Robust Optimization of Unitary Tensors

• The paper introduces a variational Monte Carlo framework that optimizes MERA by replacing exact tensor contractions with efficient sampling and unitary manifold updates.
• It employs a causal-cone sampling strategy and gradient projection to significantly reduce computational costs in both 1D and 2D quantum systems.
• Numerical benchmarks on critical models demonstrate near-exact energy evaluations with reduced scaling exponents, highlighting practical efficiency gains.

The Variational Monte Carlo (VMC) framework for the Multi-scale Entanglement Renormalization Ansatz (MERA) is an advanced optimization and sampling methodology designed to greatly reduce the computational bottlenecks of tensor network algorithms for strongly correlated quantum systems. The approach replaces exact contractions within the MERA structure—whose cost scales rapidly with the bond dimension—with a scheme that leverages Monte Carlo sampling on a compressed, causal-cone-determined effective lattice, together with specialized optimization on the manifold of unitary and isometric tensors. This procedure yields dramatic cost reductions in both one- and two-dimensional systems, while retaining variational integrity and statistical error control (Ferris et al., 2012, Barthel et al., 2024).

1. Structure of the MERA Tensor Network

MERA is a hierarchical unitary tensor network comprised, in each layer, of disentanglers $U^{(n)}$ (unitary operators of shape $\chi\times\chi\times\chi\times\chi$) and isometries $W^{(n)}$ (isometric tensors of shape $\chi\times\chi\times\chi$) that systematically coarse-grain a lattice. The physical system with $L$ sites is mapped layer by layer to smaller effective lattices, each downsampled by factor ${\sim}2$, until the top of the network is reached, characterized by a small “top” wavefunction $\varphi$. This layered structure enables efficient renormalization of entanglement, with causality properties ensuring that only a subset of tensors (those in the “causal cone” of observables or Hamiltonian terms) contribute to expectation values (Ferris et al., 2012, Barthel et al., 2024).

2. Variational Objective and Sampling Strategy

The energy functional evaluated in MERA is

$$E[\{U,W\}] = \langle \Psi[\{U,W\}] | H | \Psi[\{U,W\}] \rangle,$$

where the state $|\Psi\rangle$ is prepared by the contraction of all MERA layers. For locality, $H = \sum_i H_i$, each $H_i$ acts on a small number $k$ of adjacent physical sites. The MERA’s unitary structure compresses the required contraction to an $O(\log L)$-site effective wavefunction $|\Psi^{\mathcal{C}}_i\rangle$ on an effective lattice $\mathcal{L}^{\mathcal{C}}$ defined by the causal cone of $H_i$. The local expectation is

$$\langle H_i\rangle = \sum_{\mathbf{n}} P(\mathbf{n}) E_{\mathrm{loc}}(\mathbf{n}),$$

with $$P(\mathbf{n}) = |\langle \mathbf{n} | \Psi^{\mathcal{C}}_i \rangle|^2$$ and $$E_{\mathrm{loc}}(\mathbf{n}) = \frac{\langle \mathbf{n} | H_i | \Psi^{\mathcal{C}}_i \rangle}{\langle \mathbf{n} | \Psi^{\mathcal{C}}_i \rangle}$$, estimated via Monte Carlo sampling over $\mathbf{n}$ (Ferris et al., 2012).

3. Efficient Sampling on the Causal Cone

Due to the circuit-like, unitary nature of MERA, it is feasible to draw independent (“perfect”) samples from $P(\mathbf{n})$ by a single sweep from the top tensor $\varphi$ through all MERA layers, sequentially measuring sites exiting the cone according to their reduced density matrices. At each layer, these reduced density matrices are diagonalized to yield sampling probabilities for exiting indices, which are then projected and iteratively propagated. In the binary 1D MERA, this sampling has total cost $O(\chi^5\log L)$ per sample, eliminating the need for Markov chains or autocorrelation corrections and greatly reducing the expense compared to exact contraction, which is $O(\chi^9\log L)$ (Ferris et al., 2012, Barthel et al., 2024).

4. Gradient-Based Optimization on the Unitary Manifold

MERA optimization requires minimizing the energy functional over the set of unitary (and isometric) tensors. The gradient of the energy with respect to the complex conjugate of a disentangler $U^{(n)}$ is given by

$$\frac{\partial E}{\partial U^{(n)*}} = \sum_i \frac{\partial \langle H_i \rangle}{\partial U^{(n)*}},$$

where each term is estimated using Monte Carlo samples from the causal cone. To maintain the unitary constraint, the gradient is projected onto the tangent space at $U$:

$$G_U = \frac{\partial E}{\partial U^*} - U \left(U^\dagger \frac{\partial E}{\partial U^*}\right).$$

This anti-Hermitian generator defines a geodesic update:

$$U \to U \exp\left[-\mu (U^\dagger \tfrac{\partial E}{\partial U^*} - (\tfrac{\partial E}{\partial U^*})^\dagger U)\right],$$

where $\mu$ is a tunable learning rate. Isometries $W$ are updated analogously (Ferris et al., 2012).

5. Asymptotic Cost Analysis and Algorithmic Phase Diagrams

The computational cost for a full optimization sweep in traditional MERA scales as a high power of bond dimension $\chi$: $O(\chi^9)$ for 1D binary MERA, and up to $O(\chi^{26})$ for 2D quaternary schemes. The VMC approach reduces the per-sample cost exponent by 2–4 (e.g., $O(\chi^6)$ in 1D binary), at the expense of requiring $N_s \sim 1/\epsilon^2 = \chi^{2\beta}$ samples to reach target energy-density precision $\epsilon \sim \chi^{-\beta}$. The optimal method (exact contraction, VMC, or Trotterized variants) depends critically on the exponents $\beta$ (energy error) and $p$ (Trotter depth scaling), which can be expressed in algorithmic phase diagrams delineating regions of asymptotic superiority for each technique (Barthel et al., 2024).

MERA Optimization Mode	Leading Cost (1D binary)	$2d$ quaternary
fMERA (exact contraction, EEG)	$\chi^9$	$\chi^{26}$
fMERA-VMC (CMB)	$\chi^{6+2\beta}$	$\chi^{16+2\beta}$
TMERA-EEG	$\chi^{8\text{--}9}$	varies
TMERA-VMC	$(\chi^{4+p})\chi^{2\beta}$	varies

A plausible implication is that while VMC gives only moderate gains in 1D critical models (where measured $\beta$ is typically $\gtrsim 2$), for 2D geometries with extremely high contraction exponents, VMC achieves asymptotic dominance over exact contraction for a broad range of $\beta$ (Barthel et al., 2024).

6. Algorithmic Implementation and Sampling Bases

The practical implementation of VMC–MERA consists of iterating the following loop: for each local Hamiltonian term, construct its causal-cone circuit, generate $N$ perfect samples for the compressed effective lattice, estimate local energies and gradients from these, project the gradient onto the unitary manifold, and perform tensor updates. Two measurement bases are employed: the Computational Measurement Basis (simply fixing tensor indices) and the Eigenstate Measurement Basis (diagonalizing reduced density matrices), with the latter reducing variance at increased computational cost. Sample reuse across terms and minibatching are practical strategies to further amortize sampling costs (Ferris et al., 2012, Barthel et al., 2024).

7. Numerical Benchmarks and Practical Guidelines

Empirical studies on critical quantum spin chains, including the Ising and Heisenberg models, confirm that the variance of the energy estimator scales as $N^{-1}$, with $\Delta E \propto N^{-1/2}$, and that energy convergence benefits from larger sample numbers. For transverse-field Ising at criticality and $\chi=8$, VMC–MERA attains energies within $10^{-3}$ of the exact contraction result as $N$ increases. For 2D, the theoretical cost savings are larger, though exhaustive benchmarks are limited by computational intractability in fMERA-EEG. Number of samples per update should be set as $N_s \sim 1/\epsilon^2$, e.g., $N_s \sim 10^6$ for $\epsilon \sim 10^{-3}$, while Trotter depth $t$ (for TMERA) should be chosen empirically $t \sim \chi^p$ with $p\approx 1.5$–$2$ for typical critical models (Ferris et al., 2012, Barthel et al., 2024).

• "Variational Monte Carlo with the Multi-Scale Entanglement Renormalization Ansatz" (Ferris et al., 2012)
• "Scaling of contraction costs for entanglement renormalization algorithms including tensor Trotterization and variational Monte Carlo" (Barthel et al., 2024)