- #338

Yaffle · 2015-04-17T03:16:31Z

Fixes #334.

Yaffle · 2015-04-17T06:56:33Z

test/math.js

-  var allowedDiff = precision || 1e-11;
-  return this.within(obj - allowedDiff, obj + allowedDiff);
+  var ok = Math.abs(1 - this.__flags.object / expected) / Number.EPSILON < MATH_EXP_ERROR + 8;
+  this.assert(ok, 'expected #{this} to be almost equal ' + expected, 'expected #{this} to not be almost equal' + expected);
 };


this is not ideal, I do not know how to access private "this.__flags.object" and I did not test it well.

I think this.flags('object') might work better

@ljharb , I fixed by using this.within

Yaffle · 2015-04-18T07:17:06Z

@ljharb , OK, what do you think now?

ljharb · 2015-04-19T02:08:55Z

es6-shim.js

      if (abs < BINARY_32_MIN_VALUE) {
        return sign * roundTiesToEven(abs / BINARY_32_MIN_VALUE / BINARY_32_EPSILON) * BINARY_32_MIN_VALUE * BINARY_32_EPSILON;
      }
      // Veltkamp's splitting (?)
      var a = (1 + BINARY_32_EPSILON / Number.EPSILON) * abs;
      var result = a - (a - abs);
-      if (result > BINARY_32_MAX_VALUE || numberIsNaN(result)) {
+      if (result > BINARY_32_MAX_VALUE || Number.isNaN(result)) {


Please revert this change - it's there intentionally, so that the ordering of shims doesn't matter. Same everywhere it used to say numberIsNaN

So, result !== result is better, because it has no dependency on both Number.isNaN and numberIsNaN

For consistency, I'd prefer numberIsNaN so that the NaN checks can all be routed through the same code path.

OK, I followed this rule

ljharb · 2015-04-19T02:14:59Z

The tests look good, and fail as expected on Chrome 42, node 0.8/0.10/0.11/0.12, and iojs 1.7.

ljharb · 2015-04-19T07:33:03Z

es6-shim.js

-      return (a - b) / (Math.exp(x) + Math.exp(-x));
+      var a = Math.expm1(2 * x);
+      var b = Math.expm1(2 * x) + 2;
+      return a === b ? 1 : a / b;


for what values is a === b && a / b !== 1?

oh snap again, good call. Is this faster enough that it's worth the far less clear version, than the previous which had explicit returns for ± Infinity?

no,
if (a === Infinity) { return 1; } could be used

I think keeping the explicit return checks for ± Infinity are a good idea - probably also the initial NaN and 0 checks too.

@ljharb , I cannot understand how explicit checks can help, I want to have number of "if" branches as small as possible. But I have replaced ternary operator with more readable if (a === Infinity && b === Infinity) { ... } .

ljharb · 2015-05-11T08:52:28Z

test/math.js

@@ -551,7 +565,11 @@ describe('Math', function () {
      expect(Math.sinh(-Infinity)).to.equal(-Infinity);
      expect(Math.sinh(-5)).to.almostEqual(-74.20321057778875);
      expect(Math.sinh(2)).to.almostEqual(3.6268604078470186);
-      expect(Math.sinh(-2e-17)).to.equal(-2e-17);
+      expect(Math.sinh(-2e-17)).to.almostEqual(-2e-17);


In which browsers is Math.sinh(-2e-17) not precisely equal to -2e-17?

@ljharb in a browser of 2150 year, which have Number.EPSILON < 2e-17 * 2e-17 / 3

ljharb · 2015-05-11T08:53:00Z

Thanks, this is looking great. I'll test it in a number of browsers before merging it.

Yaffle · 2015-05-28T06:38:23Z

es6-shim.js

-      if (b === Infinity) { return -1; }
-      return (a - b) / (Math.exp(x) + Math.exp(-x));
+      var a = Math.expm1(2 * x);
+      var b = Math.expm1(2 * x) + 2;


Why am I calculating Math.expm1(2 * x) twice?

ha, good question

Yaffle · 2015-06-04T16:01:18Z

any chance to finish this?

ljharb · 2015-06-04T16:04:22Z

Absolutely - it'll need to be freshly rebased on top of master, and then I'll do a fresh round of browser testing - but I don't think it needs much more work.

Yaffle · 2016-03-08T05:43:02Z

Math.acosh, Math.asinh, Math.cosh, Math.sinh are still not accurate ...

Math.acosh(1 + Number.EPSILON)
Math.asinh(Number.MAX_VALUE)
var EXP_NO_OVERFLOW = 709.782712893384; // Math.log(Number.MAX_VALUE)
Math.cosh(EXP_NO_OVERFLOW + 1 / Math.LOG2E);
Math.sinh(EXP_NO_OVERFLOW + 1 / Math.LOG2E);

ljharb · 2016-03-08T05:52:19Z

@Yaffle I'd love to bring this PR in if you want to freshly rebase it - it's been about 10 months :-)

Yaffle · 2017-04-06T04:58:31Z

~2 years

ljharb · 2017-04-06T05:47:44Z

@Yaffle yes, it's been 1 year since you closed the PR and deleted the branch without rebasing it, and since nobody else seems to understand this topic as well as you do, it will basically never be fixed.

I'd be delighted if you reopened the PR and rebased the branch, then it could get in very soon.

Yaffle · 2017-04-06T06:51:44Z

I have put all implementations of not precise ES6 Math functions to the comment at
compat-table/compat-table#392 (comment)

ljharb · 2017-04-06T06:53:30Z

Yes, thank you, but a dump of implementations doesn't help me change anything. What is helpful is a set of test cases, justified with evidence (browser results, spec steps, etc), and then also an implementation if a good one isn't apparent.

Yaffle · 2017-04-06T18:16:18Z

@ljharb,

a set of test cases:
Math.asinh(1e+300) should not be equal to +∞;
Math.atanh(1e-300) should not be equal to +0;
Math.cosh(710) should not be equal to +∞;
Math.sinh(710) should not be equal to +∞;
Math.acosh(1 + Number.EPSILON) should be near Math.sqrt(2 * Number.EPSILON)
in good implementations (Chrome 54+, Firefox 48+, Safari 10+, Edge):
Math.asinh(1e+300) returns ~691.4686750787736;
Math.atanh(1e-300) returns 1e-300;
Math.cosh(710) returns ~1.1169973830808557e+308;
Math.sinh(710) returns ~1.1169973830808557e+308;
Math.acosh(1 + Number.EPSILON) returns ~2.1073424255447017e-8;
And so the relative error is very big.
Note: Chrome < 54 on Windows fails tests for Math.asinh, Math.atanh and Math.acosh.
Chrome < 40 ? on Windows fails tests for Math.cosh and Math.sinh.
The implementations are available at accuracy of Math.cbrt, Math.expm1, Math.log1p compat-table/compat-table#392 (comment)

ljharb · 2017-04-06T21:54:56Z

Thank you, that is much more helpful!

ljharb · 2017-04-07T05:29:04Z

@Yaffle thanks, those first 4 are taken care of in ec76d49; i missed the edit for the 5th and will follow up with a fix for that.

ljharb · 2017-04-07T05:33:47Z

I'm unable to reproduce the acosh test failure on Chrome 53, 50, even down to 40.

Yaffle · 2017-04-07T05:43:19Z

@ljharb , the test for acosh might look similar to test for cbrt at

es6-shim/es6-shim.js

Line 2083 in ec76d49

    
           defineProperty(Math, 'cbrt', MathShims.cbrt, Math.abs(1 - (Math.cbrt(1e-300) / 1e-100)) / Number.EPSILON > 8);

:

Math.abs(1 - (Math.acosh(1 + Number.EPSILON) / Math.sqrt(2 * Number.EPSILON))) / Number.EPSILON > 8

Math.abs(1 - actual / expected) / Number.EPSILON returns something like "ulp" distance - https://en.wikipedia.org/wiki/Unit_in_the_last_place , while "almostEquals" measures only the absolute error.

ljharb · 2017-04-07T05:49:28Z

thanks, i'll update that.

Yaffle · 2017-04-07T05:53:25Z

@ljharb, thanks

- [Tests] add `to.haveULPDistance` assertion - [Refactor] add `hasULPDistance` helper - use assertion/helper for `cbrt`

ljharb · 2017-04-07T07:14:36Z

k, should be all set.

- [Tests] add `to.haveULPDistance` assertion - [Refactor] add `hasULPDistance` helper - use assertion/helper for `cbrt`

ljharb added needs followup math accuracy labels Apr 17, 2015

Yaffle mentioned this pull request Apr 17, 2015

- #337

Closed

Yaffle reviewed Apr 17, 2015
View reviewed changes

ljharb reviewed Apr 19, 2015
View reviewed changes

Yaffle changed the title ~~Math.acosh, Math.asinh, Math.atanh, ...~~ Math.acosh, Math.asinh, Math.atanh, Math.sinh, Math.cosh, Math.sinh, Math.tanh Apr 29, 2015

Yaffle changed the title ~~Math.acosh, Math.asinh, Math.atanh, Math.sinh, Math.cosh, Math.sinh, Math.tanh~~ Math.acosh, Math.asinh, Math.atanh, Math.cosh, Math.sinh, Math.tanh Apr 29, 2015

Yaffle added 2 commits May 11, 2015 09:08

Math.acosh, Math.asinh, Math.atanh, Math.cosh, Math.sinh, Math.tanh

579ff5f

FINE

6ab5226

ljharb reviewed May 11, 2015
View reviewed changes

Yaffle reviewed May 28, 2015
View reviewed changes

Yaffle added 2 commits May 28, 2015 12:06

Update es6-shim.js

e3c67ec

Update README.md

046ede0

ljharb added the needs rebase label Jun 4, 2015

Yaffle closed this Mar 8, 2016

ljharb mentioned this pull request Mar 8, 2016

- #414

Closed

Yaffle changed the title ~~Math.acosh, Math.asinh, Math.atanh, Math.cosh, Math.sinh, Math.tanh~~ - Mar 9, 2016

ljharb added a commit that referenced this pull request Apr 7, 2017

[Fix] Math.{asinh,atanh,cosh,sinh} precision (from #338)

ec76d49

ljharb added a commit that referenced this pull request Apr 7, 2017

[Tests] add Math.acosh test (per #338)

dd49e8b

ljharb added a commit that referenced this pull request Apr 7, 2017

[Tests] acosh: fix precision (#338)

9f53c0c

- [Tests] add `to.haveULPDistance` assertion - [Refactor] add `hasULPDistance` helper - use assertion/helper for `cbrt`

ljharb added a commit that referenced this pull request Apr 7, 2017

[Tests] acosh: fix precision (#338)

6d65367

- [Tests] add `to.haveULPDistance` assertion - [Refactor] add `hasULPDistance` helper - use assertion/helper for `cbrt`

hartmutobendorf mentioned this pull request Jul 3, 2024

[Snyk] Upgrade es6-shim from 0.33.13 to 0.35.8 hartmutobendorf/angular2-seed#4

Open

- #338

- #338

Conversation

Yaffle commented Apr 17, 2015 • edited by ljharb Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Yaffle commented Apr 18, 2015

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ljharb commented Apr 19, 2015

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ljharb commented May 11, 2015

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Yaffle commented Jun 4, 2015

ljharb commented Jun 4, 2015

Yaffle commented Mar 8, 2016

ljharb commented Mar 8, 2016

Yaffle commented Apr 6, 2017

ljharb commented Apr 6, 2017

Yaffle commented Apr 6, 2017

ljharb commented Apr 6, 2017

Yaffle commented Apr 6, 2017 • edited Loading

ljharb commented Apr 6, 2017

ljharb commented Apr 7, 2017

ljharb commented Apr 7, 2017

Yaffle commented Apr 7, 2017 • edited Loading

ljharb commented Apr 7, 2017

Yaffle commented Apr 7, 2017

ljharb commented Apr 7, 2017

Yaffle commented Apr 17, 2015 •

edited by ljharb

Loading

Yaffle commented Apr 6, 2017 •

edited

Loading

Yaffle commented Apr 7, 2017 •

edited

Loading