Variational Monte Carlo for MERA

• The paper introduces a stochastic optimization method that uses Monte Carlo sampling over causal-cone states to significantly reduce computational costs.
• It exploits MERA's unitary circuit structure to enable efficient tensor contractions and high-precision energy evaluations in quantum many-body simulations.
• The approach integrates gradient-based optimization on tensor manifolds with cost-efficient contraction-sequence searches, benefiting both 1D and 2D systems.

The Variational Monte Carlo (VMC) framework for the Multi-Scale Entanglement Renormalization Ansatz (MERA) is a stochastic optimization paradigm designed to address the computational bottlenecks inherent in the contraction and optimization of hierarchical tensor networks for quantum many-body systems. By leveraging Monte Carlo sampling of causal-cone states and exploiting the unitary circuit structure of MERA, this approach achieves significant reductions in cost compared to traditional exact-environment evaluations, particularly in high spatial dimensions. The framework integrates algorithmic innovations in sampling, optimization on tensor manifolds, and contraction-sequence search, yielding tunable precision and enabling studies of large and/or complex quantum systems (Ferris et al., 2012, Barthel et al., 2024).

1. Foundations of MERA and Tensor Network Structure

MERA is a hierarchical tensor network constructed from two fundamental types of tensors per layer:

• Disentanglers $U^{(n)}$: Unitary operators of rank four, acting to remove short-range entanglement between neighboring sites at each hierarchical layer.
• Isometries $W^{(n)}$: Isometric tensors that coarse-grain two sites into one, mapping larger Hilbert spaces into subspaces of dimension $\chi$.

Starting from a physical lattice $\mathcal{L}$ with $L$ sites of local dimension $d$, successive layers of disentanglers and isometries halve the number of effective sites, culminating in a small "top" wavefunction $\varphi$. The indices—carrying the MERA bond dimension $\chi$—connect layers in a directed acyclic topology, while the contraction of all layers constitutes a unitary quantum circuit preparing the variational many-body state $|\Psi[\{U,W\}]\rangle$.

The unitary nature and causal structure of MERA enable efficient computation of local observables via "causal cones," i.e., minimal sub-networks supporting each local operator. Graphical notation distinguishes unitaries (boxes with directed indices) from isometries (triangles), with specific conventions for input and output legs (Ferris et al., 2012, Barthel et al., 2024).

2. Variational Energy Functional and Monte Carlo Estimation

The variational objective is the expectation value of the Hamiltonian:

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

with $H = \sum_i H_i$ decomposed as a sum over local terms. Due to MERA normalization, the denominator $\langle\Psi|\Psi\rangle$ is unity. The efficient evaluation of $E$ exploits:

• The unitary circuit structure, which confines each $H_i$’s support to a narrow causal cone;
• The capacity to contract all tensors outside the cone, forming an effective lattice $\mathcal{L}^{\mathcal{C}}$ of size $O(\log L)$ with wavefunction $|\Psi^{\mathcal{C}}_i\rangle$.

Each local expectation value is thus:

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

where $\mathbf{n}$ is a configuration of the compressed lattice, $$P(\mathbf{n}) = |\langle \mathbf{n} |\Psi^{\mathcal{C}}_i\rangle|^2$$, and $E_{\rm loc}(\mathbf{n})$ is a local energy estimator. This expectation is approximated by drawing i.i.d. samples $\mathbf{n}^{(r)}$ from $P$, achieving unbiased stochastic estimates as $N\to\infty$ (Ferris et al., 2012).

3. Perfect Sampling and Causal-Cone State Generation

MERA's unitary circuit depth enables exact—or "perfect"—sampling of $\mathbf{n}\sim P(\mathbf{n})$ by a single top-to-bottom network sweep:

1. Initialize the state as the top tensor $\varphi$.
2. At each descending layer $\ell$, apply $U^{(\ell)}$ and $W^{(\ell)}$.
3. For each effective site $j$, compute the one-site reduced density matrix $\rho_j$, diagonalize, and sample an outcome $n_j$ with probability $\lambda_{n_j}$.
4. Project the state onto $\ket{n_j}$ and continue.

On completion, one acquires a configuration $\mathbf{n}$, corresponding to a pure state in the causal-cone basis. The computational cost per sample is $O(\chi^5\log L)$ for 1D binary MERA, markedly more efficient than full-network contraction. No Markov chains or autocorrelation corrections are required, in contrast to Markov Chain Monte Carlo-based approaches (Ferris et al., 2012, Barthel et al., 2024).

4. Optimization on Manifolds of Unitary Tensors

To maintain the unitary (or isometric) nature of MERA tensors through stochastic gradient descent, VMC-MERA employs a modified steepest descent update restricted to the appropriate manifold:

• The gradient of the energy functional with respect to each $U^{(n)*}$ is accumulated over all sampled configurations and local terms.
• The update projects this gradient onto the tangent space at $U$, yielding an anti-Hermitian generator $G_U$:

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

• The new unitary is set by a geodesic step:

$$U \longrightarrow U \exp\left[ -\mu \left(U^\dagger \frac{\partial E}{\partial U^* } - \left(\frac{\partial E}{\partial U^*}\right)^\dagger U \right) \right]$$

for step size $\mu$.

Isometries $W$ are optimized analogously. The number of Monte Carlo samples $N$ and step size $\mu$ are tunable to balance noise against convergence speed. This exact manifold-constrained update is critical for robust and unbiased VMC optimization (Ferris et al., 2012).

5. Algorithmic Workflow and Pseudocode

The canonical VMC-MERA optimization loop proceeds as follows:

1. Initialization: Randomize tensors $\{U^{(n)}, W^{(n)},\varphi\}$, enforcing isometry and unitarity constraints.
2. Causal-Cone Sampling: For each local Hamiltonian term $H_i$:
   • Build the causal-cone circuit $|\Psi^{\mathcal{C}}_i\rangle$.
   • Draw $N$ perfect samples $\{\mathbf{n}^{(r)}\}$ from $P(\mathbf{n})$.
   • For each sample, compute $E_{\rm loc}^{(r)}$ and the sample-specific environment contributions.
   • Average over samples for $E$ and stochastic gradients.
3. Gradient Projection and Update: Aggregate gradients over $i$, project to the tangent space, and update each $U$, $W$, and $\varphi$ using the manifold geodesic step.
4. Metaparameter Scheduling: Optionally decrease $\mu$ and/or increase $N$ as required.
5. Convergence Check: Iterate the sweep until target precision is achieved.

A practical pseudocode for the contraction-sequence search—essential for optimizing contraction order in Trotterized MERA—employs breadth-first search with cost-based pruning, leveraging cost exponents for both standard and Trotterized contractions (Barthel et al., 2024).

6. Cost Scaling, Phase Diagrams, and Practical Regimes

The cost of VMC-MERA decomposes into:

• Per-sample contraction: $O(\chi^5)$ in 1D fMERA, with further reduction for Trotterized MERA (TMERA) depending on Trotter depth $t\sim\chi^p$.
• Number of samples $N_s$: Scales as $N_s = O(1/\epsilon^2) = \chi^{2\beta}$ for target energy error $\epsilon \sim \chi^{-\beta}$.

Comparative cost exponents (for leading order in $\chi$) for 1D and 2D MERA geometries are summarized in the following table for a single update iteration (Barthel et al., 2024):

Algorithm	1D-binary α	2D-quaternary α
fMERA-EEG	9	26
fMERA-VMC(CMB)	6 + 2β	16 + 2β
TMERA-EEG	function of p	function of p
TMERA-VMC	4 + p + 2β	varies

The most efficient algorithm in the $(\beta, p)$-plane is determined by the smallest $\alpha$. For critical 1D models (Heisenberg, XX, BLBQ), measured $\beta$ values range from $\approx1.75$ to $3.7$ and typical $p\sim1.3$–$2$; in this regime, fMERA-EEG remains optimal. In 2D, the much higher exponents for fMERA-EEG make VMC and/or Trotterization highly advantageous for realistic $\beta$ and $p$ (Barthel et al., 2024).

Algorithmic phase diagrams, defined by thresholds such as $2\beta = \alpha_0 - \alpha_1$, demarcate the regions where each of the four schemes (fMERA-EEG, fMERA-VMC, TMERA-EEG, TMERA-VMC) is optimal.

7. Numerical Benchmarks and Outlook

Case studies for the critical Ising and various Heisenberg-type chains affirm key properties:

• Variance scaling: ${\rm Var}[E_{\rm loc}] \propto N^{-1}$, implying stochastic error $\Delta E \propto N^{-1/2}$.
• Convergence: For fixed bond dimension (e.g., $\chi=4$ or $8$) and system size (e.g., $L=24$), energies approach the exact-MERA result with increasing $N$, with accuracy improving well below $10^{-3}$.
• Criticality: The variance of the energy estimator peaks at critical points, correlating with block entanglement entropy.
• Dimensionality: In 1D, the VMC advantage is modest for large $\chi$ due to high $\beta$; in 2D, VMC and/or Trotterization provide dramatic cost reductions.

Practical guidelines include setting $N_s \propto 1/\epsilon^2$, choosing Trotter depth $t\sim\chi^p$ with empirically tuned $p$, and optimizing contraction sequences for maximally efficient usage of circuit structure (Ferris et al., 2012, Barthel et al., 2024).

---

The VMC framework for MERA thus enables scalable and tunable-precision simulation of strongly correlated quantum many-body systems, with architectural advantages that are especially prominent in two-dimensional lattice models. Continued algorithmic development—including improved contraction-sequence searches, circuit-based ansatzes, and variance-reduction techniques—remains an active direction for advancing the capacity of tensor network approaches in higher dimensions.