this in a single number between 0 and 1:
IC measures the strength of coupling, not its direction. The relationship between A and B is the same as between B and A. IC answers: how much can I learn?
A key theorem: the information lost to factorization depends only on IC, not on how large or variable the individual parts are. Relational structure, not nodal properties, determines information flow.
A critical threshold: above this value, more information lives in the connections than in the nodes. The relational structure becomes the dominant component of the system. Factorization past this point is not merely imprecise — it is wrong about what matters.
Relational Invariance Theorem C10
For a bivariate Gaussian (X, Z) with precision matrix Λ, the mutual information under the mean-field factorization q(X)q(Z) is:
Start from I(X; Z) = H(X) + H(Z) − H(X, Z) for the Gaussian joint distribution.
The joint entropy is H(X, Z) = ½ ln((2πe)² |Σ|) where |Σ| = σ²xσ²z(1 − ρ²).
Substituting and simplifying: I(X; Z) = −½ ln(1 − ρ²) = −½ ln(1 − IC²).
The individual variances σ²x and σ²z cancel completely. Only IC appears. This is the theorem: inference depends solely on relational structure.
ICcrit derivation: The threshold where relational information equals nodal information. Setting the off-diagonal contribution equal to the diagonal contribution in the Fisher information and solving yields IC² = 1/3, giving ICcrit = 1/√3 ≈ 0.577. Validated against 1,050 calibration rows (data/calibration/threshold_calibration.csv).
IC tells you how strongly two variables are linked. But who has leverage over whom? When variables differ in scale, the coupling becomes asymmetric — a small variable can be dominated by a large one, or vice versa.
CC measures intervention leverage: if z changes by one unit of its own variability, how much does x move in units of x's variability? The direction matters.
A fundamental constraint: the product of leverages in both directions always equals IC squared. If one direction has high leverage, the other must be proportionally weak. Try the variance ratio slider in the interactive below to see this in action.
Product Identity C10
CC(z→x) = IC · r and CC(x→z) = IC · (1/r) where r = σx/σz.
Their product: CC(z→x) × CC(x→z) = IC · r × IC · (1/r) = IC². QED.
The variance ratio cancels. This is why IC is the fundamental quantity: it constrains all possible CC decompositions. The total leverage is fixed; only its distribution between directions varies.
Why B = 0 predictive value: The Balance Factor B = √(2r²/(r⁴+1)) measures how evenly leverage is distributed. But in empirical testing across 50,000+ validation rows, B contributes zero additional predictive power once IC is known. The reason: B describes the shape of the CC decomposition but carries no information about the strength of coupling. Since inference outcomes depend only on coupling strength (Relational Invariance), B is architecturally interesting but diagnostically inert.
Beyond bivariate CC
In systems with three or more variables, a third form of control coupling emerges: CCmediated — influence transmitted indirectly through an intermediary. The full three-CC landscape provides the complete picture of how intervention propagates through coupled systems:
Direct leverage of z on x. The immediate effect of intervening on z.
Direct leverage of x on z. The reciprocal direction.
Indirect leverage through an intermediary m. The path z → m → x.
The mediation identity C10: CCmediated(z → m → x) = CC(z → m) · CC(m → x). Mediated influence is the product of the direct links along the path.
Ω regimes: The ratio Ω = CCmediated / CCdirect classifies system behavior:
These regimes have been validated on three empirical datasets (HCP neuroimaging, 54 observations). The Ω classification determines whether screening effects (suppression) or relay amplification (mediation) governs the system's causal structure.
Every claim in CF carries an explicit Semiotic Translation Quality Assessment (STQA) class. This is the framework's built-in overclaiming prevention: each mathematical statement, correspondence, or analogy is classified by the strength of evidence behind it. The colored badges you see throughout this site (e.g. C10) indicate the class of the claim they accompany.
Cross-domain connections are classified the same way: a Class 10 edge between two domains means a proven mathematical identity connects them; a Class 7 edge means an algebraically exact structural correspondence. Class 5 connections (analogies) are tracked but not rendered on the domain network visualization. The composite score determining each domain's tier is the minimum (bottleneck) across all its component claims.
IC measures coupling strength, but its implications depend on what the coupling is doing in the system. The same IC value can mean opposite things in different architectures. This is the Dependency Asymmetry — perhaps the most important diagnostic insight in CF.
The coupling is the mechanism. Relationships encode the system's physics, chemistry, or logic. They are structural, not incidental.
The coupling carries signal. Relationships provide learnable regularization — they help parts share information they couldn't access alone.
Real systems often have multiple layers of coupling. The archetype can differ at each interface: proximal layers (near the data) tend toward filtering; distal layers tend toward pooling.
The cardinal rule: never interpret IC without first classifying the dependency archetype. "High IC is bad" and "high IC is good" are both correct — in their respective contexts. The framework is a diagnostic, not a verdict.
When we model a system by treating its parts as independent, we discard the connections between them. Sometimes this is harmless. Sometimes the connections carry most of the information. Inference Coupling (IC) measures exactly how much is lost.
Use the slider below to increase coupling between two variables. Watch how the shape of their joint distribution stretches, how the distance in the natural geometry grows, and how the cost of ignoring their relationship accumulates.
At this coupling strength, factorization discards a moderate amount of relational structure.
The ellipse shows the shape of the joint probability distribution. A circle means independence; a stretched ellipse means the variables are coupled. The small matrix in the corner shows Λ = D + R — the bright off-diagonal region is what factorization discards.
IC maps naturally to position in hyperbolic space. The center is independence (IC = 0); the boundary is perfect coupling (IC = 1, infinitely far away). The dashed ring marks ICcrit = 1/√3 — the threshold where relational structure becomes dominant.
The cost of ignoring relational structure, measured in nats of information lost. Near zero coupling, the cost is negligible — factorization is adequate. As coupling grows past ICcrit, the cost accelerates: the connections between parts carry more information than the parts alone.
Consider three light switches wired so that the room light turns on only when an odd number of switches are up. Look at any two switches alone — they seem completely random, unrelated to each other. But all three together follow a perfect rule. The structure is real, but invisible to pairwise observation.
This phenomenon — genuine relational structure that requires collective observation to detect — is what CF calls coplexity. It is the same thing variously called synergy, interaction information, or irreducible higher-order correlation. Coplexity names it precisely: structure that appears as symmetry in its native algebraic encoding.
Compare three Boolean functions below. AND and OR have obvious pairwise structure — knowing one input tells you something about the output. XOR has zero pairwise coupling, yet the function is perfectly determined. The structure hasn't vanished; it has moved to a place that pairwise tools cannot reach.
XOR is the clean algebraic archetype. But coplexity is not confined to logic gates. It appears wherever three-way structure carries information that no pair reveals:
In HCP neuroimaging data (208 subjects), neural IC between brain regions is modulated by cardiac phase — the relationship between regions depends on a third variable (heartbeat timing) invisible to any pairwise brain-region analysis. The three-way structure (region A × region B × cardiac phase) carries information absent from any two-way slice.
In random 3-SAT problems, any two variables may appear uncorrelated in satisfying assignments. But three variables sharing a clause create an XOR-like constraint: the satisfiability of the whole depends on their joint configuration in ways invisible to pairwise variable statistics. The phase transition at α ≈ 4.27 is where this coplex structure overwhelms factorized solvers.
The metabolic epistemology explains why coplexity appears rare: organisms evolved under energy constraints that favor pairwise observation (cheap) over collective observation (expensive). The apparent scarcity of higher-order structure is a selection effect of methodology, not a fact about nature.
Because coplexity is invisible to standard methods, detecting it requires a deliberate two-stage approach. The first stage uses familiar pairwise tools; the second deploys Walsh-Hadamard spectral analysis to detect structure in the collective behavior of three or more variables simultaneously.
Compute IC₂ for all variable pairs. If IC₂ is high, you have ordinary pairwise coupling — MFVI will fail, and the failure is predictable from standard diagnostics. If IC₂ ≈ 0 everywhere, either the system is truly independent, or the structure lives higher.
When IC₂ ≈ 0, apply the Walsh-Hadamard transform to compute IC₃. This spectral method detects three-variable dependencies invisible to marginals. Non-zero IC₃ with zero IC₂ = pure coplexity.
The function is either pairwise-detectable (L₁ structure), coplex (L₂ structure), or independent. This determines which tools are adequate and where factorization will silently fail.
The two-stage detection protocol reveals a natural stratification. Some structure is accessible to pairwise observation; some requires collective observation of three or more variables simultaneously. CF formalizes this as Layer Theory v3.0:
Structure detectable by IC₂. Ordinary correlations, linear dependencies, standard pairwise coupling. Conventional statistical tools are adequate. Factorization failure here is predictable from marginal diagnostics.
Structure requiring IC₃ or higher. Invisible to pairwise observation but genuinely present. Coplex dependencies, XOR-like relations, higher-order correlations. Factorization failure here is invisible to standard diagnostics.
Encoding-relativity: Layer assignment is not absolute. A function can be L₂ in one encoding and L₁ in another (XOR is coplex in Boolean encoding but linear in GF(2) encoding). The layer describes the function as observed through a particular representational scheme — the cost of detection depends on the observer's tools, not only on the structure itself.
Which functions have pure coplexity? The answer is algebraically exact: a Boolean function has zero pairwise coupling and maximal higher-order structure if and only if it is affine over GF(2) — the finite field with two elements. In concrete terms: the XOR-like functions are exactly the coplex ones.
This is not an analogy (Class 5). It is an algebraic identity (Class 7): the group structures are the same object viewed through different encodings. Both sides detect the boundary between efficiently simulable structure and genuine complexity — the boundary where factorization breaks down in a way that no pairwise diagnostic can anticipate.
Why n = 3? Three is the minimum number of variables that can carry coplex structure. With two variables, all dependencies are pairwise by definition. At three, a new possibility opens: structure that requires all three to be observed simultaneously. This dimensional threshold (nmin = 3) recurs across the framework — in the minimal coplex dimension, the minimal CPS configuration, and the minimal PSL(2,7) presentation.
Computational Coplexity Principle C7
The Walsh-Hadamard transform decomposes any Boolean function f: {0,1}n → {0,1} into spectral coefficients indexed by subsets S ⊆ {1,...,n}:
Key insight: IC₂ measures the weight in |S| = 1 coefficients (linear/pairwise). IC₃ measures weight in |S| ≥ 2 coefficients (nonlinear/coplex). A function is affine over GF(2) if and only if all weight is in |S| ≤ 1 — meaning it is fully characterized by its linear structure, yet has zero pairwise IC because the marginal distributions are uniform.
The GF(2) ↔ Stabilizer Correspondence C7
XOR over GF(2) and CNOT in quantum circuits implement the same algebraic operation: addition in a vector space over a two-element field. Extending from gates to circuits: the affine group over GF(2)n is isomorphic to the Clifford group restricted to computational basis states. The CF coplexity detection protocol thus performs, on classical Boolean data, the same structural classification that stabilizer formalism performs on quantum states.
Encoding-relativity: XOR is coplex (L₂) in Boolean encoding but trivially linear (L₁) in GF(2) encoding. The layer assignment is not a property of the function — it is a property of the function as observed through a particular encoding. This is Layer Theory v3.0: layers are encoding-relative, not absolute.
If coplex structure can extend to arbitrary order — IC₃, IC₄, IC₅... — why doesn't complexity simply explode? The answer is a combinatorial constraint that makes higher-order coplexity self-limiting.
The number of possible coplex relationships at order m among n variables scales as C(n, m), the binomial coefficient. But the cost of observing each relationship scales exponentially with m — you need simultaneous access to all m variables. The optimal observation mode — the order where the product of available structure times detection efficiency is maximized — converges to:
Beyond mode 5, the combinatorial explosion of possible relationships is overwhelmed by the exponential cost of observing them. This is not a failure of methodology — it is a structural property of coplexity itself. The universe's relational structure is predominantly low-order, not because higher-order structure is absent, but because it is thermodynamically expensive to maintain and metabolically expensive to detect.
This explains the empirical observation that n = 3 dominates: three-way coplexity is where the balance between structural richness and detection cost is most favorable. The framework's emphasis on IC₃ as the primary coplexity diagnostic is not arbitrary — it targets the mode where coplex structure is densest and most observable.
When you map systems by their coupling strength, something unexpected emerges: the natural geometry of this space is not flat. It is hyperbolic — the same geometry as special relativity, the same geometry as the Poincaré disk. This isn't imposed by choice; it emerges from the mathematics of the precision matrix.
The key insight: as coupling grows, each increment becomes harder — it takes exponentially more information to push IC from 0.9 to 0.95 than from 0.1 to 0.15. This acceleration is what hyperbolic geometry encodes.
IC lives on the interval [0, 1], but the true distance between coupling states is measured by the geodesic: d = arctanh(IC). Near zero, IC and d are nearly the same. As IC approaches 1, the geodesic distance diverges to infinity — perfect coupling is infinitely far away because it requires infinite information.
Moving from IC = 0.5 to IC = 0.6 costs 0.12 nats of geodesic distance. Moving from IC = 0.9 to IC = 0.95 costs 0.47 nats — nearly 4× more for half the IC increment. The geometry charges exponentially more as coupling tightens.
ICcrit = 1/√3 sits at geodesic distance d ≈ 0.66, the point where the cost function's curvature shifts. Below this: factorization is cheap. Above: the cost accelerates — you're moving into territory where the relational structure dominates the system's information content.
The geometric structure is not approximate. CF possesses a set of exact mathematical identities connecting its quantities — the same way special relativity connects energy, momentum, and mass through Lorentzian geometry.
The symmetric and asymmetric leverages decompose observational effects with a Minkowski-signature metric. This is literally the same structure as E² - p² = m² in relativity.
IC is the velocity; geodesic distance is the rapidity. The same transformation that converts velocity to rapidity in special relativity converts IC to its natural coordinate on the Fisher-Rao manifold.
Mutual information is the action integral. Information accumulates along the geodesic path, with tanh(d) as the Lagrangian density. This connects statistical inference to variational mechanics.
Geff mediates between geometry (geodesic distance) and physics (information cost), converging to exactly 1/4 in the strong-coupling limit. The statistical analogue of gravitational coupling.
The cost function I(d) = log(cosh d) is not merely a mathematical curiosity. It has a direct physical interpretation: factorization cost is equal to the minimum thermodynamic dissipation required to erase the relational structure. This is the cocycle-Landauer unification — a Class 10 identity connecting CF to statistical mechanics.
Three quantities that are identically the same object: the mutual information lost to factorization, the minimum work dissipated in erasing the correlation (Landauer's principle applied to relational structure), and the entropy generated by that erasure. The cost of ignoring connections is not just informational — it is thermodynamic.
I(d) = log cosh d measures what factorization discards: the mutual information between coupled variables, expressed in the natural geodesic coordinate.
Wdiss/kT = log cosh d is the minimum dissipation to erase the correlation. Any physical process that factorizes a coupled system must pay at least this much energy.
Sgen/k = log cosh d is the entropy generated. Factorization is irreversible: the relational structure, once erased, cannot be recovered without external work.
This identity means that ICcrit = 1/√3 is simultaneously an information-theoretic, geometric, and thermodynamic threshold. Past dcrit ≈ 0.66, the energy required to erase relational structure exceeds what the system invests in maintaining its nodes — factorization becomes thermodynamically expensive, not just informationally destructive.
Cocycle-Landauer Unification C10
The derivation proceeds in two steps:
Cocycle property: The cost function I(d) = log cosh d satisfies the cocycle condition I(d₁ + d₂) = I(d₁) + I(d₂) + correction, where the correction term encodes how factorization costs compose along a geodesic path. This is not an approximation — it is a structural property of the hyperbolic geometry.
Landauer identification: Landauer's principle states that erasing one bit of correlation requires minimum dissipation kT ln 2. For Gaussian correlations of strength ρ, the mutual information is I = −½ ln(1 − ρ²) nats. Converting to the geodesic coordinate: I(d) = log cosh d = log cosh(arctanh(ρ)). The minimum dissipation to erase this correlation is Wdiss = kT · I(d), giving the triple identity directly.
Physical consequence: In statistical mechanics, the free energy of a system with coupling d is F = −kT · log cosh d. The partition function of the coupled system factorizes into the product of independent partition functions only when d = 0. Any nonzero coupling adds a thermodynamic cost to the factorization — precisely I(d) per degree of freedom.
The Gaussian MMC identity C10: For Gaussian processes, the minimum mean-field communication cost between subsystems is exactly log cosh d. This was derived independently from the thermodynamic argument and proven to be identical — a consistency check on the unification.
Curvature of the CF manifold C10
The Fisher-Rao metric on the bivariate Gaussian model with parameters (μ1, μ2, σ1, σ2, ρ) induces a 2D submanifold when we restrict to the coupling sector (ρ, r) where r = σ1/σ2.
The metric tensor components are computed from the Fisher information: gρρ = (1+ρ²)/(1−ρ²)², grr = 2/r², gρr = 0.
The Gaussian (sectional) curvature is K = −R1212/(g11g22 − g12²).
Evaluating: K(ρ) = −2(3 − ρ²)/(1 + ρ²). At ρ = 0: K = −6. As ρ → 1: K → −2.
The curvature is everywhere negative (hyperbolic) but variable. It becomes less curved as coupling strengthens — the geometry flattens toward K = −2 in the strong-coupling regime. This was corrected from an earlier error that assumed constant K = −2 by applying a 1D formula on the 2D manifold.
Physical meaning: Negative curvature means that nearby geodesics diverge — small differences in coupling produce increasingly different information costs. This divergence is maximal at independence (K = −6) and decreases toward strong coupling, consistent with the intuition that highly-coupled systems are more "rigid" in their information geometry.