diff --git a/tests/pca.py b/tests/pca.py
index 2827bb7..a3c33b2 100644
--- a/tests/pca.py
+++ b/tests/pca.py
@@ -2,11 +2,44 @@
 import sys
 import numpy as np
 from sklearn.decomposition import PCA
-from sklearn.preprocessing import StandardScaler
-import json # For original_main_logic
-import os   # For original_main_logic
-import scipy.linalg as la # For original_main_logic
-import random # For original_main_logic
+import json  # For original_main_logic
+import os    # For original_main_logic
+import scipy.linalg as la  # For original_main_logic
+import random  # For original_main_logic
+
+
+NEAR_ZERO_THRESHOLD = 1e-9
+
+
+def center_and_scale_columns_np(X):
+    """Replicates the Rust PCA preprocessing: column-wise centering and scaling
+    using the unbiased sample standard deviation (n-1 in the denominator) with
+    small-value sanitization."""
+
+    X = np.asarray(X, dtype=float)
+    if X.size == 0:
+        return X.copy(), np.array([]), np.array([])
+
+    n_samples, n_features = X.shape
+
+    mean = np.mean(X, axis=0)
+    centered = X - mean
+
+    if n_samples > 1:
+        sum_sq = np.sum(centered * centered, axis=0)
+        variance = np.maximum(sum_sq, 0.0) / float(n_samples - 1)
+    else:
+        variance = np.zeros(n_features)
+
+    std = np.sqrt(variance, dtype=float)
+    sanitized_std = np.where(
+        (~np.isfinite(std)) | (std <= NEAR_ZERO_THRESHOLD),
+        1.0,
+        std,
+    )
+
+    scaled = centered / sanitized_std
+    return scaled, mean, sanitized_std
 
 
 def print_numpy_array_for_rust(arr):
@@ -145,8 +178,7 @@ def generate_random_data_original(samples=5, features=5, random_seed=None):
     return np.random.randn(samples, features)
 
 def manual_pca_original(X, n_components=None):
-    scaler = StandardScaler()
-    X_scaled = scaler.fit_transform(X)
+    X_scaled, _, _ = center_and_scale_columns_np(X)
     n_samples, n_features = X_scaled.shape
     if n_components is None: n_components = min(n_samples, n_features)
     else: n_components = min(n_components, min(n_samples, n_features))
@@ -181,8 +213,7 @@ def manual_pca_original(X, n_components=None):
     return X_transformed, components, eigvals
 
 def library_pca_original(X, n_components=None):
-    scaler = StandardScaler()
-    X_scaled = scaler.fit_transform(X)
+    X_scaled, _, _ = center_and_scale_columns_np(X)
     n_samples, n_features = X_scaled.shape
     max_components = min(n_samples, n_features)
     if n_components is None: n_components = max_components
@@ -288,7 +319,19 @@ def parse_arguments_main():
     
     # Argument for n_components, used by both modes
     # For --generate-reference-pca, it's required. For original_main_logic, it's optional.
-    parser.add_argument("-k", "--n-components", type=int, help="Number of components for PCA.")
+    # Accept both hyphenated and underscored versions of the flag. The Rust tests
+    # currently invoke the script with "--n_components", while the original
+    # command-line interface exposed "--n-components".  Argparse treats hyphens
+    # and underscores as distinct option names, so support both spellings here
+    # by routing them to the same destination.
+    parser.add_argument(
+        "-k",
+        "--n-components",
+        "--n_components",
+        dest="n_components",
+        type=int,
+        help="Number of components for PCA.",
+    )
     
     # Argument to switch to reference generation mode
     parser.add_argument("--generate-reference-pca",