Leg Tied Tensor Network Letta

---
name: leg-tied-tensor-network-letta
description: MPS + shared physical legs encode long-range correlation.
category: ai_collection
version: "1.0.0"
trigger_words:
  - LETTA
  - leg-tied tensor
  - physical leg ties
  - DMRG generalization
  - long-range correlation tensor network
  - frustrated Heisenberg 2D
  - 3D transverse-field Ising
  - correlator product states
source: arXiv:2609.30101
source_title: Leg-Tied Tensor Network States - Entanglement Beyond Virtual Bonds
authors: Shuoyi Hu, Bing Gu (Westlake University)
published: 2026-09-24
categories: quant-ph, cond-mat.str-el
---

# Leg-Tied Tensor Ansatz (LETTA)

## Core Idea

LETTA supplements an MPS virtual chain with a **physical tie graph** G=(V,E): neighboring tensors *share physical legs* (tensor A[i] carries both s_i and the tied-neighbor tuple s_{P_i}). Physical leg ties encode long-range correlations directly in the amplitude, while the virtual MPS backbone retains short-range multipartite entanglement. Exact contraction cost is governed by the **tie-boundary width** (not a 2D virtual network like PEPS), enabling a deterministic DMRG-like sweep optimization.

Key positioning:
- Contains every compatible MPS as a submanifold; at D=1 reduces to a correlator-product / Jastrow amplitude network (NOT a product state).
- vs correlator product states: LETTA makes local correlators *matrix-valued* and contracts added virtual indices along the MPS backbone.
- vs PEPS: retains 1D ordered virtual bonds → exact contraction possible.

## Variational Optimization (DMRG-like)
1. **Active-site sweep**: contract all tensors except A[i] → effective Hamiltonian H_i and norm matrix N_i.
2. Vectorize a[i] = vec(A[i]) ∈ C^{n_i}, n_i = D_{i-1}·D_i·d_i·∏_{j∈P_i} d_j. Wavefunction is linear in a[i]: |Ψ⟩ = Σ_μ a_μ |Φ_μ⟩.
3. Solve **generalized eigenvalue problem** H_i a = ε_i N_i a (local Rayleigh quotient minimization), advance sweep.
4. Cost: norm environment grows ∏_{j∈F_i} d_j, Hamiltonian expectation ∏_{j∈F_i} d_j² — exponential in max tie-boundary width, but environments cache linearly with chain length at fixed width. MPO bond dimension χ_i enters as O(D_i² χ_i d²^{|F_i|}).
5. Gauge: QR/LQ conditional canonicalization does NOT generally give N_i = I (unlike MPS); sufficient conditions for N_i = I exist but were not met in the benchmark cases.

## Benchmark Results

**2D J1–J2 Heisenberg (6×6, snake ordering, ties across the snake path)**
- LETTA D=4 (4,008 params) beats MPS D=8/D=16 and reaches lower energy than **MPS D=32 (55,976 params) using only 7% of its variational parameters**.
- Energy-density maximum near J2/J1 ≈ 0.7, consistent with stripe-AF transition (~0.62 thermodynamic limit).

**3D transverse-field Ising (N_x×3×3 clusters, raster-scan bonds, OBC, g=1)**
- Correct even-parity ground states from symmetry-unrestricted optimization; at J/g < 0.3 the gauge condition Eq. (15) used.

General: LETTA typically matches an MPS with **one order of magnitude more variational parameters**.

## Design recipe
1. Order sites along a 1D snake/raster path through the lattice (virtual bonds follow the path).
2. Add physical ties for lattice bonds that the path does not traverse (ties across the snake).
3. Choose tie-boundary width small (keep ∏_{j∈F_i} d_j manageable) — this is the cost knob, trading against expressivity.
4. Run DMRG-style sweeps with generalized eigenproblems at each site.

## When to use
- 2D/3D lattice models mapped to a chain where mapped long-range correlations exceed MPS expressivity at feasible D.
- Frustrated magnets (J1–J2) and 3D spin models where PEPS contraction is too expensive.
- Quantum chemistry with explicitly correlated ansätze ideas (Jastrow-like direct correlation + MPS backbone).
- Any setting needing deterministic (non-VMC) optimization of correlator-product-style states.

## Pitfalls
- Contraction is exponential in **maximum tie-boundary width** — large tie degree or high local dimension d blows up cost; keep |F_i| small.
- No guaranteed N_i = I canonical gauge → local eigenproblem is genuinely generalized; conditioning issues possible.
- Tie graph design (which lattice bonds become ties vs path bonds) is currently manual; bad orderings waste the advantage.
- Benchmarks are small clusters (6×6, N_x×3×3); no large-scale comparison vs state-of-the-art PEPS yet.

## Related work
- Nonadiabatic renormalization group (shared physical legs first appeared there).
- Correlator product states / entangled-plaquette states / string-bond states (amplitude-side correlators, VMC-optimized).
- Beyond Bond Gauge (arXiv:2609.30066) — tangent-space question at fixed representation; complementary gauge analysis for WGS tensor networks.

## References
- Hu & Gu, arXiv:2609.30101 (2026).
- Schollwöck, Rev. Mod. Phys. 83 (2011) — DMRG/MPS foundations.