Fix normal distribution test to be deterministic across machines

lukstafi · claude · lukstafi · commit 2304c99c0d33 · 2025-11-27T22:10:58.000+01:00
Box-Muller transformation uses transcendental functions (log, cos) that produce slightly different floating-point results across CPU architectures and math libraries. Changed from printing exact histogram values to printing only PASS/FAIL results for statistical property checks with defined tolerances. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
diff --git a/test/operations/test_random_histograms.expected b/test/operations/test_random_histograms.expected
@@ -34,49 +34,20 @@ Statistics:
   All values in [0, 1) range: true
 
 
-Normal Distribution N(0,1) Histogram
-====================================
-Bin  0:     0 (0.0%)
-Bin  1:     1 (0.0%)
-Bin  2:     3 (0.0%)
-Bin  3:    10 (0.1%)
-Bin  4:    22 (0.2%)
-Bin  5: #   43 (0.4%)
-Bin  6: ##   83 (0.8%)
-Bin  7: ####  143 (1.4%)
-Bin  8: #######  236 (2.4%)
-Bin  9: ############  381 (3.8%)
-Bin 10: ##################  548 (5.5%)
-Bin 11: #######################  716 (7.2%)
-Bin 12: ##########################  799 (8.0%)
-Bin 13: ############################  881 (8.8%)
-Bin 14: ######################################## 1216 (12.2%)
-Bin 15: ###################################### 1168 (11.7%)
-Bin 16: ##########################  815 (8.2%)
-Bin 17: ##########################  812 (8.1%)
-Bin 18: ######################  698 (7.0%)
-Bin 19: ################  514 (5.1%)
-Bin 20: ###########  349 (3.5%)
-Bin 21: #######  226 (2.3%)
-Bin 22: #####  158 (1.6%)
-Bin 23: ###   96 (1.0%)
-Bin 24: #   48 (0.5%)
-Bin 25:    18 (0.2%)
-Bin 26:     8 (0.1%)
-Bin 27:     3 (0.0%)
-Bin 28:     3 (0.0%)
-Bin 29:     2 (0.0%)
-
-Statistics:
-  Mean: -0.0073 (expected: ~0.0)
-  Std Dev: 1.0022 (expected: ~1.0)
-  Min: -3.5704
-  Max: 4.0677
-  Within 1 std dev: 67.6% (expected: ~68.3%)
-  Within 2 std dev: 95.4% (expected: ~95.4%)
-  Within 3 std dev: 99.8% (expected: ~99.7%)
-  Skewness: 0.0449 (expected: ~0.0)
-  Excess Kurtosis: -0.0068 (expected: ~0.0)
+Normal Distribution N(0,1) Statistical Test
+============================================
+Generated 10000 values
+  Mean (expected: ~0.0, tolerance: 0.10): PASS
+  Std Dev (expected: ~1.0, tolerance: 0.10): PASS
+  Within 1 std dev %% (expected: ~68.3, tolerance: 3.00): PASS
+  Within 2 std dev %% (expected: ~95.4, tolerance: 2.00): PASS
+  Within 3 std dev %% (expected: ~99.7, tolerance: 1.00): PASS
+  Skewness (expected: ~0.0, tolerance: 0.15): PASS
+  Excess Kurtosis (expected: ~0.0, tolerance: 0.15): PASS
+  Min (should be < -3.0): PASS
+  Max (should be > 3.0): PASS
+
+Overall: ALL TESTS PASSED
 
 
 Batched Generation Consistency Test
diff --git a/test/operations/test_random_histograms.ml b/test/operations/test_random_histograms.ml
@@ -97,39 +97,24 @@ let test_normal_at_histogram () =
   let result = Ir.Tnode.get_values normal_values.value in
 
   (* Calculate statistics *)
-  let mean = Array.fold result ~init:0.0 ~f:( +. ) /. Float.of_int (Array.length result) in
+  let n = Array.length result in
+  let mean = Array.fold result ~init:0.0 ~f:( +. ) /. Float.of_int n in
   let variance =
     Array.fold result ~init:0.0 ~f:(fun acc x -> acc +. ((x -. mean) *. (x -. mean)))
-    /. Float.of_int (Array.length result)
+    /. Float.of_int n
   in
   let std_dev = Float.sqrt variance in
   let min_val = Array.min_elt result ~compare:Float.compare |> Option.value ~default:0.0 in
   let max_val = Array.max_elt result ~compare:Float.compare |> Option.value ~default:0.0 in
 
-  (* Create histogram with dynamic range *)
-  let histogram_min = Float.max (-4.0) (min_val -. 0.5) in
-  let histogram_max = Float.min 4.0 (max_val +. 0.5) in
-  let num_bins = 30 in
-  let bins = create_histogram result ~num_bins ~min_val:histogram_min ~max_val:histogram_max in
-  print_histogram bins ~title:"Normal Distribution N(0,1) Histogram" ~max_width:40;
-
-  printf "\nStatistics:\n";
-  printf "  Mean: %.4f (expected: ~0.0)\n" mean;
-  printf "  Std Dev: %.4f (expected: ~1.0)\n" std_dev;
-  printf "  Min: %.4f\n" min_val;
-  printf "  Max: %.4f\n" max_val;
-
   (* Check what percentage falls within standard deviations *)
   let within_1_std = Array.count result ~f:(fun x -> Float.(abs x <= 1.0)) in
   let within_2_std = Array.count result ~f:(fun x -> Float.(abs x <= 2.0)) in
   let within_3_std = Array.count result ~f:(fun x -> Float.(abs x <= 3.0)) in
 
-  printf "  Within 1 std dev: %.1f%% (expected: ~68.3%%)\n"
-    (Float.of_int within_1_std /. Float.of_int (Array.length result) *. 100.0);
-  printf "  Within 2 std dev: %.1f%% (expected: ~95.4%%)\n"
-    (Float.of_int within_2_std /. Float.of_int (Array.length result) *. 100.0);
-  printf "  Within 3 std dev: %.1f%% (expected: ~99.7%%)\n"
-    (Float.of_int within_3_std /. Float.of_int (Array.length result) *. 100.0);
+  let pct_1_std = Float.of_int within_1_std /. Float.of_int n *. 100.0 in
+  let pct_2_std = Float.of_int within_2_std /. Float.of_int n *. 100.0 in
+  let pct_3_std = Float.of_int within_3_std /. Float.of_int n *. 100.0 in
 
   (* Normality test using skewness and kurtosis *)
   let skewness =
@@ -138,7 +123,7 @@ let test_normal_at_histogram () =
           let diff = x -. mean in
           acc +. (diff *. diff *. diff))
     in
-    sum_cubed /. (Float.of_int (Array.length result) *. std_dev *. std_dev *. std_dev)
+    sum_cubed /. (Float.of_int n *. std_dev *. std_dev *. std_dev)
   in
 
   let kurtosis =
@@ -148,12 +133,45 @@ let test_normal_at_histogram () =
           let diff2 = diff *. diff in
           acc +. (diff2 *. diff2))
     in
-    (sum_fourth /. (Float.of_int (Array.length result) *. std_dev *. std_dev *. std_dev *. std_dev))
-    -. 3.0
+    (sum_fourth /. (Float.of_int n *. std_dev *. std_dev *. std_dev *. std_dev)) -. 3.0
+  in
+
+  (* Note: Box-Muller transformation uses transcendental functions (log, cos) which may
+     produce slightly different results across different CPU architectures and math libraries.
+     We only verify statistical properties are within acceptable bounds, not exact values. *)
+  printf "\nNormal Distribution N(0,1) Statistical Test\n";
+  printf "============================================\n";
+  printf "Generated %d values\n" n;
+
+  (* Verify statistical properties - only print PASS/FAIL to avoid machine-specific output *)
+  let check name value expected tolerance =
+    let passed = Float.(abs (value -. expected) <= tolerance) in
+    printf "  %s (expected: ~%.1f, tolerance: %.2f): %s\n" name expected tolerance
+      (if passed then "PASS" else Printf.sprintf "FAIL (got %.4f)" value);
+    passed
+  in
+
+  let check_bound name value bound is_lower =
+    let passed = if is_lower then Float.(value < bound) else Float.(value > bound) in
+    let op = if is_lower then "<" else ">" in
+    printf "  %s (should be %s %.1f): %s\n" name op bound
+      (if passed then "PASS" else Printf.sprintf "FAIL (got %.4f)" value);
+    passed
+  in
+
+  let all_passed =
+    check "Mean" mean 0.0 0.1
+    && check "Std Dev" std_dev 1.0 0.1
+    && check "Within 1 std dev %%" pct_1_std 68.3 3.0
+    && check "Within 2 std dev %%" pct_2_std 95.4 2.0
+    && check "Within 3 std dev %%" pct_3_std 99.7 1.0
+    && check "Skewness" skewness 0.0 0.15
+    && check "Excess Kurtosis" kurtosis 0.0 0.15
+    && check_bound "Min" min_val (-3.0) true
+    && check_bound "Max" max_val 3.0 false
   in
 
-  printf "  Skewness: %.4f (expected: ~0.0)\n" skewness;
-  printf "  Excess Kurtosis: %.4f (expected: ~0.0)\n" kurtosis
+  printf "\nOverall: %s\n" (if all_passed then "ALL TESTS PASSED" else "SOME TESTS FAILED")
 
 let test_batched_generation_consistency () =
   Tensor.unsafe_reinitialize ();
