Interaction Tensor: Methods & Applications

• Interaction Tensor is a higher-order construct that compactly encodes multiparticle or multivariable interactions using algebraic and tensor-network approaches.
• It enables systematic truncation, compression, and evaluation of complex systems in quantum many-body physics, geometric affordance analysis, explainable AI, and nuclear structure.
• Its applications range from matrix product operator representations and variational tensor truncation to spatial affordance mapping and multilinear feature attribution in machine learning.

An interaction tensor is a higher-order tensor construct encoding multiparticle or multivariable interactions, formulated in various domains—including quantum many-body Hamiltonians, geometric affordance analysis, machine learning explainability, and nuclear structure—in order to compactly and algebraically represent complex systems or processes governed by pairwise or higher-order couplings. The “interaction tensor” achieves computational and conceptual efficiency by leveraging the intrinsic structure of interactions, enabling systematic truncation, compression, and efficient evaluation, particularly when coupled with tensor-network methods or explicit analytical decompositions. While widely used terminologies and specific instantiations differ by field, the unifying thread is the use of tensor object(s) to encode, manipulate, or analyze the interdependencies among system components.

1. Interaction Tensor in Quantum Many-Body Hamiltonians

Within quantum many-body physics, the interaction tensor is synonymous with the matrix product operator (MPO) formalism, in which an arbitrary operator (e.g., Hamiltonian) on $N$ sites of local dimension $d$ is encoded as a product of local site tensors: $$H = \sum_{i_1,\dots,i_N \atop j_1,\dots,j_N} \mathrm{Tr}\left[ W^{[1]}_{i_1j_1} W^{[2]}_{i_2j_2} \cdots W^{[N]}_{i_Nj_N} \right] |i_1... i_N\rangle \langle j_1... j_N|$$ Here, $W^{[k]}$ is a rank-4 tensor for site $k$, with two “virtual” and two “physical” indices. For arbitrary two-body Hamiltonians $H = \sum_{i < j} h_{ij}$, an exact MPO construction with bond dimension at most linear in $N$ is always possible; for restricted couplings or specific symmetry structures, the bond dimension can often be reduced to $\mathcal{O}(1)$ or $\mathcal{O}(\log N)$, as in distance-limited or symmetric models.

For Hamiltonians on non-1D geometries, the interaction tensor generalizes naturally to tree tensor networks (TTNOs) and two-dimensional (PEPO) tensor networks, encoding physical interactions and their couplings in a locality-aware tensor structure. These constructs enable efficient state evolution, ground-state searches, and operator compressions by exploiting algebraic properties such as low entanglement scaling or decay of interactions (Fröwis et al., 2010).

2. Compression and Optimization of Interaction Tensors

Analytic and numerical compression schemes are integral to the practical use of interaction tensors. Analytically, decaying or distance-dependent interaction coefficients can often be approximated by a sum of exponentials, each of which admits a low-bond-dimension MPO. The total truncated operator is then the sum (or concatenation) of compressed sub-tensors, controlling operator-norm error. Numerically, variational approaches pose the tensor truncation as a minimization in Hilbert-Schmidt norm between the exact and approximated MPOs, leveraging DMRG-like sweeping, mixed-canonical forms, and iterative SVD updates for rapid convergence even in large systems. These strategies ensure efficient representation and allow scalable simulation of complex quantum dynamics and ground-state properties (Fröwis et al., 2010).

3. Geometry and Affordance: The Interaction Tensor Field

In geometric scene understanding, the interaction tensor is a sparse, weight-driven 3D vector field encoding the “affordance” of an object with respect to another (e.g., where to place a mug under a faucet for “filling”). This tensor field is built by obtaining the bisector surface $B$ (the locus of points equidistant from two objects), and enriching each sample $d$0 with a “provenance” vector $d$1 pointing toward the interacting scene-object surface. Scalar weights $d$2 further quantify the importance of each locus, emphasizing potential or critical contact or action points.

To generate compact, computation-friendly descriptors, a subset of affine-invariant keypoints $d$3 are sampled (preferentially by weight), forming the interaction tensor signature for an affordance. Matching these tensors across previously unseen scenes enables robust generalization and high correlation with human affordance judgments, outperforming nearest-neighbor bisector or naïve alignment baselines (Ruiz et al., 2017).

4. Interaction Tensors in Explainable Machine Learning

In high-dimensional model explainability, the interaction tensor arises in the AI interpretability context as an explicit multilinear representation of feature interactions. The Shapley–Taylor Interaction Index (STII) provides an axiomatic foundation for multi-order feature attributions but is classically intractable due to exponential cost. The IT-SHAP framework recasts STII as a contraction of a Value Tensor and a Weight Tensor (each encoding model outputs and combinatorial weights on all feature subsets), enabling evaluation of all main effects and higher-order interactions.

Importantly, introducing a tensor train (TT) decomposition—where tensors are represented as chain-wise contractions of small “core” tensors—reduces exponential cost to polynomial in input dimension and polylogarithmic parallel depth (NC$d$4), provided the model and distribution tensors admit such structure. This enables scalable, exact calculation of interaction attributions for classes of large black-box models previously infeasible for interaction-aware explainable AI (Hasegawa et al., 5 Dec 2025).

5. Interaction Tensors in Nuclear and Particle Physics

In nuclear structure theory, tensor interactions describe components of the nucleon-nucleon force with explicit tensor operator dependence, often written as

$d$5

where $d$6 is the rank-2 tensor operator in spin-space and $d$7 a radial function. In mean-field EDFs (e.g., Skyrme or Gogny), the tensor interaction contributes terms quadratic in spin–current densities, leading to pronounced effects such as tensorial shell gaps, magic numbers, deformation effects, and level splittings. Ab initio and variational methods leverage optimized basis sets and explicit tensor couplings (as in TOSM+UCOM or AQCM/iSMT frameworks) to capture the dynamically generated high-momentum and pairing correlations driven by tensor terms (Myo et al., 2011, 0811.0279, 0811.1139, Ishizuka et al., 2022).

For coherent elastic neutrino-nucleus scattering, tensor interactions are parameterized as effective operators at the nucleon or quark level and enter the cross section through interference and nuclear response functions. While spin-dependent pieces are suppressed, certain parity-odd tensor contributions exhibit coherent enhancement, dominating the cross section and significantly tightening experimental constraints on new tensor physics in the neutrino sector (Liao et al., 15 Feb 2025).

6. Theoretical and Practical Implications

Interaction tensors unify disparate threads in many-body quantum theory, geometric analysis, feature attribution, and nuclear structure by encapsulating the complexity of multiparticle or multifeature couplings into a compact, algebraically tractable object. Their utility hinges on the existence of efficient decomposition (e.g., MPO, PEPO, TT), the ability to compress high-order dependencies, and their amenability to analytic or parallel computation.

In quantum simulation and condensed matter, optimal interaction tensor construction and truncation directly control algorithmic scaling, enabling simulations of systems with long-range or structured interactions. In scene affordance and machine learning, interaction tensors enable robust generalization, explicit modeling of context-sensitive interaction patterns, and tractable exploration of the exponentially large combinatorics of feature interplay—provided suitable tensor decompositions exist.

The tensor formalism is agnostic to the specific physical or mathematical field, requiring only that the system's interaction structure be interpretable as a multilinear or tensorial map. Advances in tensor network representations, decomposition theory, and computational methods continue to expand the scope and tractability of problems accessible to interaction tensor-based approaches.