Alice 0.1.5

Alice 0.1.5 — Cache Isolation and Thermal MPO

Release Date: June 8, 2026

Version 0.1.5 brings two targeted improvements: DMRG environment-block caching is now
safe for concurrent runs sharing the same configuration directory, and the early-stopping
criterion in thermal_mpo is corrected to account for the operator norm of H.
No breaking changes.

📦 DMRG Environment Cache Isolation

• dmrg.run() now creates a unique subdirectory inside env_cache_dir for each
  invocation — the first 8 hex characters of a freshly generated UUID4
  (e.g. {env_cache_dir}/a1b2c3d4/). Previously, all runs sharing the same
  env_cache_dir wrote to the same left/ and right/ subdirectories, so launching
  two DMRG jobs with the same configuration could corrupt each other's cached blocks.
• The unique subdirectory and all its contents are removed automatically in a finally
  block when run() returns or raises, so no stale files are left behind even after an
  exception.
• Options.env_cache_dir documentation updated to describe the subdirectory scheme and
  the automatic cleanup.

🌡️ thermal_mpo Early-Stopping Correction

• The early-stopping test in thermal_mpo previously compared the bare Taylor
  coefficient |β^n / n!| against coeff_thresh. This is insufficient for Hamiltonians
  with a large operator norm: ‖H^n‖_F can be many orders of magnitude larger than 1,
  so the actual contribution |β^n / n!| × ‖H^n‖_F can remain significant long after
  the bare coefficient falls below coeff_thresh.
• The criterion is now |β^n / n!| × H_pow.scale < coeff_thresh, where H_pow.scale
  is the running Frobenius norm of H^n tracked by NormalMPO. This ensures the loop
  exits only when the contribution of the next term is genuinely negligible.
• A related redundant guard that conditionally skipped the power-update step has been
  removed; H_pow is now always advanced when n < order, which is correct because the
  decision to exit is made at the top of the next iteration.
• Docstrings for the order and coeff_thresh parameters updated to describe the
  corrected criterion.

📊 Statistics

• ~795 tests across 23 test modules (up from ~740 / 23 modules in v0.1.4)
• 3 commits since v0.1.4
• 4 files changed, 141 insertions, 19 deletions
• 22 source modules in three subpackages: alice.network, alice.physics,
  alice.algorithm.dmrg

✅ Compatibility

Breaking Changes: None — fully backward compatible with v0.1.4. All import paths,
function signatures, TOML configuration keys, and geometry APIs are unchanged.

Requirements:

• Python ≥ 3.11
• PyTorch ≥ 2.5
• Nicole ≥ 0.3.7

📝 Notes

The subdirectory isolation in env_cache_dir is a strict improvement: existing
configurations that set env_cache_dir to a shared path will now work correctly across
parallel DMRG jobs without any change to their TOML files or call sites.

The thermal_mpo fix matters whenever ‖H‖_F > 1, which is the common case for
physical Hamiltonians on chains of more than a handful of sites. With the old criterion
the loop could exit prematurely — stopping before the series had converged — producing
a silent accuracy regression for larger systems or higher temperatures.