This is an automated email from the ASF dual-hosted git repository.
aherbert pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/commons-statistics.git
commit 5a7c980dc56aa3ccb9e621a495651fd699f3af17
Author: aherbert <aherb...@apache.org>
AuthorDate: Tue Feb 21 12:35:23 2023 +0000
Lower relative epsilon for fastPow test.
The value of 2^-93 is still below the limit of 2^-90.7 for a compensated
pow function.
---
.../org/apache/commons/statistics/inference/DDTest.java | 16 ++++++++++------
1 file changed, 10 insertions(+), 6 deletions(-)
diff --git
a/commons-statistics-inference/src/test/java/org/apache/commons/statistics/inference/DDTest.java
b/commons-statistics-inference/src/test/java/org/apache/commons/statistics/inference/DDTest.java
index 431dc00..a43758f 100644
---
a/commons-statistics-inference/src/test/java/org/apache/commons/statistics/inference/DDTest.java
+++
b/commons-statistics-inference/src/test/java/org/apache/commons/statistics/inference/DDTest.java
@@ -1335,12 +1335,12 @@ class DDTest {
// and we can use the same relative error for both.
// This method uses two epsilon values.
- // The first is for simplePowScaled, the second for powScaled.
+ // The first is for simplePowScaled, the second for fastPowScaled.
// van Mulbregt (2018) pp 22: Error of a compensated pow is ~ 16(n-1)
eps^2.
// The limit is:
// Math.log(16.0 * (Integer.MAX_VALUE-1) * 0x1.0p-106) / Math.log(2) =
-71.0
// For this test the thresholds are slightly above this limit.
- // Note: slowPowScaled does not require an epsilon as it is ~ eps^2.
+ // Note: powScaled does not require an epsilon as it is ~ eps^2.
// Powers that approach and do overflow.
// Here the fractional representation does not overflow.
@@ -1366,8 +1366,8 @@ class DDTest {
DD.fastTwoSum(x, xx, s);
final int n = rng.nextInt(2400, 2500);
// Math.log(16.0 * (2499) * 0x1.0p-106) / Math.log(2) = -90.7
- builder.add(Arguments.of(s.hi(), s.lo(), n, 0x1.0p-52, 0x1.0p-94));
- builder.add(Arguments.of(s.hi(), s.lo(), -n, 0x1.0p-52,
0x1.0p-94));
+ builder.add(Arguments.of(s.hi(), s.lo(), n, 0x1.0p-52, 0x1.0p-93));
+ builder.add(Arguments.of(s.hi(), s.lo(), -n, 0x1.0p-52,
0x1.0p-93));
}
// Powers where the fractional representation overflow/underflow
@@ -1460,6 +1460,10 @@ class DDTest {
// that could be used to correctly round the triple-double to a
double-double.
builder.add(Arguments.of(0.5319568842468022, -3.190137112420756E-17,
2473, 0x1.0p-51, 0x1.0p-94));
+ // Fails fastPow at 2^-94
+ builder.add(Arguments.of(0.5014627401015759, 4.9149107900633496E-17,
2424, 0x1.0p-52, 0x1.0p-93));
+ builder.add(Arguments.of(0.5014627401015759, 4.9149107900633496E-17,
-2424, 0x1.0p-52, 0x1.0p-93));
+
return builder.build();
}
@@ -1469,8 +1473,8 @@ class DDTest {
// Results are obtained from the debugging assertion
// message in TestUtils and thus the BigDecimal is rounded to
DECIMAL128 format.
// simplePowScaled loses ~ 67-bits from a double-double (14-bits from
a double).
- // powScaled loses ~ 26-bits from a double-double.
- // slowPowScaled loses ~ 1-bit from a double-double.
+ // fastPowScaled loses ~ 26-bits from a double-double.
+ // powScaled loses ~ 1-bit from a double-double.
builder.add(Arguments.of(1.402774996679172, 4.203934137477261E-17,
58162209, 28399655, "0.5069511623667528687158515355802548", 0x1.0p-39,
0x1.0p-80));
builder.add(Arguments.of(1.4024304626662112, -1.4084179645855846E-17,
55066019, 26868321, "0.8324073012126417513056910315887745", 0x1.0p-39,
0x1.0p-80));
builder.add(Arguments.of(1.4125582593027008, -3.545476880711939E-17,
50869441, 25348771, "0.5062665858255789519032946906819150", 0x1.0p-39,
0x1.0p-80));
