Relative Synergy Gap

• Relative Synergy Gap is a measure that quantifies how far a joint configuration departs from expected baselines, such as additive performance or best individual outcomes.
• Its formulations vary by domain, using baselines like average Shapley values, maximum individual performance, or normalized mutual information, with sign and interpretation tailored to context.
• Understanding these gaps helps optimize strategies in areas like human–AI decision-making, economic modeling, and network analysis by revealing hidden inefficiencies and performance discrepancies.

Searching arXiv for the primary paper and related uses of “synergy gap” / “relative synergy gap” across domains. “Relative synergy gap” denotes a family of baseline-relative measures for quantifying how far a joint configuration departs from an expected, additive, redundant, or individually best reference. In the literature, the reference varies by formalism: in characteristic-function games it is a coalition’s “normative expectation” built from average Shapley values; in human–AI decision-making it is the better of human-alone and AI-alone performance; in Triple-Helix regional analysis it is the share of total synergy attributable to foreign participation; in multiplex networks it is a normalized discrepancy between total synergy and total redundancy; in entropy-based statistics it is a normalized higher-order forward difference; and in formal cognitive architectures it is the resource saving obtained by translating a task through another cognitive process rather than executing it directly (Rahwan et al., 2014, Turchi et al., 20 May 2026, Ivanova et al., 2016, Luppi et al., 2023, Whittaker et al., 2015, Goertzel, 2017). Taken together, these formulations suggest a common conceptual pattern—joint effect minus baseline, often followed by normalization—but not a single domain-independent invariant.

1. Common formal pattern

Across the cited literatures, the object being measured is always a deviation from a reference rather than joint performance in isolation. In cooperative games, the deviation is $v(C)-\sum_{i\in C}\overline{\phi}_i(N,v)$; in human–AI systems it is $P_{H+A}-\max\{P_H,P_A\}$; in regional economics it is the share $$g_{\mathrm{int}}=\Delta I_{\mathrm{int}}/\Delta I_{\mathrm{total}}$$, together with the reported difference $s_R-s_I$; in multiplex networks it is $\Delta=|S-R|$ and the normalized quantity $\Gamma=\Delta/F$; in entropy forward differences it is $\Delta_{ijk}$ normalized either by pairwise mutual information or by triple entropy; and in cognitive synergy it is the gap between the direct cost of a transformation and the much smaller indirect cost obtained via a second process (Rahwan et al., 2014, Turchi et al., 20 May 2026, Ivanova et al., 2016, Luppi et al., 2023, Whittaker et al., 2015, Goertzel, 2017).

A second shared feature is that sign and interpretation are domain-specific. In the coalition and human–AI formulations, positive values denote outperformance relative to the baseline. In the forward-difference and Triple-Helix formalisms, synergy is associated with negative quantities: $\Delta_{ijk}<0$ in the entropy expansion, and $\Delta I(G,T,O)<0$ in the regional-information setting (Whittaker et al., 2015, Ivanova et al., 2016). In multiplex networks, the principal “gap” is absolute, $|S-R|$, so the magnitude alone does not indicate whether synergy or redundancy dominates without inspecting $P_{H+A}-\max\{P_H,P_A\}$0 and $P_{H+A}-\max\{P_H,P_A\}$1 separately (Luppi et al., 2023).

This suggests that “relative synergy gap” is best understood as a comparative measurement schema. The baseline may be additive expectation, best individual performance, total system communicability, or a direct computational route; the resulting number is only interpretable relative to that chosen reference.

2. Cooperative-game formulation

In "A Measure of Synergy in Coalitions" (Rahwan et al., 2014), a cooperative game is written as $P_{H+A}-\max\{P_H,P_A\}$2, with $P_{H+A}-\max\{P_H,P_A\}$3 and $P_{H+A}-\max\{P_H,P_A\}$4, and a coalition by $P_{H+A}-\max\{P_H,P_A\}$5. Let $P_{H+A}-\max\{P_H,P_A\}$6 be the Shapley value of player $P_{H+A}-\max\{P_H,P_A\}$7 in game $P_{H+A}-\max\{P_H,P_A\}$8. The average Shapley value of player $P_{H+A}-\max\{P_H,P_A\}$9 over all subgames is defined as

$$g_{\mathrm{int}}=\Delta I_{\mathrm{int}}/\Delta I_{\mathrm{total}}$$0

The synergy value $$g_{\mathrm{int}}=\Delta I_{\mathrm{int}}/\Delta I_{\mathrm{total}}$$1 is then

$$g_{\mathrm{int}}=\Delta I_{\mathrm{int}}/\Delta I_{\mathrm{total}}$$2

Here $$g_{\mathrm{int}}=\Delta I_{\mathrm{int}}/\Delta I_{\mathrm{total}}$$3 is the coalition’s “normative expectation” in the absence of unusual interaction, and $$g_{\mathrm{int}}=\Delta I_{\mathrm{int}}/\Delta I_{\mathrm{total}}$$4 is the gap between actual worth and that expectation.

The paper proves that $$g_{\mathrm{int}}=\Delta I_{\mathrm{int}}/\Delta I_{\mathrm{total}}$$5 is the unique synergy measure $$g_{\mathrm{int}}=\Delta I_{\mathrm{int}}/\Delta I_{\mathrm{total}}$$6 satisfying five axioms. These are P$$g_{\mathrm{int}}=\Delta I_{\mathrm{int}}/\Delta I_{\mathrm{total}}$$7 Symmetric-Synergy, P$$g_{\mathrm{int}}=\Delta I_{\mathrm{int}}/\Delta I_{\mathrm{total}}$$8 Null-Synergy, P$$g_{\mathrm{int}}=\Delta I_{\mathrm{int}}/\Delta I_{\mathrm{total}}$$9 Dummy-Synergy, P$s_R-s_I$0 Normalized-Synergy, and P$s_R-s_I$1 Additive-Synergy. P$s_R-s_I$2 requires that if one defines $s_R-s_I$3, then $s_R-s_I$4 is a game of dummies with $s_R-s_I$5. P$s_R-s_I$6 requires $s_R-s_I$7, so that positive and negative synergy cancel over the power set. P$s_R-s_I$8 requires linearity in the underlying game. An alternative but equivalent axiomatization replaces P$s_R-s_I$9+P$\Delta=|S-R|$0 by a marginal-synergy axiom P$\Delta=|S-R|$1, under which $\Delta=|S-R|$2 depends only on the pattern of marginal contributions of members of $\Delta=|S-R|$3 across subcoalitions.

The derivation proceeds by decomposing $\Delta=|S-R|$4 into an additive part and a synergy part. P$\Delta=|S-R|$5 forces the additive part to be of the form $\Delta=|S-R|$6; P$\Delta=|S-R|$7 pins down $\Delta=|S-R|$8 as average contribution over all coalitions; P$\Delta=|S-R|$9 and P$\Gamma=\Delta/F$0 force $\Gamma=\Delta/F$1; and P$\Gamma=\Delta/F$2 fixes the overall form. Uniqueness is obtained by expanding $\Gamma=\Delta/F$3 in the basis of carrier games $\Gamma=\Delta/F$4 and showing that any admissible $\Gamma=\Delta/F$5 must agree with $\Gamma=\Delta/F$6 on each basis element.

The paper also gives two alternative expressions. One introduces an average-impact quantity $\Gamma=\Delta/F$7, leading to

$\Gamma=\Delta/F$8

Another writes

$\Gamma=\Delta/F$9

for known coefficients $\Delta_{ijk}$0 that sum to zero for each $\Delta_{ijk}$1.

A two-player example illustrates the construction. For $\Delta_{ijk}$2, with $\Delta_{ijk}$3, $\Delta_{ijk}$4, and $\Delta_{ijk}$5, the Shapley values in the 2-player subgame are $\Delta_{ijk}$6 and $\Delta_{ijk}$7. The average Shapley values are $\Delta_{ijk}$8 and $\Delta_{ijk}$9, and therefore

$\Delta_{ijk}<0$0

The paper interprets this as a $\Delta_{ijk}<0$1 positive synergy gap beyond what would be normatively expected from the players’ average stand-alone impact. Zero synergy means $\Delta_{ijk}<0$2; positive synergy means the team outperforms norms; negative synergy means underperformance. The summary also notes that $\Delta_{ijk}<0$3 or $\Delta_{ijk}<0$4 can be used as per-member or fractional synergy gaps, although these are not part of the original axioms.

3. Human–AI decision-making

"Addressing the Synergy Gap: The Six Elements of the Design Space" (Turchi et al., 20 May 2026) introduces the synergy gap as the shortfall observed when human–AI teams improve over humans alone or AI alone but rarely exceed the better of the two. The consolidated exposition formalizes this with an absolute synergy gap

$\Delta_{ijk}<0$5

where $\Delta_{ijk}<0$6 is combined performance, $\Delta_{ijk}<0$7 human-alone performance, and $\Delta_{ijk}<0$8 AI-alone performance. The corresponding relative synergy gap is a dimensionless ratio comparing the same deviation to $\Delta_{ijk}<0$9. Positive values indicate true synergy; negative values indicate a persistent gap. The exposition explicitly notes that no closed-form derivation of the relative quantity appears in the paper itself, but that these metrics underlie its conceptual framing.

The paper’s main contribution is a six-element design space describing what determines whether $\Delta I(G,T,O)<0$0 or the relative gap can be driven above zero. The six elements are sociotechnical context, decision-making frameworks, human decision participants, AI capabilities, interaction, and holistic evaluation. In sociotechnical context, task definability, goal flexibility, time pressure, and stakeholder structure affect the baseline $\Delta I(G,T,O)<0$1, $\Delta I(G,T,O)<0$2, and the available headroom for combination. In decision-making frameworks, normative framing versus descriptive framing alters how combined performance is optimized and evaluated. Human decision participants contribute expertise distribution, working memory, trust disposition, and risk tolerance; poorly calibrated trust drives the relative gap downward. AI capabilities include raw accuracy, transparency, adaptability, and informational scope. Interaction covers roles, proactivity, initiative, and communication modality, which govern when and how contributions are merged. Holistic evaluation extends beyond accuracy to workload, trust calibration, and related side effects.

The paper cites empirical illustrations rather than introducing new data. Vaccaro et al. (2024) is reported with an average effect size for human–AI combination versus the better single agent of Hedges’ $\Delta I(G,T,O)<0$3 with 95% CI $\Delta I(G,T,O)<0$4, consistent with $\Delta I(G,T,O)<0$5. Jacobs et al. (2021) is cited as a case where clinicians receiving machine-learning antidepressant recommendations did not improve accuracy versus baseline. Schemmer et al. (2023) proposed an “Appropriateness of Reliance” metric, with poor AoR correlated with strongly negative synergy gaps.

The design implications are organized into three prescriptions: Build from Context, Not Technology; Combine Carefully; and Iterate & Evaluate Holistically. The exposition adds that prior work such as Inkpen et al. (2023) shows that tuning AI error rates to complement user strengths can recover up to 10–15 percentage points of accuracy in $\Delta I(G,T,O)<0$6, and that longitudinal evaluation can reveal swings in the relative gap from $\Delta I(G,T,O)<0$7 up to $\Delta I(G,T,O)<0$8 as users internalize the AI’s error boundary, as reported for Bansal et al. (2019). The conclusion calls for explicit relative synergy-gap metrics across contexts, quantitative tests of design knobs such as transparency and interaction pacing, domain-specific benchmarks, and adaptive AI ensembles that switch between aligned and complementary models in real time.

4. Information-theoretic and statistical forms

In "Synergy, suppression and immorality: forward differences of the entropy function" (Whittaker et al., 2015), the relevant quantity arises from the forward-difference expansion of entropy. For variables $\Delta I(G,T,O)<0$9, and subset $|S-R|$0, the joint entropy is $|S-R|$1, and the forward differences $|S-R|$2 satisfy

$|S-R|$3

For a triple $|S-R|$4,

$|S-R|$5

and direct algebra yields

$|S-R|$6

The sign of $|S-R|$7 defines the phenomenon. Synergy corresponds to $|S-R|$8, meaning the combined explanatory power of $|S-R|$9 and $P_{H+A}-\max\{P_H,P_A\}$00 for $P_{H+A}-\max\{P_H,P_A\}$01 exceeds the sum of their individual effects. Suppression corresponds to $P_{H+A}-\max\{P_H,P_A\}$02, where $P_{H+A}-\max\{P_H,P_A\}$03 and $P_{H+A}-\max\{P_H,P_A\}$04 provide alternative explanations for $P_{H+A}-\max\{P_H,P_A\}$05. Classical suppression is the special case $P_{H+A}-\max\{P_H,P_A\}$06 but $P_{H+A}-\max\{P_H,P_A\}$07, so that variables marginally independent become dependent once $P_{H+A}-\max\{P_H,P_A\}$08 is introduced.

The summary then introduces two normalized relative quantities. The first is

$P_{H+A}-\max\{P_H,P_A\}$09

which measures the relative synergy as a fraction of the total pairwise information in $P_{H+A}-\max\{P_H,P_A\}$10. The second is

$P_{H+A}-\max\{P_H,P_A\}$11

which measures synergy as a proportion of the total uncertainty in $P_{H+A}-\max\{P_H,P_A\}$12. Both are dimensionless; the sign of $P_{H+A}-\max\{P_H,P_A\}$13 matches the sign of $P_{H+A}-\max\{P_H,P_A\}$14; and $P_{H+A}-\max\{P_H,P_A\}$15.

The paper also gives a computational perspective. Low-order forward differences vanish on disconnected subsets: $P_{H+A}-\max\{P_H,P_A\}$16 whenever the induced subgraph on $P_{H+A}-\max\{P_H,P_A\}$17 is disconnected. Consequently, one can compute only connected node clusters up to a fixed order $P_{H+A}-\max\{P_H,P_A\}$18, using a breadth-first node-cluster algorithm over sparse graphs. This turns the relative synergy gap from a purely analytic notion into a practical statistic for empirical graphical-model analysis.

5. Network topology and cognitive-process cost

In "Quantifying synergy and redundancy in multiplex networks" (Luppi et al., 2023), the framework begins with two undirected, unweighted graphs $P_{H+A}-\max\{P_H,P_A\}$19 and $P_{H+A}-\max\{P_H,P_A\}$20 on the same node set $P_{H+A}-\max\{P_H,P_A\}$21. For each ordered pair of distinct nodes $P_{H+A}-\max\{P_H,P_A\}$22, pairwise efficiency is

$P_{H+A}-\max\{P_H,P_A\}$23

with $P_{H+A}-\max\{P_H,P_A\}$24. The decomposition is

$P_{H+A}-\max\{P_H,P_A\}$25

$P_{H+A}-\max\{P_H,P_A\}$26

$P_{H+A}-\max\{P_H,P_A\}$27

$P_{H+A}-\max\{P_H,P_A\}$28

These satisfy

$P_{H+A}-\max\{P_H,P_A\}$29

Averaging uniformly over node pairs gives global quantities $P_{H+A}-\max\{P_H,P_A\}$30, $P_{H+A}-\max\{P_H,P_A\}$31, $P_{H+A}-\max\{P_H,P_A\}$32, and $P_{H+A}-\max\{P_H,P_A\}$33, with

$P_{H+A}-\max\{P_H,P_A\}$34

The absolute synergy gap is then

$P_{H+A}-\max\{P_H,P_A\}$35

and the relative synergy gap is

$P_{H+A}-\max\{P_H,P_A\}$36

so that $P_{H+A}-\max\{P_H,P_A\}$37. A null-corrected version compares $P_{H+A}-\max\{P_H,P_A\}$38 to degree-preserving random rewiring. In the human structural connectome case, the paper reports $P_{H+A}-\max\{P_H,P_A\}$39, $P_{H+A}-\max\{P_H,P_A\}$40, and therefore $P_{H+A}-\max\{P_H,P_A\}$41 and $P_{H+A}-\max\{P_H,P_A\}$42. In mammalian ex-vivo connectomes, it reports $P_{H+A}-\max\{P_H,P_A\}$43, $P_{H+A}-\max\{P_H,P_A\}$44, $P_{H+A}-\max\{P_H,P_A\}$45, and $P_{H+A}-\max\{P_H,P_A\}$46. The London transport case is described with $P_{H+A}-\max\{P_H,P_A\}$47 and $P_{H+A}-\max\{P_H,P_A\}$48, indicating a major role for synergy in medium-length trips.

A distinct but related cost-based formulation appears in "Toward a Formal Model of Cognitive Synergy" (Goertzel, 2017). There, a cognitive process $P_{H+A}-\max\{P_H,P_A\}$49 is associated with a functor $P_{H+A}-\max\{P_H,P_A\}$50 from a category of subgraphs of the system’s state-transition hypergraph $P_{H+A}-\max\{P_H,P_A\}$51, and two processes $P_{H+A}-\max\{P_H,P_A\}$52 and $P_{H+A}-\max\{P_H,P_A\}$53 are linked by natural transformations $P_{H+A}-\max\{P_H,P_A\}$54 and $P_{H+A}-\max\{P_H,P_A\}$55. The paper defines a process’s confidence in making goal progress and then its stuckness as

$P_{H+A}-\max\{P_H,P_A\}$56

Conjecture 5 proposes that for synergetic processes the commutative diagram induced by the natural transformation is accompanied by the inequality

$P_{H+A}-\max\{P_H,P_A\}$57

The left-hand side is the indirect route through process $P_{H+A}-\max\{P_H,P_A\}$58; the right-hand side is the direct route in process $P_{H+A}-\max\{P_H,P_A\}$59. The summary explicitly interprets this inequality as formalizing a synergy gap: the relative advantage arises when it is much cheaper to translate into another process’s representational mode, perform the task there, and translate back.

These two formalisms differ in object—shortest-path efficiency versus categorical translation cost—but they share a route-based logic. Synergy is not a property of isolated components; it is the gain achieved by combining pathways that are not available, or not efficient, when components are considered separately.

6. Regional-economic usage and comparative interpretation

In "What is the effect of synergy in international collaboration on regional economies?" (Ivanova et al., 2016), synergy is defined through multivariate mutual information among geographical, technological, and organizational distributions of firms. With $P_{H+A}-\max\{P_H,P_A\}$60 denoting the joint probability of geographical cell $P_{H+A}-\max\{P_H,P_A\}$61, organizational class $P_{H+A}-\max\{P_H,P_A\}$62, and technological sector $P_{H+A}-\max\{P_H,P_A\}$63, the Triple-Helix synergy is

$P_{H+A}-\max\{P_H,P_A\}$64

If $P_{H+A}-\max\{P_H,P_A\}$65, the configuration exhibits synergy in the sense of a net reduction of uncertainty. The paper partitions total synergy into domestic and international components: $P_{H+A}-\max\{P_H,P_A\}$66 The “relative synergy gap” $P_{H+A}-\max\{P_H,P_A\}$67 is defined as

$P_{H+A}-\max\{P_H,P_A\}$68

with $P_{H+A}-\max\{P_H,P_A\}$69, and the residual domestic share $P_{H+A}-\max\{P_H,P_A\}$70.

The paper then maps this quantity into economic terms. Let $P_{H+A}-\max\{P_H,P_A\}$71 be aggregate regional turnover and $P_{H+A}-\max\{P_H,P_A\}$72 turnover from foreign-participated firms. Define

$P_{H+A}-\max\{P_H,P_A\}$73

The study posits an invertible function $P_{H+A}-\max\{P_H,P_A\}$74, with $P_{H+A}-\max\{P_H,P_A\}$75 and $P_{H+A}-\max\{P_H,P_A\}$76, and uses the linear approximation

$P_{H+A}-\max\{P_H,P_A\}$77

interpreted as a “returns to synergy” or “effectiveness coefficient.”

The empirical application compares two Norwegian counties. In Møre og Romsdal, $P_{H+A}-\max\{P_H,P_A\}$78, $P_{H+A}-\max\{P_H,P_A\}$79, so $P_{H+A}-\max\{P_H,P_A\}$80; the turnover share is $P_{H+A}-\max\{P_H,P_A\}$81, giving $P_{H+A}-\max\{P_H,P_A\}$82. In Sør-Trøndelag, $P_{H+A}-\max\{P_H,P_A\}$83, $P_{H+A}-\max\{P_H,P_A\}$84, so $P_{H+A}-\max\{P_H,P_A\}$85; the turnover share is $P_{H+A}-\max\{P_H,P_A\}$86, giving $P_{H+A}-\max\{P_H,P_A\}$87. The results section also reports the “relative synergy gap” $P_{H+A}-\max\{P_H,P_A\}$88, equal to $P_{H+A}-\max\{P_H,P_A\}$89 in Møre og Romsdal and $P_{H+A}-\max\{P_H,P_A\}$90 in Sør-Trøndelag. In this formulation, the same term therefore refers both to the foreign share of total regional synergy and to the discrepancy between turnover share and synergy share.

This regional-economic usage makes explicit a point that remains implicit in other literatures: relative synergy gaps are meaningful only with respect to the conversion process being studied. In the Triple-Helix setting the issue is conversion of informational synergy into turnover; in cooperative games it is deviation from additive expectation; in human–AI systems it is failure or success relative to the best single agent; in multiplex networks it is balance between synergy and redundancy; in entropy decompositions it is higher-order interaction relative to lower-order informational quantities; and in cognitive synergy it is the resource advantage of indirect computation. A plausible implication is that numerical values drawn from different formulations are not directly comparable unless their baselines, sign conventions, and normalization constants are aligned first.