Tensor Disentangling: Methods & Applications

• Tensor disentangling is a family of operations that separates interacting degrees of freedom in tensorial representations, making high-dimensional data more factorized and interpretable.
• It employs techniques from multilinear algebra, quantum theory, and neural representation learning to achieve low-rank approximations and disentangled latent factors.
• These methods enable efficient tensor compression, faster computations, and improved insights in applications ranging from chemometrics and quantum many-body physics to multimodal learning.

Searching arXiv for papers on tensor disentangling and closely related formulations. Tensor disentangling is a family of operations that separates interacting degrees of freedom in tensorial representations. In classical tensor analysis, it denotes the approximation of a tensor by rank-$1$ or low-rank structure, so that variables become separated or only weakly coupled. In neural representation learning, it refers to isolating latent factors or disentangling higher-order interactions from individual effects. In tensor-network and quantum settings, it denotes transformations—typically unitary or orthogonal—that reduce entanglement, lower bond dimensions, or expose a tensor product structure in which a state becomes separable. Taken together, these usages indicate a common objective: replacing an entangled, highly coupled, or opaque description by one that is more factorized, compressible, or interpretable (Franc, 2022, Slagle, 2021, Soulas, 26 Jun 2025).

1. Terminology and conceptual scope

The term has distinct technical meanings across subfields. In multilinear algebra, a tensor of rank $1$ is an elementary tensor $x_1 \otimes x_2 \otimes \ldots \otimes x_d$, and this is the case in which variables are separated; low rank means the variables are weakly coupled (Franc, 2022). In tensor-network algorithms, disentangling usually means applying unitary operators to parts of a tensor network in order to reduce entanglement, often as a preprocessing step for truncation or compression (Slagle, 2021). In quantum information, the notion can shift from changing the state to changing the subsystem decomposition itself: a bipartite tensor product structure (TPS) of $\mathcal{H}$ is an equivalence class of isomorphisms $$\Phi : \mathcal{H} \to \mathcal{H}_1 \otimes \mathcal{H}_2$$, where equivalence is up to local unitaries, and a TPS is said to disentangle a trajectory $\ket{\Psi(t)}$ if $$\ket{\Psi(t)} = \ket{\Psi_1(t)} \otimes \ket{\Psi_2(t)}$$ for all $t$ (Soulas, 26 Jun 2025).

A related but narrower usage appears in topological many-body physics. There, tensor disentangling is the possibility to factorize a topological ground state wavefunction into a tensor product of wavefunctions on spatially separated subregions, with entanglement negativity serving as the key diagnostic (Lim et al., 2021). In multimodal learning, the same phrase can denote disentangling higher-order multiplicative interactions from unimodal effects, as in interpretable tensor fusion models that separately encode additive and tensor-product terms (Varshneya et al., 2024).

This multiplicity of meanings is not a contradiction. It reflects a shared technical motif: identifying a transformation, factorization, or coordinate system in which couplings that were previously represented jointly can be reassigned to smaller components, lower-rank factors, or different subsystems.

2. Low-rank separation and multilinear structure

Classical tensor disentangling begins with rank. The CP decomposition writes a $D$-way tensor element as

$$\mathcal{X}(i_1, \ldots, i_D) = \sum_{r=1}^R U^1_{i_1, r} U^2_{i_2, r} \cdots U^D_{i_D, r},$$

while Tucker decomposition represents a tensor by a core and mode-specific factor matrices (Liu et al., 2016). In the rank-$1$0 case, the variables are completely separated; for small rank, the variables are only weakly coupled (Franc, 2022).

Several rank notions are used for disentangling interactions. CP rank is the minimal number of rank-$1$1 terms, Tucker rank is the tuple of minimal mode-wise subspace dimensions, border rank captures limits of low-CP-rank tensors, typical rank describes ranks that occur on a set of nonzero measure, and TT rank encodes sequential low-rank structure in a chain decomposition (Franc, 2022). Algorithmically, this leads to alternating least squares for best rank-$1$2 and rank-$1$3 CP approximation, HOSVD and HOOI for Tucker approximation, and TT-SVD for tensor trains (Franc, 2022).

In tensor-network language, these decompositions become graphical statements about cuts. CP, Tucker, and TT decompositions can be represented as tensor network diagrams, and the rank of any matricization is at most the product of the dimensions of the edges in any cut that separates rows from columns in the diagram (Rakhshan et al., 15 May 2026). This makes mode separation visually explicit: cutting between a Tucker core and a factor isolates one mode, while cutting a TT bond exposes the rank of the corresponding flattening (Rakhshan et al., 15 May 2026).

A persistent limitation of classical CP and Tucker models is the multi-linearity assumption. In applications with nonlinear high-order interactions, that assumption can make classical decompositions insufficiently expressive, even when the target object is still meaningfully “disentangled” in a broader latent-variable sense (Liu et al., 2016).

3. Probabilistic and neural tensor disentangling

A direct response to the limits of multilinearity is to replace linear interaction models by nonlinear latent-variable generators. The VAECP model, “Variational Auto-Encoder CP,” is a Bayesian generative model for tensor decomposition in which tensor entries are generated by a nonlinear process dependent on latent factors for each tensor mode (Liu et al., 2016). Instead of the CP inner product, it models

$1$4

with

$1$5

The model places Gaussian priors on latent factors, uses variational inference and the reparameterization trick, and is trained end-to-end with stochastic gradient descent such as Adam (Liu et al., 2016). On synthetic $1$6 tensors and on the Amino Acid, Flow Injection Analysis, and Sugar Process chemometrics tensors, it achieved the lowest or among the lowest RMSE for missing-data prediction, and performance did not deteriorate sharply when the latent dimension exceeded typical values (Liu et al., 2016).

A more explicitly multilinear neural approach appears in the adversarial neuro-tensorial model for face images. There, facial texture is modeled as

$1$7

with separate latents for illumination, expression, and identity, plus a pose latent applied as a $1$8 rotation matrix (Wang et al., 2017). The method combines reconstruction, adversarial, verification, pseudo-supervision, and multilinear tensor losses, using Khatri-Rao products in batch form to enforce a latent factorization that supports expression transfer, pose editing, illumination editing, 3D reconstruction, and classification (Wang et al., 2017).

Another line replaces Euclidean latent vectors by tensor-product latent spaces. In “Unsupervised Disentanglement with Tensor Product Representations on the Torus,” each latent factor is represented by a unit vector on a circle,

$1$9

and the full latent code is the vectorized tensor product

$x_1 \otimes x_2 \otimes \ldots \otimes x_d$0

This gives a latent space distributed uniformly over a set of unit circles, i.e. a torus $x_1 \otimes x_2 \otimes \ldots \otimes x_d$1, and experiments reported higher disentanglement, completeness, and informativeness, summarized by the DC-score, than several vector-latent baselines on Teapots, 2dshapes, 3dshapes, Cars3D, and dSprites (Rotman et al., 2022).

In multimodal learning, Interpretable Tensor Fusion (InTense) separates linear fusion from multiplicative fusion over modality subsets. Its prediction has the form

$x_1 \otimes x_2 \otimes \ldots \otimes x_d$2

where $x_1 \otimes x_2 \otimes \ldots \otimes x_d$3 is a tensor product of modality-specific representations. The method uses generalized centering and normalization to remove higher-order interaction bias, so that learned relevance scores reflect unique unimodal and multimodal effects rather than spurious absorption of lower-order terms into higher-order components (Varshneya et al., 2024).

4. Tensor-network algorithms and disentanglers

In tensor-network computation, disentangling is often a local optimization problem. The “Fast Tensor Disentangling Algorithm” takes a tensor $x_1 \otimes x_2 \otimes \ldots \otimes x_d$4 and constructs a unitary $x_1 \otimes x_2 \otimes \ldots \otimes x_d$5 non-iteratively: it uses a random vector to break degeneracies, extracts dominant singular vectors, performs truncated SVDs to obtain mode-specific features, forms a reduced tensor $x_1 \otimes x_2 \otimes \ldots \otimes x_d$6, and then uses Gram–Schmidt orthogonalization on $x_1 \otimes x_2 \otimes \ldots \otimes x_d$7 to obtain the unitary (Slagle, 2021). For $x_1 \otimes x_2 \otimes \ldots \otimes x_d$8 and $x_1 \otimes x_2 \otimes \ldots \otimes x_d$9, its complexity is

$\mathcal{H}$0

and on random and structured order-$\mathcal{H}$1 tensors it produced residual entanglement entropy typically within $\mathcal{H}$2 of the minimum while being $\mathcal{H}$3–$\mathcal{H}$4 times faster than iterative methods (Slagle, 2021). In the symmetric order-$\mathcal{H}$5 case, it guarantees nearly half the singular values across the cut are zero: $\mathcal{H}$6

A more general formulation treats tensor disentangling as optimization over orthogonal transformations. For a tensor $\mathcal{H}$7, with unfolding $\mathcal{H}$8, the aim is to find $\mathcal{H}$9 minimizing the tail sum of squared singular values

$$\Phi : \mathcal{H} \to \mathcal{H}_1 \otimes \mathcal{H}_2$$0

or more generally

$$\Phi : \mathcal{H} \to \mathcal{H}_1 \otimes \mathcal{H}_2$$1

with choices of $$\Phi : \mathcal{H} \to \mathcal{H}_1 \otimes \mathcal{H}_2$$2 corresponding to fixed-rank truncation, von Neumann entropy, or Rényi entropy (Wei et al., 26 Aug 2025). The paper develops Riemannian conjugate gradient, Riemannian trust-region Newton, alternating minimization over $$\Phi : \mathcal{H} \to \mathcal{H}_1 \otimes \mathcal{H}_2$$3, and a binary search procedure for the often unknown optimal rank (Wei et al., 26 Aug 2025).

Neural tensor-network hybrids use disentangling in a more global sense. In the $$\Phi : \mathcal{H} \to \mathcal{H}_1 \otimes \mathcal{H}_2$$4TNS framework, a neural network serves as a disentangler of the wave-function, transforming the physical degrees of freedom into renormalized variables with much less entanglement, after which a back-flow tensor network such as an MPS compresses the remaining correlations (Fan et al., 15 Mar 2026). The ansatz is written schematically as

$$\Phi : \mathcal{H} \to \mathcal{H}_1 \otimes \mathcal{H}_2$$5

and in a CNN-MPS implementation it achieved state-of-the-art variational energies for the spin-$$\Phi : \mathcal{H} \to \mathcal{H}_1 \otimes \mathcal{H}_2$$6 $$\Phi : \mathcal{H} \to \mathcal{H}_1 \otimes \mathcal{H}_2$$7-$$\Phi : \mathcal{H} \to \mathcal{H}_1 \otimes \mathcal{H}_2$$8 Heisenberg model at $$\Phi : \mathcal{H} \to \mathcal{H}_1 \otimes \mathcal{H}_2$$9, with MPS bond dimensions as small as $\ket{\Psi(t)}$0, where a pure MPS would need $\ket{\Psi(t)}$1 for comparable accuracy (Fan et al., 15 Mar 2026).

A related compression problem appears in hybrid quantum-classical machine learning. There, a pre-trained dense linear layer is first represented as an effective MPO, then disentangled into a more compact form,

$\ket{\Psi(t)}$2

using either an explicitly disentangling variational method or an implicitly disentangling gradient-descent method (Aizpurua et al., 8 Sep 2025). The resulting split assigns the disentangling circuits to quantum hardware and the disentangled MPO to classical hardware.

5. Tensor product structures, topology, and limits

In finite-dimensional quantum systems, entanglement is TPS-dependent. The paper “Disentangling tensor product structures” gives a constructive example in which the C-NOT evolution of two qubits appears entangled in the standard TPS but is product for all $\ket{\Psi(t)}$3 in a different, explicitly constructed TPS (Soulas, 26 Jun 2025). At the same time, Proposition 2 establishes a generic non-existence result: if the products $\ket{\Psi(t)}$4 are linearly independent as functions of $\ket{\Psi(t)}$5 in $\ket{\Psi(t)}$6, then no fixed TPS can make $\ket{\Psi(t)}$7 product for all $\ket{\Psi(t)}$8 (Soulas, 26 Jun 2025). This rules out any universal expectation that entanglement can always be removed by redefining subsystems.

Topological phases introduce additional obstructions. For $\ket{\Psi(t)}$9-dimensional Laughlin and Moore-Read states, the disentangling condition

$$\ket{\Psi(t)} = \ket{\Psi_1(t)} \otimes \ket{\Psi_2(t)}$$0

is necessary for factorization into tensor products on cylinder subregions (Lim et al., 2021). For the Laughlin state, being in a single definite topological sector makes this condition sufficient; for twisted Moore-Read sectors, it does not, because the zero-mode sector remains intrinsically entangled through Majorana constraints (Lim et al., 2021).

Clifford-based disentangling provides another example of both power and limitation. States from deep random Clifford circuits doped with non-Clifford phase gates can be completely disentangled, provided the number of non-Clifford gates is smaller or approximately equal to the number of qubits, and such states can typically be written as

$$\ket{\Psi(t)} = \ket{\Psi_1(t)} \otimes \ket{\Psi_2(t)}$$1

up to the threshold characterized by $$\ket{\Psi(t)} = \ket{\Psi_1(t)} \otimes \ket{\Psi_2(t)}$$2 (Fux et al., 2024). However, beyond stabilizer settings, no Clifford operation can universally disentangle even a single qubit from an arbitrary non-Clifford rotation, unless that qubit is initially in a stabilizer state (Masot-Llima et al., 17 Feb 2026). Heuristic entanglement cooling with local Clifford sweeps can help in sparse-magic regimes, but increasing locality or sweep depth does not remove the breakdown once non-Clifford resources accumulate (Masot-Llima et al., 17 Feb 2026).

6. Diagnostics, applications, and recurrent issues

Because tensor disentangling spans several problem classes, its diagnostics are correspondingly heterogeneous. In tensor-network optimization, the standard measures are the entanglement spectrum, the von Neumann entanglement entropy

$$\ket{\Psi(t)} = \ket{\Psi_1(t)} \otimes \ket{\Psi_2(t)}$$3

and the truncation error

$$\ket{\Psi(t)} = \ket{\Psi_1(t)} \otimes \ket{\Psi_2(t)}$$4

(Slagle, 2021). In topological settings, entanglement negativity is the relevant mixed-state quantity (Lim et al., 2021). In tensor completion and nonlinear decomposition, prediction RMSE on held-out entries is the reported criterion, as in VAECP’s synthetic and chemometrics experiments (Liu et al., 2016). In unsupervised representation learning on the torus, evaluation is based on disentanglement, completeness, informativeness, and the DC-score (Rotman et al., 2022). In interpretable multimodal tensor fusion, relevance scores $$\ket{\Psi(t)} = \ket{\Psi_1(t)} \otimes \ket{\Psi_2(t)}$$5 quantify unimodal and interaction importance directly from learned weights (Varshneya et al., 2024). In diffusion tensor cardiovascular magnetic resonance, disentangling diffusion contrast, respiratory motion, and cardiac motion is evaluated with the Negative Eigenvalue Percentage (NE%) and the Helix Angle Gradient (HAG) linear profile, including $$\ket{\Psi(t)} = \ket{\Psi_1(t)} \otimes \ket{\Psi_2(t)}$$6 and RMSE (Wang et al., 2024).

The applications are equally broad. They include missing-data prediction in chemometrics tensors (Liu et al., 2016), disentangled face editing and 3D reconstruction (Wang et al., 2017), multimodal interpretability (Varshneya et al., 2024), quantum many-body variational ansätze (Fan et al., 15 Mar 2026), hybrid classical-quantum implementations of neural bottlenecks (Aizpurua et al., 8 Sep 2025), and physics-informed medical image registration (Wang et al., 2024).

A recurrent misconception is that “tensor disentangling” names a single method. The literature instead supports three distinct claims. First, exact disentangling is often available only in special algebraic, low-rank, or stabilizer-compatible cases (Franc, 2022, Fux et al., 2024). Second, generic dynamics or topology can obstruct any fixed disentangling transformation, whether by linear-independence criteria for TPS changes, by non-Abelian zero-mode structure, or by the nonuniversality of Clifford cooling (Soulas, 26 Jun 2025, Lim et al., 2021, Masot-Llima et al., 17 Feb 2026). Third, nonlinear or neural models do not simply supersede multilinear ones; they change the notion of disentangling from exact variable separation to learned factorization of complex interactions, often trading strict identifiability for expressivity and empirical robustness (Liu et al., 2016, Varshneya et al., 2024).

In that sense, tensor disentangling is best understood not as a single algorithmic recipe but as a technical program: to identify tensorial representations, transformations, or subsystem decompositions in which interaction structure becomes simpler than in the original coordinates.