bench

Author: sdat2

cle15/bench_precision.py

@@ -4,7 +4,12 @@
 Sweeps each resolution / iteration knob independently while holding the
 others at their current default values.  Reports timing (ms/call) and
 accuracy (% rmax error relative to the pure-Python reference) for the
-75-case benchmark grid.
+192-case benchmark grid (Vmax 20-90 m/s, r0 300-2000 km,
+f in {3, 5, 7} x 1e-5 s^-1).
+
+Note: r0 = 200 km is excluded because the high-Rossby degenerate regime
+produces large bisection sensitivity that inflates the apparent error
+without reflecting normal solver behaviour (see cle15.md).
 
 Usage::
 
@@ -54,9 +59,14 @@
     ALPHA_EYE_DEFAULT,
 )
 
-# ── Parameter grid (identical to bench_cle15.py) ─────────────────────────────
+# ── Parameter grid ───────────────────────────────────────────────────────────
+# r0 = 200 km is excluded from the benchmark grid because the high-Rossby
+# degenerate regime (very small r0, very high Vmax, very low f) produces
+# large sensitivity to bisection termination side and inflates the apparent
+# error without reflecting normal solver behaviour.  The instability is
+# documented separately (see cle15.md, MATLAB cross-validation section).
 VMAX_VALS = [20.0, 30.0, 40.0, 50.0, 60.0, 70.0, 80.0, 90.0]
-R0_VALS = [200e3, 400e3, 600e3, 800e3, 1000e3, 1200e3, 1600e3, 2000e3]
+R0_VALS = [300e3, 400e3, 600e3, 800e3, 1000e3, 1200e3, 1600e3, 2000e3]
 FCOR_VALS = [3e-5, 5e-5, 7e-5]
 
 ALL_CASES: List[Tuple[float, float, float]] = [

cle15/cle15.md

@@ -106,7 +106,9 @@ res = chavas_et_al_2015_profile(..., solver=SolverConfig.precise())
 
 ### Accuracy vs. cost sweep (`bench_precision.py`)
 
-The four knobs were swept independently over a 192-case benchmark grid ($V_\text{max} \in \{20, 30, \ldots, 90\}$ m/s, $r_0 \in \{200, 400, \ldots, 2000\}$ km, $f \in \{3, 5, 7\} \times 10^{-5}$ s$^{-1}$), with rmax error measured relative to the pure-Python reference.  The grid was expanded from the original 75-case grid ($V_\text{max} \in \{30\text{–}70\}$ m/s, $r_0 \in \{400\text{–}1200\}$ km) to cover the full physically meaningful range including tropical-depression intensities and very large or very compact storms.
+The four knobs were swept independently over a 192-case benchmark grid ($V_\text{max} \in \{20, 30, \ldots, 90\}$ m/s, $r_0 \in \{300, 400, \ldots, 2000\}$ km, $f \in \{3, 5, 7\} \times 10^{-5}$ s$^{-1}$), with rmax error measured relative to the pure-Python reference.  The grid was expanded from the original 75-case grid ($V_\text{max} \in \{30\text{–}70\}$ m/s, $r_0 \in \{400\text{–}1200\}$ km) to cover the full physically meaningful range including tropical-depression intensities and very large or very compact storms.
+
+> **Note on $r_0 = 200$ km:** this value is excluded from the benchmark grid because the high-Rossby degenerate regime ($\mathrm{Ro} \gg 1$) produces large sensitivity to bisection termination side, inflating the apparent error without reflecting normal solver behaviour.  The regime exists and is documented in the `TestMatlabRegression` cross-validation section below.
 
 #### `Nr_e04` — E04 Euler grid points (default 200 000)