Sie sind auf Seite 1von 16
6.5 Integral Transforms 383 6.5 Integral Transforms ‘The eigenfunction expansion method is, for the most part, applicable to prob- lems on bounded spatial domains. Transform methods, on the other hand, are usually applied to problems on infinite or semi-infinite spatial domains. In this, section we introduce two fundamental integral transforms, the Laplace trans- form and the Fourier transform. Basically, the idea is to transform a partial differential equation into an ordinary differential equation in a transform do- main, solve the ordinary differential equation, and then return to the original domain by an inverse transform. 6.5.1 Laplace Transforms Laplace transforms are encountered in elementary differential equations courses as a technique for solving linear ordinary differential equations with constant coefficients. They are particularly useful for nonhomogeneous problems where the source term is piecewise continuous or is an impulse function (a delta fune- tion), We assume the reader is familiar with these elementary methods. Here we show that they easily extend to partial differential equations. First, we recall some ideas for functions of a single variable. If u = u(t) isa piecewise continuous function on ¢ > 0 that does not grow too fast (for example, assume that u is of exponential order, i.e., |u(t)| < cexp(at) for t sufficiently large, where a,c > 0), then the Laplace transform of u is defined by (Cu)(s) = U(s) = a u(the** de (6.41) We often write this as Cu = U. The Laplace transform is an example of an integral transform; it takes a given function u(t) in the time domain and maps it toa new function U(s) in the so-called transform domain. U and s are called the transform variables. The transform is linear in that C(cyu+opv) = Luton, where c; and cz are constants. If the transform U(s) is known, then u(t) is called the inverse transform of U(s) and we write (£~'UV)(t) = w(t) , or just £71 = u. Pairs of Laplace transforms and their inverses have been tabulated in extensive tables (see Table 6.1 for a brief list), and they are available on most computer algebra systems. ‘The importance of the Laplace transform, like other transforms, is that it changes derivative operations in the time domain to multiplication operations in the transform domain. In fact, we have the important operational formulas (Lu')(s) = sU(s) ~ u(0), (6.42) (Lu)(s) = ?U(s) — su(0) ~ u'(0), (6.43) 384 6. Partial Differential Equations Table 6.1 Laplace Transforms u(t) U(s) 1 st, s>0 et a s>a t", na positive integer 34, s>0 sin at and cos at gig and xi, s>0 sinh at and cosh at wig and zig, > |al ce sin bt cata sD e** cos bt jays >a t* exp(at) wit, s>a H(t-a) exp(-as), s>0 6(t—a) exp(—as) H(t —a)f(t—a) F(s) exp(—as) fen" P(s+a) H(t—a)f(t) e-*L(f(t+a)) erfyt sT(1+s)-¥?, s>0 va/sexp(-ays), (8 > 0) stexp(-ava), s>0 Virexp(-ays), s>0 s"U(s) — s*~!u(0) — s*2u/(0) ~ «= u-D(0) U(s) V(s) Formulae (6.42) and (6.43) are readily proved using the definition (6.41) and integration by parts. The strategy for solving differential equations is simple. ‘Taking the Laplace transform of an ordinary differential equation for u(t) re- sults in an algebraic equation for U(s) in the transformed domain. Solve the algebraic equation for U(s) and then recover u(t) by inversion. In general, to determine u(t) from knowledge of its transform U(s) would take us into the realm of complex variables (which is not a prerequisite for this book). However, we can indicate the general formula for the inverse transform. ‘The inversion formula is tie ult) = CW = 3 [ Uohet a. 6.5 Integral Transforms 385 ‘The integral in this formula is a complex contour integral taken over the infinite vertical line, called a Bromwich path, in the complex plane from a ~ ico to a+ ico. The number a is any real number for which the resulting Bromwich path lies to the right of any singularities (poles, essential singular points, or branch points and cuts) of the function U(s). Calculating inverse transforms using the Bromwich integral usually involves a difficult contour integration in the complex plane and using the residue theorem. In this text we only use a table or a computer algebra system. A result that is useful in calculations is the convolution theorem. Often, solving a differential equation results in having to invert the product of two transforms in the transform domain. The convolution theorem tells what the inverse transformation is. Theorem 6.28 (Convolution theorem) Let u and v be piecewise continuous on t > 0 and of exponential order. Then L(u+v)(s) = U(s)V(s), where (ut v)(t) = f u(t — y)u(y) dy is the convolution of u and v, and U = Lu, V = Lv. Furthermore, LUV)(s) = (ux v)(t) ‘Therefore, the Laplace transform is additive, but not multiplicative. That is, the Laplace transform of a product is not the product of the Laplace transforms. ‘The convolution theorem tells what to take the transform of in order to get a product of Laplace transforms, namely the convolution. ‘The same strategy applies to partial differential equations where the un- known is a function of two variables, for example, u = u(z,t). Now we trans- form on t, as before, with the variable being a parameter unaffected by the transform. In particular, we define the Laplace transform of u(sx,t) by (Lu)(x, s) = U(x,s) = f u(x, the** dt. 0 ‘Then time derivatives transform as in (6.42) and (6.43); for example (Cu(z,s) = sU(2,s)—u(x,0), (Luu)(e,s) = s°U(z,s) ~ su(z,0) ~ w(x, 0). 386 6, Partial Differential Equations On the other hand, spatial derivatives are left unaffected, for example (Cuz) (a, 8) = f 2 (2, t)e-* dt = a u(x, ten" dt = Us (x, ), and (Luse)(2,8) = Ure(v,s) On taking Laplace transforms, a partial differ- ential equation in x and ¢ is reduced to an ordinary differential equation in x with s asa parameter; all the t derivatives are turned into multiplication in the transform domain. Example 6.29 Let u = u(x,t) denote the concentration of a chemical contaminant, and let x > 0 be a semi-infinite region that initially contains no contaminant. For times £ > 0 we impose a constant, unit concentration on the boundary x = 0, and we ask how the contaminant diffuses into the region. Assuming a unit diffusion constant (the problem can be rescaled to make the diffusion constant unity), the mathematical model is uy ~ tee =0, 2 >0,t>0, u(e,0) =0, x >0, u(0,t) =1, t > 0, u(x, t) bounded. ‘Taking Laplace transforms of both sides of the equation yields sU (x, 8) ~ Use (2,8) = 0. This is an ordinary differential equation with x as the independent. variable, and the solution is U(x, 8) = a(s)e~V* + b(s)ev™ Because we require bounded solutions we set b(s) = 0. Then U(z, 8) = a(s)e"¥**. Next we take Laplace transforms of the boundary condition to get U(0, 8) = 1/s, where we have used £(1) = 1/s. Therefore a(s) = 1/s and the solution in the transform domain is U(a,s) = fev Consulting Table 5.1, we find that the solution is u(x,t) =1—erf (#) Figure 6.5 shows several time snapshots of the concentration of the contaminant as it diffuses into the medium. 6.5 Integral Transforms 387 Figure 6.5 Concentration profiles for different times t. ‘The Laplace transform of a distribution can be defined as well, and we may formally carry out transformations of delta functions and solve equations with point source functions. See Table 6.1 for the relevant entries. The theoretical basis for these calculations is contained in Section 6.7.2. 6.5.2 Fourier Transforms It is impossible to overestimate the important role that Fourier analysis played in the evolution of mathematical analysis over the last two centuries. Much of this development was motivated theoretical issues associated with problems in engineering and the sciences. Now, Fourier analysis is a monument of applied mathematics, forming a base of signal processing, imaging, probability theory, and differential equations, only to mention a few such areas. From a partial differential equations viewpoint, the Fourier transform is in- tegral operator with properties similar to the Laplace transform in that deriva- tives are turned into multiplication operations in the transform domain. ‘Thus the Fourier transform, like the Laplace transform, is useful as a computational tool in solving differential equations. The Fourier transform of a function in one dimension u = u(z), x € R, is defined by the equation a) f. a)et de. (Full ‘The variable € is called the frequency (or wave number) variable. There is 388 6. Partial Differential Equations one important remark about notation. There is no standard convention on how to define the Fourier transform; some put a factor of 1/2m or 1//2m in front of the integral, and some have a negative power in the exponential or even a factor of 2x. Therefore one needs to be aware of these differences when consulting other sources because the formulas change depending upon the con- vention. Another issue is to decide what set of functions on which to define the transform. Certainly, if w is integrable on R, that is, f°, |ul de < oo, then la(@)| = [JR w(x)et€* de] < JP, lu] dx < 00, and @ exists. But, as the ex- ample below shows, if u is integrable then d is not necessarily integrable; this causes problems, therefore, with inversion of the transform. Consequently, in the theory of Fourier transforms one often takes the domain to be a much smaller class of functions, for example, the square-integrable functions L?(R), or the Schwartz class of functions S. Schwartz, class consists of very smooth functions that decay, along with all their derivatives, very rapidly at infinity. Specifically, u € S if u € C°(R) and w and alll its derivatives u(*) have the prop- erty limjs|—sco (x? u(x) = 0, for any positive integer p. This assumption makes the theory go through in a highly aesthetic way (see, for example, Strichartz (1994). Example 6.30 Compute the Fourier transform of the characteristic function ye _ 1 iel 0, using a differential equation technique, 390 6, Partial Differential Equations Example 6.31 We want to calculate & where ii) = f. ent?” elt dy, Differentiating with respect to € and then integrating by parts gives Therefore we have a differential equation 42 = (—€)/(2a)4 for a. Separating variables and integrating gives the general solution a(€) = Cee ‘The constant C can be determined by noticing a= ot de VE = [Eee (6.46) So, the Fourier transform of a Gaussian function is a Gaussian; conversely, the inverse transform of a Gaussian is a Gaussian, Observe that a Gaussian with large spread (variance) gets transformed to a Gaussian with a narrow spread, and conversely. Consequently we have Fe A convolution relation also holds for Fourier transforms. If u and v are absolutely integrable, then we define the convolution of u and v by (ut v)(2) = fiw ~ y)o(y) dy. ‘Then we have the following theorem. 6.5 Integral Transforms 301 Theorem 6.32 (Convolution theorem) If u and v are in S, then F (wx v)(€) = a(6)6(6) Then (us v)(2) = FM(a(g)0(8) This formula, which can also be stated under less strenuous conditions, as- serts that the inverse transform of a product is a convolution. As with Laplace transforms, this is a useful relationship in solving differential equations because we frequently end up with a produet of transforms in the frequency domain. ‘The convolution theorem follows from formally interchanging the order of in- tegration in the following string of equalities. Fouevye)= (J we—vuloray) ode [ se= wpotne® dedy = [ ~ f ~ u(r)u(y)e"* e4 drdy = f. une dr fo v(y)eY dy = a(6)6(6) Example 6.33 Consider the ordinary differential equation u!-u=f(2), reR, Assume that w and its derivatives, and f, are sufficiently smooth and decay rapidly at infinity. Proceeding formally, we take the transform of both sides to obtain (-ig'a-a=f, Hence 1. u0) = -7yp Al In the transform domain the solution is a product of transforms, and so we apply the convolution theorem. From Table 6.2 Fe) =3 +B 392 6. Partial Differential Equations ‘Therefore fel f(a) = [iene u(a) We can show, in fact, that this is a solution if f is continuous and integrable. Example 6.34 Use Fourier transforms to solve the pure initial value problem for the diffusion equation: ty kuze=0, TER t>0; u(2,0)=f(x)reR. (6.47) We assume the functions u and f are in S. Taking Fourier transforms of the equation gives ay = -€7ka, which is an ordinary differential equation in ¢ for ai(€,t), with € as a parameter. Its solution is : a(E,t) = C(e™. ‘The initial condition gives a(€,0) = f(€), and so C(€) = f(€). Therefore HG.) = eM FE) Replacing a by 1/4kt in formula (6.46) gives eke 1 2 F 2 vt) = ( vankt By the convolution theorem we have eo 1 2 ~(emu)*/ake u(z,t)= [| =e dy. 6.48) (ext) = [~ Tee Fu) (648) This solution was derived under the assumption that u and f are in the Schwartz class of functions. But, now that we have it, we can show that it is a solution under milder restrictions on f. For example, one can prove that (6.48) is a solution to (5.5.7) if f is a continuous, bounded function on R. If f is only piecewise continuous, then u remains a solution for ¢ > 0, but at point, of discontinuity of f, say a, we have lim,.o+ u(a0,t) = 3(f (ap) + f(ad)). Example 6.35 Next we work a problem involving Laplace’s equation in the upper half- plane. Consider Ure tty = 0, reER, y>0, u(z,0) = f(z), ceR 6.5 Integral Transforms 393 We append the condition that the solution u stay bounded as y — oo. This example is similar to the last example. Taking the transform (on @, with y as a parameter) of the equation we obtain tty — Ca = 0, which has general solution (E, y) = a(g)e“™ + b(ee™ ‘The first term will grow exponentially if € <0, and the second term will grow if € > 0. To maintain boundedness for all £ we can take (Gv) = e()e“ Applying the Fourier transform to the boundary condition yields a(€) = f(€). ‘Therefore the solution in the transform domain is (6, y) = eK F(@). By the convolution theorem we obtain Yl yg_y [* _flaan Mow = pap t= */ eae After the fact, let us motivate how consideration of the Fourier transform arises naturally from examining the initial value problem (6.47) for the heat equation using a separation of variables technique and superposition. If we assume a solution of the form u(s, t) = T(t)¥ (1), then substitution gives rt) _ Ye) _ kT® ~~ Y(a) 2 where the separation constant —£? has been chosen to be negative, in an- ticipation of having a decaying solution. A solution to the time equation is T = e~®**t, while a solution of the Y equation is Y = e~***. These functions of « represent a continuum of bounded eigenfunctions. Therefore we obtain a ‘one-parameter family of solutions (one for each choice of € € R) of the heat equation of the form u(a, tig) = ete ‘We superimpose these solutions and define ulast)= [leer H a 394 6. Partial Differential Equations Table 6.2 Fourier Transforms u(x) a(x) (-ig)na(e) use F(§) Gg) where c(E) represents a continuum of coefficients. By superposition, u satisfies the heat equation provided c is nice enough. The initial condition implies ua0) = ste) =f eee" a, which, up to a constant multiple 3, is the inverse Fourier transform of c(€). So, in this manner, the transform integral arises naturally. This being the case, the Fourier coefficients (€) are given by e(€) = a f(€), and ula) = ef fees eM ag ‘This integral can be calculated to obtain the solution (6.48) of the Cauchy problem. We again notice that the solution resembles a continuous Fourier expansion, As in the case of the Laplace transform, the Fourier transform can be ex- tended to distributions (see Table 6.2). We end this section with some comments on eigenvalue problems on infinite domains. For example, consider the eigenvalue problem -Au=u, xER". On a bounded domain there were accompanying boundary conditions (Dirich- let, Neumann, Robin) on u. Now we must decide what kind of boundary con- ditions to impose at infinity. For example, do we assume u is square-integrable, 6.5 Integral Transforms 395 bounded, or goes to zero as |x| —» 00? The one-dimensional problem reveals the issues, Consider "=du, 2eR 0, ule) = ae~¥* + bev? ‘The general solution is: u(x) = ax +b if \ if \ <0, and u(x) = ae*¥* + bet¥¥, if X > 0. None of these solutions is square-integrable on R or goes to zero at & = +00; 50 in these cases there would be no eigenvalues. However, if we impose a boundedness condition on all of R, then any A > 0 gives a nontrivial solution u(r, A). In this case we say belongs to the continuous spectrum of the operator —<,, and the associated eigenfunctions are e~*¥%*, e'V3#, ) > 0. These are sometimes given the unfortunate misnomers improper eigenvalues and improper eigenfunctions. We can write this set of eigenfunctions as e~**, where —o0 < k < 00. Then, if f is a given function, we can think of representing it as an as eigenfunction expansion, but this time not as a sum (as for the case of discrete eigenvalues), but as an integral Lf aqye-ik a | elke ak foe But this means that the continuum ¢(k) of Fourier coefficients is given by the Fourier transform of f. In a distributional sense, the eigenfunctions e~'** are orthogonal. Therefore, the entire theory of eigenfunctions expansions seems to remain valid for operators with continuous spectrum. Indeed, this is true and f(x) there is a close connection to eigenfunction expansions and transform theory. ‘These ideas have their origin in quantum theory. The eigenvalue problem for the Schrodinger operator -Au+V(xju= du, xR", where V is the potential function, can have both eigenvalues (with square- integrable eigenfunctions, called bound states) and continuous spectrum (with bounded eigenfunctions, called unbound states). ‘The eigenfunction expansion then involves both Fourier series (sums) and Fourier integrals. These topics are beyond our scope and we refer the reader to Keener (2000) or Stakgold (1998). EXERCISES 1. Use a transform method to solve the diffusion problem uy = Duse, O0, u(0,t) = u(lt)=1, t>0, (me u(z,0) = 1+sin(), O 0 from its bound- ary 2 =0, where the concentration is maintained at g(#). The model is ty =Duze, 2>0,t>0, u(2,0)=0, ©>0, u(0,t) = g(t), t>0. Determine the concentration u = u(z,t) and write the solution in the form of u(2,t) = fi K(x,t—r)g(r)dr, identifying the kernel K. 3. The small, transverse deflections from equilibrium of an elastic string under the influence of gravity satisfy the forced wave equation Ue = Cure — a where c? is a constant and g is the acceleration due to gravity; u(c, t) is the actual deflection at location « at time t. (Wave equations are derived in Chapter 7.) At t= 0 a semi-infinite string (x > 0), supported underneath, lies motionless along the axis. Suddenly the support is removed and the string falls under the force of gravity with the deflection at x = 0 is held fixed at the origin; the initial velocity u(x, 0) of the string is zero. Determine the displacement of the string and sketch several time snapshots of the solution. 4, Show that the inverse Fourier transform of e~*!¢l, a > 0, is aii na +a? 5. Verify the following properties of the Fourier transform. a) (Fu)(£) = 2n(F-"u)(-8). b) Fleltu)(é) ©) Flu(e+a)) 6. Find the Fourier transform of the following functions. a) u(x) = H(c)e~*, where H is the heaviside function. b) u(x) = 2en™* 7. Use Fourier transforms to find the solution to the initial value problem for the advection-diffusion equation ty — cls tse = 0, TER, £>0; u(z,0) = f(z), reR. 6.5 Integral Transforms 307 8. Solve the Cauchy problem for the nonhomogeneous heat equation: ue = tee + F(a,t), TER, t>0; u(x,0)=0, ceR. 9. The Fourier transform and its inverse in R" are given by ae) = [ wegetax, x€eR", 1 us) = ae [ago 0) = Gaye f Boden de a) Show that the Fourier transform of the Laplacian is F(Aw) = —l¢/°a(€) b) Find the Fourier transform of the n-dimensional Gaussian u(x) = eval? ¢c) Solve the Cauchy problem for the diffusion equation in R”. um =DAu xeER", u(x,0)= f(x), xeR™ 10. Prove the Plancherel relation f°. |u(x)|?dx = 3 f°, \i(€) Padé. 11. Use the Plancherel relation to evaluate the integral [veer 12. Find a formula for the solution to the initial value problem * 2eR u, = iter, TER, t>0; u(x,0)= 13, Solve the initial boundary value problem ty = Ditzs, 2 >0,t>0, u(2,0) = f(z), 2>0, u(0,t)=0, t>0, by extending f to the entire real line as an odd function and then using Fourier transforms. 14, Solve the following initial value problem for the heat equation. ue =kuse, TER, t>0, Iz] >a, uo, |z| 0, u,(z,0) = f(z), reER, is given by oo ule) = 5 f Inlee 0)? + vend (Hint: let v(2,y) = uy(2,y).) 6.6 Stability of Solutions 6.6.1 Reaction—Diffusion Equations We already noted the broad occurrence of diffusion problems in biological sys- tems. In this section we investigate another aspect of such problems, namely the persistence of equilibrium states in systems governed by reaction-diffusion systems. At issue is the stability of those states: if a system is in an equilibrium state and it is slightly perturbed, does the system return to that state, or does it evolve to a completely different state? Underpinned by the seminal work of Alan Turing® in 1952 on the chemical basis of morphogenesis, it has been shown in extensive research that diffusion induced instabilities can give rise to spatial patterns in all sorts of biological sys- tems. Reaction-diffusion models have been used to explain the evolution of form and structure in developmental biology (morphogenesis), tumor growth, eco- logical spatial patterns, aggregation in slime molds, patterns on animal coats, and many other observed processes in molecular and cellular biology. In this section we introduce the basic idea of stability in reaction-diffusion models and we observe that such systems can have instabilities that lead to density variations and patterns. A similar theory of local stability of equilibrium solutions can be devel- oped for partial differential equations. In the interval 0 < « < L consider the reaction-diffusion equation Duzz + f(u), (6.49) with no-flux boundary conditions u,(0,t) = ue(m,t) = 0. Let u(x,t) = te be a constant equilibrium solution (so that f(ue) = 0). To fix the idea let * A. Turing was a British mathematician who was instrumental in breaking the German codes during World War IL

Das könnte Ihnen auch gefallen