Invariant Tensor Method

Updated 7 July 2026

Invariant Tensor Method is a framework that builds invariant scalars and equivariant maps through systematic tensor contractions using key elements like δ and ε.
It employs symmetric tensor networks and diagrammatic approaches to generate invariant polynomials in SO(3) and O(3) settings.
The method has practical applications in deep learning, gauge theories, and curvature canonicalization, offering finite generation and explicit generator theorems.

Searching arXiv for the cited work and closely related papers on invariant tensor methods. Invariant tensor method denotes a family of constructions in which invariant scalars, equivariant maps, or invariant operators are built by contracting tensorial inputs with group-invariant tensors, or by organizing those contractions as tensor networks. In the $SO(3)$ and $O(3)$ setting, a systematic formulation shows that the space of all real-valued invariants is generated by connected symmetric tensor-network contractions of the inputs together with $\delta$ and at most one $\epsilon$ , and that all equivariant functions arise by differentiating such invariants with respect to auxiliary outputs (Zhang et al., 18 Aug 2025). The same expression, or closely related constructions, appears in invariant feature coding for finite orthogonal groups, local-unitary invariant theory, tensor models, gauge theories, Weyl-invariant gravity, tensor data analysis, and computer algebra for Riemann invariants (Mukuta et al., 2019).

1. Definitions, representations, and invariant tensors

Let $G$ be a group, with the principal geometric setting taking $G=SO(3)$ or $O(3)$ . A real vector space $V$ carries a linear representation $\rho:G\to GL(V)$ , and for $\mathbf x\in V$ one writes $O(3)$ 0. A multi-input function

$O(3)$ 1

is invariant if

$O(3)$ 2

and equivariant if $O(3)$ 3 itself carries a representation $O(3)$ 4 of $O(3)$ 5 and

$O(3)$ 6

In the $O(3)$ 7 formulation, a Cartesian tensor of rank $O(3)$ 8 is an array $O(3)$ 9 whose indices run over the standard $\delta$ 0-dimensional vector representation, while a spherical or irreducible tensor of type $\delta$ 1 is a $\delta$ 2-component object $\delta$ 3 transforming in the spin- $\delta$ 4 representation (Zhang et al., 18 Aug 2025).

Two special invariant tensors play a central role. The Kronecker delta $\delta$ 5 is rank $\delta$ 6 and invariant under rotations, and the Levi-Civita tensor $\delta$ 7 is rank $\delta$ 8 and invariant for $\delta$ 9. These tensors supply the elementary contraction mechanisms from which invariant scalars and equivariant outputs are assembled. In related discrete-group formulations, the invariant subspace can also be characterized as the fixed-point range of the group-averaging projector

$\epsilon$ 0

or, equivalently, as the kernel of stacked linear constraints of the form $\epsilon$ 1 when one works with a generating set of the group (Mukuta et al., 2019).

2. Symmetric tensor networks as a constructive language

A central 2025 formulation replaces coordinate-heavy manipulations by a graphical language. A tensor $\epsilon$ 2 is drawn as a node with $\epsilon$ 3 legs labeled by indices, and connecting two legs means summing over the shared index. A network of nodes and legs therefore represents a multilinear contraction yielding a new tensor on the unattached legs. A tensor is $\epsilon$ 4-symmetric when the representation matrices can be attached to its legs and slid through the node without changing the tensor; a symmetric tensor network is then a contraction of symmetric tensors, and contraction preserves overall symmetry (Zhang et al., 18 Aug 2025).

This graphical viewpoint has close analogues in other invariant-theoretic settings. In local-unitary invariant theory, Penrose string notation represents density operators, identities, cups, and caps as wires and boxes, and degree- $\epsilon$ 5 polynomial invariants are written as

$\epsilon$ 6

with $\epsilon$ 7 drawn as a permutation network. By the Brauer–Procesi statement quoted there, the commutant of $\epsilon$ 8 is generated by permutation operators of the $\epsilon$ 9 factors, which yields a complete tensor-network recipe for homogeneous invariants and gives direct graphical expressions for quantities such as $G$ 0 and hence Rényi entropies (Biamonte et al., 2012).

For tensor data with mode-wise orthogonal symmetry, the analogous combinatorial object is the colored Brauer diagram. Here $G$ 1 acts on $G$ 2, and a colored Brauer diagram overlays three perfect matchings, one for each mode. The corresponding multilinear map $G$ 3 and homogeneous polynomial $G$ 4 generate the entire ring of $G$ 5-invariant polynomials (Tokcan et al., 2020). This suggests that tensor-network, string-diagram, and Brauer-diagram formalisms are different realizations of the same structural idea: encode invariance by admissible contractions.

3. Generator theorems for invariants and the passage to equivariants

The modern $G$ 6 construction is organized around finite generation and explicit generators. Hilbert’s finiteness theorem is invoked in the form: for a compact group $G$ 7 acting linearly on a finite-dimensional real space $G$ 8, the algebra of invariant real polynomials $G$ 9 is finitely generated. For vector inputs $G=SO(3)$ 0, Weyl’s classical result is stated as Lemma 3.1: the ring of $G=SO(3)$ 1-invariant polynomials is generated by

$G=SO(3)$ 2

For Cartesian-tensor inputs $G=SO(3)$ 3, Theorem 3.2 states that every invariant polynomial can be written as a linear combination of contractions of a connected network built from the input tensors, copies of $G=SO(3)$ 4, and at most one $G=SO(3)$ 5. For spherical-tensor inputs, Theorem 3.3 inserts the standard projector $G=SO(3)$ 6 from Cartesian tensor powers onto the spin- $G=SO(3)$ 7 subspace and obtains the same statement after replacing the inputs by the $G=SO(3)$ 8 blocks (Zhang et al., 18 Aug 2025).

The transition from invariants to equivariants is given by the Up-derivative trick. If

$G=SO(3)$ 9

is invariant, then

$O(3)$ 0

defines an equivariant map from $O(3)$ 1 to $O(3)$ 2. Conversely, every equivariant map arises this way from some invariant with extra dummy outputs. Diagrammatically, one starts from a generator network for $O(3)$ 3 and removes the single $O(3)$ 4 node, leaving an open leg whose transformation law matches the target space. The stated consequence is strong: all equivariant functions arise by differentiating such invariants with respect to auxiliary outputs, yielding a complete and systematic method for constructing all polynomial, and by approximation smooth, invariants and equivariants under $O(3)$ 5 (Zhang et al., 18 Aug 2025).

In practical parameterizations, one fixes a finite set $O(3)$ 6 of invariant generators and $O(3)$ 7 of equivariant bases, truncates to a maximum polynomial degree $O(3)$ 8, and writes

$O(3)$ 9

where $V$ 0 and the $V$ 1 are ordinary neural networks such as MLPs (Zhang et al., 18 Aug 2025).

4. Representative constructions in geometric deep learning and invariant feature design

Elementary examples already display the logic of the method. For two vectors $V$ 2, the connected invariant network is the dot product $V$ 3; with a third vector $V$ 4, the triple product $V$ 5 appears. Starting from the invariant $V$ 6 and differentiating with respect to the dummy output $V$ 7 yields the equivariant vector

$V$ 8

For matrix-valued node features $V$ 9 and message $\rho:G\to GL(V)$ 0, invariant quantities include traces such as $\rho:G\to GL(V)$ 1 and $\rho:G\to GL(V)$ 2, while equivariant bases include $\rho:G\to GL(V)$ 3, $\rho:G\to GL(V)$ 4, and the antisymmetric part $\rho:G\to GL(V)$ 5. In spherical $\rho:G\to GL(V)$ 6 graph neural networks, the tensor-product layer with inputs of spins $\rho:G\to GL(V)$ 7 and output spin $\rho:G\to GL(V)$ 8 is written with Clebsch–Gordan coefficients as

$\rho:G\to GL(V)$ 9

and is represented by a triangle of projectors $\mathbf x\in V$ 0 joined by Cartesian legs of appropriate multiplicities (Zhang et al., 18 Aug 2025).

These constructions are inserted directly into message passing. In a geometry GNN layer

$\mathbf x\in V$ 1

each of $\mathbf x\in V$ 2 and $\mathbf x\in V$ 3 is assembled from equivariant building blocks, which ensures overall $\mathbf x\in V$ 4-equivariance. The same architecture is used for material constitutive modeling with deformation gradient $\mathbf x\in V$ 5, where invariant inputs include $\mathbf x\in V$ 6, $\mathbf x\in V$ 7, and $\mathbf x\in V$ 8, equivariant bases include $\mathbf x\in V$ 9, and stress is modeled as

$O(3)$ 00

The reported numerical experiments on Neo-Hookean synthetic data show dramatic sample-efficiency gains over unconstrained MLPs (Zhang et al., 18 Aug 2025).

A related machine-learning line treats invariance through explicit averaging or invariant subspace construction. For a finite group of orthogonal matrices, a canonical invariant feature map is

$O(3)$ 01

and the associated discriminative result states that when the data law is $O(3)$ 02-invariant and one minimizes an $O(3)$ 03-regularized convex risk, any minimizer lies in the trivial subspace, so restricting to $O(3)$ 04 costs no extra loss but lowers model complexity (Mukuta et al., 2019). In tensor-train networks, the invariant subspace can be computed by a dedicated basis-construction algorithm for normal matrix representations of arbitrary discrete groups; the resulting group-invariant tensor train network was applied to transcription-factor binding with reverse-complement symmetry and obtained prediction accuracy in line with state-of-the-art deep learning approaches, while the basis-construction step was reported to run in seconds and to be up to several orders of magnitude faster than previous approaches (Sprangers et al., 2022).

The phrase has specialized meanings in several neighboring literatures. This suggests that “Invariant Tensor Method” is best treated as a family of constructive techniques rather than a single universally standardized algorithm.

In general tensor models with symmetry group

$O(3)$ 05

an invariant operator is a polynomial $O(3)$ 06 built from $O(3)$ 07 copies of $O(3)$ 08 and $O(3)$ 09 copies of $O(3)$ 10 with all indices contracted to form a $O(3)$ 11-singlet. Representation theory of $O(3)$ 12, Schur–Weyl duality, and Kronecker coefficients are used to count invariants, construct an orthogonal basis

$O(3)$ 13

and diagonalize the Gaussian two-point function. The construction is presented as the tensor-model analogue of the restricted Schur basis, with Kronecker coefficients playing the role occupied by Littlewood–Richardson numbers in multi-matrix models (Diaz et al., 2018). A related $O(3)$ 14-invariant Gaussian theory for real three-index tensors uses Young-diagram and partition-algebra techniques to diagonalize the two-point function in a representation basis and then reconstruct it in the original tensor basis (Barnes et al., 2023).

In gauge theories, invariant tensors generate the ring of gauge-invariant polynomials, sometimes called the chiral ring. A symmetry-breaking plus lifting procedure is given for constructing a Hilbert basis: choose a generic vev $O(3)$ 15, decompose the representation under the unbroken subgroup $O(3)$ 16, list the $O(3)$ 17-invariant monomials, and lift each one to a $O(3)$ 18-invariant polynomial of the same degree. For $O(3)$ 19 and $O(3)$ 20, the reported complete minimal generating set consists of five basic invariants $O(3)$ 21 of degrees $O(3)$ 22, with no invariant of degree $O(3)$ 23 needed and no relations among $O(3)$ 24 before degree $O(3)$ 25 (Berger et al., 2018).

In Weyl-invariant scalar-tensor gravity, the invariant-tensor idea takes the form of an invariant metric

$O(3)$ 26

with $O(3)$ 27 a compensator of weight $O(3)$ 28. Since $O(3)$ 29 has weight zero, any purely metric diffeomorphism-invariant action can be “Weyl uplifted” by the literal substitution $O(3)$ 30, $O(3)$ 31, $O(3)$ 32. This prescription is applied to higher-curvature theories, Lovelock terms, Einsteinian cubic gravity, and conformal renormalization in Einstein–AdS (Anastasiou et al., 2023).

6. Canonicalization, syzygies, and computational discovery

A major computational branch of invariant-tensor methodology is symbolic reduction to a minimal basis. The Invar package addresses scalar polynomial expressions formed from the Riemann tensor of a four-dimensional metric-compatible connection. Its reduction pipeline combines the algebraic symmetries of the Riemann tensor, the first Bianchi identity, dimension- and signature-dependent identities, and permutation-group algorithms. The four-step canonicalization algorithm is: Step A, permutation symmetries and double-coset canonicalization; Step B, the cyclic identity $O(3)$ 33; Step C, four-dimensional Lovelock identities from antisymmetrization over five indices; and Step D, signature-dependent $O(3)$ 34- $O(3)$ 35 identities. The package stores precomputed syzygies up to degree $O(3)$ 36 for non-dual invariants and up to degree $O(3)$ 37 for duals, with the stated benchmark that degree- $O(3)$ 38 monomials simplify in $O(3)$ 39 ms on a $O(3)$ 40 GHz PC with $O(3)$ 41 GB RAM (0704.1756).

A contrasting line is explicitly data-driven. One recent algorithm enumerates inequivalent contraction graphs, computes the associated scalar contractions on randomly generated tensors, forms a data matrix

$O(3)$ 42

and detects syzygies through singular-value decomposition or rank-revealing QR. For an antisymmetric $O(3)$ 43-form $O(3)$ 44 in six dimensions, this procedure finds exactly five independent scalar invariants, and further computations at higher orders report no new independent invariants beyond those five (Elamaran et al., 26 Dec 2025). The same paper presents explicit formulas expressing other admissible order- $O(3)$ 45 and order- $O(3)$ 46 contractions as polynomials in the generators.

Colored Brauer-diagram methods occupy an intermediate position between symbolic and numerical approaches. They provide explicit invariant polynomials $O(3)$ 47, complexity estimates for their computation, degree- $O(3)$ 48 formulas for spectral-norm approximation, and polynomial amplification maps $O(3)$ 49 and $O(3)$ 50 obtained as gradients of invariant norms. In the reported numerical tests, these amplification maps improve the performance of ALS for low-rank tensor approximation and yield consistently higher fit than repeated random ALS restarts on noisy tensors up to size $O(3)$ 51 (Tokcan et al., 2020).

Taken together, these developments define invariant tensor method as a constructive program rather than a single formalism. Its recurring elements are invariant building blocks such as $O(3)$ 52, $O(3)$ 53, metrics, cups, caps, traces, and projectors; admissible contraction rules encoded by tensor networks, permutation diagrams, or Brauer diagrams; finite generation or basis reduction; and, where required, explicit recovery of equivariants from invariant scalars. In the $O(3)$ 54 setting this program is complete in the stated polynomial sense (Zhang et al., 18 Aug 2025), while in other domains it appears as projector-based feature design, orthogonal-basis construction, Hilbert-basis generation, Weyl uplift, canonicalization of curvature invariants, and numerical discovery of syzygies.