Skip to content

TN/MPS Baseline Design

This is the CPU-first design artifact for the N=30-40 tensor-network baseline path. It is planning and preregistration evidence only.

Boundary

Design and preregistration evidence only; no N>=30 tensor-network row, hardware advantage, or broad quantum-advantage claim is established.

Decision

Use the Python/quimb CPU MPS adapter as the first execution path, with the bounded native Schmidt/resource model as a deterministic fallback and validation scaffold. Keep ITensor/Julia and GPU TN as explicitly blocked follow-ups until owner-gated toolchain work.

Adapters

adapter language dependency status role max N
quimb_mps_cpu Python quimb[tensor] optional extra optional_dependency CPU DMRG/TEBD execution adapter for nearest-neighbour or explicitly truncated K_nm 40
bounded_native_schmidt Python/NumPy none ready Deterministic resource and discarded-weight scaffold for row validation 40
itensor_julia Julia ITensor.jl blocked Future independent MPS parity adapter after Julia ownership is assigned 40
gpu_tn CUDA/Python cuTensorNet or equivalent owner_gated Future GPU tensor-network comparison lane 40

Size Plan

N class CPU-first adapter CPU execution enabled GPU follow-up
30 pilot quimb_mps_cpu True defer to the owner-gated GPU lane
32 pilot quimb_mps_cpu True defer to the owner-gated GPU lane
36 extension quimb_mps_cpu True defer to the owner-gated GPU lane
40 extension quimb_mps_cpu True defer to the owner-gated GPU lane

Acceptance Gates

  • Every N=30-40 row must record status, wall_time_ms, memory_bytes, max_bond, discarded_weight, entropy proxy, command, machine, dependencies, and git_commit.
  • Skipped rows must carry explicit size, dependency, or resource-gate reasons.
  • TN/MPS rows must be compared against the registered scaling-protocol matrix before any execution evidence is promoted.

Blocked Claims

  • No broad quantum advantage claim from this design artifact.
  • No tensor-network hardness claim until measured N=30-40 TN rows exist.
  • No GPU TN comparison until the owner-gated GPU lane promotes it.
  • No Julia/ITensor parity claim until a maintained Julia adapter is measured.

CPU Execution Prerequisites

The CPU-first TN/MPS rows can execute once the quimb dependency, resource caps, and scaling-row schema are pinned by this manifest.

Regeneration

scpn-bench s2-tn-mps-baseline-design