518:
non-zero coefficients, respectively, so for fewer terms than your 20-little-chunks approximation we got to full double precision. Anyway, sure, if you limit yourself to 5th degree polynomials you can pretty well use whatever basis you want and more or less whatever set of nodes you want. This would fall into the "trivially small" category I mentioned before. But you should not pass off advice that only works out in the degree ~5 case as generically relevant, or people who follow it will end up shooting themselves in the foot when they try to apply it to a higher-degree situation. –
95:
26:
342:
that this method gives pretty good results; maybe there's a paper analyzing the numerical stability etc. (Aside: It logically seems to me like the 2nd derivative matrix should just be the square of the 1st derivative matrix, and the matrix for an arbitrary other derivative should be some higher power. But I haven't played around with it.) –
254:
want to improve this section, maybe you can find some good sources discussing the topic? (The current section is entirely unsourced and I'm not 100% convinced it is exactly correct.) I imagine some paper or another has gone into more detail about both abstract formulas and concrete numerical implementation. –
331:.) I'm fairly convinced sources could be found explicitly supporting the previous version, but if not it should be modified/removed. Secondly, the recommendation to convert to monomial basis still seems like a poor choice to me; Knowledge (XXG) should especially not make this kind of claim without a source.
567:
By the way, you can do much (much) better for such functions with a rational approximation instead of a polynomial. Obviously Runge's function in particular is directly a low-degree rational function (so that's not so impressive to approximate), but the same also applies to arbitrary other functions
551:
Unless you can cite a clear source, I don't think any of that discussion can really stay in
Knowledge (XXG) (in theory, the same standard applies to any other topic). I'll leave you to work on the article for a bit if you want, but when I get a chance I'll come back and rewrite this section based on
517:
which gets you about 10 digits of precision. If you just need 6 digits, you can do with a degree ~60 polynomial. (You need a 188-degree polynomial to get full double precision, ~16 digits.) Note that for this even function, all of the odd-degree terms have coefficient zero, so this is 50 or 30 or 94
357:
Have you implemented in computer code the derivative formulas given in this article and also conversion of
Lagrange polynomials to power basis form and computation of derivatives using the power basis form? I have done this and it is clear that the power basis method is very accurate when the nodes
536:
I will add some text saying that the converting to power basis method for computing derivatives is fine for relatively small polynomial degrees, but for larger degrees, some more sophisticated method is necessary. Then maybe you could clarify what that method is, since I don't know anything about
341:
who describe how to make a matrix for 1st or 2nd derivatives, with matrix multiplication taking values at the nodes to derivatives at the nodes. The derivatives at the nodes can then be thrown into the barycentric formula to compute derivatives at some arbitrary other point(s). My understanding is
253:
which shouldn't really be too big a problem to just simplify to 1 where applicable, in practical implementation. I don't think your proposed alternative suggestion of converting to monomial basis is a good general idea: it's likely to be numerically unstable and give poor practical results. If you
572:
to be at infinity, which makes it very hard to approximate any function which has a pole near the interval. A rational function can put poles anywhere in the extended complex plane. Even for functions with other kinds of singularities like branch cuts, etc. rational functions can make much better
495:
This is why the finite element method works so well. The finite element method is based on polynomial interpolation of a smooth function (the PDE solution) at equally spaced points in the polynomial domain. But the function domain is split up into many smaller regions (the cells of the finite
417:
Polynomial interpolation at equally spaced points in the polynomial domain is an excellent method of approximating any smooth function very accurately, even for fairly high polynomial degrees and functions with large higher-order derivatives like the Runge
270:
Power basis form is fine for many practical applications of
Lagrange interpolation (such as the finite element method) where there are a fairly small number of nodes and the domain is the unit interval or the interval
140:
354:"Elementary numerical analysis" by Conte and de Boor does not even discuss derivatives of Lagrange polynomials. So they are clearly assuming that people will convert to power basis form to compute the derivatives.
53:" error. References show this error when the name of the magazine or journal is not given. Please edit the article to add the name of the magazine/journal to the reference, or use a different citation template. (
380:
citing
Wilkinson (1984) "The perfidious polynomial") Indeed you should avoid equispaced data in general if you can help it: it's a provably bad point distribution; see Platte, Trefethen & Kuijlaars (2011)
496:
element mesh). A good finite element mesh will also have smaller cells in regions where the PDE solution has larger higher-order derivatives in order to minimize the error of the approximate solution.
124:. It has the authority to impose binding solutions to disputes between editors, primarily for serious conduct disputes the community has been unable to resolve. This includes the authority to impose
401:
To answer your other points: (1) I have not implemented the derivative formulas given in this article. As I said, I didn't add that section (though I did clean up the LaTeX a bit.) And (2)
109:
184:, I haven't ever thought too extensively about the subject. I reverted your change though, as from what I can tell all of the problems near nodes boil down to the inclusion of terms like
251:
389:, which greedily adds to a collection of nodes to exactly interpolate one at a time, at each step fitting an optimized rational function minimizing error at the remaining nodes. –
385:. If you need to approximate a function and all you can get is equispaced data, there are are a variety of methods you can try. One recent one is Huybrechs & Trefethen (2023)
490:
456:
of the Runge function into 20 equal length intervals and using a degree 5 polynomial interpolant over each interval results in a maximum approximation error of only
454:
304:
552:
what I can find directly stated in published sources, which will involve wiping out any unsourced material, including the formulas that are currently there. –
372:
You should generally avoid polynomial interpolation of (more than a trivial number of) uniformly spaced nodes, because the result is very badly behaved; see
530:
That is amazing accuracy. I didn't know it was possible to get that kind of accuracy approximating the Runge function using a single degree 100 polynomial.
358:
are uniformly spaced points on the unit interval. This also shows that the formulas given in the article are correct at points not very close to a node.
533:
For my purposes (mainly curve and surface interpolation and the finite element method), degree six is a "large" degree for polynomial interpolation.
50:
573:
approximations than polynomials. The downside is that rational functions are trickier and not quite as uniform to deal with as polynomials. –
405:– speculation about what authors might have been assuming is, for better or worse, not a good enough source for claims in Knowledge (XXG). –
163:
267:
Those derivative formulas are getting very expensive for higher order derivatives. They will also always be inaccurate close to a node.
565:"I didn't know it was possible to get that kind of accuracy approximating the Runge function using a single degree 100 polynomial."
180:
To be honest, while I somewhat cleaned up the formulas about higher derivatives of lagrange polynomials that someone else added to
159:
133:
324:
145:
376:. (And converting to monomial basis is among the worst possible things you can do after that. See Trefethen (2011)
421:
You just need to split up the domain of the function you are approximating. For example, splitting up the domain
328:
78:
121:
38:
323:
First of all, your addition is unsourced. Knowledge (XXG) should only make claims backed by sources. (See
187:
382:
459:
103:
577:
556:
522:
409:
403:"they are clearly assuming that people will convert to power basis form to compute the derivatives"
393:
373:
346:
258:
181:
74:
42:
30:
18:
386:
117:
129:
513:– this is not a very good approximation compared to a single degree 100 interpolant through
152:
132:, editing restrictions, and other measures needed to maintain our editing environment. The
125:
569:
514:
383:"Impossibility of fast stable approximation of analytic functions from equispaced samples"
338:
377:
511:
20 equal length intervals and using a degree 5 polynomial interpolant over each interval
574:
553:
538:
519:
497:
406:
390:
359:
343:
310:
255:
66:
424:
274:
337:
In my opinion we should rewrite this section (I didn't initially add it), following
94:
65:
Please check this page and fix the errors highlighted. If you think this is a
309:
So your reversion is definitely not helpful for readers of this article.
136:
describes the
Committee's roles and responsibilities in greater detail.
580:
559:
546:
525:
505:
412:
396:
367:
349:
318:
261:
167:
113:
are allowed to vote. Users with alternate accounts may only vote once.
82:
334:
However, I agree these derivative formulas get practically ridiculous.
568:
of a pretty broad class. Polynomials necessarily force all of their
149:. If you no longer wish to receive these messages, you may add
139:
If you wish to participate in the 2023 election, please review
107:
is now open until 23:59 (UTC) on Monday, 11 December 2023. All
45:, may have introduced referencing errors. They are as follows:
120:
is the panel of editors responsible for conducting the
70:
58:
54:
378:"Six Myths of Polynomial Interpolation and Quadrature"
462:
427:
277:
190:
484:
448:
298:
245:
8:
473:
461:
426:
276:
231:
213:
204:
189:
564:
510:
402:
387:"AAA interpolation of equispaced data"
7:
246:{\displaystyle (x-x_{i})/(x-x_{i}),}
104:2023 Arbitration Committee elections
176:derivatives of Lagrange polynomials
122:Knowledge (XXG) arbitration process
88:ArbCom 2023 Elections voter message
14:
485:{\displaystyle 3.5\times 10^{-6}}
325:Knowledge (XXG):Original Research
93:
24:
143:and submit your choices on the
443:
428:
293:
278:
237:
218:
210:
191:
41:performed by you, on the page
1:
339:Berrut & Trefethen (2004)
329:Knowledge (XXG):Verifiability
168:00:46, 28 November 2023 (UTC)
581:06:13, 19 January 2024 (UTC)
560:06:05, 19 January 2024 (UTC)
547:05:14, 19 January 2024 (UTC)
526:23:53, 18 January 2024 (UTC)
506:22:40, 18 January 2024 (UTC)
413:19:15, 18 January 2024 (UTC)
397:19:12, 18 January 2024 (UTC)
368:18:42, 18 January 2024 (UTC)
350:18:08, 18 January 2024 (UTC)
319:17:46, 18 January 2024 (UTC)
262:03:05, 17 January 2024 (UTC)
537:that. Does that sound OK?
611:
160:MediaWiki message delivery
542:
501:
363:
314:
157:to your user talk page.
83:03:53, 18 May 2023 (UTC)
71:report it to my operator
486:
450:
300:
247:
35:automatically detected
487:
451:
301:
248:
118:Arbitration Committee
101:Hello! Voting in the
460:
425:
275:
188:
182:Lagrange polynomial
43:Firehole Composites
19:Firehole Composites
482:
446:
374:Runge's phenomenon
296:
243:
134:arbitration policy
51:missing periodical
170:
602:
491:
489:
488:
483:
481:
480:
455:
453:
452:
449:{\displaystyle }
447:
305:
303:
302:
299:{\displaystyle }
297:
252:
250:
249:
244:
236:
235:
217:
209:
208:
158:
156:
97:
28:
27:
610:
609:
605:
604:
603:
601:
600:
599:
515:Chebyshev nodes
469:
458:
457:
423:
422:
273:
272:
227:
200:
186:
185:
178:
173:
172:
150:
98:
90:
25:
22:
12:
11:
5:
608:
606:
598:
597:
596:
595:
594:
593:
592:
591:
590:
589:
588:
587:
586:
585:
584:
583:
562:
534:
531:
493:
479:
476:
472:
468:
465:
445:
442:
439:
436:
433:
430:
419:
399:
355:
335:
332:
307:
295:
292:
289:
286:
283:
280:
268:
242:
239:
234:
230:
226:
223:
220:
216:
212:
207:
203:
199:
196:
193:
177:
174:
141:the candidates
110:eligible users
99:
92:
91:
89:
86:
75:Qwerfjkl (bot)
67:false positive
63:
62:
31:Qwerfjkl (bot)
21:
15:
13:
10:
9:
6:
4:
3:
2:
607:
582:
579:
576:
571:
566:
563:
561:
558:
555:
550:
549:
548:
544:
540:
535:
532:
529:
528:
527:
524:
521:
516:
512:
509:
508:
507:
503:
499:
494:
477:
474:
470:
466:
463:
440:
437:
434:
431:
420:
416:
415:
414:
411:
408:
404:
400:
398:
395:
392:
388:
384:
379:
375:
371:
370:
369:
365:
361:
356:
353:
352:
351:
348:
345:
340:
336:
333:
330:
326:
322:
321:
320:
316:
312:
308:
290:
287:
284:
281:
269:
266:
265:
264:
263:
260:
257:
240:
232:
228:
224:
221:
214:
205:
201:
197:
194:
183:
175:
171:
169:
165:
161:
154:
148:
147:
142:
137:
135:
131:
127:
123:
119:
114:
112:
111:
106:
105:
96:
87:
85:
84:
80:
76:
72:
68:
60:
56:
52:
48:
47:
46:
44:
40:
36:
32:
20:
17:CS1 error on
16:
179:
144:
138:
115:
108:
102:
100:
64:
59:Ask for help
34:
23:
146:voting page
29:Hello, I'm
130:topic bans
73:. Thanks,
69:, you can
575:jacobolus
554:jacobolus
539:Jrheller1
520:jacobolus
498:Jrheller1
418:function.
407:jacobolus
391:jacobolus
360:Jrheller1
344:jacobolus
311:Jrheller1
256:jacobolus
126:site bans
39:this edit
33:. I have
153:NoACEMM
570:poles
37:that
543:talk
502:talk
364:talk
327:and
315:talk
164:talk
116:The
79:talk
578:(t)
557:(t)
523:(t)
464:3.5
410:(t)
394:(t)
347:(t)
259:(t)
55:Fix
49:A "
545:)
504:)
475:−
471:10
467:×
432:−
366:)
317:)
282:−
225:−
198:−
166:)
155:}}
151:{{
128:,
81:)
57:|
541:(
500:(
492:.
478:6
444:]
441:1
438:,
435:1
429:[
362:(
313:(
306:.
294:]
291:1
288:,
285:1
279:[
241:,
238:)
233:i
229:x
222:x
219:(
215:/
211:)
206:i
202:x
195:x
192:(
162:(
77:(
61:)
Text is available under the Creative Commons Attribution-ShareAlike License. Additional terms may apply.
