Revert "calculate in log space for better accuracy"

This reverts commit 390a087.

Author: taoliu

MACS3/Signal/ScoreTrackMLX.py

@@ -186,24 +186,14 @@ def compute_pvalue(self):
             if self.approx_pvalue:
                 lam = ctrl + pc
                 k = treat + pc
-                tiny = mx.array(np.finfo(np.float32).tiny, dtype=FLOAT32_MX)
-                log2 = mx.log(mx.array(2.0, dtype=FLOAT32_MX))
-                log10_const = mx.log(mx.array(10.0, dtype=FLOAT32_MX))
                 sqrt2 = mx.sqrt(mx.array(2.0, dtype=FLOAT32_MX))
                 z = (k + 0.5 - lam) / mx.sqrt(lam + 1e-9)
-                erfc_arg = z / sqrt2
-                erf_val = mx.erf(erfc_arg)
-                # Clamp away from 1 to keep log1p well-defined for extreme tails.
-                erf_clamped = mx.minimum(
-                    erf_val,
-                    mx.array(1.0 - np.finfo(np.float32).eps, dtype=FLOAT32_MX),
-                )
-                # Compute -log10(tail) in log space to preserve precision for tiny tails.
-                log_tail = mx.log1p(-erf_clamped)
-                log_tail = mx.where(mx.isfinite(log_tail), log_tail, mx.log(tiny))
-                log_tail = log_tail - log2
-                score_backend = -(log_tail / log10_const)
+                tail = 0.5 * (1.0 - mx.erf(z / sqrt2))
+                tail = mx.maximum(tail, mx.array(1e-30, dtype=FLOAT32_MX))
+                score_backend = -_log10(tail)
                 score_np = _to_numpy(score_backend)
+                lam_np = _to_numpy(lam)
+                k_np = _to_numpy(k)
                 score = mx.array(score_np, dtype=FLOAT32_MX)
             else:
                 # exact Poisson on CPU