Interacting Matrix Model Overview

• Interacting Matrix Models are frameworks that compress complex many-body interactions into matrix entries, products, and commutators to enable exact or variational reformulations.
• They employ diverse formulations—ranging from transfer matrices in lattice systems and cMPS in quantum fields to Lax pair dynamics in integrable models—to capture both static and dynamic properties.
• Applications span statistical mechanics, high-energy theory, and recommender systems, demonstrating the models’ practical significance in organizing and analyzing interaction data.

Searching arXiv for recent and relevant papers on interacting matrix models to ground the article in published work. In the literature represented here, an interacting matrix model is not a single formalism but a family of constructions in which matrices carry the interaction structure of a system. Depending on context, the matrices may be exact transfer operators for a generalized Ising partition function, finite auxiliary parameters of a continuous matrix product state, dynamical variables in integrable or string-theoretic Hamiltonians, multi-linear interaction maps in recommender systems, or the basic degrees of freedom of large-$N$ matrix theories (Sakhno et al., 2020, Eichler et al., 2015, Grekov et al., 2019, Yu et al., 2014, Čubrović, 2022). This range of usage matters: in some works the phrase denotes an exact reformulation of another many-body problem, whereas in others it denotes a theory whose fundamental variables are themselves matrices.

1. Conceptual range and common structure

A useful way to organize the term is by asking what the matrices do. In one class of works, matrices encode interactions of an underlying spin or field system; in another, they are themselves the interacting dynamical variables. A plausible unifying characterization is that the model compresses many-body interaction data into matrix entries, matrix products, or matrix commutators.

Domain	Representative formulation	Paper
Generalized Ising systems	Exact partition-function matrix product	(Sakhno et al., 2020)
Continuous quantum fields	cMPS with finite matrices $Q,R$	(Eichler et al., 2015)
Reversible lattice dynamics	Time-dependent matrix product ansatz	(Klobas et al., 2018)
Integrable classical dynamics	$M$ interacting ${\rm gl}_N$ tops	(Grekov et al., 2019)
AdS/CFT and D-branes	Four-matrix quantum mechanics for giant gravitons	(Harmark, 2016)
Large-$N$ string matrix theory	IKKT matrices $X_\mu,\Psi_\alpha$ with replica questions	(Čubrović, 2022)
Recommender systems	Multi-linear interactive matrix factorization	(Yu et al., 2014)
Constructive field theory	Loop vertex expansion for $\operatorname{Tr}(M M^\dagger)^p$ or $\operatorname{Tr}H^{2p}$	(Krajewski et al., 2017, Krajewski et al., 2019)

A recurring misconception is that “interacting matrix model” automatically refers to random matrix theory or large-$N$ matrix integrals. That identification is too narrow. The generalized Ising construction explicitly remains in the field of transfer-matrix-like exact rewritings rather than random matrix theory (Sakhno et al., 2020). Conversely, the IKKT and constructive matrix-theory papers do study matrix models in the large-$N$ and functional-integral sense (Čubrović, 2022, Krajewski et al., 2017).

2. Exact matrix reformulations of interacting lattice systems

In the generalized Ising setting, the model begins from a decomposition into cells whose energies satisfy

$Q,R$0

where $Q,R$1 are local spins and $Q,R$2 are shared spins. The crucial structural assumption is that cells interact exclusively by the existence of shared spins. Once local spins are summed out,

$Q,R$3

the partition function can be rewritten as a nested sum over shared spins only. By replacing shared-spin values with integer spin-numbers and organizing the resulting frames into incoming and outgoing index sets, one obtains the exact matrix model

$Q,R$4

This construction is exact, systematic, and applies to arbitrary intracell interactions, including higher-order terms, provided the cells can be ordered so that shared spins are eliminated step by step (Sakhno et al., 2020).

The same paper makes the two-dimensional example fully explicit. Internal cell $Q,R$5 carries four shared 2-magnitude spins,

$Q,R$6

and the resulting transfer objects are $Q,R$7 matrices. Their block structure is clarified באמצעות cyclic-shift matrices $Q,R$8 and $Q,R$9, with

$M$0

The “interacting” character is not that matrices interact with one another as dynamical variables, but that multi-spin intracell couplings generate non-factorizable matrix elements (Sakhno et al., 2020).

A different exact matrix-product realization appears for the rule 54 reversible cellular automaton. There the time evolution of any local observable is represented by an explicit time-dependent matrix product ansatz. The auxiliary space is

$M$1

where $M$2 counts collisions, $M$3 is a scattering width, $M$4 encodes local occupancy history, and $M$5 is an activation bit. This tMPA yields exact finite-time expressions for the dynamic structure factor and for an inhomogeneous quench, and those exact results demonstrate coexistence of ballistic and diffusive transport in an interacting deterministic model (Klobas et al., 2018). Here again, the matrix product is exact and dynamical rather than variational.

3. Matrix-product states and interacting quantum fields

In one-dimensional quantum field theory, the phrase takes a more literal variational meaning through continuous matrix product states. A translation-invariant cMPS is written as

$M$6

with $M$7 and $M$8 acting on a finite-dimensional auxiliary Hilbert space $M$9. In this formulation the interacting field theory is converted into a finite-dimensional matrix variational problem: observables of the field are functionals of ${\rm gl}_N$0 and ${\rm gl}_N$1, while interaction terms of the Hamiltonian become traces of matrix products such as ${\rm gl}_N$2 (Eichler et al., 2015).

The experimental realization in circuit QED targets the Lieb–Liniger Hamiltonian,

${\rm gl}_N$3

with

${\rm gl}_N$4

A key conceptual point is that the simulated Hamiltonian is not implemented as the circuit Hamiltonian. Rather, the output field ${\rm gl}_N$5 is identified with the Bose field by

${\rm gl}_N$6

and the energy functional is reconstructed from measured correlators,

${\rm gl}_N$7

The interacting matrix model is therefore embodied in the pair ${\rm gl}_N$8, while the interaction physics of the Lieb–Liniger gas is accessed through first- and second-order correlation functions (Eichler et al., 2015).

The same work also clarifies a frequent misunderstanding about finite-bond-dimension matrix descriptions. Because finite-${\rm gl}_N$9 cMPS do not strictly enforce global $N$0 symmetry, the measured first-order correlator can exhibit a nonzero large-$N$1 plateau. The authors explicitly interpret this as a feature of the variational class, not as true Bose–Einstein condensation in one dimension, which is forbidden by Hohenberg–Mermin–Wagner (Eichler et al., 2015). The interacting matrix description is therefore powerful but not synonymous with an exact representation of all long-distance constraints.

4. Integrable tops and giant-graviton matrix dynamics

In classical integrable systems, the term refers to models whose phase-space variables are matrices. The generalized model of interacting integrable tops introduces coordinates $N$2 and matrix variables $N$3, assembled into a single $N$4 matrix

$N$5

The Hamiltonian is

$N$6

where $N$7 is extracted from $N$8-matrix data and the interaction potential involves

$N$9

The system admits a $X_\mu,\Psi_\alpha$0-valued Lax pair with spectral parameter and a classical dynamical $X_\mu,\Psi_\alpha$1-matrix. In the rank-one sector $X_\mu,\Psi_\alpha$2, one obtains $X_\mu,\Psi_\alpha$3 interacting matrix variables $X_\mu,\Psi_\alpha$4 with pairwise matrix interaction

$X_\mu,\Psi_\alpha$5

which is the paper’s clearest interacting matrix model in the strict dynamical sense (Grekov et al., 2019).

A related but physically distinct example arises in AdS/CFT. For weakly interacting AdS giant gravitons wrapped on $X_\mu,\Psi_\alpha$6, and simultaneously in the strong-coupling, large-energy limit of SU(2) Spin Matrix Theory, the effective dynamics is a four-matrix quantum mechanics with Hermitian matrices $X_\mu,\Psi_\alpha$7,

$X_\mu,\Psi_\alpha$8

subject to the Gauss constraint

$X_\mu,\Psi_\alpha$9

On the D3-brane side this Hamiltonian is derived from the non-abelian DBI action in a high-energy regime with $\operatorname{Tr}(M M^\dagger)^p$0 and $\operatorname{Tr}(M M^\dagger)^p$1; on the gauge-theory side the exact same matrix model emerges from the classical limit of SU(2) Spin Matrix Theory near the unitarity bound $\operatorname{Tr}(M M^\dagger)^p$2 (Harmark, 2016). Here the commutator quartics encode genuine interactions among matrix degrees of freedom rather than a bookkeeping device for another theory.

5. Large-$\operatorname{Tr}(M M^\dagger)^p$3 matrix theory, factorization, and constructive control

The IKKT or type IIB matrix model is a paradigmatic interacting large-$\operatorname{Tr}(M M^\dagger)^p$4 matrix theory with action

$\operatorname{Tr}(M M^\dagger)^p$5

where $\operatorname{Tr}(M M^\dagger)^p$6 are $\operatorname{Tr}(M M^\dagger)^p$7 Hermitian matrices and $\operatorname{Tr}(M M^\dagger)^p$8 are fermionic matrices. In the quenched regime, one studies fluctuations around D-brane backgrounds and then averages over the background ensemble. The replica question is whether averaged partition functions satisfy $\operatorname{Tr}(M M^\dagger)^p$9. The paper shows a sharp contrast: non-interacting BPS configurations may fail to factorize at low replica number and require replica-symmetry-breaking half-wormhole saddles, while interacting D-branes factorize in a simple way. The authors argue that this simple factorization mechanism of interacting branes carries signatures of quantum gravity absent from disordered field theories like SYK (Čubrović, 2022).

Constructive matrix theory studies another sense of interaction: higher-order single-trace or positive matrix potentials. For complex matrices one considers

$\operatorname{Tr}H^{2p}$0

and for Hermitian matrices

$\operatorname{Tr}H^{2p}$1

with $\operatorname{Tr}H^{2p}$2. The main technical tool is a loop vertex representation built from the Fuss–Catalan function $\operatorname{Tr}H^{2p}$3, followed by a forest expansion. In both the complex rectangular case and the Hermitian/real symmetric case, the free energy is shown to be analytic in a “pacman” domain

$\operatorname{Tr}H^{2p}$4

uniformly in the matrix size $\operatorname{Tr}H^{2p}$5 (Krajewski et al., 2017, Krajewski et al., 2019). This literature uses “interacting matrix model” in the strict functional-integral sense: the matrices are the fundamental variables, and the objective is non-perturbative control of their free energy.

6. Data-analytic and algebraic meanings of interaction

In recommender systems, “interactive” or “interacting” matrix models refer to low-rank factorizations that go beyond a single user–item matrix. Multi-Linear Interactive Matrix Factorization extends the standard prediction

$\operatorname{Tr}H^{2p}$6

to

$\operatorname{Tr}H^{2p}$7

The added terms model pairwise interactions between a user and contextual or decision factors such as release date, genre group, genre group size, or time. In experiments on MovieLens 100k and 1M, the method outperforms user-based collaborative filtering, item-based collaborative filtering, and regularized matrix factorization under RMSE evaluation, while retaining per-epoch complexity $\operatorname{Tr}H^{2p}$8 linear in the number of ratings and total latent dimension (Yu et al., 2014). In this setting the “interacting matrix model” is neither dynamical nor large-$\operatorname{Tr}H^{2p}$9; it is a multi-linear statistical model whose interaction structure is carried by coupled latent matrices.

A more specialized algebraic usage appears in the study of the matrix equation

$N$0

Here “interaction” refers to block interactions in the congruence canonical form of $N$1. For the Type $N$2–Type I case,

$N$3

the off-diagonal blocks $N$4 and $N$5 satisfy the coupled equations

$N$6

The difficult case is odd $N$7, for which the paper derives an explicit closed-form solution and shows that the $N$8–I interaction space is $N$9-dimensional; for even $N$0, the interaction is trivial in the sense that $N$1 (Chan et al., 2013). This is a different but legitimate use of the term: interaction means the coupling between canonical matrix blocks inside a linear-algebraic orbit problem.

Across these literatures, the term therefore names a structural role rather than a single theory. Sometimes the matrices summarize interactions of another system exactly; sometimes they are the finite variational parameters of an interacting field; sometimes they are the interacting dynamical variables themselves; and sometimes they encode pairwise latent relations or block couplings. What remains constant is that interaction data are organized into matrix form so that partition functions, correlation functions, equations of motion, optimization objectives, or orbit dimensions can be analyzed through matrix products, matrix spectra, matrix commutators, or matrix factorizations (Sakhno et al., 2020, Eichler et al., 2015, Grekov et al., 2019, Yu et al., 2014, Čubrović, 2022).