Basis Interaction Operator Overview

• Basis Interaction Operator is an operator-level construct that represents a system's action through interactions among basis elements, highlighting sparse and interpretable structures.
• In Hilbert-space operator learning, it is implemented as a coefficient map between learned input and output bases, achieving significant accuracy improvements over traditional methods.
• It also facilitates Hamiltonian symmetry recovery and optical circuit simulation by encoding mode transformations via structured tensor-network representations.

Searching arXiv for recent and directly relevant papers on “Basis Interaction Operator” and adjacent formulations. A Basis Interaction Operator is an operator-level construct in which the essential action of a system is represented through interactions among basis elements rather than through raw coordinates, fixed sensor values, or an unstructured ambient representation. In recent arXiv usage, the term and closely related constructions appear in several distinct settings: as a coefficient-space map between learned input and output bases in Hilbert-space operator learning; as a sparse generator extracted from a degenerate parent operator subspace to reveal Hamiltonian structure, symmetries, and interaction geometry; and as an operator-basis tensor-network object encoding mode transformations in linear optical circuits (Ingebrand et al., 2024, Ye-Ming et al., 4 Mar 2026, Cilluffo et al., 3 Feb 2025). The literature does not supply a single universal formal definition; instead, it exhibits a common structural pattern in which a carefully chosen basis makes the physically or computationally relevant interaction law explicit.

1. Conceptual scope and defining pattern

In Basis-to-Basis (B2B) operator learning, the operative object is a finite-dimensional map between coefficient vectors after separate bases have been learned for the input and output Hilbert spaces. For linear operators, that map is a matrix $A$ obtained from paired coefficient data $(\alpha_n,\beta_n)$ through

$$\min_A \frac{1}{N}\sum_{n=1}^{N}\|\beta_n-A\alpha_n\|^2,$$

and the reconstructed output is

$$\hat{\mathcal{T}}f(y)=\sum_{j=1}^{\ell}(A\alpha)_j\, h_j(y).$$

In that setting, the finite-dimensional coefficient-space transformation is the object that plays the role of a basis interaction operator (Ingebrand et al., 2024).

In O-Sensing, the phrase basis interaction operator or basis interaction structure refers to sparse generators extracted from a large degenerate operator subspace determined by low-energy eigenstate constraints. The distinguished generator is then identified as the Hamiltonian by maximizing spectral entropy, while the remaining sparse generators encode intrinsic symmetries, geometric symmetries, and other conserved quantities (Ye-Ming et al., 4 Mar 2026).

In the operator-basis Matrix Product State formalism for optical circuits, the interaction is encoded directly at the operator level. The transformed ladder operator

$\mathcal{U}(a_i)=\sum_{k=1}^M u_k^i\, a_k$

is represented in a reduced local operator basis, and the interferometer matrix entries $u_k^i$ specify how basis operators propagate and combine. The paper does not define a separate named object called a Basis Interaction Operator, but the interaction law is explicit in the operator-basis tensors and their contraction rules (Cilluffo et al., 3 Feb 2025).

Setting	Basis interaction object	Role
B2B operator learning	Matrix $A$	Maps input coefficients $\alpha$ to output coefficients $\beta$
O-Sensing	Sparse basis of $\mathcal K$ and entropy-selected Hamiltonian	Separates Hamiltonian from symmetries and reveals interaction geometry
Optical operator-basis MPS	Local operator-basis tensors with entries $(\alpha_n,\beta_n)$0	Encodes mode-to-mode propagation and amplitude contraction

The unifying idea is that the operator is not treated as an opaque transformation. Instead, its action is factorized through a basis in which the relevant interaction structure becomes sparse, diagonal, local, or otherwise computationally tractable.

2. Coefficient-space basis interaction in Hilbert-space operator learning

B2B operator learning formulates operator learning on Hilbert spaces of functions as a two-stage problem: first learn basis functions for the input space $(\alpha_n,\beta_n)$1 and output space $(\alpha_n,\beta_n)$2, then learn the map between the corresponding coefficient vectors (Ingebrand et al., 2024). For the input space, the representation is

$(\alpha_n,\beta_n)$3

while the output is written as

$(\alpha_n,\beta_n)$4

The central reduction is therefore

$(\alpha_n,\beta_n)$5

A notable feature is that coefficients are not inferred by an auxiliary neural network. They are computed by least squares. Given samples $(\alpha_n,\beta_n)$6, the input coefficients solve

$(\alpha_n,\beta_n)$7

with closed-form solution

$(\alpha_n,\beta_n)$8

where $(\alpha_n,\beta_n)$9. The same construction is used for the output coefficients $$\min_A \frac{1}{N}\sum_{n=1}^{N}\|\beta_n-A\alpha_n\|^2,$$0. Because the coefficients are obtained by projection, sample locations $$\min_A \frac{1}{N}\sum_{n=1}^{N}\|\beta_n-A\alpha_n\|^2,$$1 and $$\min_A \frac{1}{N}\sum_{n=1}^{N}\|\beta_n-A\alpha_n\|^2,$$2 do not need to be fixed across functions. This is the basis-level mechanism behind the claim that B2B circumvents fixed-sensor requirements.

For linear operators, the matrix $$\min_A \frac{1}{N}\sum_{n=1}^{N}\|\beta_n-A\alpha_n\|^2,$$3 is the cleanest realization of a basis interaction operator. It summarizes how coordinates in the learned input basis are transformed into coordinates in the learned output basis. The paper further proves a linearity result: if $$\min_A \frac{1}{N}\sum_{n=1}^{N}\|\beta_n-A\alpha_n\|^2,$$4 and $$\min_A \frac{1}{N}\sum_{n=1}^{N}\|\beta_n-A\alpha_n\|^2,$$5 is linear, then

$$\min_A \frac{1}{N}\sum_{n=1}^{N}\|\beta_n-A\alpha_n\|^2,$$6

This gives a direct operator-theoretic justification for representing the learned transformation by a single coefficient-space matrix.

The same framework extends to nonlinear operators by keeping the learned bases fixed and replacing the matrix $$\min_A \frac{1}{N}\sum_{n=1}^{N}\|\beta_n-A\alpha_n\|^2,$$7 by a neural coefficient-to-coefficient map,

$$\min_A \frac{1}{N}\sum_{n=1}^{N}\|\beta_n-A\alpha_n\|^2,$$8

The decomposition into basis learning and coefficient learning is preserved; only the interaction law in coefficient space ceases to be linear.

The paper also derives variants directly analogous to spectral factorizations. For compact linear operators,

$$\min_A \frac{1}{N}\sum_{n=1}^{N}\|\beta_n-A\alpha_n\|^2,$$9

and for compact self-adjoint operators,

$$\hat{\mathcal{T}}f(y)=\sum_{j=1}^{\ell}(A\alpha)_j\, h_j(y).$$0

In these cases the basis interaction becomes diagonal in the singular or eigen basis. This suggests that the most analytically transparent form of a basis interaction operator is often the one in which the chosen basis aligns with the operator’s spectral structure.

Empirically, B2B is evaluated on seven benchmark tasks: anti-derivative operator, derivative operator, 1D Darcy flow, 2D Darcy flow on an L-shaped domain, elastic plate / plane stress elasticity, parameterized heat equation, and Burger’s equation. The paper reports that it demonstrates a two-orders-of-magnitude improvement in accuracy over existing approaches on several benchmark tasks, including about $$\hat{\mathcal{T}}f(y)=\sum_{j=1}^{\ell}(A\alpha)_j\, h_j(y).$$1 MSE on the anti-derivative problem versus $$\hat{\mathcal{T}}f(y)=\sum_{j=1}^{\ell}(A\alpha)_j\, h_j(y).$$2 for the best DeepONet baseline shown, and $$\hat{\mathcal{T}}f(y)=\sum_{j=1}^{\ell}(A\alpha)_j\, h_j(y).$$3 MSE on the derivative problem (Ingebrand et al., 2024).

3. Sparse basis interaction structure in Hamiltonian and symmetry recovery

O-Sensing treats Hamiltonian learning as an operator-identification problem in a large, geometry-agnostic local basis (Ye-Ming et al., 4 Mar 2026). Hermitian operators are expanded as

$$\hat{\mathcal{T}}f(y)=\sum_{j=1}^{\ell}(A\alpha)_j\, h_j(y).$$4

where $$\hat{\mathcal{T}}f(y)=\sum_{j=1}^{\ell}(A\alpha)_j\, h_j(y).$$5 contains few-body Pauli strings up to three sites. Given sampled low-energy eigenstates $$\hat{\mathcal{T}}f(y)=\sum_{j=1}^{\ell}(A\alpha)_j\, h_j(y).$$6, the eigenvalue condition is imposed through the zero-variance constraint

$$\hat{\mathcal{T}}f(y)=\sum_{j=1}^{\ell}(A\alpha)_j\, h_j(y).$$7

This yields quadratic forms

$$\hat{\mathcal{T}}f(y)=\sum_{j=1}^{\ell}(A\alpha)_j\, h_j(y).$$8

with covariance matrices

$$\hat{\mathcal{T}}f(y)=\sum_{j=1}^{\ell}(A\alpha)_j\, h_j(y).$$9

The resulting parent operator subspace

$\mathcal{U}(a_i)=\sum_{k=1}^M u_k^i\, a_k$0

is generally highly degenerate. It contains the parent Hamiltonian, symmetries, and dense mixtures of many local terms. The key problem is therefore not existence of solutions but selection of a physically meaningful basis within $\mathcal{U}(a_i)=\sum_{k=1}^M u_k^i\, a_k$1.

O-Sensing resolves this by searching for a maximally sparse operator basis. If $\mathcal{U}(a_i)=\sum_{k=1}^M u_k^i\, a_k$2 is an initial basis of $\mathcal{U}(a_i)=\sum_{k=1}^M u_k^i\, a_k$3, the method seeks an invertible change of basis $\mathcal{U}(a_i)=\sum_{k=1}^M u_k^i\, a_k$4 such that $\mathcal{U}(a_i)=\sum_{k=1}^M u_k^i\, a_k$5 minimizes support: $\mathcal{U}(a_i)=\sum_{k=1}^M u_k^i\, a_k$6 Because the exact $\mathcal{U}(a_i)=\sum_{k=1}^M u_k^i\, a_k$7 problem is NP-hard, the implemented algorithm uses a two-stage relaxation: global exploration via $\mathcal{U}(a_i)=\sum_{k=1}^M u_k^i\, a_k$8-spikiness,

$\mathcal{U}(a_i)=\sum_{k=1}^M u_k^i\, a_k$9

followed by local refinement through $u_k^i$0-minimization,

$u_k^i$1

In this framework, a basis interaction operator is not merely a matrix between coefficient vectors. It is a sparse generator in the recovered basis of conserved operators. The Hamiltonian is then chosen by maximizing spectral entropy,

$u_k^i$2

where $u_k^i$3 is the multiplicity of the $u_k^i$4-th distinct eigenvalue of $u_k^i$5 within the sampled low-energy subspace. The intuition stated in the paper is that symmetry operators often have lower entropy because of larger degeneracies, whereas the Hamiltonian produces a more resolved spectrum.

For Heisenberg models on connected Erdős–Rényi graphs, the recovered sparse basis includes the parent Hamiltonian

$u_k^i$6

global symmetry or redundancy operators such as

$u_k^i$7

families $u_k^i$8, and, when graph automorphisms are present, geometric conserved operators such as

$u_k^i$9

The support of the recovered Hamiltonian in the local basis reveals the interaction graph directly.

The paper also establishes a learnability phase diagram across graph densities, including a pronounced confusion regime. There, parsimony can favor a dual description on the complement graph. The corresponding alternative operator is

$A$0

and the analytic parsimony crossover occurs around

$A$1

This is an important limitation: sparsity fixes the basis ambiguity only when the physical Hamiltonian is also the parsimonious representative.

4. Operator-basis interaction in linear optical circuits

The operator-basis MPS formalism for optical circuits is a Heisenberg-picture tensor-network representation of bosonic ladder operators (Cilluffo et al., 3 Feb 2025). Instead of working in a local Fock-state basis, it represents the evolved output operator directly: $A$2 where $A$3 is the interferometer matrix. The paper interprets this as a basis-interaction idea because the entries $A$4 encode how input-mode operators are routed into output-mode operators.

Each transformed ladder operator is written as a product of augmented $A$5 blocks and recast as

$A$6

Here $A$7 distinguishes identity and creation-operator basis elements. The tensors $A$8 are explicitly specified in the paper, and for intermediate sites $A$9 is nilpotent of index $\alpha$0. This nilpotence enforces the rule that incompatible basis-operator configurations vanish.

The graphical language introduced in the paper makes the interaction law combinatorial. White squares represent identity components, green squares represent non-identity components, horizontal concatenation represents matrix multiplication, stacking represents tensor products, and merging corresponds to multiplication of the corresponding $\alpha$1 tensors. If two green squares overlap at the same position, the product vanishes. If they are compatible, their occupations add. The interferometer matrix therefore becomes the local coefficient system governing how operator-basis lines propagate and merge.

For a product of input operators $\alpha$2, the evolution is

$\alpha$3

A key algebraic statement is that nonzero entries of tensor products of local $\alpha$4-matrices are products of selected interferometer coefficients: $\alpha$5

For Boson Sampling-like input states, the photon-counting amplitude is

$\alpha$6

and in the one-photon-per-input case this contraction reproduces the permanent of the relevant submatrix of $\alpha$7. The total cost of the graphical contraction is

$\alpha$8

matching the best known classical complexity for the permanent via Ryser’s algorithm. The result is not a sub-Ryser algorithm; it is a tensor-network formulation whose complexity coincides with the best known classical benchmark.

The same basis-level formalism extends to partial distinguishability and photon loss. For distinguishability, each photon acquires an internal two-level label,

$\alpha$9

and the local operator basis is enlarged by direct sums without changing bond dimension. For loss, interferometer entries are modified as

$\beta$0

with an extra environmental output channel. The same operator-basis machinery then computes amplitudes conditioned on a fixed number of lost photons.

5. Broader operator-basis constructions related to interaction structure

Several adjacent literatures do not use the exact phrase Basis Interaction Operator, but they develop closely related operator-basis architectures in which interaction content is encoded by a distinguished basis.

In chiral perturbation theory, the amplitude/operator basis is built from on-shell soft blocks, defined as local contact amplitudes satisfying Adler’s zero,

$\beta$1

together with flavor tensors. The paper establishes a direct correspondence

$\beta$2

and constructs explicit purely mesonic even- and odd-parity bases at $\beta$3 and $\beta$4 for $\beta$5 and arbitrary $\beta$6 (Low et al., 2022). In that setting, the basis is interaction-resolving because each independent on-shell amplitude corresponds to one independent low-energy operator coefficient.

In heavy-particle EFT, an operator basis is the set of effective Lagrangian operators that remain after removing redundancies from integration by parts, heavy-fermion equations of motion, gauge-field equations of motion, and total derivatives. One line of work shows that the independent basis is spanned by highest-weight operators of the Schrödinger algebra that are neutral scalars after tensor products are combined, while another constructs the basis with Hilbert-series methods through mass dimension $\beta$7 for NRQED and NRQCD/HQET, and a further refinement shows how reparameterization invariance recovers Lorentz constraints through order $\beta$8 (Kobach et al., 2018, Kobach et al., 2017, Kobach et al., 2018). In these EFT applications, the “interaction operator” is a basis element of the effective Lagrangian rather than a coefficient-space map.

A more representation-theoretic example is the covariant basis induced by parity for the $\beta$9 Lorentz representation. There the basis always contains the identity, chirality, the Lorentz generators $\mathcal K$0, symmetric traceless tensors $\mathcal K$1 and $\mathcal K$2, and for $\mathcal K$3 higher-rank Weyl-like tensors. The resulting inventory of covariant operators shows that the unique antisymmetric rank-2 structure is $\mathcal K$4, implying a single Pauli term and a single gyromagnetic factor $\mathcal K$5 in the Poincaré projector formalism (Gómez-Ávila et al., 2013).

In mathematical physics, the Bender–Dunne basis operators furnish a different kind of basis-centric operator construction. For

$\mathcal K$6

the relevant question is not basis learning or sparsification but whether the formal Weyl-ordered expressions with inverse momentum powers define genuine Hilbert-space operators. The paper proves, by explicit construction of a dense domain, that for fixed positive integers $\mathcal K$7 and $\mathcal K$8, $\mathcal K$9 is densely defined on $(\alpha_n,\beta_n)$00 (Bunao et al., 2015).

This suggests a broader taxonomy: basis interaction operators are part of a family of basis-dependent operator constructions in which the chosen basis is used to expose locality, symmetry, analyticity, spectral structure, or computational sparsity.

6. Common themes, limitations, and recurring misconceptions

A first misconception is that Basis Interaction Operator denotes a universally standardized mathematical object. The surveyed literature does not support that interpretation. In B2B learning, the central object is the coefficient map $(\alpha_n,\beta_n)$01; in O-Sensing, it is a sparse conserved generator inside a parent operator subspace; and in optical tensor networks, the interaction is encoded by local operator-basis tensors and interferometer coefficients rather than by a separately named operator (Ingebrand et al., 2024, Ye-Ming et al., 4 Mar 2026, Cilluffo et al., 3 Feb 2025).

A second misconception is that a basis formulation removes ambiguity automatically. O-Sensing shows the opposite: the eigenstate constraints determine a highly degenerate parent operator subspace $(\alpha_n,\beta_n)$02, and any basis rotation within that subspace is admissible before sparsity is imposed. The sparse basis therefore functions as a kind of gauge fixing, and the confusion regime near the complement-graph crossover shows that parsimony itself can be ambiguous (Ye-Ming et al., 4 Mar 2026).

A third misconception is that basis interaction formulations imply unrestricted functional-analytic regularity. The Bender–Dunne analysis establishes only that $(\alpha_n,\beta_n)$03 is a bona fide densely defined linear operator on $(\alpha_n,\beta_n)$04. It is not shown there to be bounded, self-adjoint, or even closable (Bunao et al., 2015). More generally, changing to a basis-adapted representation does not by itself settle questions of domain, closure, or spectrum.

A fourth misconception is that basis interaction methods necessarily exceed existing computational complexity limits. The optical operator-basis MPS achieves a significant complexity bridge, but its central result is equality with the best known classical permanent algorithm: $(\alpha_n,\beta_n)$05 not an asymptotic improvement beyond that benchmark (Cilluffo et al., 3 Feb 2025).

Finally, interpretability and accuracy need not coincide. In B2B learning, the SVD and eigen-decomposition variants are more interpretable but generally do not match B2B’s accuracy, while DeepONet can be competitive on the parameterized heat equation, where end-to-end training may help when there is no clean input-function-space structure (Ingebrand et al., 2024). Basis interaction operators are therefore best viewed as structured representations whose value depends on how well the selected basis aligns with the operator’s actual geometry, symmetry, or spectrum.