节点文献

几类反问题的偏差原则及收敛阶

Discrepancy Principle and Orders of Convergence for Inverse Problems

【作者】 王光明

【导师】 张玉海;

【作者基本信息】 山东大学 , 应用数学, 2008, 硕士

【摘要】 反问题的一个特别重要的属性就是它通常是“不适定”的数学问题,使得它无论在进行理论分析还是在进行数值计算时都有特定的困难。其算子形式为:求解这类问题的普遍方法是正则化方法:用一族与原不适定问题相“邻近”的适定问题的解去逼近原问题的解,如何建立有效的正则化方法是反问题领域中不适定问题研究的重要内容.其中Tikllonov提出求解不适定问题的变分正则化方法包括解如下问题F(u):=||Au-fδ||2+a||u||2=min,其中a是大于0的常数,称为正则化参数.如何有效地选取正则化参数是此方法的关键,偏差原则是选取正则化参数的一类重要方法.首先,本文讨论了基于一类动力系统方法的偏差原则及收敛阶.假设(1)是Hilbert空间中的可解方程,||A||<∞并且R(A)是非闭的,所以问题(1)是不适定的.可解方程(1)等价于Bu=A*f,B:=A*A令(?)(t)是一个单调递减函数,满足(?)(t)>0,limt→0(?)(t)=0,limt→∞supt/2≤s≤t|(?)|(?)-2(t)=0.解决(1)的动力系统方法,包含解下面的Cauchy问题:此处u0是任意的.而且可以证明对任意的u0:上述问题对所有大于零的t都有唯一解.存在y:=u(∞):=limt→∞u(t),Ay=f.并且y是(1)的唯一极小范数解.这些结论在[3]中被证明.如果给定fδ,使得||f-fδ||≤δ,那么在f被fδ代替情况下,定义uδ(t)为上述问题的解.定义tδ为停止时刻,tδ满足limδ→0||uδ(tδ)-y||=0且limδ→0tδ=∞.假设fδ⊥N(A*),基于此类动力系统方法的偏差原则如下:定理1如果A是Hilbert空间的有界线性算子,方程(1)可解,||fδ||>δ,fδ⊥N(A*),并且c(t)是一个单调递减函数,满足那么方程有唯一的解tδ,并且成立,此处y是(1)的唯一极小范数解.进一步地,对由该偏差原则得到的解收敛性作出如下估计:定理2如果A是Hilbert空间的有界线性算子,方程(1)可解,y是(1)的唯一极小范数解,uδ,(?)(tσ)由动力系统方程解得.那么令y=A*z,||z||≤E,可以选取(?)(tδ),使得本文第二部分讨论了求解不适定问题的几类动力系统方法及第二种偏差原则:(1)自伴算子的动力系统方法关于自伴算子的DSM方法可如下构造.考虑方程其中,a是大于0的常数.第一个结果由定理3给出:定理3假设(1)中的A是自伴算子,即A=A*.如果Ay=f且y⊥V,那么第二个结果基于定理3,假设||fδ-f||≤δ,则可以得到方程(1)的稳定解ua,δ(t)定理4存在t=tδ,limδ→0tδ=∞,且a=a(δ),limδ→0a(δ)=0,使得uδ:=ua(δ).δ(tδ)满足在定理3和定理4证明之后,我们给出a(δ),tδ的选取方法.(2)改进的自伴算子的动力系统方法我们重新考察一下将其中的a换为a(t).定理5假设a(t)>0是单调递减的连续函数,且满足此问题的解为那么进一步地,定理5可得到厅程(1)的稳定解.定理6存在一个停止时刻tδ,limδ→0tδ=∞,使得其中,uδ=uδ(tδ)是的解,其中||fδ-f||≤δ.(3)动力系统方法的第二种偏差原则在前面使用的动力系统方法的偏差原则中,我们假设fδ⊥N(A*).如果此假设不成立,我们可使用如下的偏差原则:其中,1<C<2为一常数.定理7如果其中f=Ay,y⊥N,a(t)>0是单调递减连续函数,使得那么DSM方法u=-u+Ta(t)-1A*f,u(0)=0,可得到方程(1)的稳定解,陈述如下:定理8如果选取tδ使得成立,那么的解满足定理9假设a(t)>0是单调递减二阶可导函数,且满足:方程其中,1<C<2为一常数.方程有解t=tδ,limδ→0tδ=∞,使得:其中,uδ是的解,||fδ||>Cδ.本文第三部分计论了一种新的偏差原则及收敛阶:定理10假设A是Hilbcrt空间H中线性有界算子,f=Ay,y⊥N,||fδ-f||≤δ,||fδ||≥Cδ,C是常数,C∈(1,2),且ua,δ满足不等式:其中b是大于0的常数,且C2>1+b.那么对任意给定的δ>0,方程有解ua,δ满足以及定理10可叙述为:定理10’假定iA足Hilbcrt空间的有界线性算子,ii)方程Au=f可解,y是极小范数解,iii)||fδ-f||≤δ,||fδ||≥Cδ,此处C为大于1的常数.那么a)方程对任意δ>0可解,其中uδ.(?)满足不等式F(uδ.(?))≤m+(C2-1-b)δ2,F(u):=||A(u)-fδ||2+(?)||u||2,m=m(δ.(?)):=infuF(u),b为大于0的常数,C2>1+b,b)如果(?)=c(δ)是(3)的解,记uδ:=uδ.c(δ),那么limδ→0||uδ-y||=0.在此基础上,对由该偏差原则得到的解收敛性作出估计.进一步得到;定理11假定i)A是Hilbert空间的有界线性算子,ii)方程Au=f可解,y是极小范数解,iii)||fδ-f||≤δ,||fδ||≥Cδ.此处C为大于1的常数.那么c)如果c=c(δ)是(3)的解,记uδ:=uδ.c(δ),那么||uδ-y||=O(δ1/2).d)若||uδ-y||=o(δ1/2),则R(A)为有限维,即收敛阶O(δ1/2)是最优的.

【Abstract】 A character of Inverse Problem is that it is an ill-posed problem usually. Operator formula:Method used usually of solving ill-posed problems is regularization strategy:a family of well-posed problem converge ill-posed problem. Tikhonov Variational regularization method for stable solution of (1) consists of solving the problem F(u) := ||Au-fδ||2+ a||u||2=min,where a is regularization parameters > 0,α= const. An important method for choosing regularization parameter is discrepancy principle.Firstly.this paper study a discrepancy principle and orders of convergence based Dynamical Systems Method (DSM).Assume (l)is a solvable linear equation in a Hilbert space. ||A||<∞, and R(A) is not closed, so problem (1) is ill-posed. A solvable quation (1) is equivalent toBu = A*f,B :=A*A. Let (?)(t) be a monotone, decreasing function,(?)(t)>0,limt→0(?)(t)= 0,limt→∞sup1/2≤s≤t|(?)|(?)-2(t) = 0.A DSM (dynamical systems method) for solving (1), consists of solving the followingCauchy problem:where u0 is arbitrary, and proving that, for any u0.Cauchy problem has a unique solution for all t>0, there exists y:=u(∞) := lirnt→∞u(t), Ay =f, and y is the unique minimal-norm solution to (1). These results are proved in [3]. If fδis given. such that ||f-fδ||≤δ, then uδ(t) is denned as the solution to Cauchy problem with f replaced by fg. The stopping time is defined as a number tδ为such that limδ→0||uδ(tδ)-y||=0,and limδ→0tδ=∞.Let us assume fδ⊥N(A*). Then this discrepancy principle:theorem 1 If A is a bounded linear operator in a Hilbert space H, equation (1) is solvable,||fδ||>δ,fδ⊥N(A*) , and (?)(t) satisfies the assumptions stated above, then equationhas a unique solution tδ, andholds, where y is the unique minimalnorm solution to (1).Furthermore,we estimate the convergence of solution for the discrepancy principle: theorem 2 If A is a bounded linear operator in a Hilbert space H, equation (1)is solvable, uδ,∈(tδl) from DSMobtained.thenLety=A*z,||z||≤E , can choose t(tδ),such thatThe second part of this paper study some DSM solving ill-posed problem and another discrepancy principle:(1)DSM of selfadjoint operatorThe DSM for solving equation (1) with a linear selfadjoint operator can be constructedas follows. Consider the problemwhere a=const > 0. Our first result is formulated as Theorem 3.theorem 3 If Ay = f and y⊥N , then Our second result shows that the method, based on Theorem 1, gives a stable solution of the equation Au = f. Assume that ||fδ-f||≤δ, and let ua,δ(t)be the solution towith fs in place of f.theorem 4 There exist t = tδ, limδ→0tδ=∞:and a=a(δ),limδ→0a(δ)=0 , such that uδ:= ua(δ),δ(tδ) satisfiesWe will discuss the ways to choose a(δ) and tδafter the proofs of these theorems are given.(2)Improved DSM of selfadjoint operatorConsider problem u=i(A + ia)uα-if,u(0)=0;u =((du)/(dt)), with a=a(t).theorem 5 Assume that a(t) > 0 is a continuous function monotonically decayingto zero as t→∞andThe solution to this problem isthenFurtlicrmorc,Thoorem 5 yields a stable solution to equation (1).theorem 6 There exists a stopping time tδ,limδ→0tδ=∞, such thathold , where uδ=uδ(tδ) solving||fδ-f||≤δ.(3) Another discrepancy principle based DSM Suppose that the assumption fδ⊥N(A*) does not hold. Then one can use the discrepancy principle of the form:theorem 7 Ifwhere f = Ay,y⊥N ,that a(t) > 0 is a continuous function monotonically decaying to zero as t→∞andthenDSM method u = -u + Ta(t)-1A*f, u(0) = 0, yields a stable solution to problem (1).theorem 8 If tδis chosen so thatholds,then the solution ofsatisfiestheorem 9 Assume that a(t)>0 is a monotonically decaying twice continuously differentiate function.The equationhas a solution t = tδ,limδ→0tδ=∞,such that holds,where uδis the solution to||fδll>Cδ.The third part of this paper study a new discrepancy principle and orders of convergence :theorem 10 Assume that A is a bounded linear operator in a Hilbert space H, that f = Ay,y⊥N,||fδ-f||≤6,||fδ||≥Cδ,C=const,C∈(1.2),and uα,βis any element which satisfies the inequality:whereb=const > 0 , and C2>1+b. Then equationhas a solution for any fixedδ>0 , limδ→0a(δ)=0, andAnother version of theorem 10 :theorem 10 ’ Assume thati)A is a bounded linear operator in a Hilbert spaceⅡ,ii)Thc equation Au = f solvable , y is the unique minimalnorm solution ,iii) ||fδ-f||≤δ,||fδ||≥Cδ.C = const>1.thena)The equationhas a solution tor any fixedδ>0, where uδ,(?) satisfied F(uδ,(?))≤m + (C2-1-b)δ2,F(u) := ||A(u)-fδ||2+∈||u||2,m=m(δ.∈) := infuF(u), b=canst>0,C2>1+b,b) If (?) = (?)(δ) solved ||Auδ,(?)-fδ||=Cδ,uδ:=uδ.(?)(δ),then limδ→0||uδ-y||=0.Furthermore,we estimate the convergence of solution for the discrepancy principle: theorem 11 Assume thati)A is a bounded linear operator in a Hilbert space H,ii)The equation Au = f solvable , y is the unique minimalnorm solution ,iii) ||(?)δ-(?)||≤δ,||(?)δ||≥Cδ,C = const>1.thenc)If t = (?)(δ) solved ||Auδ,(?)-fδ||= Cδ,uδ:= uδ,(?)(δ),then ||uδ-y||=O(δ1/2).d)If ||uδ-y||=o(δ1/2), then R(A) has to be finite dimension, that is, orders of convergence O(δ1/2)is optimal.

  • 【网络出版投稿人】 山东大学
  • 【网络出版年期】2009年 01期
节点文献中: 

本文链接的文献网络图示:

本文的引文网络