Relaxed Balance Theory in Signed Networks

Updated 7 July 2026

Relaxed Balance Theory is a framework that generalizes structural balance by relaxing assumptions like binary sign dichotomy and global partitioning.
It incorporates methods such as frustration indices, conditional probabilities, and walk-based measures to assess network balance under noise and heterogeneity.
Empirical findings demonstrate its ability to capture complex triadic, multi-relational, and neutral dynamics in political and social networks.

Searching arXiv for recent and foundational papers on relaxed balance theory and related balance-theoretic frameworks. Relaxed Balance Theory denotes a family of generalizations of structural balance theory in signed and multi-relational networks. Across these formulations, the common move is to retain the intuition that some local configurations are more stable or prevalent than others, while relaxing at least one classical assumption: exact positive/negative dichotomy, a globally balanced two-group partition, homogeneous node behavior, symmetry of antagonism, independence of triads, or deterministic absence of noise (Dekker et al., 2023, Kirkley et al., 2018, Gallo et al., 2024). In the most specific contemporary formulation, it is a multi-relational, heterogeneity-aware framework that operationalizes Heider’s ideas at the triadic level without imposing a global signed partition, explicitly treating neutrality as a distinct relational state and evaluating mechanism-specific balance via conditional probabilities and dyad–2-path correlations (Dekker et al., 2023).

1. From structural balance to relaxed balance

Classical structural balance theory is defined for signed networks whose edges are positive or negative. At the triad level, a triangle is balanced when the product of its signs is positive, equivalently when it contains an even number of negative edges. The balanced patterns are $(+++)$ and $(+--)$ ; the unbalanced patterns are $(++-)$ and $(---)$ . At the graph level, the Cartwright–Harary structure theorem states that a complete signed graph is balanced if and only if its vertices can be partitioned into two groups with positive edges within groups and negative edges between groups (Kirkley et al., 2018, Gallo et al., 2023).

Relaxed balance begins by weakening this binary and global conception. In Davis’s weak structural balance, vertices may be partitioned into $k \ge 2$ groups with positive intra-group edges and negative inter-group edges. Under this criterion, only $(++-)$ is forbidden; $(---)$ is allowed because its three nodes may belong to three different groups (Gallo et al., 2023, Kirkley et al., 2018). This is the canonical relaxation of strong balance and remains central in later work.

Subsequent formulations relax additional assumptions. One line treats balance as a continuum rather than a strict binary property, using graded proportions of balanced triads or distance-from-balance measures such as the line or frustration index (Aref et al., 2020). Another line argues that deterministic ideal patterns are too brittle for empirical signed graphs and replaces them with statistical models in which balanced structure is inferred under noise and heterogeneity (Gallo et al., 2024). A more recent line abandons the signed dichotomy altogether and models a mutually exclusive and exhaustive set of relation types, including neutrality, at the triadic level (Dekker et al., 2023).

These variants do not replace one another so much as they target different limitations of the classical theory. Strong balance preserves the strictest two-faction interpretation; weak balance admits multiple antagonistic clusters; multilevel and statistical formulations permit partial, noisy, or heterogeneous balance; and the multi-relational formulation makes neutrality and other non-binary ties first-class states.

2. Weak balance, partial balance, and frustration

Weak balance is the historically standard relaxed formulation. Its partition characterization states that a graph is weakly balanced if and only if its vertices can be partitioned into $k \ge 2$ clusters with positive edges within clusters and negative edges between clusters (Kirkley et al., 2018). In triadic terms, this means that $(+++)$ , $(+--)$ , and $(+--)$ 0 are permitted, while $(+--)$ 1 remains the only forbidden type (Gallo et al., 2023).

A related but distinct relaxation treats balance as approximate rather than exact. In this view, a network may contain imbalanced motifs while still being “more or less” balanced. A central quantity is the frustration or line index, which counts the minimum number of edges whose removal, sign change, or placement relative to a partition would be needed to obtain a balanced pattern. For a directed signed graph with $(+--)$ 2, the normalized line index is

$(+--)$ 3

where $(+--)$ 4 is the frustration index. The proportion of edges whose position suits balance is

$(+--)$ 5

These measures place network-level balance on a continuum rather than treating it as all-or-nothing (Aref et al., 2020).

A compact comparison of major balance formulations is useful.

Formulation	Relaxation	Core diagnostic
Strong structural balance	None beyond signed graphs	Triad sign product; two-group partition
Weak structural balance	Allows $(+--)$ 6 groups; permits $(+--)$ 7	Forbids only $(+--)$ 8
Partial/graded balance	Allows local imbalance	Frustration index, line index, triad proportions
Multi-relational relaxed balance	Goes beyond signed dichotomy	$(+--)$ 9, $(++-)$ 0
Statistical balance	Allows noise and heterogeneous block structure	Likelihood, BIC, blockwise sign dominance

The walk-based measures proposed for weak and strong balance provide another route to graded assessment. For weak balance, the measure

$(++-)$ 1

counts exactly-one-negative closures through a resolvent of the positive adjacency matrix, with geometric downweighting of long closed walks. For strong balance, an analogous walk-weighted quantity admits a log-determinant form (Kirkley et al., 2018). These measures operationalize the idea that short loops are more socially salient than long ones.

The practical implication is that “relaxed” can mean either a change in what counts as balanced, as in weak balance, or a change in how strictly balance is required, as in frustration-based and walk-based approaches. The two senses are analytically distinct but frequently conflated in the literature.

3. Multi-relational relaxed balance and agentic zeros

The most explicit recent formulation of Relaxed Balance Theory generalizes Heider’s triadic logic to a set of mutually exclusive and exhaustive relation categories $(++-)$ 2, rather than a binary sign set. Each dyad $(++-)$ 3 is assigned exactly one relation type $(++-)$ 4 at a given time, with adjacency indicators

$(++-)$ 5

This relaxes the signed graph dichotomy and explicitly distinguishes neutral ties from both positive and negative ones (Dekker et al., 2023).

For relation types $(++-)$ 6, the number of $(++-)$ 7– $(++-)$ 8 2-paths from $(++-)$ 9 to $(---)$ 0 is

$(---)$ 1

where $(---)$ 2 and $(---)$ 3. Balance is then formulated not as a global partition property but as a family of mechanism-specific triadic statements. For a given signature $(---)$ 4, one compares

$(---)$ 5

and defines

$(---)$ 6

A positive $(---)$ 7 encodes a balanced tendency for the corresponding motif; a negative $(---)$ 8 encodes an unbalanced tendency (Dekker et al., 2023).

The complementary scale-free statistic is the point-biserial “balance correlation”

$(---)$ 9

where $k \ge 2$ 0 is binary. Because it normalizes by the standard deviation of the 2-path count and by $k \ge 2$ 1, it supports comparison across time slices and across networks with different sizes and densities (Dekker et al., 2023).

In this framework, Heiderian predictions become motif signatures. “Friend-of-friend is friend” is written as $k \ge 2$ 2 and predicts $k \ge 2$ 3 and $k \ge 2$ 4. “Enemy-of-enemy is friend” becomes $k \ge 2$ 5. Hostile balanced responses correspond to $k \ge 2$ 6 or $k \ge 2$ 7, while the classic unbalanced claims correspond to negative $k \ge 2$ 8 and negative $k \ge 2$ 9 for signatures such as $(++-)$ 0 or $(++-)$ 1 in the opposite sense (Dekker et al., 2023).

A distinctive feature is the treatment of “agentic zeros,” or neutral ties. These are neither missing data nor inferred negative ties; in the international-relations application they are explicit states denoting “no alliance and no hostile act” in a given year (Dekker et al., 2023). This matters because neutrality can itself function as a balancing response. Configurations such as $(++-)$ 2, $(++-)$ 3, and $(++-)$ 4 test whether actors systematically choose neutrality under specific relational contexts, something classical signed balance cannot represent.

This formulation therefore replaces global partition tests with direct tests of specific triadic mechanisms. A plausible implication is that it separates distinct behaviors that classical signed balance aggregates into a single partition score.

4. Inference, null models, and the problem of heterogeneity

Whether a network appears weakly balanced, strongly balanced, or not balanced at all depends critically on the benchmark against which it is compared. This point is explicit in work on heterogeneous signed networks, where homogeneous nulls preserve only global counts of positive and negative ties, while heterogeneous nulls preserve node-level positive and negative degrees (Gallo et al., 2023).

Using signed Exponential Random Graph Models, homogeneous benchmarks such as SRGM and SRGM-FT tend to support weak balance: $(++-)$ 5 is under-represented, while $(++-)$ 6 is near expectation or not significant. Heterogeneous benchmarks such as SCM and SCM-FT instead shift evidence toward strong balance, under which both $(++-)$ 7 and $(++-)$ 8 are under-represented and the balanced motifs $(++-)$ 9 and $(---)$ 0 are over-represented (Gallo et al., 2023). The substantive claim is that controlling for actor heterogeneity changes the baseline expectation of fully negative triangles and can reveal stronger avoidance of them.

A related critique targets the classical permutation test that uniformly shuffles signs over the observed edge set. That procedure assumes that positive and negative ties are exchangeable. In empirical social networks, however, positive ties typically exhibit clustering and high embeddedness, whereas negative ties are much sparser and less embedded. A stratified-permutation test that permutes signs only within embeddedness strata preserves this asymmetry and yields markedly different inferences; in the Honduran village networks used as a case study, the old test reported 19 significant villages at 5%, whereas the new stratified test reduced this to approximately 10 villages (Feng et al., 2018). This suggests that some earlier evidence for balance was partly evidence for structural differences between positive and negative ties.

Statistical balance theory reframes the issue further by modeling noisy deviations from an ideal block structure (Gallo et al., 2024). In a simple flip-noise formulation, an unobserved balanced template $(---)$ 1 is corrupted independently with probability $(---)$ 2, giving likelihood

$(---)$ 3

where $(---)$ 4 is the frustration relative to partition $(---)$ 5 and $(---)$ 6 is the number of observed signed edges. For fixed $(---)$ 7, the maximum-likelihood estimate is $(---)$ 8, so minimizing frustration is equivalent to maximizing the likelihood (Gallo et al., 2024).

The same work advocates a signed stochastic block model with block-specific probabilities of positive and negative ties. Rather than asking whether a graph exactly fits an ideal sign pattern, it asks whether estimated blockwise sign probabilities satisfy the dominance inequalities expected under balance. Model selection is performed by BIC, and the outcome can be “statistical traditional balance,” “statistical relaxed balance,” or “no balance beyond chance” (Gallo et al., 2024).

These results make the main methodological controversy clear. Balance is not observable without a null model or generative benchmark, and different benchmarks encode different substantive assumptions about exchangeability, topology, and actor heterogeneity.

5. Directed, non-symmetric, and higher-order generalizations

Relaxed balance has also been extended beyond undirected signed graphs. For directed signed networks, one multilevel framework distinguishes triad-level, subgroup-level, and network-level balance (Aref et al., 2020). At the triad level, balance is defined on transitive semicycles. A transitive semicycle with directed signs $(---)$ 9 is balanced iff

$k \ge 2$ 0

If $k \ge 2$ 1 is the set of transitive triads and $k \ge 2$ 2 the transitive semicycles in triad $k \ge 2$ 3, then

$k \ge 2$ 4

Subgroup balance is quantified by cohesiveness

$k \ge 2$ 5

and divisiveness

$k \ge 2$ 6

evaluated on the optimal bipartition $k \ge 2$ 7; network-level balance is summarized by the normalized line index $k \ge 2$ 8 (Aref et al., 2020). This generalization shows that directionality can be incorporated without abandoning the partition-based language of balance.

A more radical extension dispenses with symmetry of antagonism. In the collusion-based framework, the primitive relation is a possibly non-symmetric attack relation $k \ge 2$ 9, read as $(+++)$ 0 meaning “ $(+++)$ 1 attacks $(+++)$ 2.” The key quadrangular property is collusivity,

$(+++)$ 3

together with irreflexivity and, in some results, totality and surjectivity (Joinet et al., 3 May 2026). Positive relations are then induced from $(+++)$ 4 via the equivalence relations of being attacked by the same agents or attacking the same targets. If $(+++)$ 5 is an irreflexive collusion, the signed frames $(+++)$ 6 and $(+++)$ 7 are weakly balanced even though the negative relation need not be symmetric (Joinet et al., 3 May 2026). This is a genuine relaxation of a condition built into standard balance theory.

Higher-order interaction models relax a different assumption: the independence of triads. In the hybrid model with both triangle and square interactions,

$(+++)$ 8

the quartic term couples neighboring triangles through shared edges (Siboni et al., 2021). On a complete graph with $(+++)$ 9 nodes, the threshold

$(+--)$ 0

separates a triangle-dominated regime from a square-dominated regime. Below $(+--)$ 1, low-temperature states are Heider-dominated and nearly all triangles are balanced; above $(+--)$ 2, quartic interactions can stabilize states with many Heider-imbalanced triangles, including “hell,” in which all edges are negative (Siboni et al., 2021). This does not coincide with Davis weak balance, but it provides a mechanism by which triadic imbalance can persist under higher-order coherence constraints.

Taken together, these generalizations show that “relaxed” does not have a single technical meaning. It can refer to directionality, non-symmetry, higher-order dependence, or multi-level measurement, each relaxing a different constraint of the classical theory.

6. Empirical findings, interpretation, and limitations

The most detailed empirical application of multi-relational Relaxed Balance Theory uses annual Correlates of War data from 1816 to 2007, with positive ties coded as formal defense alliances, negative ties as militarized interstate actions codes 2–5, and neutral ties as dyads with neither positive nor negative relation in a given year (Dekker et al., 2023). The principal result is strong support for “friend-of-friend is friend,” represented by $(+--)$ 3: the balance correlation $(+--)$ 4 is typically between $(+--)$ 5 and $(+--)$ 6 outside the LOA period (1867–1936). “Enemy-of-enemy is friend” receives moderate support, and the hostile balanced motifs $(+--)$ 7 and $(+--)$ 8 are positive and significant over many years (Dekker et al., 2023).

By contrast, classic unbalanced predictions receive little support. The correlation for “friend-of-friend is enemy,” $(+--)$ 9, hovers near zero across almost all years rather than becoming significantly negative. “Enemy-of-enemy is enemy,” $(+--)$ 00, is likewise typically near zero. One surprising exception is that $(+--)$ 01 becomes positive and significant from the late 1950s onward, suggesting a tertius gaudens tendency in which friendships form across conflictual structures, even though the balanced hostile response $(+--)$ 02 remains stronger on average (Dekker et al., 2023).

Neutral ties are central rather than residual. The signature $(+--)$ 03 is strongly negative, with mean around $(+--)$ 04, indicating that mutual allies strongly avoid remaining neutral with each other. Neutrality is also suppressed in $(+--)$ 05, whereas $(+--)$ 06 tends to be positive and significant outside the LOA era. This indicates that neutrality can be either a balancing or an unbalancing response, depending on the surrounding configuration (Dekker et al., 2023).

Other empirical literatures complicate any single interpretation of balance. In directed signed networks, triad-level balance scores are generally high across static datasets, with average $(+--)$ 07, while subgroup cohesiveness and divisiveness are also high on average; yet different levels can diverge, and some systems do not move monotonically toward greater balance over time (Aref et al., 2020). In heterogeneous signed networks, evidence for weak versus strong balance changes when null models preserve node-level signed degrees (Gallo et al., 2023). In the Honduran village data, once positive and negative ties are no longer treated as exchangeable, only marginal evidence for balance remains (Feng et al., 2018). In the statistical block-model framework, some datasets are best interpreted as statistically relaxedly balanced, while others support no modular balance beyond chance (Gallo et al., 2024).

The limitations are correspondingly varied. Multi-relational analyses based on annual time slices are descriptive rather than causal, and coding choices such as giving hostility precedence over alliance can affect some signatures (Dekker et al., 2023). Directed and non-symmetric extensions enlarge the configuration space substantially and can create non-unique optimal partitions (Aref et al., 2020, Joinet et al., 3 May 2026). Weak-balance and frustration-based methods can overfit by returning many clusters unless penalized or regularized (Gallo et al., 2024). Tests based on sign permutation can severely inflate Type I error when negative ties are structurally unlike positive ones (Feng et al., 2018). Higher-order models demonstrate that local triadic imbalance may be stable once square-level interactions are included, but they have mainly been studied on complete graphs under mean-field assumptions (Siboni et al., 2021).

The overall trajectory of the literature is therefore not a simple replacement of classical balance theory by a single relaxed alternative. Rather, it is a progressive decomposition of the classical doctrine into separable assumptions—binary signs, bipartition, homogeneity, symmetry, independence, and exactness—and an attempt to test or relax each assumption in a controlled way. In that sense, Relaxed Balance Theory is best understood as a research program rather than a single theorem.