Correct figures in the comments and docs added by #543

benchmarks/README.md

@@ -10,7 +10,7 @@ and a change is only accepted if the fingerprints still match.
 | `surface_bitflip` | 5x5 surface code, bit-flip noise, 6 shots, χ = 64, natural qubit order | verdicts exact, posterior entries within 1e-2 |
 | `css_optimised` | the same code under `qubit_order_strategy="Optimised"` (reverse Cuthill-McKee) | as above |
 | `shor_depolarising` | Shor code, depolarising noise, 40 shots, χ = 128: the dense-readout path | as above |
-| `classical_ldpc` | random (3,4) LDPC code, XOR constraints + dephasing-DMRG readout | overlaps exact to 1e-10 |
+| `classical_ldpc` | three random (3,4) LDPC codes, XOR constraints + dephasing-DMRG readout | overlaps exact to 1e-10 |
 | `dmrg_ground_state` | DMRG on a 24-site transverse-field Ising chain | energy exact to 1e-10 |
 | `dem_d3` | circuit-level detector error model, d=3 r=3 p=0.8%, eight busiest syndromes, χ = 32 | verdicts exact, class masses within 1e-2 |
 | `dem_d5` | the same at d=5 r=5 p=0.5%, busiest syndrome, χ = 32 | as above |

mdopt/contractor/contractor.py

@@ -23,8 +23,9 @@ def _contract_cached(subscripts, path, backend, *tensors):
     rebuilds path metadata on every call even when ``optimize`` is explicit.
     An expression with an explicit path is shape-independent, so it is
     cached per (subscripts, path) only -- keying on operand shapes made
-    truncation's data-dependent bond dimensions miss 7-14% of calls on
-    large codes -- and built from whatever shapes the first call carries.
+    truncation's data-dependent bond dimensions miss the cache (1-2% of the
+    zip-up calls on the circuit-level DEM workloads, 23% on the RCM
+    surface-code one) -- and built from whatever shapes the first call carries.
     """
     key = (subscripts, path)
     expression = _EXPRESSIONS.get(key)

mdopt/utils/utils.py

@@ -16,7 +16,8 @@ def _to_numpy(a):
 
     Resolved through the backend's own transfer, which honours MDOPT_BACKEND
     and the CUDA device probe; a per-call ``import cupy`` here once cost
-    ~18% of a decoding run.
+    about 11,000 failed imports per decode, between roughly a tenth and
+    a third of the decode depending on the workload.
     """
     host = xp.to_host(a)
     try:
@@ -93,12 +94,14 @@ def svd(
                 #   (the macOS 14+ arm64 wheels) it corrupted memory on the
                 #   decoders' rank-deficient matrices: numpy.linalg.qr died
                 #   with SIGBUS, and a [[72,12,6]] decode at chi_max=400
-                #   returned wrong verdicts on 24 of 24 shots while every
-                #   unit test and benchmark fingerprint still passed;
+                #   returned wrong verdicts on 21 of 22 shots with a
+                #   non-trivial error, while every unit test and benchmark
+                #   fingerprint still passed;
                 # - with OpenBLAS it is memory-safe but not equivalent under
                 #   truncation: it moved a chi_max=64 surface-code posterior
                 #   entry by 0.017, beyond the benchmark suite's tolerance;
-                # - it saved 0-7% on the benchmark workloads.
+                # - it saved at most 7% on the benchmark workloads, and
+                #   nothing measurable on most of them.
                 # Reintroducing it needs `benchmarks/bench_suite.py --check`
                 # and tests/decoding/test_convergence.py to pass.
                 # No finiteness pre-scan: a non-finite input makes the SVD

tests/decoding/test_convergence.py

@@ -29,9 +29,10 @@ def test_bb_72_12_6_natural_order_single_error_converges_in_chi():
 
     This instance exposed the QR pre-reduction of ``svd`` (since removed):
     with it, the decode returned a flat or wrongly peaked
-    posterior at chi_max=400 on NumPy/Accelerate builds -- 24 of 24 shots
-    -- while every unit test and benchmark fingerprint still passed. Each
-    chi_max=400 decode takes 30-60 minutes on a laptop.
+    posterior at chi_max=400 on NumPy/Accelerate builds, as did 21 of the 22
+    shots with a non-trivial error in the natural-order experiment, while
+    every unit test and benchmark fingerprint still passed. The test takes
+    about 20 minutes on an idle laptop.
     """
     code = create_bb_code(6, 6, "x**3 + y + y**2", "y**3 + x + x**2")
     error = "I" * 58 + "Z" + "I" * 13