Bilinear Relational Structure Overview

• Bilinear relational structure is a formalism that uses bilinear maps or operators to encode interactions between two domains, unifying algebraic, geometric, and learning frameworks.
• It underpins methodologies from knowledge-graph embedding to bilinear control, with applications including rank-one feasibility problems, error estimation in reinforcement learning, and tensor decompositions.
• Its broad utility is demonstrated in areas such as symmetric matrix attention, multirelation modeling, and the analysis of structured dynamics in control and graph inference.

Searching arXiv for the cited papers and closely related work on bilinear relational structure. Bilinear relational structure denotes a family of mathematical and computational formalisms in which a relation between two argument domains is encoded through a bilinear map, a bilinear form, or a bilinear operator. Across the literature, this idea appears in several non-equivalent but closely related senses: polynomial relations can induce nondegenerate bilinear spaces, knowledge-graph triples can be scored by relation-specific matrices, Bellman errors can factor as inner products of hypothesis-side and rollout-side representations, and covariance-based graph inference can be organized through bilinear attention on symmetric positive definite matrices (Jouve et al., 2010, Wang et al., 2017, Du et al., 2021, Froehlich et al., 2024). A broader, but explicitly non-strict, extension also appears in the study of multirelations, where the relevant structure is two-layered rather than bilinear in the algebraic sense (Furusawa et al., 2023). This suggests that the term is best understood as a structural motif: relations are not treated as primitive labels alone, but as operators that mediate interactions between two organized spaces.

1. Conceptual scope and canonical forms

In strict algebraic usage, a bilinear relation is a map that is linear in each argument separately. Representative formulas in the cited literature include the scalar equations

$y^T A_i x = g_i,$

used to define systems of bilinear equations (Johnson et al., 2013), the knowledge-graph score

$s_k(i,j)=a_i^\top R_k a_j,$

used in multi-relational link prediction (Wang et al., 2017), the bandit reward

$y_t = x_t^\top \Theta^* z_t + \eta_t,$

used for pairwise decision problems with two entity types (Jun et al., 2019), and the probe score

$f_r(s_l, o_l) = s_l^\top M_r o_l,$

used to test whether language-model hidden states carry relation-specific bilinear geometry (Kim et al., 26 Sep 2025).

A second recurring pattern is that a bilinear structure often mediates between an error term and an observable statistic. In reinforcement learning, the defining factorization of a Bilinear Class is

$\langle W_h(g)-W_h(f^\star), X_h(f)\rangle,$

which simultaneously upper bounds Bellman error and is estimable from data collected under the rollout induced by $f$ (Du et al., 2021). In combinatorial tensor theory, the same structural intuition appears through the equivalence, up to constants, of slice rank, geometric rank, and analytic rank for $3$-tensors associated with bilinear maps (Cohen et al., 2021).

The common feature is not a single universal syntax but a common architecture: one side encodes a candidate relation or hypothesis, the other side encodes the argument pair, state distribution, or dependency object on which that relation acts. This interpretation is explicit in some papers and implicit in others. Where the term is stretched beyond strict bilinearity, as in binary multirelations with angelic and demonic choice, the literature itself marks the distinction and treats the construction as only partially analogous to bilinear structure (Furusawa et al., 2023).

2. Algebraic and geometric constructions

One classical algebraic realization of bilinear relational structure is the skew Bezoutian. For $w=p/q$ with $p,q$ coprime reciprocal or skew-reciprocal polynomials of the same degree, the construction defines

$V:=k[T,T^{-1}]/(q^*)$

with basis $s_k(i,j)=a_i^\top R_k a_j,$0 and bilinear form

$s_k(i,j)=a_i^\top R_k a_j,$1

whose Gram matrix is the skew Bezoutian $s_k(i,j)=a_i^\top R_k a_j,$2 (Jouve et al., 2010). Under

$s_k(i,j)=a_i^\top R_k a_j,$3

the resulting space is $s_k(i,j)=a_i^\top R_k a_j,$4-symmetric: symmetric when $s_k(i,j)=a_i^\top R_k a_j,$5, alternating when $s_k(i,j)=a_i^\top R_k a_j,$6. The construction is nondegenerate exactly when $s_k(i,j)=a_i^\top R_k a_j,$7 and $s_k(i,j)=a_i^\top R_k a_j,$8 are coprime, since

$s_k(i,j)=a_i^\top R_k a_j,$9

The same formalism yields explicit isometries with prescribed characteristic polynomial, discriminant constraints, spinor norm, and Jordan form, and it identifies the invariant form of certain hypergeometric groups (Jouve et al., 2010).

A different but complementary viewpoint treats a bilinear relation as a linear constraint on an outer-product matrix. For a system

$y_t = x_t^\top \Theta^* z_t + \eta_t,$0

introducing

$y_t = x_t^\top \Theta^* z_t + \eta_t,$1

turns the bilinear system into the linear system

$y_t = x_t^\top \Theta^* z_t + \eta_t,$2

together with the nonlinear condition that $y_t = x_t^\top \Theta^* z_t + \eta_t,$3 have rank one (Johnson et al., 2013). The paper’s central message is that the essential difficulty is therefore not the linear constraints themselves but the rank-one feasibility problem inside an affine matrix space. Universal solvability is sharply restricted: over $y_t = x_t^\top \Theta^* z_t + \eta_t,$4, $y_t = x_t^\top \Theta^* z_t + \eta_t,$5, or a finite field, if a system is solvable for every right-hand side, then

$y_t = x_t^\top \Theta^* z_t + \eta_t,$6

while every system with $y_t = x_t^\top \Theta^* z_t + \eta_t,$7 is always solvable (Johnson et al., 2013).

A moduli-theoretic generalization appears in the 2026 study of the Bilinear scheme

$y_t = x_t^\top \Theta^* z_t + \eta_t,$8

which parameterizes families of quotient modules together with a surjection

$y_t = x_t^\top \Theta^* z_t + \eta_t,$9

and is realized as a closed subfunctor of a product of Quot schemes (Obcowska, 4 Jan 2026). Its tangent space is described by compatible triples

$f_r(s_l, o_l) = s_l^\top M_r o_l,$0

and the geometry is typically reducible: the paper proves that

$f_r(s_l, o_l) = s_l^\top M_r o_l,$1

is reducible for all $f_r(s_l, o_l) = s_l^\top M_r o_l,$2 whenever $f_r(s_l, o_l) = s_l^\top M_r o_l,$3 (Obcowska, 4 Jan 2026). This places bilinear relational structure in the same geometric orbit as Hilbert and Quot functors, but with tensor-product quotient data as the basic object.

3. Categories, symmetries, and rigid identities

A categorical treatment is developed in the theory of adjoint-morphisms between bimaps. For bimaps $f_r(s_l, o_l) = s_l^\top M_r o_l,$4 and $f_r(s_l, o_l) = s_l^\top M_r o_l,$5, an adjoint-morphism

$f_r(s_l, o_l) = s_l^\top M_r o_l,$6

satisfies

$f_r(s_l, o_l) = s_l^\top M_r o_l,$7

These morphisms form the category $f_r(s_l, o_l) = s_l^\top M_r o_l,$8, in which transpose defines a duality, products are orthogonal sums, and kernels and cokernels are constructed from module-theoretic kernels together with orthogonality quotients (Wilson, 2010). The paper proves that $f_r(s_l, o_l) = s_l^\top M_r o_l,$9 is a complete and cocomplete abelian category and that adjoint-isomorphism coincides with principal isotopism (Wilson, 2010). In this setting, bilinear division maps become the simple objects relative to nondegenerate adjoint-morphisms.

On the symmetry side, the group preserving an arbitrary bilinear form

$\langle W_h(g)-W_h(f^\star), X_h(f)\rangle,$0

is governed by a decomposition of the bilinear space into odd degenerate, even degenerate, and nondegenerate pieces (Szechtman, 2013). The odd part contributes a unipotent radical $\langle W_h(g)-W_h(f^\star), X_h(f)\rangle,$1 and a Levi-type factor

$\langle W_h(g)-W_h(f^\star), X_h(f)\rangle,$2

the even part yields a centralizer $\langle W_h(g)-W_h(f^\star), X_h(f)\rangle,$3 of an explicit nilpotent endomorphism, and the nondegenerate part reduces via the asymmetry operator to centralizers in $\langle W_h(g)-W_h(f^\star), X_h(f)\rangle,$4, $\langle W_h(g)-W_h(f^\star), X_h(f)\rangle,$5, or $\langle W_h(g)-W_h(f^\star), X_h(f)\rangle,$6 (Szechtman, 2013). The result shows that preserving a bilinear relation can produce a symmetry group that is neither purely classical nor purely linear, but a structured extension mixing reductive and unipotent pieces.

A much more rigid functional-identity problem appears for strictly upper triangular matrix rings. For bilinear maps

$\langle W_h(g)-W_h(f^\star), X_h(f)\rangle,$7

satisfying

$\langle W_h(g)-W_h(f^\star), X_h(f)\rangle,$8

the diagonal trace is forced into the form

$\langle W_h(g)-W_h(f^\star), X_h(f)\rangle,$9

where $f$0 is central-valued and $f$1 lands in the small corner space

$f$2

(Bounds et al., 16 Feb 2025). A sharper corollary writes

$f$3

This is a particularly explicit instance of bilinear relational rigidity: almost all candidate interactions are excluded by a commuting-trace condition.

4. Relational representation in learning systems

In knowledge-graph completion, bilinear relational structure is the organizing template behind a family of embedding models. The generic score

$f$4

specializes to RESCAL with full $f$5, DISTMULT with diagonal $f$6, HolE with circulant $f$7, and ComplEx with complex diagonal structure (Wang et al., 2017). The paper studies universality and subsumption at the ranking level, proving, for example, that $f$8 is universal, that DISTMULT is not universal because it can only represent symmetric relation score matrices, and that $f$9 subsumes $3$0 (Wang et al., 2017). The central structural parameter is the constraint class imposed on the relation operator $3$1.

For compositional analogy detection from word embeddings, the question is narrower: whether second-order cross-coordinate interactions are needed at all. The generalized relation operator

$3$2

contains a bilinear tensor term and linear terms (Hakami et al., 2017). Under standardized, uncorrelated embeddings and relational independence, the paper’s Theorem 1 states that the expected loss

$3$3

is independent of $3$4; with regularization, the optimum collapses to a linear form in which PairDiff,

$3$5

is a special case (Hakami et al., 2017). Here bilinear relational structure is not absent from the formalism, but it becomes redundant under the analyzed assumptions.

An online version of the same theme appears in bilinear bandits with low-rank structure. Actions are pairs $3$6, rewards satisfy

$3$7

and the unknown relation matrix $3$8 has rank $3$9 (Jun et al., 2019). The algorithm ESTR first estimates the row and column spaces of $w=p/q$0, then runs an almost-low-dimensional linear bandit with anisotropic regularization. The resulting regret bound

$w=p/q$1

improves on the naive reduction

$w=p/q$2

(Jun et al., 2019). In this setting, the relation matrix is the primary object, and low rank is the intrinsic complexity measure.

The same operator form has recently been used to interpret language-model behavior on synthetic relational knowledge graphs. A bilinear probe scores facts by

$w=p/q$3

and successful models exhibit approximate inverse and composition laws

$w=p/q$4

(Kim et al., 26 Sep 2025). Models in which this structure emerges largely escape the reversal curse and support logically consistent model editing, with the paper reporting

$w=p/q$5

between best bilinear probe accuracy and best post-edit logical generalization (Kim et al., 26 Sep 2025). A plausible implication is that bilinear internal geometry can couple fact retrieval and edit propagation in a single relational algebra.

5. Structured dynamics, control, and graph inference

In bilinear control theory, the defining object is often not a bare first-order state equation but a hierarchy of structured subsystem transfer functions. For structured bilinear systems, the $w=p/q$6-th subsystem transfer function is written

$w=p/q$7

with reductions obtained by projecting $w=p/q$8 directly rather than flattening the model to an unstructured first-order system (Benner et al., 2020). This framework preserves second-order mechanical structure and time-delay structure, and two-sided projection yields interpolation of subsystem transfer functions together with mixed higher-order conditions (Benner et al., 2020).

The parametric extension introduces explicit dependence on a parameter vector $w=p/q$9,

$p,q$0

and defines corresponding structured subsystem transfer functions

$p,q$1

from matrix-valued functions $p,q$2 (Benner et al., 2020). The main interpolation theorems specify recursive basis constructions that match selected frequency points and parameter values, and when the same left and right interpolation points are used, parameter sensitivities

$p,q$3

are matched implicitly (Benner et al., 2020). In this literature, bilinear relational structure is the ordered interaction of input injection, resolvent propagation, bilinear coupling, and observation.

A statistically different but conceptually parallel use appears in graph structure inference with the Bilinear Attention Mechanism. BAM constructs channelwise covariance matrices $p,q$4 from transformed data, forms SPD key/query objects by channel mixing, and computes attention/output through matrix sandwiches of the form

$p,q$5

channelwise (Froehlich et al., 2024). The associated SPD softmax

$p,q$6

preserves manifold structure, and ablation results show a clear degradation when the bilinear layer is removed (Froehlich et al., 2024). Here the relation is not between latent entity vectors but between variable-variable covariance descriptors, and bilinearity supplies a richer receptive field than direct pairwise similarity.

6. Rank, randomness, and generalization

At the tensor level, bilinear relational structure is encoded by a $p,q$7-tensor

$p,q$8

or equivalently a trilinear form

$p,q$9

For the associated bilinear map $V:=k[T,T^{-1}]/(q^*)$0, analytic rank satisfies

$V:=k[T,T^{-1}]/(q^*)$1

while geometric rank is

$V:=k[T,T^{-1}]/(q^*)$2

and slice rank measures decomposition into simple slices (Cohen et al., 2021). The main theorem

$V:=k[T,T^{-1}]/(q^*)$3

shows that combinatorial structure, algebraic degeneracy, and Fourier bias are equivalent up to constants for $V:=k[T,T^{-1}]/(q^*)$4-tensors (Cohen et al., 2021). In this sense, bilinear relational structure is the precise algebraic content of non-randomness.

In reinforcement learning, the Bilinear Class framework uses a different factorization. For each stage $V:=k[T,T^{-1}]/(q^*)$5, there exist maps $V:=k[T,T^{-1}]/(q^*)$6 and $V:=k[T,T^{-1}]/(q^*)$7 such that Bellman error is controlled by

$V:=k[T,T^{-1}]/(q^*)$8

and the same quantity is estimable from data through a discrepancy function $V:=k[T,T^{-1}]/(q^*)$9 (Du et al., 2021). The paper proves polynomial sample complexity in terms of a supervised-learning generalization term $s_k(i,j)=a_i^\top R_k a_j,$00, and extends the theory to infinite-dimensional RKHS settings using information gain rather than explicit feature dimension (Du et al., 2021). This is an especially general formulation of bilinear relational structure: one factor represents hypothesis-side error, the other rollout-side coverage, and their inner product governs both estimation and control.

Taken together, these results show that bilinear relational structure has become a unifying language for problems in algebra, geometry, learning theory, control, and combinatorics. In some contexts it encodes exact orthogonality and nondegeneracy; in others it measures low-rank compatibility, latent interaction, or estimable error. What remains constant is the organizing principle that a relation between two domains is best captured not by isolated labels or unrestricted nonlinearities, but by a structured interaction law that is linear in each side separately.