---
title: Tensor Network Variational Diagonalization (TNVD)
url: https://www.emergentmind.com/topics/tensor-network-variational-diagonalization-tnvd
type: topic
---

# Tensor Network Variational Diagonalization (TNVD)

Searching arXiv for the primary TNVD paper and any directly related references mentioned in the provided data.
Tensor Network Variational Diagonalization (TNVD) is a variational framework for diagonalizing quantum many-body Hamiltonians by combining tensor-network compression of the full eigenvalue spectrum with a variational quantum circuit (VQC) representation of eigenstates. In the formulation introduced in "Diagonalizing large-scale quantum many-body Hamiltonians using variational quantum circuit and tensor network" [2508.06159], the complete set of eigenenergies is encoded as a matrix product state (MPS), while the corresponding eigenstates are represented as the images of computational-basis product states under a shared circuit $\hat U(\theta_U)$. The method is designed to replace the exponentially scaled computational complexity of exact diagonalization (ED) by a polynomial-in-$N$ variational procedure, and was benchmarked for spin chains up to $N=100$ [2508.06159].

## 1. Definition and representational structure

TNVD starts from the observation that a spin-$\tfrac12$ chain of length $N$ has a $2^N$-dimensional spectrum $\{E_\alpha\}_{\alpha=0}^{2^N-1}$, which can be reindexed as the entries of an $N$th-order tensor $E[r_1\cdots r_N]$ with $r_i\in\{0,1\}$. In TNVD, this tensor is approximated by a right-canonical MPS of bond dimension $\chi_a$,
$$
E_{r_1 r_2\cdots r_N}
\;=\;\sum_{a_1,\dots,a_{N-1}=1}^{\chi_a}
A^{[1]}_{r_1,a_1}\;
A^{[2]}_{r_2,a_1,a_2}\;
\cdots\;
A^{[N]}_{r_N,a_{N-1}}.
$$
The storage cost of this MPS scales as $O(2\,N\,\chi_a^2)$, which the paper identifies as polynomial in system size [2508.06159].

The eigenstates are represented differently. Rather than storing each $|\alpha\rangle$ independently, TNVD assumes that all eigenstates are generated by applying the same VQC to different product states:
$$
|\alpha\rangle
\;=\;\hat U(\theta_U)\;|r_\alpha\rangle,
$$
where $|r_\alpha\rangle\equiv |r_1r_2\cdots r_N\rangle$ is the binary encoding of $\alpha$. The circuit uses a brick-wall layout of nearest-neighbor two-qubit gates with depth $N_L$, and its total number of real parameters scales as $O(N\,N_L)$ [2508.06159].

This factorization separates spectral information from basis transformation. The MPS carries the full set of eigenvalues, while the VQC defines a global change of basis from product states to approximate eigenstates. A plausible implication is that TNVD may be viewed as a compressed eigendecomposition ansatz tailored to one-dimensional many-body systems, with efficiency determined jointly by the compressibility of the spectrum tensor and the circuit complexity of the eigenbasis.

## 2. Variational objective and optimization problem

The general tensor-network state ansatz in the paper is written as
$$
|\Psi(\boldsymbol\theta)\rangle
\;=\;\sum_{i_1,\dots,i_N}
A^{i_1}(\theta_1)\,
A^{i_2}(\theta_2)\,\cdots\,
A^{i_N}(\theta_N)\;
|i_1\,i_2\cdots i_N\rangle.
$$
For TNVD itself, the variational parameters are collected into $\theta$, comprising both the MPS tensors for the spectrum and the VQC parameters for the eigenbasis [2508.06159].

The target of optimization is an ansatz operator $\tilde H(\boldsymbol\theta)$ intended to approximate the Hamiltonian $H$. The paper measures the discrepancy through the logarithmic Schmidt distance
$$
C(\boldsymbol\theta)
\;=\;
F
\;=\;
\log_2\bigl\|H - \tilde H(\boldsymbol\theta)\bigr\|^2
\;-\;N,
$$
where $\|\cdot\|$ is the Frobenius norm. The subtraction of $N$ is chosen so that a perfect decomposition gives $F\to -\infty$ [2508.06159]. The norm is expanded as
$$
\|H-\tilde H\|^2
=\mathrm{Tr}[H^2]+\mathrm{Tr}[\tilde H^2]
-2\,\mathrm{Tr}[H\,\tilde H].
$$

Each contribution is reported to be computable in either $O(N)$ or $O(N\,N_L\,\chi_t^3)$ time, where $\chi_t$ is the TEBD bond cutoff used while evolving $H$ through the circuit [2508.06159]. This places the computational bottleneck in tensor contractions and TEBD-based propagation through $\hat U$ and $\hat U^\dagger$, rather than in explicit matrix diagonalization.

Within this formulation, the optimization problem is not cast as minimizing individual eigenstate residuals or a Rayleigh quotient for selected states. Instead, it targets an operator-level approximation to the full Hamiltonian decomposition. This suggests that TNVD is intrinsically a full-spectrum variational method rather than a low-energy variational solver.

## 3. Algorithmic workflow

TNVD takes as inputs an MPO representation of the Hamiltonian $H$, a maximum MPS bond dimension $\chi_a$, circuit depth $N_L$, TEBD cutoff $\chi_t$, learning rate $\eta$, and convergence tolerance $\epsilon$ [2508.06159]. The paper specifies the following workflow.

First, one initializes random MPS tensors $\{A[n]\}$ with bond dimension at most $\chi_a$, together with random VQC parameters $\theta_U$. The MPO of $H$ is then constructed via automata. The optimization loop proceeds until $|\Delta C|<\epsilon$.

At each iteration, the ansatz operator $\tilde H$ is reconstructed by contracting the spectrum MPS, $\delta$-tensors, the circuit $U(\theta_U)$, additional $\delta$-tensors, and the Hermitian conjugate MPS. After computing the cost $C=F$, gradients $\partial F/\partial A[n]$ and $\partial F/\partial \theta_U$ are obtained either by automatic differentiation or by analytic tensor-network contractions combined with TEBD. Parameters are updated according to gradient descent,
$$
A[n]\leftarrow A[n] - \eta\,\partial F/\partial A[n],\qquad
\theta_U\leftarrow \theta_U - \eta\,\partial F/\partial \theta_U.
$$
An optional re-orthonormalization step may be applied to maintain canonical form of the MPS [2508.06159].

After convergence, the MPS tensors are used to recover the full eigenvalue set $E_\alpha$, while the optimized circuit generates eigenstates by acting on all product states. The explicit extraction rule is to contract the MPS at physical indices $r_1\ldots r_N$ to obtain a scalar $E_r$, and to build each eigenstate as $|\alpha\rangle=U(\theta_U)|r\rangle$ [2508.06159].

The paper also notes that, in practice, each two-qubit gate is parameterized via a small latent tensor network, citing a prior reference identified there as Ref. [ZHR21ADQC], so that gradients can be computed uniformly. Since no arXiv identifier is supplied in the provided material, the role of that reference can only be stated at this level of specificity.

## 4. Computational scaling and comparison with exact diagonalization

The computational complexity reported for TNVD is organized by subroutine. MPO-to-MPO contractions, including terms such as $\mathrm{Tr}[H^2]$, scale as $O(N)$. MPS inner products scale as $O(N\,\chi_a^3)$. TEBD evolution of the MPO through the circuit and its conjugate scales as $O(N\,N_L\,\chi_t^3)$. The total per-iteration cost is therefore approximately
$$
O(N\,\chi_a^3 + N\,N_L\,\chi_t^3).
$$
By contrast, ED is reported to cost $O(2^{3N})$ to store and $O(2^{3N})$ or worse to diagonalize [2508.06159].

The central claim of TNVD is therefore a change in asymptotic scaling from exponential to polynomial in $N$, contingent on fixed or controlled growth of the variational bond and truncation parameters. In the presentation of the method, this polynomial scaling is not merely a heuristic compression statement but an explicit accounting of the dominant contraction costs.

The significance of this contrast is practical as well as formal. Because TNVD stores the spectrum as an MPS rather than as an explicit vector of $2^N$ energies, downstream operations such as sampling from the density of states become tractable at sizes beyond the ED limit. This suggests a different computational regime from conventional diagonalization: one in which global spectral access is retained, but only in compressed variational form.

## 5. Benchmarks on the transverse-field Ising chain

The numerical benchmarks reported in the paper use the transverse-field Ising chain
$$
H=-\sum_{n=1}^{N-1} S^z_n S^z_{n+1} - h_x\sum_n S^x_n,
$$
with $h_x\in\{0.2,0.5,0.8\}$ [2508.06159].

For small systems with $N\le 16$, comparison to ED is given through the mean absolute eigenenergy error
$$
\epsilon
= \frac1{2^N}\sum_{\alpha=0}^{2^N-1}\bigl|E_\alpha^{\rm ED}-E_\alpha^{\rm TNVD}\bigr|.
$$
For $\chi_a=8$, $N_L=10$, and $\chi_t=16$, this error stayed below $10^{-2}$, and the logarithmic Schmidt distance $F$ tracked $\epsilon$ closely, with sublinear growth versus $N$ [2508.06159].

For larger systems up to $N=100$, the paper reports that $F$ grew only slowly, reaching approximately $0.1$ to $0.2$, which is interpreted there as indicating controlled global error. The authors further sampled $10^6$ eigenenergies from the MPS spectrum and constructed the density-of-states histogram $N_s(E)$. For both $N=16$ and $N=100$, the histograms were fit by a Gaussian,
$$
N_s(E)\propto \exp\!\Bigl[-\tfrac{(E-\mu)^2}{2\sigma^2}\Bigr],
$$
with identical $\sigma$ to within $10^{-3}$ [2508.06159].

These benchmarks establish the main empirical profile of TNVD given in the paper: accurate small-system agreement with ED, followed by extension to system sizes inaccessible to ED while retaining full-spectrum observables in compressed form. The data do not claim exact recovery at $N=100$; rather, they document controlled variational approximation measured by $F$ and consistency of sampled spectral statistics.

## 6. Dependence on entanglement structure

A major component of the TNVD study concerns the random-field Ising chain
$$
H^R = H_0 + \sum_n w_n S^z_n,\qquad w_n\in[-W,W],
$$
used to probe how TNVD efficiency depends on the entanglement structure of eigenstates [2508.06159].

The paper evaluates the normalized level-spacing ratio
$$
r = \Bigl\langle \frac{\min(\Delta_n,\Delta_{n+1})}{\max(\Delta_n,\Delta_{n+1})}\Bigr\rangle,
\qquad \Delta_n=E_n-E_{n-1},
$$
and reports $r\approx 0.53$ in the thermal ("GOE") phase and $r\approx 0.39$ in deep many-body localized ("Poisson") behavior. TNVD error $\epsilon$ and $F$ are stated to be smallest in deep MBL, where eigenstates satisfy area-law entanglement entropy, and to increase when eigenstates acquire volume-law entanglement entropy [2508.06159].

Within the thermal regime, the paper distinguishes two subregions. In Region I, at small $W$, low-lying states obey area law while mid-spectrum states display increasing entanglement entropy, described there as an AL$\to$VL transition; TNVD error rises rapidly in this regime. In Region II, at larger $W$ but below the disorder threshold, the spectrum is described as full volume-law, and the errors plateau [2508.06159].

The paper further identifies entanglement-entropy signatures correlated with TNVD performance. In the MBL phase, the distribution of entanglement entropy versus normalized eigenenergy $\epsilon$ forms a broad triangular cloud, and the density of states versus entanglement entropy is Gaussian. In the deep volume-law regime, the entanglement-entropy distribution collapses onto a "slender" Gaussian arc $S_E(\epsilon)$, while the density of states versus entanglement entropy becomes a shifted Poisson,
$$
N_s(S)\propto
\exp\!\bigl[-\tfrac{\Omega\,(S-\tilde S)}{\delta}\bigr].
$$
In deep MBL, the arc collapses and nearly all states have very small area-law entanglement entropy, with the density of states versus $S$ peaking at $S\approx 0$; the paper states that TNVD recovers these trivially [2508.06159].

The interpretation advanced in the work is that area-law or weakly violating entanglement structures are favorable for TNVD, whereas volume-law eigenstates degrade efficiency. This suggests that TNVD’s practical domain is governed not only by Hamiltonian locality but also by the entanglement geometry of the full eigenbasis.

## 7. Position within large-scale many-body diagonalization

TNVD is presented as a diagonalization strategy that combines tensor-network compression and quantum-circuit parametrization to access the full spectrum of large-scale quantum many-body Hamiltonians. In the paper’s formulation, its distinctive feature is that the eigenenergy spectrum is encoded as an MPS while the eigenstates are encoded as circuit evolutions of product states [2508.06159]. That is different from ED, which stores and diagonalizes explicit matrices, and it is also different from variational methods aimed only at ground states or a small number of excited states.

The method’s benchmarks up to $N=100$ are used in the paper to argue that TNVD can operate far beyond the computational limit of ED [2508.06159]. The paper further states that typical signs, including the distribution of entanglement entropy versus eigenenergy and the density of state versus entanglement entropy, indicate area law of entanglement entropy or its violation, and that these features are essential to TNVD efficiency. In this sense, TNVD is not simply a compression scheme; it is a diagnostic framework in which diagonalization accuracy, level statistics, and entanglement profiles are analyzed together.

The work also assigns a specific role to the VQC component. It states that the incorporation of VQC "lays a promising pathway to applying quantum computation to address the volume-law-EE Hamiltonians that lack efficient classical approaches" [2508.06159]. As phrased, this is a forward-looking implication rather than a demonstrated capability of the present benchmarks. A plausible implication is that hybrid tensor-network and circuit-based diagonalization may be especially relevant in regimes where classical tensor networks remain useful for spectral compression but the eigenbasis becomes too entangled for purely classical state representations to remain efficient.

Within the scope of the reported results, TNVD is therefore best understood as a full-spectrum variational diagonalization method whose success is tied to the compressibility of energies as an MPS and to the entanglement complexity of the eigenstates represented by a shared brick-wall VQC.

Source: https://www.emergentmind.com/topics/tensor-network-variational-diagonalization-tnvd