齊 靜
(重慶師范大學(xué)涉外商貿(mào)學(xué)院數(shù)學(xué)與計(jì)算機(jī)系,重慶401520)
徑向基函數(shù)插值逼近的誤差分析
齊 靜
(重慶師范大學(xué)涉外商貿(mào)學(xué)院數(shù)學(xué)與計(jì)算機(jī)系,重慶401520)
函數(shù)逼近是數(shù)學(xué)規(guī)劃中一個(gè)基本的問(wèn)題,近年來(lái),國(guó)內(nèi)外的一些學(xué)者對(duì)徑向基函數(shù)插值逼近問(wèn)題進(jìn)行了廣泛的研究,對(duì)于某些測(cè)試函數(shù)來(lái)說(shuō),徑向基插值相對(duì)于經(jīng)典的插值方法,如牛頓插值、拉格朗日插值來(lái)說(shuō),在CPU時(shí)間、逼近程度等方面有著一定的優(yōu)勢(shì),因此徑向基函數(shù)插值成為解決散亂數(shù)據(jù)插值的一種新的有效的方法.將采用幾種常見(jiàn)的徑向基函數(shù)來(lái)逼近一元函數(shù)、二元函數(shù),進(jìn)行數(shù)值試驗(yàn)以及誤差分析,并對(duì)徑向基函數(shù)中的參數(shù)進(jìn)行分析,獲得了良好的誤差分析結(jié)果.
徑向基函數(shù);散亂數(shù)據(jù)插值;函數(shù)逼近;誤差分析;響應(yīng)面模型
徑向基函數(shù)具有計(jì)算格式簡(jiǎn)單、計(jì)算工作量?。?-2]等特點(diǎn),在實(shí)際科研領(lǐng)域和工業(yè)中有著廣泛的應(yīng)用.常見(jiàn)的徑向基函數(shù)有等.
給定n個(gè)不同的點(diǎn)x1,x2,…,xn∈Rd,并且它們的函數(shù)值f(x1),f(x2),…,f(xn)是已知的,取定徑向基函數(shù)?,構(gòu)建響應(yīng)面模型,尋找具有如下形式的函數(shù):
其中:f(x)是一個(gè)確定性的連續(xù)函數(shù),‖·‖是歐幾里得范數(shù),‖x-xi‖表示x與中心點(diǎn)xi之間的歐氏距離,λi∈R,i=1,2,…,n,φ就是徑向基函數(shù),選定一個(gè)徑向基函數(shù)?之后,定義矩陣Φij:=φ(‖xi-xj‖),i,j=1,2,…,n.
xi 0.5000 1.0000 1.5000 2.0000 2.5000 3.0000 f(xi) 0.4794 0.8415 0.9975 0.9093 0.5985 0.1411
2.1 已知數(shù)據(jù)
①用 Mat1ab語(yǔ)言編寫(xiě) Newton插值法、radia1basis插值法的程序,對(duì)以上數(shù)據(jù)進(jìn)行插值;②利用Mat1ab畫(huà)出用牛頓插值和徑向基插值所得函數(shù)在區(qū)間[0.2∶0.01∶0.8]的函數(shù)值、誤差值的CPU時(shí)間.
1)牛頓插值法所得的解析式為:-0.001413333333333*x^5+0.050400000000001*x^4-0.279833333333338*x^3+0.128200000000006*x^2+0.929846666666663*x+0.014300000000001;
徑向基插值所得的解析式為:
1186084396974121/(18014398509481984*e^((2*x-5)^2/4))-8838835290143603/ (2251799813685248*e^((2*x-3)^2/4))-(680434345985903*x)/4503599627370496+4370545598947323/ (1125899906842624*e^(x^2-x+1/4))+3426394797347797/(144115188075855872*e^((x-1)^2))-1096960021468651/(281474976710656*e^((x-2)^2))+8671626295667839/(2251799813685248*e^((x-3)^2))+2682551975049329/1125899906842624.
2)圖1為Newton插值和radia1 basis插值的CPU時(shí)間
從圖1的牛頓插值計(jì)算CPU時(shí)間和徑向基插值的CPU時(shí)間圖中可以明顯看出,對(duì)于某些問(wèn)題來(lái)說(shuō),徑向基可以在一定程度上節(jié)約時(shí)間,提高工作效率,這也是近年來(lái)徑向基引起人們廣泛關(guān)注的原因.現(xiàn)已知道,徑向基函數(shù)有5種,其中的MQ函數(shù)和高斯函數(shù)中都含有參數(shù),針對(duì)不同的問(wèn)題參數(shù)應(yīng)當(dāng)取不同的值,下面通過(guò)數(shù)值試驗(yàn)來(lái)研究其中參數(shù)的選取.
圖1 Newton插值和radia1 basis插值的CPU時(shí)間Fig.1 The CPU time of Newton interpo1ationand radia1 basis interpo1ation
圖2 當(dāng)c=0.001和c=0.01時(shí),MQ函數(shù)的誤差圖像Fig.2 The error map of MQ function when c=0.001 and c=0.01
表1 MQ函數(shù)參數(shù)c=0.001時(shí)的插值Tab.1 MQ interpo1ation function of parameter c=0.001
表2 MQ函數(shù)參數(shù)c=0.01時(shí)的插值Tab.2 MQ interpo1ation function of parameter c=0.01
2.2 徑向基函數(shù)中參數(shù)C的數(shù)值試驗(yàn)
以下數(shù)值試驗(yàn)采用Mat1ab2001b進(jìn)行編程計(jì)算.下面將利用Mu1tiquadric(MQ)函數(shù)、Gaussian函數(shù)、Cubic函數(shù)、Linear函數(shù)等進(jìn)行插值數(shù)值試驗(yàn).對(duì)于MQ函數(shù)和Gaussian函數(shù),當(dāng)參數(shù)C取不同的值時(shí),得到的誤差效果也將不一樣.
例1 利用徑向基函數(shù)對(duì)原函數(shù)f(x)= sin(2x)在區(qū)間[0,1]上作插值,進(jìn)行數(shù)值試驗(yàn)及誤差分析.
首先將區(qū)間[ 0,1]進(jìn)行100等分,將得到101個(gè)點(diǎn):(xi,f(xi)),i=0,0.01,…,1.選定徑向基函數(shù)?,定義矩陣Φij:=?(‖xi-xj‖),i,j=1,2,…,101[4].
λi,b,a通過(guò)求解下面的線性方程組得
將解出的λi,b,a代入式(1)即可得到Sn(x)函數(shù),以及誤差函數(shù)W(x)=Sn(x)-f(x).
當(dāng)c取不同值時(shí),MQ函數(shù)插值誤差圖像如下:
圖2為MQ函數(shù)c=0.001和c=0.01的誤差圖像.
表3 MQ函數(shù)參數(shù)c=0.1時(shí)的插值Tab.3 MQ interpo1ation function of parameter c=0.1
表4 MQ函數(shù)參數(shù)c=0.1時(shí)的插值Tab.4 MQ interpo1ation function of parameter c=0.1
圖3為MQ函數(shù)c=0.1和c=1的誤差圖像.
當(dāng)c=10,100,1000時(shí)的誤差與c=1的基本相似(圖略).從MQ函數(shù)的誤差圖像上可以很明顯的看到:當(dāng)c的取值越小時(shí),誤差也越小,當(dāng)c= 0.001時(shí),誤差值集中在[-0.001,0.001],逼近效果是比較好的.形狀參數(shù)c極大的影響著徑向基函數(shù)逼近的誤差,當(dāng)形狀參數(shù)在一定范圍內(nèi)減小時(shí),逼近誤差也會(huì)減小,超出范圍時(shí)誤差又會(huì)變大[5-6].更多的相關(guān)方面的研究也往往只是通過(guò)數(shù)值實(shí)驗(yàn)研究,極少有人從理論方面進(jìn)行研究[7].
另外,其他的徑向基函數(shù),如Cubic函數(shù)、Linear函數(shù)在逼近這類函數(shù)時(shí),效果也是相當(dāng)好的.
圖4為Cubic函數(shù)、Linear函數(shù)的誤差圖像.
圖3 當(dāng)c=0.1和c=1時(shí),MQ函數(shù)的誤差圖像Fig.3 The error map of MQ function when c=0.1 and c=1
圖4 Cubic函數(shù)、Linear函數(shù)的誤差圖像Fig.4 The error map of Cubic function and Linear function
例2 利用徑向基MQ函數(shù)、Gaussian函數(shù)對(duì)原函數(shù)f(x)=sin(3x)·sin(2y)在區(qū)間[0,1]×[0,1]上作插值,進(jìn)行數(shù)值試驗(yàn)及誤差分析.
圖5為當(dāng)c=0.01時(shí),MQ函數(shù)誤差圖像.圖6為當(dāng)c=10時(shí),MQ函數(shù)誤差圖像.
圖5 當(dāng)c=0.01時(shí),MQ函數(shù)的誤差圖像Fig.5 The error map of MQ function when c=0.01
圖6 當(dāng)c=10時(shí),MQ函數(shù)的誤差圖像Fig.6 The error map of MQ function when c=10
通過(guò)上面的誤差圖像看到:MQ方法的插值誤差都是隨著c的減小而減小,即c值越大誤差越大.因此對(duì)于這樣的二元函數(shù)逼近時(shí),可以適當(dāng)?shù)臏p小c的值.
對(duì)于高斯函數(shù)來(lái)說(shuō),當(dāng)c取0.001~1時(shí),插值誤差變化不太明顯,但當(dāng)c取10甚至更大時(shí),誤差非常大. 圖7為當(dāng)c=0.1時(shí),高斯函數(shù)誤差圖像.
圖8為當(dāng)c=10時(shí),高斯函數(shù)誤差圖像:對(duì)于高斯函數(shù)來(lái)說(shuō),當(dāng)c取0.001~1時(shí),插值誤差變化不太明顯,但當(dāng)c取10甚至更大時(shí),誤差非常大.另外,對(duì)于Cubic函數(shù)、Linear函數(shù)的插值效果,與MQ函數(shù)當(dāng)c=0.01時(shí)的效果相當(dāng).
圖7 當(dāng)c=0.1時(shí),Gaussian函數(shù)的誤差圖像Fig.7 The error map of Gaussian function when c=0.1
圖8 當(dāng)c=10時(shí),Gaussian函數(shù)的誤差圖像Fig.8 The error map of Gaussian function when c=10
徑向基函數(shù)具有良好的逼近能力,MQ方法和Gaussian方法中的參數(shù)c對(duì)試驗(yàn)結(jié)果有較大的影響.徑向基函數(shù)的插值最終歸結(jié)為線性方程組的求解,當(dāng)徑向基函數(shù)是正定時(shí),它的線性組合理論上說(shuō)可以逼近任何的連續(xù)函數(shù),但是對(duì)于具體函數(shù)的逼近,該如何選取c的值目前還沒(méi)有較好的辦法,只能憑經(jīng)驗(yàn)選取,這也是未來(lái)的一個(gè)研究方向.
[1] 吳宗敏.函數(shù)的徑向基表示[J].數(shù)學(xué)進(jìn)展,1998,27:202-208.
[2] ROMMEL G Regis,CHRISTINE A Shoemaker.Improved strategies for Radia1 Basis Function methods for g1oba1 optimization[J].Journa1 of G1oba1 Optimization,2007,37:113-135.
[3] KENNETH Ho1mstr?m.An adaptive radia1 basis a1gorithm(ARBF)for expensive b1ack-box g1oba1 optimization[J].Journa1 of G1oba1 Optimization,2008,41:447-464.
[4] REGIS G,CHRISTINE A Shoemaker.Constrained G1oba1 Optimization of Expensive B1ack Box Function Using Radia1 Basis Function[J].Journa1 of G1oba1 Optimization,2005,31:153-171.
[5] SCHABACK R.Error estimates and condition numbers for radia1 basis function interpo1ation[J].Adv Comput Math,1995,3:251-264.
[6] WU Z,SCHABACK R.Loca1 error estimates for radia1 basis function interpo1ation of scattered data[J].IMAJ of Numerica1 Ana1ysis,1993,13:13-27.
[7] 熊正超.徑向基函數(shù)逼近中的若干問(wèn)題研究[D].上海:復(fù)旦大學(xué),2007.
責(zé)任編輯:時(shí) 凌
The Error AnalYsis of Radial Basis Function InterPolation and APProximation
QI Jing
(Schoo1 of Mathematics and Computer Science,Chongqing Norma1 University Foreign Trade and Business Co11ege,Chongqing 401520,China)
The function approximation is a basic prob1em in mathematica1 programming.In the past few years,some domestic and foreign scho1ars have extensive1y studied the radia1 basis function.For some test function,radia1 basis interpo1ation,compared with the c1assica1 interpo1ation methods such as Newton interpo1ation and Lagrange interpo1ation,has some advantages in terms of CPU time,and degree of approximation,so the radia1 basis function interpo1ation is a new effective method to so1ve the scattered data interpo1ation.This paper uses severa1 common radia1 basis functions to approximate a univariate function and dua1 function for numerica1 experiment and error ana1ysis and good error ana1ysis resu1ts are obtained.
radia1 basis function;scattered data interpo1ation;function approximation;error ana1ysis;response surface mode1
O241.5
A
1008-8423(2016)02-0162-04
10.13501/j.cnki.42-1569/n.2016.06.013
2016-02-24.
齊靜(1990-),女,碩士生,主要從事全局最優(yōu)化方法的研究.