ebook img

Data Mining and Business Intelligence - Artis Consulting PDF

22 Pages·2009·0.34 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Data Mining and Business Intelligence - Artis Consulting

A SPLINE PRODUCT QUASI-INTERPOLATION METHOD FOR WEAKLY SINGULAR FREDHOLM INTEGRAL EQUATIONS∗ EERO VAINIKKO AND GENNADI VAINIKKO† Abstract. A discrete method of accuracy O(hm) is constructed and justi(cid:28)ed for a class of Fredholm integral equations of the second kind with kernels that may have weak diagonal and boundarysingularities. Themethodisbasedonimprovingtheboundarybehaviourofthekernel withthehelpofachangeofvariables,andontheproductintegrationusingquasi-interpolationby smoothsplinesoforderm. Keywords. Weaklysingularintegralequations,boundarysingularities,splinequasi-interpo- lation,splineinterpolation,productintegrationmethods,Nystr(cid:246)mtypemethods AMS subject classi(cid:28)cations. 65R20,65D07,65D30 1. Introduction. In the present paper we call attention to some high order fully discrete methods for a class of Fredholm integral equations of the second kind withdi(cid:27)erentweaksingularities. Theideaofthemethodisclosetothosein[2],[4], [9]. Consider the weakly singular integral equation Z 1 (1.1) u(x)= (cid:0)a(x,y)|x−y|−ν +b(x,y)(cid:1)u(y)dy+f(x), 0≤x≤1, 0 dependingontheparameterν,0<ν <1. Atraditionalproductintegrationmethod (see [2], [4] and the literature quoted there) is based on the approximation of (1.1) by the equation Z 1 u(x)= (cid:0)|x−y|−νP (a(x,y)u(y))+P (b(x,y)u(y))(cid:1)dy+f(x), 0≤x≤1, n,m n,m 0 where P is an interpolation projector onto the space of piecewise polynomial n,m functions of degree m−1 associated with a subdivision of [0,1] into n subinter- vals by some knots 0 = x < x < ... < x = 1; P is applied to the 0,n 1,n n,n n,m products a(x,y)u(y) and b(x,y)u(y) as functions of y treating x as a parameter. This results to some Nystr(cid:246)m type methods of accuracy O(n−m) provided that the grid {x , x ,..., x } is properly graded to compensate the generic boundary 0,n 1,n n,n singularities of the derivatives of the exact solution to (1.1); even the accuracy O(n−m−1+ν) can be achieved for a skilled choice of the interpolation points, see [13]. The method of work [9] is di(cid:27)erent: the authors (cid:28)rst perform a change of variables which improves the boundary behaviour of the solution and of the coef- (cid:28)cient functions a(x,y), b(x,y), and after that, using a polynomial approximation of the solution, they apply a Gauss type quadrature formula for the integral in the transformed equation. In our method we follow [9], [10] performing a smoothing change of variables and then, instead of discontinuous piecewise polynomial interpolation by P , we n,m useaproductquasi-interpolation(oreventherealinterpolation)bysmoothsplines ofdegreem−1ontheuniformgridofthestepsizeh=1/n. Thisleadstosomere- ductionofdegreesoffreedomcomparedwiththeuseofpiecewisepolynomials. Our algorithm still has features of the Nystr(cid:246)m method. In Nystr(cid:246)m type methods for weakly singular integral equations, the computation of the quadrature coe(cid:30)cients ∗ThisworkwassupportedbytheEstonianInformationTechnologyFoundation,grantNo07- 03-01-01,andbytheEstonianScienceFoundation,grantNo7353. †Faculty of Mathematics and Computer Science, University of Tartu, Liivi 2, Tartu 50409, Estonia([email protected]@ut.ee) 1 2 E.VAINIKKOANDG.VAINIKKO is a laborious and delicate part of the algorithms. The use of quasi-interpolation (orrealinterpolation)bysmoothsplinesonuniformgridsenablesustoobtainsim- ple but still somewhat delicate formulae for the quadrature coe(cid:30)cients, and so we obtain a fully discrete method of accuracy O(hm). This is the main message of the present paper. In our considerations, the coe(cid:30)cient functions a(x,y) and b(x,y) may have boundary singularities with respect to y. Denote by T the integral operator of equation (1.1), Z 1 (Tu)(x)= (cid:0)a(x,y)|x−y|−ν +b(x,y)(cid:1)u(y)dy. 0 Lemma 1.1. Assume that a,b∈C([0,1]×(0,1)) satisfy the inequalities |a(x,y)|≤cy−λ0(1−y)−λ1, |b(x,y)|≤cy−µ0(1−y)−µ1, (x,y)∈[0,1]×(0,1), where 0<ν <1, λ ,λ ,µ ,µ ∈R, λ +ν <1, λ +ν <1, µ <1, µ <1. Then 0 1 0 1 0 1 0 1 T maps C[0,1] into C[0,1], and T : C[0,1]→C[0,1] is compact. The proof is standard, cf. [5], [7]; a detailed argument can be found in [11]. For m ∈ N, θ0,θ1 ∈ R, θ0,θ1 < 1, denote by C?m(0,1) and Cm,θ0,θ1(0,1) the weighted spaces of functions u∈C[0,1]∩Cm(0,1) such that, respectively, m X kuk := sup xk(1−x)k|u(k)(x)|<∞, Cm(0,1) ? 0<x<1 k=0 m X kuk := sup w (x)w (1−x)|u(k)(x)|<∞ Cm,θ0,θ1(0,1) k−1+θ0 k−1+θ1 0<x<1 k=0 where  1, ρ<0,  w (r)= 1/(1+|logr|), ρ=0, r,ρ∈R, r >0. ρ  rρ, ρ>0, Clearly, Cm[0,1]⊂Cm,θ0,θ1(0,1)⊂Cm(0,1). ? Lemma 1.2. ([11]). Let a,b∈Cm([0,1]×(0,1)) and (1.2)|∂k∂la(x,y)|≤cy−λ0−l(1−y)−λ1−l, |∂k∂lb(x,y)|≤cy−µ0−l(1−y)−µ1−l, x y x y (x,y)∈[0,1]×(0,1), k+l≤m, where m ∈ N, 0 < ν < 1, λ +ν < 1, λ +ν < 1, µ < 1, µ < 1. Then T maps 0 1 0 1 Cm(0,1) into Cm(0,1), T : Cm(0,1) → Cm(0,1) is bounded and T2 : Cm(0,1) → ? ? ? ? ? C?m(0,1) is compact. Further, T maps Cm,θ0,θ1(0,1) with θ0 = max{λ0 +ν,µ0}, θ1 = max{λ1+ν,µ1} into Cm,θ0,θ1(0,1) and T : Cm,θ0,θ1(0,1) → Cm,θ0,θ1(0,1) is compact. Here ∂k∂l = (∂/∂x)k(∂/∂y)l. An example of a satisfying (1.2) is given by x y a(x,y)=y−λ0(1−y)−λ1a(x,y) where a∈Cm([0,1]×[0,1]). Denote N(I − T) = {u ∈ C[0,1] : u = Tu}. The following theorem is a consequence of Lemmas 1.1 and 1.2. Theorem 1.3. Assume the conditions of Lemma 1.2 and N(I −T) = {0}. Then for f ∈Cm(0,1), equation (1.1) has a solution u∈Cm(0,1) which is unique ? ? in C[0,1], and k u k ≤ c k f k where the constant c is independent Cm(0,1) Cm(0,1) ? ? of f. Further, if f ∈ Cm,θ0,θ1(0,1), θ0 = max{λ0 +ν,µ0}, θ1 = max{λ1 +ν,µ1} PRODUCTQUASI-INTERPOLATIONMETHOD 3 then u ∈ Cm,θ0,θ1(0,1), and k u kCm,θ0,θ1(0,1)≤ cm,θ0,θ1 k f kCm,θ0,θ1(0,1) where the constant c is independent of f. m,θ0,θ1 The main results of the paper are established under assumptions of Theorem 1.3. Using the approach of [12], our treatment can be extended to the case where f has a singularity at a point x = ξ, ξ ∈ (0,1), and/or a and b have certain singularities on the lines x=ξ and y =ξ or on a system of such inner points/lines (cid:21) we can build the approximate solution on [0,ξ] and [ξ,1] in a similar manner as this is done on [0,1] in the Section 4. The rest of the paper is organised as follows. In Section 2 we recall some results about the interpolation and quasi-interpolation of functions by splines with auniformknotsetand, inparticular, abouttheoptimalityofsuchapproximations. InSection3wetransformequation(1.1)intoanewformbythechangeofvariables. Section 4 is central in the paper (cid:21) we introduce the product quasi-interpolation method for the transformed equation, examine its accuracy and present the matrix form of the method. In Section 5 we discuss di(cid:27)erent computational details of the method and in Section 6 we test the algorithms on a numerical example. Actually the method can be applied to somewhat more general class of equa- tions rather than (1.1). The case of integral equation with logarithmic diagonal singularity of the kernel will be treated in a separate work with more attention to the computational environment. In [17] the method is modi(cid:28)ed for the Volterra integral equation corresponding to (1.1). 2. Approximation tools. 2.1. The father B-spline. Thefather B-spline B of order minthetermi- m nology of [3], [15], or of degree m−1 in the terminology of [6], [16], can be de(cid:28)ned by the formula m (2.1) B (x)= 1 X(−1)iC(m)(x−i)m−1, x∈R, m∈N. m (m−1)! i + i=0 Here, as usual, 0!=1, 00 :=lim xx =1, x↓0 (cid:18) m (cid:19) m! (cid:26) (x−i)m−1, x−i≥0, C(m) = = , (x−i)m−1 = i i i!(m−i)! + 0, x−i<0. Let us recall some properties of B : B ∈C(m−2)(R), m m (2.2) suppB =[0,m], B (x)=B (m−x)>0 for 0<x<m, m m m (2.3) B(m−1)(x)=(−1)lC(m−1) for l<x<l+1, l=0,...,m−1, m l Z X (2.4) B (x−j)=1 for x∈R, B (x)dx=1. m m j∈Z R 2.2. Spline interpolation on the uniform grid in R. Introduce in R the uniform grid hZ={ih: i∈Z} of the step size h>0. Denote by S , m∈N, the h,m space of splines of order m (of degree m−1) and defect 1 with the knot set hZ. The family B-spline B (h−1x−j), j ∈ Z, belongs to S , and the same is true m h,m forP d B (h−1x−j)witharbitrarycoe(cid:30)cientsd ; therearenoproblemswith j∈Z j m j the convergence of the series since it is locally (cid:28)nite: it follows from (2.2) that 4 E.VAINIKKOANDG.VAINIKKO i X X d B (h−1x−j)= d B (h−1x−j) for x∈[ih,(i+1)h), i∈Z. j m j m j∈Z j=i−m+1 Given a function f ∈ C(R), bounded or of at most polynomial growth as |x|→∞, we determine the interpolant Q f∈S by the conditions h,m h,m X (2.5) (Q f)(x)= d B (h−1x−j), x∈R, h,m j m j∈Z m m (2.6) (Q f)((k+ )h)=f((k+ )h), k ∈Z. h,m 2 2 Form=1andm=2,Q f istheusualpiecewiseconstant,respectively,piecewise h,m linearinterpolantwhichcanbedeterminedoneverysubinterval[ih,(i+1)h),i∈Z, independentlyofothersubintervals. Form≥3,thevalueofQ f atagivenpoint h,m x ∈ R depends on the values of f at all interpolation knots (k+ m)h, k ∈ Z. It 2 occurs (see [14], [16]) that for m ≥ 3 conditions (2.5)(cid:21)(2.6) really determine d , j j ∈Z, uniquely in the space of bounded or polynomially growing bisequences (d ), j namely, X m X m (2.7) d = α f((k+ )h)= α f((j−k+ )h), j ∈Z, j j−k,m 2 k,m 2 k∈Z k∈Z where (2.8) α =Xm0 zlm,m0−1 z|k|, k ∈Z, m =(cid:26) (m−2)/2 if m is even, k,m P0 (z ) l,m 0 (m−1)/2 if m is odd, l=1 m l,m and z ∈(−1,0), l=1,...,m , are the roots of the characteristic polynomial l,m 0 X m P (z)= B (k+ )zk+m0 m m 2 |k|≤m0 (it is a polynomial of degree 2m ); it occurs that P has exactly m simple roots 0 m 0 z , l = 1,...,m , in the interval (−1,0), and the remaining m roots are of the l,m 0 0 form z =1/z ∈(−∞,−1), l=1,...,m . l+m0,m l,m 0 Denote by BC(R) the space of bounded continuous functions on R equipped with the norm k f k∞= supx∈R|f(x)|, by Vm,∞(R) the space of functions having bounded mth (distributional) derivative in R, by Wm,∞(R) the standard Sobolev spaceoffunctionsonRhavingboundedderivativesoforder≤m,andbyWm,∞(R) (0,1) the space of functions f ∈Wm,∞(R) with supports in (0,1). Lemma 2.1. ([6,21,22]). For f ∈Vm,∞(R), it holds f−Q f ∈BC(R) and h,m (2.9) kf −Q f k ≤ Φ π−mhm kf(m) k h,m ∞ m+1 ∞ where Φ = 4 P∞ (−1)km , m∈N, is the Favard constant, m π k=0 (2k+1)m Φ <Φ <Φ <...<4/π <...<Φ <Φ <Φ , lim Φ =4/π. 1 3 5 6 4 2 m m→∞ Thefollowinglemmatellsthat,insomesense,thesplineinterpolationyieldsthe bestapproximationofthefunctionclassesWm,∞(R)andVm,∞(R),asymptotically also of Wm,∞(R), compared with other methods that use the same information as (0,1) PRODUCTQUASI-INTERPOLATIONMETHOD 5 Qh,mf (cid:21) the values f|Zh,m of f on the grid Zh,m ={(j+ m2)h: j ∈Z}. Denote by C(Z ) the vector space of all (grid) functions de(cid:28)ned on Z . h,m h,m Lemma 2.2. ([21,22]). For given γ >0, we have by Lemma 2.1 sup kf −Q f k ≤Φ π−mhmγ, h,m ∞ m+1 f∈Vm,∞(R),kf(m)k∞≤γ whereas for any mapping M : C(Z ) → C(R) (linear or nonlinear, continuous h h,m or discontinuous), it holds sup kf −Mh(f|Zh,m)k∞≥Φm+1π−mhmγ, f∈Wm,∞(R),kf(m)k∞≤γ lihm→i0nff∈W(m0,,1∞)(Rs)u,pkf(m)k∞≤γ kf −Mh(f|Zh,m)k∞ /(Φm+1π−mhmγ)≥1. 2.3. Quasi-interpolation by splines. Let m ≥ 3. In a quasi-interpolant Q(p) f, the in(cid:28)nite sum (2.7) de(cid:28)ning the coe(cid:30)cients d of the spline interpolant h,m j (2.5)(cid:21)(2.6) is replaced by a (cid:28)nite sum: (2.10) (Q(p) f)(x)=Xd(p)B (h−1x−j), x∈R, h,m j m j∈Z (2.11) d(p) = X α(p) f((j−k+ m)h), p∈N. j k,m 2 |k|≤p−1 A simple truncation of the series in (2.7) does not give acceptable results. Using a special di(cid:27)erence calculus for fast decaying bisequences, the following formulae for α(p) have been proposed in [8]: k,m p−1 (cid:18) (cid:19) (2.12) α(p) = X(−1)k+q 2q γ , |k|≤p−1, k,m k+q q,m q=|k| (2.13) γ =1, γ =Xm0 (1+zl,m)zlm,m0+q−1 , q ≥1. 0,m q,m (1−z )2q+1P0 (z ) l=1 l,m m l,m Then it occurs that (2.14) kQ f −Q(p) f k ≤c h2p kf(2p) k for f ∈V2p,∞(R) h,m h,m ∞ m,p ∞ with a constant c that can be described. A consequence of (2.14) is that for m,p f ∈ Vm,∞(R) with uniformly continuous f(m) and 2p > m, it holds k Q f − h,m Q(p) f k h−m → 0 as h → 0, i.e., the quasi-interpolant Q(p) f is asymptotically h,m ∞ h,m ofthesameaccuracyastheinterpolantQ f. Itisreasonabletotakethesmallest h,m p∈N for which 2p>m; denote it by m , 1 (cid:26) m +1, m even (cid:27) (2.15) m = 2 =m−m . 1 m+1, m odd 0 2 Denote also Q0 :=Q(m1), α0 :=α(m1), |k|<m . Note that (Q0 f)(x) is well h,m h,m k,m k,m 1 h,m de(cid:28)nedforx∈[ih,(i+1)h]withani∈Ziff isgivenon[(i−m+1)h,(i+m)h]∩Z . h,m 6 E.VAINIKKOANDG.VAINIKKO Lemma 2.3. ([8]). For i∈Z, f ∈Cm[(i−m+1)h, (i+m)h], it holds (2.16) max |f(x)−(Q0 f)(x))| h,m ih≤x≤(i+1)h ≤(Φ π−m+q c0 )hm sup |f(m)(x)|. m+1 m m (i−m+1)h≤x≤(i+m)h For a relatively compact set M in C[−δ,1+δ], δ >0, it holds (2.17) sup max |f(x)−(Q0 f)(x))|→0 as h→0. h,m f∈M 0≤x≤1 Formulae for qm :=k Qh,m kBC(R)→BC(R), qm0 :=k Q0h,m kBC(R)→BC(R) and c0m from (2.16) are presented in [8] with the following numerical values: m 3 4 5 6 7 8 9 10 20 c0 0.016 0.019 0.015 0.0085 0.0060 0.0030 0.0022 0.0010 6.5e-6 m q 1.414 1.549 1.706 1.816 1.916 2.000 2.075 2.142 2.583 m q c0 0.023 0.029 0.026 0.015 0.012 0.006 0.005 0.002 1.7e-5 m m q0 1.250 1.354 1.329 1.403 1.356 1.413 1.378 1.419 1.514 m Relatively small values of the norms q and q0 tell us about good stability prop- m m erties of interpolation and quasi-interpolation processes. We assumed that m≥3. For m=1 and m=2, we may put Q0 =Q . h,m h,m 3. A smoothing change of variables. In the integral equation (1), we per- form the change of variables (3.1) x=ϕ(t), y =ϕ(s), 0≤t≤1, 0≤s≤1, where ϕ: [0,1]→[0,1] is de(cid:28)ned by the formula 1 Z t (3.2) ϕ(t)= σr0−1(1−σ)r1−1dσ, c ? 0 Z 1 Γ(r )Γ(r ) c = σr0−1(1−σ)r1−1dσ = 0 1 , ? Γ(r +r ) 0 0 1 Γ is the Euler gamma function. We assume that r ,r ∈R, r ≥1, r ≥1, but we 0 1 0 1 keep in mind that practicable algorithms correspond to the case where at least one of parameters r and r is natural. If r ∈N, (3.2) takes the form 0 1 1 ϕ(t)= 1 tr0rX1−1 (−1)kC(r1−1)tk, 0≤t≤1, c = (r1−1)! , c r +k k ? r (r +1)...(r +r −1) ? 0 0 0 0 1 k=0 butin ourcomputations, thefollowing expansionof ϕ withpositiveterms occurred to be preferable (essentially more stable numerically): (3.3) ϕ(t)=tr0"1+rX1−1r0(r0+1)...(r0+k−1)(1−t)k#, 0≤t≤1. k! k=1 If r ,r ∈N, the integral in (3.2) can be alternatively evaluated [9] in a stable way 0 1 by an exact Gauss rule, since the integrand is a polynomial of degree r +r −2. 0 1 PRODUCTQUASI-INTERPOLATIONMETHOD 7 Clearly, ϕ(0) = 0, ϕ(1) = 1 and ϕ(t) is strictly increasing in [0,1]. Hence ϕ(t)−ϕ(s) > 0, |ϕ(t)−ϕ(s)| = ϕ(t)−ϕ(s)|t−s| for s 6= t, and equation (1.1) takes t−s t−s with respect to v(t)=u(ϕ(t)) the form Z 1 (3.4) v(t)= (cid:0)A(t,s)|t−s|−ν +B(t,s)(cid:1)v(s)ds+g(t), 0≤t≤1, 0 where (cid:26) ϕ(t)−ϕ(s), t6=s, (3.5) A(t,s)=a(ϕ(t),ϕ(s))Φ(t,s)−νϕ0(s), Φ(t,s)= t−s ϕ0(s), t=s, (3.6) B(t,s)=b(ϕ(t),ϕ(s))ϕ0(s), g(t)=f(ϕ(t)). Let us characterise the boundary behaviour of functions in equation (3.4). Clearly, 0≤ϕ(t)≤ctr0, 0≤1−ϕ(t)≤c(1−t)r1, (3.7) |ϕ(k)(t)|≤ctr0−k(1−t)r1−k, 0<t<1, k =1,...m. Lemma 3.1. For u∈Cm(0,1), v(t)=u(ϕ(t)), it holds v ∈Cm(0,1) and ? ? (3.8) kv k ≤ckuk Cm(0,1) Cm(0,1) ? ? where the constant c is independent of u. Further, if u ∈ Cm,θ0,θ1(0,1), θ0 < 1, θ <1, then for j =1,...,m, 0<t<1, 1 (3.9) |v(j)(t)|≤ckuk × Cm,θ0,θ1(0,1)  tr0−j, θ0 <0  (1−t)r1−j, θ1 <0  tr0−j(1+|logt|), θ0 =0 (1−t)r1−j(1+|log(1−t)|), θ1 =0 .  t(1−θ0)r0−j, θ0 >0  (1−t)(1−θ1)r1−j, θ1 >0  Proof. Clearly k v k =k u k . The proof of (3.8) and (3.9) is based on (3.7) ∞ ∞ and the formula of Fa(cid:224) di Bruno (cid:18)d(cid:19)j X j! (cid:18)ϕ0(t)(cid:19)k1 (cid:18)ϕ(j)(t)(cid:19)kj u(ϕ(t))= u(k1+...+kj)(ϕ(t)) ... dt k !..k ! 1! j! 1 j k1+...+jkj=j where the sum is taken over all k ≥0,...,k ≥0 such that k +2k +...+jk =j. 1 j 1 2 j Here are the details for (3.9). For 0<t≤ 1, 1≤j ≤m, we have 2   1, k +...+k <1−θ X  1 j 0  |v(j)(t)|≤c 1+|logt|, k +...+k =1−θ × 1 j 0 k1+2k2...+jkj=j t(1−θ0−k1−...−kj)r0, k1+...+kj >1−θ0  ×kuk tk1(r0−1)tk2(r0−2)...tkj(r0−j) Cm,θ0,θ1 X  t(k1+...+kj)r0−j, k1+...+kj <1−θ0  =c t(1−θ0)r0−j(1+|logt|), k1+...+kj =1−θ0 kukCm,θ0,θ1 . k1+...+jkj=j t(1−θ0)r0−j, k1+...+kj >1−θ0  Theconditionk +2k ...+jk =jimpliesthefollowing. (i)Therelationk +...+k < 1 2 j 1 j 1−θ0 is possible only if θ0 < 0, and then tr0−j is the leading term as t → 0 in 8 E.VAINIKKOANDG.VAINIKKO the last sum (it corresponds to k = ... = k = 0, k = 1). (ii) The relation 1 j−1 j k +...+k = 1−θ is possible only if θ ≤ 0, θ ∈ Z; in case θ = 0, the leading 1 j 0 0 0 0 term is tr0−j(1+|logt|) whereas in case θ0 ≤ −1, similarly as in (i), the leading term is tr0−j . (iii) If 0 < θ0 < 1 then all terms in the last sum are t(1−θ0)r0−j. Theseconsiderationsprove(3.9)for0<t≤ 1. Byasymmetryargumentweobtain 2 (3.9) also for 1 ≤t<1. 2 Lemma 3.2. Let a and b satisfy the conditions of Lemma 1.2. Then for j =0,...,m, 0≤t≤1, 0<s<1, it holds (3.10) |∂ja(ϕ(t),ϕ(s))|≤cs−r0λ0−j(1−s)−r1λ1−j, s (3.11) |∂jb(ϕ(t),ϕ(s))|≤cs−r0µ0−j(1−s)−r1µ1−j. s These inequalities elementarily follow by the formula of Fa(cid:224) di Bruno. The function Φ(t,s)−ν has singularities at (0,0) and (1,1), the only zeroes of Φ(t,s) in [0,1]×[0,1]. It is easy to see that ∂kΦ(t,s)(cid:16)(t+s)r0−k−1((1−t)+(1−s))r1−k−1 as t,s→0 or as t,s→1 s that together with the formula of Fa(cid:224) di Bruno implies the following result. Lemma 3.3. For j =0,...,m, 0≤t≤1, 0<s<1, it holds (cid:12) (cid:16) (cid:17)(cid:12) (3.12) (cid:12)∂j (Φ(t,s))−ν (cid:12)≤c(t+s)−ν(r0−1)−j((1−t)+(1−s))−ν(r1−1)−j. (cid:12) s (cid:12) Next we present estimates for functions A(t,s), B(t,s) and ∂m[A(t,s)v(s)], s ∂m[B(t,s)v(s)] in somewhat speci(cid:28)c form for the needs of Section 4. s Corollary 3.4. Let a and b satisfy the conditions of Lemma 1.1. Then the following holds true: (i) if (3.13) r ,r ≥1, r >(1−ν)/(1−ν−λ ), r >(1−ν)/(1−ν−λ ), 0 1 0 0 1 1 then with δ :=(1−ν−λ )r −(1−ν)>0, δ :=(1−ν−λ )r −(1−ν)>0, it 0 0 0 1 1 1 holds (3.14) |A(t,s)|≤csδ0(1−s)δ1, (t,s)∈[0,1]×(0,1); (ii) if (3.15) r ,r ≥1, r >1/(1−µ ), r >1/(1−µ ), 0 1 0 0 1 1 then with δ :=(1−µ )r −1>0, δ :=(1−µ )r −1>0, it holds 0 0 0 1 1 1 (3.16) |B(t,s)|≤csδ0(1−s)δ1, (t,s)∈[0,1]×(0,1). Proof. These estimates are direct consequences of Lemmas 3.2 and 3.3. By (3.7), (3.10) and (3.12) we have |A(t,s)|≤cs−r0λ0+r0−1(t+s)−ν(r0−1)(1−s)−r1λ1+r1−1(2−t−s)−ν(r1−1) that for r , r satisfying (3.13) yields (3.14). Similarly, by (3.7), (3.11) and (3.12), 0 1 |B(t,s)|≤cs−r0µ0+r0−1(1−s)−r1µ1+r1−1 that for r , r satisfying (3.15) yields (3.16). 0 1 PRODUCTQUASI-INTERPOLATIONMETHOD 9 Corollary 3.5. Let a and b satisfy the conditions of Lemma 1.2, and let u ∈ Cm(0,1), v(t) = u(ϕ(t)). Then the following estimates hold true for (t,s) ∈ ? [0,1]×(0,1): (i) if (3.17) r ,r ≥1, r >m/(1−ν−λ ), r >m/(1−ν−λ ), 0 1 0 0 1 1 then with δ :=(1−ν−λ )r −m>0, δ :=(1−ν−λ )r −m>0, 0 0 0 1 1 1 (3.18) |A(t,s)|≤csm−(1−ν)+δ0(1−s)m−(1−ν)+δ1, (3.19) |∂sm[A(t,s)v(s)]|≤cs−(1−ν)+δ0(1−s)−(1−ν)+δ1 kukC?m(0,1); (ii) if (3.20) r ,r ≥1, r >m/(1−µ ), r >m/(1−µ ), 0 1 0 0 1 1 then with δ :=(1−µ )r −m>0, δ :=(1−µ )r −m>0, 0 0 0 1 1 1 (3.21) |B(t,s)|≤csm−1+δ0(1−s)m−1+δ1, (3.22) |∂sm[B(t,s)v(s)]|≤cs−1+δ0(1−s)−1+δ1 kukC?m(0,1); (iii) if (3.23) r ,r ≥1, r >(m+ν)/(1−λ ), r >(m+ν)/(1−λ ) 0 1 0 0 1 1 then with δ :=(1−λ )r −m−ν >0, δ :=(1−λ )r −m−ν >0, 0 0 0 1 1 1 (3.24) ϕ0(t)ν|A(t,s)|≤csm−(1−ν)+δ0(1−s)m−(1−ν)+δ1, (3.25) ϕ0(t)ν|∂sm[A(t,s)v(s)]|≤cs−(1−ν)+δ0(1−s)−(1−ν)+δ1 kukC?m(0,1) . Proof. These estimates are direct consequences of Lemmas 3.1(cid:21)3.3. Observe that |∂k[ϕ0(t)νΦ(t,s)−ν]|≤c(t+s)−k((1−t)+(1−s))−k. s Let us extend A(t,s) and B(t,s) with respect to s outside (0,1) by the zero value. Underconditions(3.13)and(3.15)weobtaincontinuousfunctionson[0,1]× R, see (3.14) and (3.16). 4. The product quasi-interpolation method. 4.1. The operator form of the method. Let h = 1/n, n ∈ N, n ≥ m. For 0 ≤ s ≤ 1, the quasi-interpolant Q0 w of a function w∈ C[−δ,1+δ] has the h,m expansion n−1 X X m (4.1) (Q0 w)(s)= α0 w((j−p+ )h)B (ns−j). h,m p,m 2 m j=−m+1|p|<m1 We approximate equation (3.4) by its discretised version: for 0≤t≤1, Z 1 (4.2) v (t)= [|t−s|−νQ0 (A(t,s)v (s))+Q0 (B(t,s)v (s))]ds+g(t) n h,m n h,m n 0 where Q0 is applied to the products A(t,s)v(s) and B(t,s)v(s) as functions of s. h,m This can be done for v given on [0,1] since A(t,s) = B(t,s) = 0 for s ≤ 0 and for s≥1; recall that under condition (3.13) and (3.15), A,B ∈C([0,1]×R). 10 E.VAINIKKOANDG.VAINIKKO Theorem4.1. (i)LetaandbsatisfytheconditionsofLemma1.1,N(I−T)= {0}, f ∈ C[0,1], and let r and r satisfy condition (3.13) and (3.15). Then for 0 1 su(cid:30)ciently large n, equation (4.2) has a unique solution v , and n (4.3) kv−v k := max |v(t)−v (t)|→0 as n→∞ n ∞ n 0≤t≤1 where v is the solution of equation (3.4). (ii) Under assumptions of Theorem 1.3, f ∈ Cm(0,1) and conditions (3.17) ? and (3.20) on r and r , it holds 0 1 (4.4) kv−v k ≤chm kf k . n ∞ Cm(0,1) ? (iii) Under assumptions of Theorem 1.3, f ∈ Cm(0,1) and conditions (3.20) ? and (3.23) on r and r , it holds 0 1 (4.5) max ϕ0(t)ν|v(t)−v (t)|≤chm kf k . n Cm(0,1) 0≤t≤1 ? The constant c in (4.4) and (4.5) is independent of n and f. Proof. Accept the assumptions formulated in (i). Denote by T and T the n integral operators of equations (3.4) and (4.2), Z 1 (Tv)(t)= [|t−s|−νA(t,s)+B(t,s)]v(s)ds, 0 Z 1 (T v)(t)= [|t−s|−νQ0 (A(t,s)v(s))+Q0 (B(t,s)v(s))]ds. n h,m h,m 0 We claim that T →T compactly in C[0,1] as n→∞, i.e., n (4.6) kT v−Tv k →0 for every v ∈C[0,1], n ∞ (4.7) (v )⊂C[0,1], kv k ≤1 ⇒ (T v ) is relatively compact in C[0,1]. n n ∞ n n Indeed,thesets{A(t,·): 0≤t≤1}and{B(t,·): 0≤t≤1}arerelativelycompact in C[−δ,1+δ], and by Lemma 2.3, for a (cid:28)xed v ∈ C[0,1] extended by v(t) = v(0) for −δ ≤s≤0, v(t)=v(1) for 1≤s≤1+δ, it holds sup max |A(t,s)v(s)−Q0 (A(t,s)v(s))|→0 for n→∞, h,m 0≤t≤1 0≤s≤1 sup max |B(t,s)v(s)−Q0 (B(t,s)v(s))|→0 for n→∞. h,m 0≤t≤1 0≤s≤1 This together with the equality kQ0 k=q0 implies (4.6). The proof of (4.7) can h,m m be built using the Arzela theorem. Due to the condition N(I −T) = {0}, also N(I −T) = {0}. As well known (see [1], [2], [7] or [18]), relations (4.6), (4.7) and N(I −T) = {0} imply that, for su(cid:30)cientlylargen,theoperatorsI−T areinvertibleandtheinversesareuniformly n bounded: (4.8) k(I−T )−1 k ≤c, n≥n . n C[0,1]→C[0,1] 0 Let v and v be the solutions of equations (3.4) and (4.2), respectively. Then n (4.9) v−v =(I−T )−1(Tv−T v), n n n

Description:
Data Mining and Business Intelligence Amit Basu, PhD Cox School of Business SMU
See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.