Synergy Index (SI): Concepts and Applications

• Synergy Index (SI) is a set of measures that quantify irreducible joint effects beyond pairwise interactions in diverse fields.
• Different formulations of SI include Triple-Helix configurational information, predictor–target excess information, redundancy–synergy balance, and bounded ratios in EMG studies.
• Applications of SI span innovation studies, information theory, neuromuscular coordination, and VLM explainability, enabling deeper insights into system-level interactions and modality fusion.

Synergy Index (SI) is not a single standardized quantity across the arXiv literature. The term is used for several non-equivalent measures that all seek to detect a joint effect not reducible to lower-order structure, but they do so in different mathematical settings and with different sign conventions. In the cited work, SI denotes: a signed three-way interaction information for Triple-Helix innovation systems; a target-centric excess-over-union measure in predictor–target information theory; a sign-reversed O-information or Redundancy–Synergy Index in multivariate continuous systems; a flexor–extensor balance ratio derived from NMF muscle-synergy weights; and a Harsanyi-based faithfulness score for cross-modal reasoning in vision–LLMs (Strand et al., 2011, Griffith et al., 2012, Rosas et al., 2019, Ahmadi et al., 25 Jul 2025, Ky et al., 21 May 2026).

1. Scope and semantic variation

The common thread across these usages is the attempt to operationalize a configurational effect: the whole exhibits structure that is not exhausted by pairwise, marginal, or unimodal components. However, the mathematical object called SI varies substantially by field.

In Triple-Helix studies, SI is a signed interaction information in bits, and more negative values indicate stronger synergy because the configuration reduces uncertainty at the systems level. In several multivariate information-theoretic formulations, synergy is likewise associated with a negative signed quantity, but the signed object may be the interaction information, the O-information, or the Redundancy–Synergy Index, with SI obtained either directly or by sign reversal. In EMG-informed cycling, by contrast, SI is not a redundancy–synergy balance; it is a bounded ratio,

$$SI_J=\frac{W_{\mathrm{flex}}(J)}{W_{\mathrm{flex}}(J)+W_{\mathrm{ext}}(J)},$$

so that values above or below $0.5$ indicate flexor or extensor dominance. In VLM explainability, SI is operationalized by a synergistic faithfulness score, $\mathcal{F}_{syn}$, whose positive values indicate super-additive cross-modal interaction, near-zero values indicate additivity or redundancy, and negative values indicate antagonism (Leydesdorff et al., 2012, Bounoua et al., 2024, Ahmadi et al., 25 Jul 2025, Ky et al., 21 May 2026).

A recurrent source of confusion is therefore terminological rather than mathematical. “Synergy Index” names a research objective—quantifying irreducible joint effect—rather than a universally agreed formula.

2. Triple-Helix SI as configurational uncertainty reduction

Within the Triple-Helix framework, SI is the mutual information in three dimensions among geography, technology, and organization. The underlying entropy is

$H(X)=-\sum_x p(x)\log_2 p(x),$

and the configurational quantity is

$$T_{XYZ}=H(X)+H(Y)+H(Z)-H(X,Y)-H(X,Z)-H(Y,Z)+H(X,Y,Z).$$

For Triple-Helix applications, $X=G$, $Y=T$, and $Z=O$; the paper interprets $T_{GTO}<0$ as synergy, that is, a net reduction of uncertainty at the systems level due to configurational interactions, while $T_{GTO}>0$ indicates redundancy or dividedness. In this literature, the more negative $0.5$0, the stronger the innovation-system “systemness” (Strand et al., 2011).

The Norwegian study operationalized $0.5$1 by firm locations aggregated along the NUTS hierarchy, $0.5$2 by 2-digit NACE sector codes, and $0.5$3 by eight firm-size classes, using 481,819 firms from Statistics Norway for Q4 2008. Counts $0.5$4 were converted into empirical marginals and joint distributions, entropies were computed in base-2 logarithms, and SI was reported in bits or millibits. Geographic decomposition followed a Theil-style expression,

$0.5$5

where $0.5$6 is the between-group component and $0.5$7 are within-group synergies. Nationally, the pairwise couplings were $0.5$8, $0.5$9, and $\mathcal{F}_{syn}$0 bits, so geography–technology was the strongest pairwise relation and geography–organization the weakest. The national Triple-Helix synergy was $\mathcal{F}_{syn}$1 bits $\mathcal{F}_{syn}$2. Decomposition showed that $\mathcal{F}_{syn}$3 of the national synergy arose as an in-between counties component and only $\mathcal{F}_{syn}$4 was added by aggregation from regions to the nation, so most synergy was realized within regions. Hordaland, Møre og Romsdal, Nordland, and Rogaland exhibited the strongest negative county contributions, whereas Oslo og Akershus and Trøndelag were comparatively weaker despite major university infrastructures. The paper interpreted west-coast patterns in relation to oil, gas, offshore, chemistry, and marine industries, and described Northern Norway as showing a deviant pattern under substantial government intervention (Strand et al., 2011).

A related methodological paper generalized the same interaction-information logic to Triple-Helix and Quadruple-Helix settings and provided the open-source routine th4.exe. For three variables it computes the same $\mathcal{F}_{syn}$5; for four variables it applies the inclusion–exclusion form

$\mathcal{F}_{syn}$6

The routine accepts case-based nominal data, computes entropies and mutual informations in two, three, and four dimensions, and writes outputs to th4.dbf, enabling regional or sectoral subset analysis at scale (Leydesdorff et al., 2012).

The U.S. study extended the same Triple-Helix SI to 8,121,301 firms from ORBIS, with organization measured by 11 size classes, geography by three-digit ZIPs mapped to CBSA and CSA units, and technology by NACE Rev. 2 codes. At the national scale, aggregation of within-state synergies accounted for $\mathcal{F}_{syn}$7 of national synergy and the between-state term added only $\mathcal{F}_{syn}$8, supporting the conclusion that U.S. innovation-systemness operates primarily at the state level. CBSA units were described as too small because $\mathcal{F}_{syn}$9 of synergy was realized above the CBSA scale, whereas CSA units retained more systemness, with $H(X)=-\sum_x p(x)\log_2 p(x),$0 realized above CSA. Sectoral filtering into HTM, MHTM, KIS, and HTKIS refined the geography without overturning the state-centric pattern; California dominated HTM and MHTM, Texas stood out in HTKIS, and KIS displayed strong cross-boundary spillovers (Leydesdorff et al., 2017).

3. Target-centric SI in predictor–target information theory

A second major usage treats SI as information that a coalition of predictors provides about a target beyond what is available from singleton predictors or other lower-order constructions. The 2012 “synergistic mutual information” formulation defines

$H(X)=-\sum_x p(x)\log_2 p(x),$1

where the union term is the constrained minimum

$H(X)=-\sum_x p(x)\log_2 p(x),$2

The intuition is explicit: among all joint distributions consistent with what each singleton predictor tells us about $H(X)=-\sum_x p(x)\log_2 p(x),$3, union-information is the minimal possible whole-to-target mutual information, and synergy is the excess of the actual whole over that union. The paper places $H(X)=-\sum_x p(x)\log_2 p(x),$4 between $H(X)=-\sum_x p(x)\log_2 p(x),$5 and $H(X)=-\sum_x p(x)\log_2 p(x),$6, proves invariance to duplicate predictors, and shows that independent predictors can still have positive redundant information about a target. On the canonical circuits used in the paper, XOR has synergy $H(X)=-\sum_x p(x)\log_2 p(x),$7 bit, the “Unq” example has synergy $H(X)=-\sum_x p(x)\log_2 p(x),$8, RdnXor has synergy $H(X)=-\sum_x p(x)\log_2 p(x),$9, and the AND gate yields bounds $$T_{XYZ}=H(X)+H(Y)+H(Z)-H(X,Y)-H(X,Z)-H(Y,Z)+H(X,Y,Z).$$0 bits (Griffith et al., 2012).

A later union-information formulation adopts a communication-channel perspective. In the bivariate case it constructs

$$T_{XYZ}=H(X)+H(Y)+H(Z)-H(X,Y)-H(X,Z)-H(Y,Z)+H(X,Y,Z).$$1

and defines

$$T_{XYZ}=H(X)+H(Y)+H(Z)-H(X,Y)-H(X,Z)-H(Y,Z)+H(X,Y,Z).$$2

The associated synergy is

$$T_{XYZ}=H(X)+H(Y)+H(Z)-H(X,Y)-H(X,Z)-H(Y,Z)+H(X,Y,Z).$$3

The multivariate extension replaces the single $$T_{XYZ}=H(X)+H(Y)+H(Z)-H(X,Y)-H(X,Z)-H(Y,Z)+H(X,Y,Z).$$4 by a set $$T_{XYZ}=H(X)+H(Y)+H(Z)-H(X,Y)-H(X,Z)-H(Y,Z)+H(X,Y,Z).$$5 of admissible conditionally independent reconstructions and uses

$$T_{XYZ}=H(X)+H(Y)+H(Z)-H(X,Y)-H(X,Z)-H(Y,Z)+H(X,Y,Z).$$6

This measure satisfies the extended Williams–Beer axioms for union information, including symmetry, self-redundancy, monotonicity, and equality for monotonicity. The benchmark values emphasized in the paper are XOR $$T_{XYZ}=H(X)+H(Y)+H(Z)-H(X,Y)-H(X,Z)-H(Y,Z)+H(X,Y,Z).$$7 bit, Copy $$T_{XYZ}=H(X)+H(Y)+H(Z)-H(X,Y)-H(X,Z)-H(Y,Z)+H(X,Y,Z).$$8, and AND/OR $$T_{XYZ}=H(X)+H(Y)+H(Z)-H(X,Y)-H(X,Z)-H(Y,Z)+H(X,Y,Z).$$9 bits (Gomes et al., 2024).

A different program constructs synergy through synergistic random variables (SRVs). An SRV $X=G$0 of a source set $X=G$1 satisfies

$X=G$2

After removing redundancy among SRVs and orthogonalizing them into mutually independent OSRVs $X=G$3, the realized synergistic information is

$X=G$4

The same framework defines a source-side synergy capacity $X=G$5. The paper proves non-negativity, the upper bound $X=G$6, zero synergy for a single source, and maximal realized synergy under identity. It also proves coexistence of different synergistic types, illustrates the construction with XOR and mod-3 examples, and reports numerically that synergy is associated with resilience to noise (Quax et al., 2016).

These target-centric SI formulations are often grouped with PID-adjacent work, but they do not coincide. The data repeatedly emphasize that union-information measures, SRV-based measures, and whole-minus-union measures embody different axiomatic commitments and can return different answers on the same distribution.

4. O-information, score-based estimation, and Gaussian log-det SI

A third family characterizes the net redundancy–synergy balance of a multivariate system without designating a privileged target. The central quantity is the O-information,

$X=G$7

with the entropy form

$X=G$8

Its sign convention is opposite to the Triple-Helix literature’s direct use of $X=G$9: $Y=T$0 indicates a redundancy-dominated system, $Y=T$1 a synergy-dominated system, and $Y=T$2 no net dominance. For $Y=T$3, $Y=T$4 reduces to the co-information / interaction information. The paper derives the tight bounds

$Y=T$5

with the $Y=T$6-bit copy attaining the positive extreme and the $Y=T$7-bit XOR the negative extreme, and proves additivity over independent blocks (Rosas et al., 2019).

The 2024 score-based estimator $Y=T$8 addresses the practical estimation problem for continuous variables without restrictive assumptions. It uses the decomposition $Y=T$9 together with a KL-through-scores identity under Gaussian corruption. A single amortized denoiser is trained on three task families—joint, marginal, and conditional—corresponding to learning $Z=O$0, $Z=O$1, and $Z=O$2. Inference estimates

$Z=O$3

and then forms $Z=O$4. The paper does not define a separate SI, but explicitly states that a natural mapping is $Z=O$5 or the nonnegative version $Z=O$6. It reports accurate recovery on synthetic redundancy-only, synergy-only, and mixed systems, and applies the estimator to Allen Brain Visual Behavior Neuropixels data, where $Z=O$7 is higher during “change” flashes than “no-change” flashes (Bounoua et al., 2024).

A closely related 2025 treatment specializes these ideas to real and complex Gaussian, and more generally elliptical, data. With covariance $Z=O$8, precision $Z=O$9, coherence $T_{GTO}<0$0, and standardized precision $T_{GTO}<0$1, the real-valued closed forms are

$T_{GTO}<0$2

The paper identifies two exact sign-based re-expressions as practical “Synergy Index” forms: a system-level $T_{GTO}<0$3, and a local target-centric version based on

$T_{GTO}<0$4

namely $T_{GTO}<0$5. It also introduces structured between-group measures $T_{GTO}<0$6, $T_{GTO}<0$7, and $T_{GTO}<0$8, arguing that standard $T_{GTO}<0$9 can miss between-group synergy when within-group redundancy is predominant. In this framework, DTC is shown to be the KL divergence for the inverse-Wishart pair $T_{GTO}>0$0, which gives it the interpretation of total partial correlation or coherence (Pascual-Marqui et al., 11 Jul 2025).

Across these O-information-based papers, SI therefore becomes a shorthand for the magnitude and sign of synergy-dominance rather than a direct decomposition into unique, redundant, and synergistic atoms.

5. EMG-informed SI in lower-limb muscle coordination

In neuromuscular analysis, SI has a different status. The cycling study models bilateral lower-limb EMG from 14 channels with non-negative matrix factorization,

$T_{GTO}>0$1

where $T_{GTO}>0$2 contains synergy weights and $T_{GTO}>0$3 activation time courses. Preprocessing included normalization to a dynamic MVC, high-pass filtering at $T_{GTO}>0$4 Hz, detrending, rectification, low-pass filtering at $T_{GTO}>0$5 Hz, downsampling from $T_{GTO}>0$6 kHz to $T_{GTO}>0$7 Hz, cycle normalization with crank-angle synchronization, and unit-variance channel scaling before NMF. Four synergies satisfied the selection criteria across power levels, with $T_{GTO}>0$8 and typical local reconstruction quality $T_{GTO}>0$9 (Ahmadi et al., 25 Jul 2025).

The paper’s SI is derived from the largest synergy weight per muscle,

$0.5$00

and then forms, for each joint $0.5$01,

$0.5$02

Thus $0.5$03, $0.5$04 indicates flexor dominance, $0.5$05 extensor dominance, and $0.5$06 balance. The joint-specific formulas are:

$0.5$07

$0.5$08

$0.5$09

Empirically, hip SI remained near $0.5$10 across low, middle, and high power levels, with a significant side asymmetry at middle power. Knee SI shifted from flexor-dominant or near-balanced values at low power to clearly extensor-dominant values at higher power in both limbs, consistent with increased quadriceps recruitment during the downstroke. Ankle SI remained extensor-dominant at all power levels but increased with power, indicating relatively greater dorsiflexor contribution for stabilization under higher loads. The paper interpreted these changes jointly with the Coactivation Index and the Synergy Coordination Index, the latter increasing from $0.5$11 at low power to $0.5$12 at high power, which it described as a reduction in synergy-space size and improved neuromuscular coordination (Ahmadi et al., 25 Jul 2025).

This SI therefore does not measure uncertainty reduction or redundancy–synergy balance in the Shannon sense. It is a joint-level ratio on NMF-derived modular muscle weights.

6. Cross-modal SI in VLM explainability

The most recent usage treats SI as a faithfulness metric for multimodal explanation. The 2026 VLM paper argues that unimodal perturbation metrics collapse in the presence of language priors and modality biases because models can often answer with text alone, leading visual and textual rankings to contradict one another. The reported global Kendall correlation between visual and textual unimodal rankings is $0.5$13 with $0.5$14, and dataset-level correlations remain weak to marginal (Ky et al., 21 May 2026).

The formal basis is the Harsanyi dividend on modality subsets $0.5$15. With utility $0.5$16, the two-modality interaction is

$0.5$17

The paper defines a practical perturbation-trajectory surrogate, the Synergistic Faithfulness score $0.5$18, from joint, image-only, and text-only insertion/deletion curves. For each threshold $0.5$19,

$0.5$20

$0.5$21

These are integrated,

$0.5$22

and averaged:

$0.5$23

The interpretation is direct: $0.5$24 indicates super-additive cross-modal synergy, $0.5$25 indicates independence, additivity, or redundancy, and $0.5$26 indicates antagonistic interference (Ky et al., 21 May 2026).

Computation requires six forward passes per threshold $0.5$27, plus two constants, for total complexity $0.5$28. The paper compares $0.5$29 with an exact Shapley Interaction Index baseline built from an eight-player macro-game and reports Spearman $0.5$30, Kendall $0.5$31, and a $0.5$32 speedup. Evaluating 8 XAI methods across 3 VLMs and 3 datasets, it finds that attention-based methods dominate the synergy-centric leaderboard: AttnLRP has mean $0.5$33, Rollout $0.5$34, and Grad$0.5$35Rollout $0.5$36, whereas VLM-native explainers such as LLaVA-CAM and TAM are lower at $0.5$37. The methodological point is that SI here isolates the joint Harsanyi dividend between modalities rather than rewarding unimodal salience (Ky et al., 21 May 2026).

Taken together, these formulations show that “Synergy Index” functions as a family resemblance term. In innovation studies it is a signed configurational reduction of uncertainty; in target-centric information theory it is excess whole-over-union or information carried by orthogonalized synergistic variables; in O-information work it is the sign-reversed magnitude of synergy-dominance; in EMG it is a bounded modular balance ratio; and in VLM auditing it is a Shapley-interaction surrogate for cross-modal faithfulness. The shared objective is stable, but the quantity itself is field-specific.