AD 646 APPLIED MECHANICS REVIEWS Vol. 19, NO. 9 SEPTEMBER 1966 The Method of Weighted Residuals—A Review B.A. FINLAYSON* AND L.E. SCRIVEN UNIVERSITY OF MINNESOTA, MINNEAPOLIS, MINNESOTA Abstract T: method of weighted residuals unifies many ap- proximate methods of solution of differential equa tions that are being used cumtently. This review presents the basic method in its historical context and shows some of the many possible modifications that have been used throughout the past fifty years. The re lationship between the Galeckin method, which is one version of the method of weighted residuals, and varia tional methods is outlined. Also included is an exten- sive listing of published applications of the method of weighted residuals, Introduction ‘The method of weighted residuals is an engineer's tool for finding approximate solutions to the equations of change of distributed systems. Experience and tuition can be distilled into a reasonable and some- times quite accurate first guess, fom which it ix possi+ ble to proceed to successively improved approxima tions. The analytical form of the approximate solution is often mote useful than solutions generated by qumer= ical integration, and the approximate solution vsually requires less computation time to generace. The method is applicable co nonlinear and non-self-adjoine problems—one of its most attractive Features. ‘The method of weighted cesiduals (MWR) includes ‘many approximation methods that are being used cur~ rently. It provides a vantage point from which it is ‘easy (0 see the unity of these methods as well as the relationships between them, This review, after outlin= ing application of the basic method to initial-value, boundary-value, and cigenvalve problems, surveys the history of majer contibutions to the subject and dis- cusses some of che many modifications of the basic method. The review concludes with a listing of appli= cations of weighted residual methods to problems atis~ ing in applied mechanics and related fields. Four prac= tical aspects of MWK in need of further research are identified be. Tidcers: Olfice of Naval Research, Washington, ‘A. Basic Met ‘The best available treatments of MWR have been those by Crandall [1], who coined the name method of weighted residuals, Ames [2], and Collatz [3], who calls these methods error-distribution principles. The following outline pasallels theie treatments, in places contrasting them and elaborating on them. Given a system of differential of integro-differential equations of change ahd constitutive relations, bound= ary conditions representing the interactions between the system and ies surroSndings, and initial conditions representing some base state of interest, the general approach is to assume a trial solution whose functional dependence on position is chosen, but which includes undetermined functions of time. The latter are found by requiring that the trial solution satisfy the differential equation in some specified approximate sense Initio! Value Problem Consider the differential equation for (u(t in wey -% -0 Q) or xinV,1>0 aw where N() denotes a general differential operator in- volving spatial derivatives of u, Vis » three-dimen- sional domain with boundary S, and f represents time Suppose the initial and boundary conditions are 4% O) © 9fx), Xin V @ 208, = (08. oS Assume a cial solution of the form HO NaH DoS HD — OD at where the approximating functions, ui, are prescribed and satisfy the boundary conditions j= 0, Kons “ ‘Then u* satisfies the boundary conditions for all fane~ tions (A). Ie is not necessary that the trial solution 735 Heebe linear in the cj but such a choice is usually made for simplicity; no systematic study of altematives has been reporced, so far as the authors know. The difece ential equation residval and initial residual, a = Hf0) ~ ss, 0) es(Oduels, 0) (6) Ru) = Nw © Ryu) fire measures of the extent to which the function ut satisfies the differential equation and initial condi- tions, respectively. As the number ¥ of approximating functions u, is increased in successive approximations, fone hopes the residvals will become smaller; the exact solution is obtained when both residuals are identically zero. As an approximation to this ideal, the weighted integrals of the residuals are set equal to’ zero: Kus ROM)= 0 foot o (ep Rdur)y=0 eos fone represents a spatial average of inner product and w, is prescribed weighting function. If'u* is the exact solution, Equations (7) are satistied regardless of the choice of weighting functions. ‘The weighting functions can be chosen in several different ways, and each choice corresponds to a dif- ferent criterion in MUR. Once the choice is made, Equations (7) become a set of N first-order ordinary differential equations in the N unknowns cj(t).For the linear problem where: @ w o a with approximating functions uj and xy that do not themselves depend on tine, Equations (7) become simply 1 Quy Llu) + di bw) (00) ~Be+k, an ‘The solution to these equations is substituted into Equation (3) to give the approximate solution to the problem. Successive approximations are obtained by increasing N and solving Equation (10) anew. |The convergeace of successive approximasions gives & clue, but not necessarily a definitive one, to the tea sonableness of the approximation, Boundary Value and Eigenvalue Problems The method is equally applicable to steady-state and eigenvalue problems. For steady-state problems, the cj ate constants rather chan fonctions of times for linear problems they are determined as solutions to a2) For nonlinear boundary-value problems it may be useful to assume tial solutions of a more general form than Equation (3), viz, wax) = (teil, uso) 3) For the linear eigenvalue problem La) ~ Au a4) the approximate solution is decermined by x foctmp reno} Die@a-aap-0 as and this set of equations has a noa-tivial solution only if det(Aje~ABy) = 0 «as ‘The values of A for which this is tue are the approxi= rations to the first N eigenvalues Ay Weighting Functions ‘The choice of ‘he weighting functions, wj in (7), conesponds to vatious criteria in MWR: the historical ‘relationship of the criteria is portrayed in Table I. In the collocation method, due to Frazer, Jones, and Skan [4], the weighting functions ace the Dirac delea functions y= BOK; ~ 8); a TABLE | HISTORY OF APPROXIMATE METHODS Dae —lnvestgacon Method 1915 Gatti) Galerkin method 1921 Potlhausen (18) Integral nechod 1923, Biezeao and Koch (5) Subdomain method 1928 Pieone (9) Method of least squares Be rw Method 2f moments 1933 Kantorovich [30] Method of reduction to oni= ary differential equations 1937 Frazee, Jones, and ‘Seas (4 Collocation mechod 1938 Poviaky [31] Method of reduction to ordi= ‘nay differential equations 1940. Repman [55] Coaveigence of Gairkn'e 1941 Bickley [12] Collocation, Galerkin, least ‘squates for initialcalve problems 1942 Keldysh (52] Convergence of Galerkia's Tmethod, steady-state 1947 Yamada [16] Metiod of momeate oe Fandoh 9 Convergence of Galetkin’s 1953, Green [6a ‘method, ansrendy state 1956 Crandall (1) Unification as method of ‘weighted residual 736the differential equation is then satistied exactly at the 1 collocation points, x). AS N is increased, the resid ual vanishes at more and more points and presumably. ‘approaches zero throughout V. Ifthe weighting functions are 1 xin Vy aye cas) 0 x noe ia vj then the differential equation is satisfied on the aver~ age in each of the N subdomains, Vj; this is the sub- domain method [5, 6]. Ifthe V, are disjoint (which they need not be), the size of one’ or more subdomains de~ creases as 1 is incveased, with the result that the dif- ferential equation is satisfied on the average in smaller and smaller regions, and presumably the residual ap- proaches zero everywhere. It was Biezeno’s presenta~ on (7) of the subdomain method at the Fiest lnrerna tional Congress of Applied Mechanics which prompeed Courant's remark {s] that led Crandall to choose the name, “method of weighted residvals."” The authors’ translation of Courane’s remark reads: ‘Mr. Courant (Géttingen) indicated afterward that the method advanced by Mr. Biezeno can be viewed from the standpoimt of the calculus of variations in the following manner. If a differential equation, as i arises for example in a variational problem, must be satisfied, then we can express it so that the left side of the differential equation, multiplied with an arbitrary function and then integrated, must give us the value zero (vanishing of the first variation). Ia- stead of taking an arbitrary function, we can also take infinitely many determined functions, if these only form a so-called complete function system for the region in question. The piecewise constant fune~ tions advanced by Me, Biezeno ate indeed just an especially simple special case of such a complete function system.” The least-squares method, which seems eo have been first presented for this type of application by Picone in 1928 [9], uses the weighting functions 9R(w*)/de). The coresponding interpretation is that the mean ‘square residual te f enav us) is minimized with respect to the constants cj. In the Galerkin method [10], developed in 1915 as the fies criterion of what is now known as the method of weighted residuals," the weighting functions w; are just the approximating functions of u,.. The approximat- ing functions are often members of complete aystem of fonctions, although this property, requited for mathe matical purposes, is sometimes ignoced in practice. The Galerkin method then can be interpreted as making the residual orthogonal to members of the complete set, See MR Bubaor in 191 fnethod {Mikblin and others in eecent Ruselan literatare Te the BubnowGalerkin method}, ie was Caletkia who de, oped the mothol independently of any variational principle. Li] for « discussion of the contribution by ‘hile Ris method ie the same athe Galerie ‘A fundamental property (sometimes the definition) of a complete system of functions is that a piecewise con- tinuous funetion can be omhogonal to each and every member only if the function is identically zero, In the approximation scheme outlined above, the residual is usually continuous (depending upon’ the differential operator and the choice of approximating functions), and hence the residual can vanish only if it is ortho: gonal to cach member of a complete system of fune= fons. Of course in practice dhe residual is made oftho- gonal to 99 more than a modest, finite aumber of the embers of a complete set, In the original Galedkin method, develoged in the study of elastic equilibrium fand stability of rods and plates, Galerkin used tial solutions with unknown constant coefficients. Now rnany similar techniques are often referred to as the leskia or generalized Galerkin method: (i) the one sven above in which c,~ c4(t) for time-dependent probe ems (12, 1; (i) one in which trial solueions are of the more general form u* = /(X, leg!) with weighting fanc- tions 0//de; L13h and (it) one in which weighting func= tons ate of the form K(x) rather than w, where K is a specified differeatial operator (14, 15]. ‘The method of moments is similar co the Galerkin method except that the residual is made orthogonal £0 members of a system of functions which need not be the same as the approximating functions. Both methods Sue combined under the single name of orthogonaliza- tion methods by Collatz [3}, Yamada [16] and Krav- cchuk [17] applied the method of moments to ordinary differential equations by using the weighting functions Ls} regardless of the choice of approximating func- tions. For the fist approximation, the weighting func= tion is unity, and the method of moments in this case Js equivalent to the subdomain methed and is usually called the integral method, or von Kérméa-Pohlhausen imethod [18, 19}. For the integral method, reviewed in Secail by Godan [20], the differential equation is Satisfied on the average over the domain of interest. Boundary Methods In the foregoing it is presumed that the ial solution satisfies the boundary conditions but not the differen tial equation. The converse situation can also be treated: the differential equation is satisfied but the boundary conditions are not. Trial solutions of this ‘sort lead to boundary methods, as they are called by Collatz {3}; the procedures ace analogous to those above, but with the spatial average, Equation (8), re placed by an average over the boundary, Mixed Methods ‘The intermediate situation can also be handled: in so-called mixed methods the tial solution satisfies neither the differential equations nor boundary condi- tioas. In Schuleshko's weatment of mixed methods [21], the differential-equation residyal is made orthogonal to fone set of weighting functions, using (8) as the inner product, while the boundary residual is simultancously ‘made ofthogonal to another set of weighting functions, using an appropriate surface integral as the inner prod- uct. IFN weighting functions are used, this leads to 7372N conditions, yet in general only N conditions can be satisfied by the N independent cj. For this procedure to work, some of the conditions must be discarded, as was noticed by Soyder, Spriggs and Stewart [22] in their discussion of the Galerkin method. On the other hand, Bolotin [23], Mikblin [11], and Finlayson (24) have pointed out that for the Galerkin ‘method the dilemma can be resolved by adding the dif- ferential-equation residuals to the boundary residuals. ‘The combination is made in such a way that the differ- -equation residual, when integrated by parts, can- els identical terms of the boundary eesidual. The sit- uation is analogous to the treatment of natural boundary conditions in the calculus of variations, and indeed only boundary conditions analogous to natural boundary conditions can be handled in this way. Sach a combina tion of equation and boundary residuals represents a ‘generalization beyond the weatment given by Cran~ dall [1] (page 235), who states that MWR cannot be ap- plied unless the tial solution satisfies all boundary conditions. However, Crandall [1] (page 321) does com- bine the residuals for eigeavalue problems in which an eigenvalue appears in both the differential equation and boundary conditions. The combination of residuals for more general problems is important in establishing the equivalence between the Galeskia method and several so-called variational methods [25, 24] (see below). B. Refinements on the Basic Method Other modifications ace possible: Duncan [26] makes the approximating functions satisfy derived (or second= ary) boundary conditions which are determined by ze- quiting that the differential equation be satisfied on the boundary. Derived boundary conditions ate also used in boundary-layer theory in the von Kérmdn-Pohl~ hhavsen method; other compatibility conditions—such as continuity of the velocity and certain of its deriva tives at the edge of the boundary layer—are employed fas well, Recently it has been shown [27, 28] that ad- ditional compatibility conditions are required to assure good results when the integral method is. applied to Imagnetohydrodynamic boundary-layer problems. la these cases the additional conditions are found by dif- ferentiating the differential equation in the dicection nommal to the surface; all tial solutions must then satisfy this equation at both the solid surface and the ‘edge of the boundary layer. A variation of the colloca- tion method is given by Collatz [3], who differentiates fan ordinary differential equation and applies the collo- cation method to the residual of the resulting equation, Kantorovich and Krylov (29] outline @ method for two-dimensional problems in which the residual is te- quired to be zero along a line in the domain (such as ¥ =y). The method of reduction to ordinary differential equations, as developed by Kantorovich [30] and inde- peqdencly’ by Poritsky [31], reduces a partial differen~ tial equation to a system of ordinary differential equa tions. This is the procedure described above for initial= value problems but it can be applied equally well to boundary-value or eigenvalue problems. The spatial averages (Equation (8)) are taken over all the inde~ pendent variables except one, and the approximate so- lution is found by solving a set of ordinary differential equations involving this xemaining independent vari- fable. While this semi-direct method was originally pro= posed in the context of variational principles, Kantoro+ vieh (32] ia 1942 showed its equivalence to the Gale kkin method, Even earlier Bickley [12] had applied the Galeskin method to unsteady-state problems in a manner equivalent to the method of reduction to ordinary dif- ferential equations. In general, MWR can be used to reduce the sumber of independent variables in any par- tial differential equation. The resulting system of equations is simpler (it may be algebraic or ordinary differential equations ot even a set of partial diffecen- tial equations), bur its solution zemains only an approx- imate solution to the original problem. Tn the collocation method a critical problem is the choice of collocation points. For ordinary differential equations Wright [33] has shown that the residual is ninimized if the collocation points are given by the roots of the Chebyshev polynomials, Naturally the method of weighted residuals ean be combined with other methods, "Collatz [3] presents combination of che iteration method and MUR. Yang (34, 35, 36] uses the approximate solution generated by the integral method as the first step in the following procedure for time-dependent problems: the result of the Integral method is substitsted into those terms involy- jing time-derivatives and equation thereby obtained is solved as a steady-state, nonhomogeneous, patial dif- ferential equation. * An advantageous coupling of MWR and sumerical finite difference methods has been em- ployed by Kaplan (37), Kaplan and Bewick [38], and Kaplan, Marlowe, and Bewick [59] to reduce the com= puter tine necessary to solve certain nuclear reactor problems; the number of independent variables was re- duced from four to three or two by using MUR. Other modifications and hybrid schemes are possible and will ‘undoubeedly be proposed as needs arise. Choice of Approximating Functions ‘The choice of approximating functions can be crucial in applying MWR, How to arrive at a good, if not the best, selection is an outstanding problem. Certainly any ‘symmetry properties of the system should be e: ploited but there seems to be no way available at pres- feat to do this systematically for all probleme. In prob- lems of conventional types it is usually convenient to have the approximating functions satisfy the boundary conditions, and Kantorovich and Keylov [29] show how to consteuct complete sets of functions which vanish on ‘2 boundary of complicated shape. Sayder and Stewart [40] combine this scheme and. symmetry arguments to find approximating functions for the velocity vector field in fluid flowing through regularly packed beds of spheres. Derived boundary conditions can also be used to place restrictions on the approximating functions ad- mitted, and improvement sometimes results (26, 148]. Usually, however, several sets of approximating func tions are admissible and it is aot possible to choose fone as the "best." Heywood and Moffatt [41] even suggest as a qualitative criterion chat the approximate 738solution be relatively insensitive to different but reae sonable choices of the approximating functions, Methods have been devised for constructing approxi- mating functions especially for eigenvalue problems in volving high-order ordinary differential equations of the sort that arise in the theory of convective instability (42, 43, 44]. The approximating functions are just eigenfunctions of one or another lower-order, simpler yet related eigenvalue problem on the same domain. Polynomials are popular approximating functions; they hhave even been used in cylindrical and spherical do- ‘mains [45] where proper regard must be taken of possi- ble singularities. Falk (46] uses Hermite polynomials, which are orthogonal og a semi-infinite domain, Other authors [39, 47] emphasize chat numerical difficulties (or large N) can be avoided in the Galerkin method if the approximating functions are orthonormalized, Selecting approximating functions remains somewhat dependent on the user's intuition and experience, and this is often regarded as major disadvantage of MUR. Clearly, the question of methods for arriving at optimal choices of approximating functions wactants thorough investigation. Leads may exist in the local solutions ‘and regional expansions used in perturbation methe ods [47a Comparison of Different Criteria Comparisons of different criteria as applied to the same problem exist only for relatively simple, linear, initial-value problems (12, 25, 48] and boundary-value problems {4, 13, I]. In the literature on eigenvalue problems the Galerkin method predominates, although there are a few comparisons with the collocation and least-squares methods (1, 4]. ‘The cesules of these comparisons may be summarized by Crandall’s re mark [1] (page 375): "The variation between results ob- tained by applying different criteria to the same tial family...is much less significant than the variations that can result from the choice of different tial fami- lies."" However, there may be a great difference in the work necessary to obtain the appcoximate solution when using different criteria, Crandall’s experience evi= dently is based entirely on linear problems. The only comparison for nonlinear problems appears to be the unr Published thesis by Collings [49], ax referenced by Ames [2]. Ames comes to the conclusion that the Galetkin method is superior, but cautions that this stand is based on limited experience and may not hold in general [2}. For linear, ordinary differential equations Frazer, Jones and Skan {4} argue that the collocation, least Squares, and Galerkin methods are equivalent in the limi as Ns, Other similarities exist between the methods [1]; e.g. whea the approximating functions are chosen to be the eigenfunctions of the linear opecator, ie. L(uj) = d uj, then the least-squares and Galerkin method coincide, For selisadjoine (hence necessarily linear) eigen value problems, the eigenvalues are real, and Cran all [1] emphasizes that the Galerkin method leads to symmetsic matrices in Equation (15)—and hence real- valued approximations—whereas the other methods may give complex eigenvalues as approximations to the exact real eigenvalues, The least-squares method is particularly unsuited for linear eigenvalue problems be- cause it turns the linear problem into a sonlinear one {1}. ‘The Teastesquares procedure for eigenvalue problems as outlined by Becker [50] differs somewhat from that of Crandall [1]; Becker does nat have such « difficulty in the first approximation, Whereas Crandall uses the weighting function dI/dc, for the first approximation, Becker uses al/dA, where A is the eigenvalue. Con sider the linear eigenvalue problem Liu) + Au =o (20) with «= 0on the boundary, For the first approximation with a trial solution v* = cju, the eesidval is Rew.) = e(L(u,) + Aa) en The mean square error is then tee fltapeantav (22) Crandall appatently would determine \ from Hae, = 02, fila) rulav — (3) which is a quadatic in’ A and may lead to complex values of A. Becker, on te other hand, would deter- imine A from H/dd~0- 24 fla) srulae — (26 which is linear in X and gives real values as long as the equation and u, are teal, Becker's procedure ap= pears co be simpler for che fiest approximation, For higher approximations both procedues lead to nonlinear equations for this linear problem. Crandall would use as weighting functions al/3ei, 1 = 1, 2. +, Ny and Becker would use al/AA, al/dej i ~ 203) Lo. Ns The later is thus using the eigenvalue \ as one of the paramecers and is also exploiting the fact that the mean square ezror can be minimized as a function of 10) = ey/eyL eather than tej} since He, era, dy, A (25) For initial-value problems, the least-squares method ‘must be applied carefully and has certain disadvan- tages. The method is applicable if the time depend- ence of che approximate solution is specified—in other words, semicdirect methods cannot be used in the method of least squares. Consider the problem du /2t = L(q) and assume a trial solution of the formu ~ uy + % any(x, ). Then the functional I, representing the mean square residual, can be minimized: ef [Fe] ae 26) Of course the solution depends on the value of T. If the upper limit of integration is infinite, the solution may no longer have this ambiguity. This was the ap= yy ee EN 739proach taken by Bickley [12] in his leastesquates cal- ulations for time-dependent problems. Oftentines, however, the time dependence of the so- lution is difficult to guess and che ial solution must involve undetermined functions of tine, YX etuy(. The mean square residual is . STR Now, however, {depends on time and involves time derivatives, Consequently ie cannot in general be made ‘2 minimum for all time by any set of functions c;(0); this was shown by Citron [51] and Finlayson and Scriven (25]. Consequently, if 2 semi-direet method is used to solve this type of problem che term least squares is a misnomer Because the mean square resid- ual is not being minimized. ‘The least-squares method is discussed at length in @ monograph by Becker [30]. Listing criteria which he maintains a good variational method must satisfy, he concludes that the least-equares method is the best general criterion of MWR. Becker's list includes the following points: (i) ecors should be minimized in some sense; (i) the functional should be positive definite; ii) the procedure should be eapable of treating init value problems, as well as others. These seem some- what slanted toward the least-squares method; indeed, items (i) and (ii) cannot be fealized for all problems except in the least-squares method. Yet no one has shown that a solution is necessarily best because its mean square residual is smallest; such a definition ot proof will certainly depend on the particular applica tion, In addition, the least-squares method can be used to treat initial-value problems in only a limited way, as shown above, Furthermore, an important point should be added to the list of desiderata—the method should be simple to apply. As already shown, this criterion immediately eliminates the least-squares method for Linear eigenvalue problems because it tums a linear problem into a more difficult noalinear one. Becker realized that his conclusion may aor always be valid [50] (page 61): "While the least-squares method seems to be the most suitable general approach, in specific applications (ia which some specific criteria may be ‘added to our “general” list) other methods may be pref- crable."” Becker illustrates the advantages of the method of least-squares by solving » set of aonlinear, time-dependent partial differential equatioas which ‘model che fuel depletion in a nuclear reactor; he finds results that compare well with che more lengthy numeri cal solutions In this discussion of the various criteria of MER, the Galetkin method has been distinguished feom' che method of moments by means of the weighting functions used in the two. In the Galerkin method, the weighting functions must be the same set of functions which ave used for the ‘trial solution, whereas in the method of ‘moments the weighting functions can be some other set of functions. This distinction is not always made [52] fand is probably unimportane in practice, although the tn] en ‘two methods have different histories and may have dif ferent convergence properties, There are inconsisten- cies of terminology in che licerature; for example, Kawaguic (53] used the method of moments eather than the Galerkin method as he claimed, for the weighting functions differ from the approximating functions in his work. Another example of confusing nomenclature is the name method of integral zelations, which refers to a generalization of the subdomain method; it is ade~ {quately reviewed by Belotserkovskii and Chushkin (54) C. Convergence Theorems Golerkin Method After introdaction of the Galerkin method in 1915 some twenty-five years elapsed before the convergence fof the method was studied. Even today much remains to be done; only a few theorems have been proved, and these pertain exclusively to linear problems. Rep 10 (55] was the first to prove convergence of solo- tions obcained by the Galerkin method though only for a certain Fredholm-ype integral equation. Petrov [36] thea studied the convergence of the Galerkin method for eigenvalue problems of fourth-order ordinary differential ‘equations—in particular, the Orr-Sommerfeld equation of hydrodynamic stability theory. Keldysh (57] treated general ordinary differential equations and also second- order elliptic partial differential equations, Mikhlio [11] later simplified Keldysh's proofs. The equations are of the form Ue 4 Cu=f (28) Both Keldysh and Mikhlin prove chat che first deriva tives of the Galerkin approximate solution converge in the mean to the first derivatives of the exact solution Whenever the Rayleigh-Ritz and Galerkin methods coincide (see below), the convergence proofs for the Rayleigh-Ritz method imply convergence of the Galerkin ‘method, too. ‘Thus the Galerkin-convergence proofs given by Kantorovich and Keylov (29] apply only co Specific problems with a minimum or maximum princi~ ple, wheseas the convergence proofs mentioned here are Applicable to problems whether or not they have a cor- responding variational principle. It has been claimed [40] that completeness of the set of approximating func tions is sufficient to assure convergence, but the proofs given by Mikhlin and others show clearly that this is not enough. Recently convergence proofs have become available for certain eigenvalue problems associated with hydo- dynamic stability investigations (58, 15]- Resulee applicable to unsteady-state problems are less extensive. Faedo [59] applied the Galerkin method to a hypetbolic differential equation and in spired che important work of Green (60), who proved the tuniform convergence of the Galerkin method when ap- 740plied to the following equation au de Se Hate, um fox, 0 29) Sen Fe aes Oem I ‘The Galetkin method has been used to prove the exist- ence of weak solutions to (i) the Navier-Stokes equa tions with time dependence [61], and (ii) equations representing the unsteady-seate teansport equation with a known velocity field [6212 ae Bs Pee, HF 0 G0) Recently the Galerkin method has been applied to the ‘aylor problem with time-dependent disturbances [63], sind a method bas been developed to generate improv able, pointwise upper and lower bounds—and hence ferrof bounds—for the solution 9 Equation (30) (12a Other Meth Convergence proofs are rarely available for the other criteria of MWR, The notable exception is the least- squares method, which is well-treated (for boundary problems) in the notable cext by Mikblin (11). Mikhlin proves conditions which insuce that the method of least Squares gives a sequence of approximate solutions which converge in the mean to the exact solution. Fur thermore, the mean-square-error of the approximate so- lution can be determined. He also points out that the least-squares method converges mote slowly than the Ritz method (when ehe latter can be applied) bue may give uniform convergence eather than convergence in the mean. Some results for the collocation method are given by Kadner [64], while the method of moments is treated by Kravehuk [65]. Very litle is known about the convergence of MVR for nonlinear problems without a corresponding vatia~ tional principle. Krasnosel’ski [66] Presents theo- fems—aostly without proofs—for the Galerkin method ‘applied to nonlinear integral equations. Glansdoctf [67] mentions a forthcoming proof af the convergence of the Tocal potential method, which is identical in applica tion to the Galerkin method; he teats the steady-state heat ‘conduction equation, with ‘temperavure-dependent thermal “conductiviy. Of course, as Ames [2] has pointed our, convergence proofs are not a2 useful as er fer bounds.” Even « computer docs net make it possible to calculate infinitely many terme and when wuneatiog the series one always wonders how good the resulting spprosinate solution is. Compatison of successive Proxinations ie an aid in suck a cese, but even an Proximate solution that seems co be converging may not be converging to the exact solution, The available convergence theorems and etror bounds ace so eeatce that engineer aod applied scientist must usually extrap- late from previously tested results for ethet problems to new sistations when applying approximate methodor D. Comperison to Other Methods Separation of Variables ‘The Galerkin method is related to a wide variewy of other approximate methods as well as to some exact methods of analysis. In particular, it can be shown [29, 22, 24] chat if a problem yields to the method of ‘separation of variables and if the Galerkin method is, applied in a certain way*, then the two solutions are the same, provided the Galerkin method is catried through to completion. Of course in numerical calcula tions, after obtaining an exact solution in the form of an infinite series, one calculates only a finite number of terms as a matter of practical necessity onship between the Galerkin method and the Riez or Rayleigh-Ritz! method when the latter ean be applied [14, 69, 70, 1, 21, 3, 29, 71, 72). In particular, if the same tial functions are used, the resulting calculations are identical, Contrary to a currently prevalent opinion, this equivalence still persists when the trial functions do not satisfy the ‘atural boundary conditions [23, 11, 25, 24], which they faced not do in the Rayleigh-Ritz method. The boundary residual is either added or suberacted to the differential ‘equation residual, and the calculations are again equiv alent to the Ritz or Rayleigh-Ritz method, ‘The choice ‘of adding or subtracting is dictated either by mathemae- ical convenience—part of the differential-equation residual can be integrated by pacts to cancel part of the boundary residual—or by the physics—the differential ‘equation and boundary conditions both come from macro- ‘scopic balances taken over the volume and surface, respectively; these macroscopic balances can be com: bined in only one way, and the residvals are combined in exactly the same way. A very important difference between the Rayleigh-Riez method and the Galerkin method is that in the former some funetional—possibly representing an eigenvalue—is being minimized of maximized. Consequently che approximate values of the functional represent either upper or lower bounds. In the Galerkin method this information ia. missing: exactly the same values would be obtained, but one would not know that these were upper or lower bounds. However, when the variational incegral is of no signifi= cance, the Galerkin method, because of its generality, may be peeferced, ‘The variational and Galeskin meth ‘ods are compared schematically in Figure 1 Most variational principles are merely stationary principles, rather than minimum or maximum principles. TThe approximating function’ in the Galerkin method muse be the cigeafunctions found by che sepatstion of variables snd the Galeskin method ust be applied to the iniial conde tions as well as to she differential equation. Such a serule ly that if the exact solution fx contained in the ‘ial function, che Galerkin method will find “Though there is basically bat single method, ic is com venient to follow the custom (scarcely universal) of distin sulshing beoween the “Rayleigh-Rie methot"” when Ie lied to ‘minimum cr maximum ponciples sad the. “Rite hnethod™” when it is applied to merely sationacy principles. 741a rine oue cate cehtT METAB) wee foe caavusiontirz caLeRKIN wero) “tera In such cases, the Ritz method is again equivalent to the Galerkin method. The caleulations are identical; the results are identical; but in the vatiational method fone knows that the variational integral is being made stationary, i.e., insensitive to changes in the tial so- lotion. If che variational integral has physical signifi- cance and is the quantity of interest, then the varia- tonal methods have an advantage over the Galerkin ‘method even though the answers are the same. Adjoint Var yal Methods Variational principles exist for linear problems only if they are selfadjoine. For non-selfadjoine linear problems, variational principles can be formulated for the original equations and theit adjoints, and again MUR ia related to the corresponding tatiational meth- ods. ‘The impetus for using the adjoin operator in vai- ‘tional formulations seems to stem from Morse and Feshbach [73], who gave a variational principle for the unsteady-state heat conduction equation; Rosssopoulos 74) also gave a variational peineiple for any linear on-selfadjaint problem. Schait (75) and Washiza [76] have applied uch a principle to the unsteady-scare heat conduction equation, while Selengut (77, 78] de~ veloped the idea (or nuclear reactor problems. Macy other examples exist, such as those of Nichols and Bankott (79) for convective diffision of heat, Finlay son [24] for convective diffusion of « maleicomponent mixture of chemical species; Lewins (80), Slattery [81), Flumetfele and Slattery (82) for estensions to nonlinent problems; and many authors (83-94, 38) for nuclear fe- factor and associated problems. In applications to Tinear non-self-adjoin problems, the method of weighted residuals yields the same results as any of these vatia- tional principles as long as the weighting functions for the origiaal equations are taken as the approximating functions for the adjoint, and vice versa (75, 24]. The question then arises as to whether this variational method, which can be regarded as an application of one form of MWR, is preferable to Galerkin's method, which does not require the complication of an auxiliary ad- joint system. There is some evidence that the adjoint ‘variational principle leads to slightly better resules[95~ 97], and Clark and Hansen [98] imply thar the use of adjoint weighting functions might speed convergence. Kaplan and Bewick [38] claim that the variational method is the best strategy in that it gives better an- swers more often. However, they go on to say: “OF course, practical considerations may intro- duce yet another meaning of the work “best’’; namely,’ “‘nost economical.” In this sense we find that the Galerkin method (which uses the trial functions also as weighting functions) is, in most inseances, preferable co the variational’ method, since ie gives results which are almost as good but does not require separate calculations of the weighting functions."" In essence, the adjoint variational method trades in- creased complexity for possibly better results: there still is no clear-cut answer to the question of whether the Galerkin method or the adjoint-variational method is best. For certain initial-value peoblems there may be no difference between the variational method and Galeckia’s method if the semi-direct approach is used in both. Whenever the coresponding steady-state problem is self-adjoint, it is reasonable to expand the unsteady- state solution and its adjoint in terms of the same func tions of position with unknown functions of time as co- elficients: weet Denne) 2) Because the approximating functions are the same for both w and u*, che weighting functions in the adjoint method are the same as those in the Galerkin method. Consequently, the solutions ate identical, whether derived by the adjoint method or the more disect Galerkin method. The adjoint system is also useful for eigenvalue problems. Roberts [99] presents che general theory, and examples can be found in the works of Chandrasekhar [43] and DiPrima (100] as well as others. While intro duction of the agjoine system does increase the com plexity of the problem—particularly the boundary con- ditions—some advantage is gained over the sttaight- forward application of MWR because the eigenvalue is made stationary, and hence insensitive to changes in the wial function. This advantage docs not usually ap- ply to boundary and initial-value problems since the variational incegral is seldom of interest in those cases. 742Method of Least Squares Mika (11 poiats out that for boundary-ralue probe lems the Least-aquaces method for Lu=f inv 33) Bu=0 ons oa is equivalent to applying the calculus of variations co the equation er) G9) ‘eta fos-ma fans 00 ‘among all functions » having the appropriate continuity and differeatiability requirements. The natural bound- ary conditions comesponding to this variational princi ple are of the form [50] Ni(Lu~f=o0 en where Ny ace differential operators. The Euler equation is just equation (35) and the equivalence with equa tion (33) rests with the premise that the equation Lovo G8) Bu =o G9) hhas only the trivial solution, Note that the natural houndary conditions (37) are similar to the compatibility conditions mentioned above in connection with the inte= gral method for magnetohydrodynamic boundary-layer problems. Method of the Local Potential ‘A procedure based on the so-called local potential of Prigogine and Glansdosff [101-104] has been proposed as a variational method for determining approximate so lutions to boundary-value [105-108], eigenvalue [105], land more recently initial-value problems [108], Rosen [110-113] used the same type of computational scheme ccarlier, The actual applications of these methods have been shown to be equivalent to the Galerkin method L114, 241. Moreover, it has been demonstrated that the variational integral is aot stationary in the local poten= tial method and that no minimum principle exists in ap- plications (114, 24]. Consequently, the advantages usually associated with variational principles are miss- ing from the local potential method, which can be re= garded as a disguised application of the Galerkin Imethod, See Ref. 129b for « more detailed critique: Lagrangian Thermodynamics ‘The sovcalled variational methods due to Biot (115 121) and others [122-128, 51) ace also equivalent to the Galerkin method [25, 24}. In these Lagrangian thermo= dynamic methods there is ao variational integral which is being made stationary [25, 79]; their sole signifi- ccance appears to be as means for generating a computa tional scheme, That scheme is, however, identical co the Galetkin method, which is moce straightforward and applicable to a broader range of situations. There is no reason that the Galerkin method should not be pre= ferred, so far as the authors know. See Ref. 129b for a smote detailed critique, E. Applications ‘The general features of MWR in its numerous ver~ sions and various refinements have been presented, and its relationships to certain other approximation methods have been sketched. Which of all these methods are superior, and over just what ranges of circumstances the superiority exists, are matters that can be sectled finally only on the basis of representative applications. More systematic comparative studies and evaluations are needed than have bees reported to date. Until they are forthcoming the investigator of a new problem can expect little more help than he can get out of seeing how others have handed more or less similar problems. References 130-187 have been selected as much to il- lustrate pitfalls, shorecomings, and failures as to cite the attractive features and successes of different ver~ sions of MUR. The preponderance of recent papers ac- curately reflects the upsurge of applications of these methods in one field after another; the emphasis on problems of flow and transport is conditioned by inter ests of the authors. ‘The popularity of the. integral Imethods which originated in,boundary-layer studies {18, 20, 34-36, 136-153, 160-163, 166-169] can be dis Counted in part as a tradition perpetuated by formal in- struction beginning with elementary texts in fluid me- chanics and heat transfer. Beyond any guidance he can get from past experience the problem-solver can look for seassurances in com parisons of different forms of crial solutions and of sue~ cessive approximations in any one form: the appear= ances of convergence with more numerous adjustable parameters and of insensitivity co form of approximating functions do lend confidence to results. So do close ‘matches with established information on special cases tnd limiting cases. Ie is also true that MWR, like vari= ational methods, may yield better estimates of proper ties of the solution at large, such as an integral of igeavalue, than of the solution itself. ‘The main ade vantage and disadvantage of MWR are contained in the same fearure, namely, that the results depend on moce or less arbitrary decisions by the user. latuition, ex perience, any available information all can be rapidly exploited but the reliability of the resules is frequently hazy. Hopefully this review sheds light on the basic issues and will be useful co those interested in apply- ing weighted residual methods and related techniques. F. Areos for Further Research OF the unsolved problems concerning MWR the follow- ing ate most important in che opinion of the authors: (1) Choice of eriterion in MWR. Systematic compacae tive studies using representative (nonlinear) problems are needed. The least-squares procedure for nonlinear problems particularly warrants attention. The Galerkin method and adjoint variational method for linear prob lems need co be compared. 7483(2) Development of rational methods for selecting ap- proximating fonctions, fer example by systematically applying symmetry properties of the problem. A related (qoestion is whether linear combinations of functions fare more effective than nonlinear forms, especially for nonlinear problems (G) Definition of mathematical or engineering criteria for identifying optimal approaches in (1) and (2). Op timization theory should be brought to beat on the de- cisions that have to be made by users of weighted residual and other approximation methods. (@) Derivation of error bounds for approximate solu- tions by MUR, And though they are not as directly wse- ful, convergence proofs are needed, especially for non linear problems. ‘Acknowledgment ‘This review stems from research supported in part by the National Science Foundation chrough a Graduate Fellowship received by B. A. Finlayson and in part by the Air Force Office of Scientific Research chrough AFOSK Grant Number 219-63. References 1 Crandall, S, Hoy “Engineering. analy itl, 1956; AMR 1201955), Rev. 1122. * MeGram- 2 Ames onlineat partial differential equations in engineering,” ‘Academie Press, 1965; AMR 19(I9c0) Reve 724. 3.Collatz, Lay “The numerical teatment of differestial equations," Springer, 1960; AMR T4961), Rev. 1738. 1 Frazer, Re Aw, Jones, W. P. and Ska, 5s, “Approxi= rations to functions and to he solutions of diferctial equa tons,"" Gte Brits Aeto. Rear Council Rept. and. Memos 199 (G937}; reprneed in Ge, Brie, Air Ministry Acro. Ree, Comm. Tech. Rept. 1, 317-549 (1937). 5 Biezeno, CB, and Koch, Js Joy Over eon niovene methode tet tetetcning van vlokee latca met coepassing op Cakcle voor de techaiek belangriike belastingsgevallen, ng. Gran. 38, 25-36(1923). 's Blezeno, C. Bay Over een vereenvoutiging en over een iebeiding van de methode van Rite, Civitiaan Huygens 3, 89 (1925~1924). Blezeno, C. B., Graphical and aumerical methods for solving stress problemi, in roc. First Interns Conge- Apple Meche, Delft 1924, pp. 3-17. Counni, Re Remark on “weighted averagen of the residual” in discussion following Biezeno's paper in. Pros. ese Inger, Congr. Apple Mech Delf 1924, pg- = ‘9 Picone, My Sul ctode delle miaime potenze ponderace «sul metedo di Rts per il calcolo appronsimato net problem ella Hsicarmateratiea, Rend. Circ. Mat, Palermo 52, 225- 253 (1928). LO Galerkin, Be G., Seetshoi i plastinys Ryady ¥ neko- tarykh‘voprossich uprogogo tenoveniys serahaei i plastin (Reds) and plates. "Series sccuting in some, probleme of Slaotic ‘equlibtiam of rods and. plates}, Vestn, Inchen # Tekh. Petrograd 19, 897-908 (1915); Translation 63-18924, Clearinghouse Fed. Sci. Tech. Info. ‘See. algo: On the sevensieth anniversary of che bisth of B. G. Galerkin (ia English wich a Tist of his publications), Prikl, Mat, Wok. 5, 337341 19D). 11 Mikdlin, 'S.'Gy, "Vasiacional mechods in mathematical physics,” Perganon Macmillan, 1954 12 Bickley, WU. Gy. Experiments in approximating to the solution of a partial diferensial equation, Phil Bags (7) 32, 50-56(1941). T3Eringen, A. Cu, Transverse impact on beans and plates, J Appl Mech. 20, 461466 1953). 1A'Dancan, W Ju, "the principles of Galerkin's methods Gu, Brie Aero. Resr Council Rept. and Memos 1848 (1938); 744 teprinted in Ge. Brit. Ais Ministry Aero, Res. Comm. Tech. epee 2, 569=612(1938). 15 Sani, Re L, "Convergence of a generalized Galeekia method for censin fluid stability. problems,” Rensselaer Polyeechaie Inst. Meth. Repts ne. 63 (1964). 16 Yamada, Hoy Repte Res Ins Fluid Engngry Kyushu Unies 3°29 (194); refered to by H. Fujita in’ Rel. 129. Also: 8 method of approsimate incegration of the lamiaat boundary-layer equation, Repl. Ress Inst. Fluid Engnsy Kyushu Univ, 6, 87298 (1950); AMR $(1952), Revs 488. T7Krehal, Mo F., Application of the method of mo iments to the solution of lige diferential and integral equ Hone (in Ukrainian), Kew. Soobebeb. Akad. Nauk URSSR 1, {6H i952); eelereed to by B. Shuleshko ip Ret, 21 and Le V+ Kenrorovich sad V. I Keylov in Ref, 29. Another citation io sven by S.G. Miki in Rete I. ‘le Pohlhusen, Ke, Zar ndherangeweizen Integration der Ditferentalysleichong det laminae ca Grenzschichey ZAMS Ty 252626801921), 19 Von Kéemén, They Uber laminare und curbulente Rei- bung, ZANT, 253-253 (1921), NACA ‘Tech. Memo. 1092 38. 30 Goodawa, T. Ru, Application of integral methods co transient nonlinear heat transfer, io "Ade. Heat Transfer” 1, Ss122(1964); AMR 1811965), Rev. 3009. Zi Shuleshico, Pa,_A new method of solving boundacy- valve problems of mathematical physics, Generelizations of Frevious methods "of “solving. boundary-alve problems of Rathematical physics, Ausinal Je Appl. Sel, W, 1-7, 816 (1959) AME 13(1960), Rew. 3 22 Snyder, Le Jn Sprigae, T. W. and Stewart, W. F., Solue toa of the equations of change by Galerkin's method, AICHE J. 10, 535-540 (1964). 25 Bolotin, Ve ‘Theory of Elastic. Stability, 1. 60; AMR. 16(1963), devs 46. 24 Finlayson, By An, "Appeoximate solutions of equar tions of changer Convective instability by active streas,” PLD. thesis, Uaiv. of Mianesora, 1955. 25 Finlayson, B. A. and Scriven, L. E., The method of weighted residuals and ite relation to ceria. variational Principles "for the analysis of taasport processee, Chem. Engng. Sei. 20, 395-404 (1965), ‘2Epancan, W. Jo MGaleckia's method in mechanics aod diferent equations," Gr. Brit. Aero. Res. Council Repts find. Memo. 1798 (1987); reprinted in Ge. Beit Air Miaistey ‘Reto. Res. Comm. Tech. Repe. 1, 4847516 (1937). 27 Hugelman, R. D. and Haworth, D. R., Aa MHD bound- axyelayer “compatibility condition, “AIAA Je 3, 1367-1369 (1965), AMI. T9(1966), Rev. 2539. 28 Hugelana, Rs Dey Application of Poblbausen’s method to stagnation poioe How, ALAA Je 3, 215862159. (1963): ‘AMM 19(1966), Revs 3014. 29 Kantorovich, LV. and Keylory Vs by rethods “in higher anaiysie,” lotetscience, 1211959), Rew. 4830. 30 Kantsroviehy Ls Vin A direct method of solving the problem of che minimum of a double integral (in Russian), Fees Abed Nauk SSSH 5, 647-552 (1953). 31 Poriaky, a, the seduction of the solution of cet= dilterential equations to ordinary. differeatal fqusclons,"" in Trans, Fifth lacers Congre Apple Meche, Cambridge, Mess., 1938) ppe 00-707. 32 Kestorovich, Vor Application of Galeckin’s method to she sovcalled procedure of teductien to ordinary differen tial "equations (la Russiag with English summafy). Prikl Mat Mek 6, 5110 (1942). SB Wright, Kip Chebyshev collocation methods for etdix sary diferential equations, Comput. Jo 6, 358-365 (1964). 34 Yang, Ke T-, "Calculation of unsteedy heat condve- tion im a single layer and composite finite slabs with and Without property variations by ea improved integral proce Sure," in Intern. Developments in Heat Teaaster, ASAE (1961), pp. 18-27. 35'Yang, K Ts, An improved integral procedure for com presrible laminae boundary layer analy, J- Appts Meche 28, 5520 (1961); AMR 14(1961), Revs 6188. Nonconservative Problems of the Pergamos/Macrillan, 1963, ‘Approsiimace 1958; ANT36 Yang, K. To, Laminas forced convection of liguids ia aber with variable viscosity, Trans: ASHE, J. Heat Frenajor 84, 3532362 (1962). 37 Kaplan, So, Some new methods of flax syathesis, Nucl, Set, Engne, 13, 23-31 (1962). 38 Kaplan, 3. aad Bewick, J. A. "'Space and tine syathee sis by the variational method.” Beis Techns Review, Rez fctor” Technology, 27=44, WAPD-BT-28, (1963), Cleatings houre Fed. Sel feeb. Info, 39 Kaplan, Su, Maciowe, O. Ja, and Bewick, J., Applica tion of synthesis techniques to pcblems involving tae pendence, Nucl. Seu Engng. 18, 163-176 (1964). 40 Snjder, Ly J. and Stewart, W. fy Velocity and pres sure profiles for Neweonian creeping flow in a regulat pocked bed of spheres, AICBE J. 12, 187A173 (1966), AMR TPUIDGC), Rev. 6002. “U1 Heywood, J. Bs and Motfate, W. Cu) Validity of inte= ‘gral methols in MHD boundary layer analyses, AUAA Jo'3, Tsos-1567 1963). 42 Reid, W. Hh and Hanis, D. Ha On orthogonal fanc~ tions which satisfy four boundary conditions, Astphys. Jo, Suppl’, 429-452 (1958). “43 Chandrasehas, S., "Hydsodynamie and hydromagnesic sebbiliey,” Clarendon Press, 1561. 44 Riueger, E- R. and Dibtimay Re Co, The stability of « viscous fluld between rotating cyliniess with an atial Cow, J, Fluid Mech. 19, 528-538 (1564 45 Lardner, . J. and Pobley F.W., Application of the heatcalanece integral co probleme of yipdefeal peomety Appl. Mech, 28, 310-312 (961); AMR 13(1962), Rew. 2660. 46 Falk, Su, Das Verfahren von Rayleigh Rice mie heamieschen Inverpolationspolysomen, ZAMM 43, 149=166 (1963. 47 Caldwell, W. R., Kaplan, S. end Marlowe, 0. J) Equations” and programs for synthesis approxitat ions, WAPD-T¥.-377 (1963), Clearinghouse Fed. Seis Techs Infor 47a Van Dyke, No, “Perturbation methods. in fia. me: chanics,” Academie Bressy 1964 AMR 19(1966), Rev. 1710. “8 Tharsen, Mand Landis, F., Integra! methods in transient heat’ conduction problem with nemuniferm initial conditions, Toternat J. Meat Mass Transfer 8, 185-192 (1965); A 190966}, Rew, 3145. 49 Collings, Ws 2) "The method of undetermined fane~ tions as. applied to nonlinear difvasion problema, MAME. thesis, Univ. of Delaware, 1962. 50'Becker, Ma, "The principles and applications of vari- ational methods," Res. Monogr. no. 27, MIT Press, 19¢i. 51 Citron, S.J. A’ gate on the relation of Bist"s method in heat conduction tom least-squares. procedure, J. fero/, shace Sei, 27, 317318 (1960); AMR T3960), Rev. 5951, 52 Hildeteind, FB., "Methods of applied mathematien Prentice-Hall, 1954. 53 Kawaguti, M., The crivical Reynolds nunber for the low past a sphere, Rept. Inst. Sci Technol, Tobyo 2, 66 (1948); English version in Je Phys. Socs Japs Vp 604-699 4955), 54 Beloteerkovskii, 0. ML and Chashki Pe Ie, "The aur ierical method of integral relations," NASA Tech, Transl, 8356 (1964); AMR 18(1965), Rew. 2578. 55 Repman, Yu. Vo, A problem in the mathematical foun- dstions of Galerkin's ‘method for solving, problems on the stability of elastic systems (ia Russian with French sume faary), Prikl Mat. Mekh, 4,2, So6 1940). 556 Petrov, G. 1, Application of Galetkin’s method 10 & problem involving sicady How of a viscous fluid, (in Russian ith English surmaty), Prikl, Mat. Motor 4, 3, 3-12 (1940), 57 Keldysh, Ie Vor On By Gr Galeikin's method for the solution of boundary-value problems,” I2us Akad Nauk SSSR, Ser ‘Mat 6," 309-830 (1942), NASA Tech. ‘Traele F195 1964. SS'DIPrina, R. C. and Sani, Re La, The convergence of ‘he Galeskin method for the Tayler-Dean stability problem, Quart Apple Matb. 23, 183-187 (1965). 59 Faedo, S., Un nuovo mecalo per Ysnalisi esistenziale f¢ quantitative dei problem! di peepagacioni, Anne Scuote Normale Superiore ai Pesa (3) Y, I=ll 919), GOGreen, J. We, An expansion method for parabolic par- tial differential equations, J. Res 51, 1270132 (1953) 61 Hopf, Es, Uber die Anfangswerauigae fir dle hydeom Ayaamschen Grundgleichuagen, Math ‘Nach. 213-231 950/195). 62,IVin, A. Me, Kalashaikoy, A. S. and Oleinik, 0. A Second-order linear equations of parabolic sype, Russian Mathe Surveys 17, 1144 962). 63 Meister, Bh, Die Antangswercaufgabe fr die Ssiungs- Aifferetiagleichanges des Taylorschen. Stabiieatsprotlem, ‘Arch. Rational Mech. Anal. 4, 81-107 (1963), AMR T7964), Rev. 3964. 4 Kadner, H., Unsersuchungen zur Kollckationsmethade, ZAMM 40, 99-113 960}, AMR T4(1961), Reve 173° 65 Keavehuk, M. Fo, On the convergence of the method of mmoments for partial diffetental equations (in Ukrainian), Z Ita Mat.» Abad Nauk URSSR 1, 25-27 (1936); telesed to by LIN, Kantorovich and V. 1 Krylov in Rete 29. 66Kesnosel'ski, M. Ay, "Topological methods in the theory of nonlinear integral equations," Pergamon Macmillan 1964, 67 Glansdertt, P., “Mechanical flow processes and varie ational methods based on fluctuation thecry.” Fhe 45-54 in Rete 6 ‘66 Donnelly, R Jo, Hermao, R. and Prigogine, “Non-eqllibritie ‘thermodynamics, earlational tod seability,” Univ, of Chicago Press, 1966. © Biezeno, ‘Cz B- and. Granmel, Rey “Engincesing Gynaniis, Vel. 1: Theery of elasticity,” Biackie and Som, 1555; pg 173. To'Sckolnikof, 1. So, "Mathematical theory of elastic- lays MeGraw-Hill, 1946, pg. 413. 71 Singer, Joy On the equivalence of the Galerkin ani RayleighRice methods, Jv Roy. Aeros Soc. 66, 592 (1962), ‘AMR 18(1963), Rew. 1331. 72 Reiss; "E. La, Supplement hound with “Problems. of mathematical physics!" by Lebedev, Ny Noy et aly Preatce= Hall, 1965, entitled "Kantorovieh's Mechot," pp. 401-413, 73 Mocse, P- Me and Feshbach, H, "Methods of theoretl- cal physics,” MeGrawHill, 1955, Chapt. 3+ TA Roussopoulos, P., Mthodes variacionnelles en ehgeele Ades collisions, Compe rondus 736, 18sb=1860 (1953). 15 Sebmit, Ly Aa, “Application of the variational method, the Galeskin technique, and normal coordigates iat ttansiend temperature distribation problem,”” WADI Tech. Repts 36-207 (1956), 4D:97326, Clearinghouse Fed. Sei. Teck, Infos 76 Washisu, K., “Application ofthe variational method co masient heat condvetion problems,"" MIT Techy Repes 25-17 (1955), AD-39951, Clearinghouse Fed, Sci Techy Infos "7 Selengut, De So,'*Wasiationsl analysis of mulidimen- sional systems,” ‘in Nucl. Physies Res. Quart. Rept. fr OctDecsy 89=124 (1958); HW59126 (20 Jan 1959). ‘alSelengut, D. Ss» On the desivation of a variational principle for linear systems, Nucl. Sei. Engng. 17, 310-311 Gis63. 79 Nichols, R. A. and Bankolf,S. Ge, Adjoine variational inciples oe conrective diffusion, Internat, . Heat Mass Transfer 8, 329-335 (1965); AM 18(1965), Rew, 5880. ‘0 Lewing, Ju, A variational peineiple for nonlinear 5y3° Kem, Nucl. Sci Bnpngs 12, 10-14 (1962) 81 Slatiry, J. Co, A widely applicable type of variational negra. ly Chem: Ehgngy Sei 18. 801-806 (1968). 82 Flamerfelt, Ri W. and Slattery, Ju Coy A widely ap plicable type of variational integra Il, Chem. Engne., Set 30, 157=163 (1965}- 83 Lippmann, B. A. and Schwinger, Jo, Vasiational peinci= ples for scattering processes. I" Phys, Reus 79, 40480 (i950. 14 Comgotd, homogeneous media, Proc. Phys. Soe cass. 85 Lewins, Jo, The timedepentent importance of nese trong and precursors, Nucl, Sets Engng: 7, 268-274 (1960). 16 Lewina, J Variational repreacatitions ia. teactor Physics derived irom a physical principle, Nucls Sei, Enenge 8, 95=104 1960). 87 Dougherty, D. B. and Shen, C. N, Natt, Bur Side. (53) editors, Na Slowing down of neutrons in infiaite ‘Londo A70,_ 793-801 The spacetine 745eutton Kinetic equations obtained by the semindivect varis~ Sonal method, Nucl. Set Enns 13, 141=148 (1962) 18 Goldstein, R- and Coben, Kz Re, Theory of resonance absoeption of neurons, Nucl, Sei Engng. V3, 132-140 (1962). 1 Pommaning, Ge Cx and Clay Mer Jon The variational method applied €0 the nonoenergetic Boltzmann equation. ‘and Il, Nuele Sets Engng. 16, 147-154, 155-164 01963). 90'Lewine, Jy The weighting functions of reactor phys ies, J. Brit Nucl, Energy Soc. 2, 326-334 (1963). ‘Sl Gast, P. F., A variational peinciple fee heterogeneous resonance ‘capture. as a basis. foe approximation methods, [Nate Seis Engng. 19, 196-202 (1964). 52 Mika, Ju R,, On the vatiational method applied to the rronoenergetie Boltzman equation, Nucl Selr Engng. 19, 377> Ss (1964). 53 Lewins, Jo, Tiecdependent variational peiaciples foe onconservative systems, Neck Sch Engng. 20, 517-520 96a. 94 Lewios, Jo, “Importance: the adjoint function," Per sane, 1965. 3S Fischer, P. Ga Group parameters for few-eroup pet turbation caleulationsy Trans, Amer: Nuch Socr 2 a0. 2, 164 (1939); Summary 16-4. ‘30 Calames Ge Be aad Fedetighi, Fs Ds, A variational procedure for determining spatially dependent chermal spec Pim Nel. Ses, Engng, Ws 190-201 (1961)- 97 Kaplas, Sur On the best method for choosing the ncighting furctidas in the method of ‘weighted residuals, Transs Aimer Nucl. Soa, 6 3-4 Grane 1963) ‘96 Clark, Me'and Hansen, Ke Fey "Numerical methods of reactor analyaies” Academic Pree, 1964. 99 Roberts, P. Ty Characteristic value problems posed by differential eguations arising ia hydrodynamics aad hyo ragacticn,J- Mat. Anal. Applic. I, 195-214 (1960) 100 DiPrima, R- Cy Some variational peincipler for peobe lems in bydeodymamic and hydromagaetic. stability, Quart Apple Math. 18, 375-385 (1961), AME TA{L961),, Rev. 4462. TOL Glansderf, Pay Prigogine, ly and Hays, De Foy Vario sional properties ofa’ viscous ligld at « nomunifora tempera ture; Phynies of Flutds 5, 144-149 (1982). 102 Glansdortf, P. and Prigogine, In, On a general evolu tion criterion in macroscopic physics, Physica 30, 351-374 (1964), 103 Prigogine, 1 ties and fluctuation thecey, Physica 3i 104 Prigogine, I., Evolttion ceteris ter, and fuetaation’, ppr 16 in Rel. 68. 13 Schechter, Re'Ssy/ On vasiaional principle foe the ReinersRivlin fluids’ Chem. Prem. Sei 17, 805-606 (1962). Too Hayes, Dr E., A. variational formuiation applied 20 Cosette and'Poiseuille fom, Acai Roy. Belge, Bull. Classe Seu. 49, 576-602 (1963), AMR 17(1964), Res. 4702. ‘io? Glansdertt, Px, Dekeuster, Es and Haezerdonck, Yo, Application du potentiel lacal de production deneropie & ix fetherche de distributions statioanaites de cempératares et Se concentrations, aca Roy. Belg., Bull: Classe Set. 5, 0-73 960. 108 Hays, D. F., Application of « variational formulation fcc time-dependent’ and. steady-state solutions. of the heat ‘Squations temperataceesependent thermal conductivity, ppe17= Bin Ret. 6s. 109 Schecheer, Ry and Himmelblau, D. M., Local poten- tial and syetert stability, Phyaice of Fluide 6, La3IC1437 (56s). THO Rosen, Ps, On variational principles for iereversible processes, J. Chem. Phys. 21, 1220-1221 (1953) Tii Rosen, Ber Use of reetsicted vasiational principles for the solution of differential equations, Je APPL. Pbys: 25, 5360838 (1954. 712 Rosen, P., The solution of the Holtsmana equation for a shock wave using a ieetieved vatiational priniple, J Chem Physr 22, 1045-1049 (1954). 113 Rowen, Ply. Vatlationt! approtch to magnerohydro- aynatiies, Physice of Fluide 1, 291 (1958), AMR 120939), Rev. 1031. 114 Finlayson, B, As and Scriven, L- Ey Galetkia’s method and the local potential, pp. 291-294 ia Rel. 68. 10d Glansdortt, Psy Variational proper- 124201256 (1965). varisclonal_ proper 115 Biot, M. Au, New methods in heat flow analysis with application’ «0 flight stevetures, Jo Aero. Sei 24, 857-873, 957. T16 Biot, Me Ay Further developments of new methods in heaelow analysis, J. Aer0/Space Set 26, 367-381 (1959); AMEC131960), Ree. 1376. 11? Biot, Nk As, Theemedyeamice and heat-flow analysis by Lagrangian methods,"" in Seventh Anglo-American Acco ‘astical Conference, Inst. Aeso. Sei. 1960, Dp. 418~131. 118 Bir, M. As, Variational aad Lageangian thermodynane ice of theseal coavection’ fundamestal shortcomings of tbe hheaceranster coefficient, J. Aero/Space Sct 29," 103-106 (1562). Ti9"Bioe, M. A., Lagrangian thermodynamics of heat trans~ fer in systema including Fluid motion, J. Aero/Space Set. 29, 560-579 (1962); AM 16(1963), Res. 3413. 120 Blot, Me A. and Daughaday, ty Variational analysis of ‘ablation, Ji. ‘Aero/ Space’ Sci. 39,"227-229 (1962); AMR. 15(1962), Rev. 5487. 121 Biot, Me Ar and Agrawal, H. Co, Variational analysis of ablation for variable properties, Trans. ASME, J Heat Transfer 66, 437-142 (1964); AMR 16(1965), Rev. 4989. 123 Nigttm, S.D. and Agrawal, He Cu, A'vatiational peinci= ple foe coavection of heat. ly Je Math Mech, 9, 869-872 &50y; ut 1411961), Rew. 4983. 123 Agsawaly He Ci A atlational principle for convection of heaty Il, J. Math. flack 9, 872-884 (1960); AMR 14(1961), Rew. 4982, 134 Gupea, S. C., Variational principle fr fally developed laminar heat eransler ia uniform channels, Appl. Sel. Res AIO, B5o10r (196). 125 Gupea, Ss C.y A variational principle fr convection of heat ia anisotropic. media, Appl. Se. Res. AND, 2299234 (196195 AMR T5(1962), Revs 1680. 126 Guptay S. Cu, Avlenecaquites procedure forthe varia~ tional Principle’ for héat conduction” In anisotopic. medi, ‘Appl "Sei. Res. AIT, 3119312 (1963); AMR 17(1964), Rew. 2865. 12? Larder, T. Jy DBiot’s variational principle in heat conduction, AIAA j. 1" 196-206 (1983). 128 Taos Ls No, On she variational principle and Lagrange ‘equations "in "stoliee of gasdynamic’Iubrication, J. APP Mech 31, 45-46 (1964); AMR. 18(1965), Rev. 7173. 129 Fujita, Hay Diffasion with « sharp moving boundary, J. Chem, Physe 2h, 700-705 (1953). 139s Fialaysoa, B.A. and Scriven, L. E., Upper and lower bounds for solutions ro the teansport equation, AICHE J. 1298 Finlayson, B.-A. aod Seriveny Lz Ex, "On the search for vatiational principles,” to be published. Additional References to Applications. Non! Galetkin method: Refs, 13, 2, 53, 106-108, 115, 121 and 130 Fung, Y. C., On ewordinensional pane) Hutte, J Acro. Set. 25, 145160 (1938)e Ti Lowid, Av and Tebekimerey, As, Sue la solution ap= prosiaacive de quelques problenes noniinésizes de vibrations year Problems Teseces’” (in Russian with French summary), Prikl Mat Mekh 1, p. 324 (1937). 132 Banov, D. Jn, On an application of Galetkin'e method to ceria noalinen’ problems’ in the theory of elasticiey (in Russian with English sunmary), Prikl. Mar. Mokh, 3, 139-142 939). 138 Richardeon, P. Dsy Unsteady one-dimensional heat conduction with e noalineer Boundary condition, [+ Meat ‘Transfer 86, 298-299 (1964). 134 Schwesinger, Ca; On one-term appcoxiamtions of forced nocharmonie vibrations, J. Appl. Mech 17, 202-208 (1950). 135 Tanner, Re ly "Tae extination of Bearing Toads using Galerkia's method," Internat. Je Mech. Set. 3, 13927 (961), ‘AMR 151962}, Rev. 1222. Incegeal methods Refs. 20, 4-36, and 746136 Activos, Aw, Shaby M. J. and Petersen, E Es, Oo the solution of the twoslimensional boundary layer’ fisw sqeee {loos Tera non-Newtonian power law fd, Chom, Enpon, boc 20) T01=105 (1965). 137 Dizzell, G. D, and Slareey, J. Cx, NomNewtonian boundary layet flow, Chem Engng. Schv W." 7710185 tise 138 Bogue, D. C., Entrance effects and prediction of bulence in noa-Newionian flow, Ind. Engngs Chem 3), Boke 878 1955). 139 Campbell, W. D. aod Slattery, J. C., Flow in the ene trance of @ tube, Trans, ASME, J. Basic’ Engng 85, slic (1963); AMR 17(1364), Rew. 3956, 140 Goodman, T. Ry The heatcbalance integral and ite p= plicatien to_ problems involving # change int pluses ‘Trane ASME 80, 3356342 (958). 14 Goaiman, "Te Rey “The heat-bolance integral: futher goosideracions and eelinementa,™ Trans, ASME, [teat Transfer 83, 83-86 (1961); AMR T4096), Reve 4363. 142 Gosdnaay TR. and Sheay J. J. The melting of finite slabs, Te Appl Mech. 27, \on3i (i960); AMR “IKL9so8 Rew ioe. 145 Kapur, J. N. snd Gupta, R. C., Twostinensional flow of powerslaw iluids ia the inlet length of strsight channer, 4 and I, Zasil 43, 1359141. (1963), 44, 2974984 Cadets AMR 18(1965), Rew. #530, 144 Kapar, J. N. and Srivastava, A, Co, Axially symmetic ‘and “tworlineasional stagnation ‘point flows of 2 ceeeia iscorelaatic Hid, J. Phys, Son Japen 16, iets (dean AMR 17(1964), Rev. 1530. 145 Lemlich, 'R. and Steighamp, J. S.y Laminer natuenl gonvection to n isothermal flat pate mith « spatially sagre ing acceleration, AICHE J- 10, 465-447 (1964); AMR 1BC965) Rev, J02t. N46 Rajeshwari, G. K. and Rathos, S. Le, Flow of » pare ticular ‘class of non-Newtonian viscoelastic ead wikeee inelastic Hfsids near a stagnation pein; ZAM? "13, “G5 (1962); AMR 17.964), Rev. 872. 147 Rosenhead, ‘Lt, editor, "Laminar boundary layers,” Clarendon Press, 1965, esp. Chapt, 6 148 Schlichting, H., “Boundary layer theory," 4th edna, McGraw-Hill, 1960, expe Chapte 12 149 Siegel, R. aad Spactow, E. M., Simsltaneous devetop= beat of velocity and temperatute distributions ins fst deee ith uniform wall heating, AICHE J. 573-99 (1959), AMR T3960), Rev. 405. 130 Sowerby, Li, Am application of Pohlhausen’s mechod to s problem of flow along an edge, App Ses Res AVR, 417424 1964); AMR 18(1965), Reve 16476 151 Stivastava, Ay Coy The’ flow of a non-Newioniaa Higuid| nears stagnation point, ZAMP 9, 80-84 (1958). 152 Thwaices, Bu, Approximate calevlation of the laminar oundacy layers Aoronaute Quart. 1, 215°260 (1949). 155 Yang, K. T. and Saeweayk, Avy An opprenvnate teat iment of unsteady heae conduction in seminlinive solide atch variable thermal properties, ‘Trans, ASME, fr Heat Teantioe Bi, 251-252 (1959); AMR 131960)" Rev. 6304, Subdomain method: Ref. 54 and 154 Carter, L. F. and Gill, W. Noy Asymptotic solution for combined fee and forced convection in vertical sod hose, Zonal conduits with uniform suction and blowing, ATCHE 10, 330339 (1964) 135 Launder, B. Es, Aa improved Poblhausentype method of salewlating the twordimensional boundary layer fa & sees suse gradients" Trans, ASME, Jr Heat Transfer 86, 360-364 (1964) AMR 18(1965), Rew, 1004, 156 Doratnitsyn, A. AL, "On a method of mumerical sclue ion of some ‘nowlincar’ problems in serosydredyannice Actes Ninéh Inter. Congr. Appl. Mech, Brassclor 1 495 a9, 157 Doradnitsyny A. Ac, "A contribution tothe solstion of sized problems of transonie aerodynamics," in Advances st the Aeronautical Sciences, Vol. 2 of the Procy Fiest cere Congr. Acton Sein, Maddy 1958, Pergamon, 1959, ppc See Bee Leastrequaces method: Refs. 50 ad 134. Method of momenes: Refs, 16, 129 and 138 Poots, Gu, An approcinate srestment of @ heat eon luction problem involving a. two-dimensional, solidification font, Internat J. Heat Mase Transfer 3, 339-348 (190d) AMR 1601965), Rew. 433. 159 Poager, Su, Transport processes with chemical reac tons, Chem Engng. Proge, Symposs Series 55, nor 2, 11-32 1959}. Linear Initicl-value Problems Galerkin method: Refs. 12, 25, 39, 60, 63, 108, 115. Incegeal method: Refs. 34, 48, 147 and 160 Goolnan, T. Ray The heating of slabs with arbitrary heat igput, Je Aero/Space Sei. 26, 187-188 (1959), AMA 1211959), Res. 4095. 161 Lyman, F. A., Heat sranefer at a. stagnation poine when the feecstzeamtempernture is suddenly increased, ‘Ap Set. Res. ATS, 65-80 (1964)4 AMR 1B(1905), Revs 1900. 362 Sparrow, E. by Unsteady seagnacion-point heat ase: fer, NASA TN'D-79 (1959); AMR 131900), Revs 537 465 Sparow, Es M. and Siegel, R., "Thermal entance te= ica of a circular tube under tansicat heating codieens™ Proc. Thied US. Natl Congr, Appl. Mechs 1958, Providence, ROL, pp. 817-826; AMR 1201939), Rev. 5207, Leastaquaces method: Refs, 12, 50, 51. Method of momeats: Refs. 12, 48, 129. Collocation: Ret. 12, Linear Boundary-value Problems Galerkin method: Refs. 4, 37, 40, 112, 120, 124, 127, and 164 Kaplan, S. and Marlowe, 0, Jay Application of aynthe- sis approximations to theee-dimensioaal, depletion cocwoc Kons and co cell theary, Trans. cmon, Nucl Sos 6, Ssicusy (Now. 1963). 163 Wemick, R. J. and Pan, C. H. T., Static and dynamic haracteristics of scli-ncting,"patial arc, sae jouroal bear ings, Trans. ASME, J. Basic’ Enang, 88, 405=413 (Uc AMR 17(1864), Rew. 7587. lavegral method: Refs, 147, 148 and 166 Devan, L. and Obersi, Me Mey Apptoximate solution of second-order boundary-layer equationsy MAA]. % 1858 8b (360); ANIC 18965), Reve 2286. 167 Libby, P. Au, Monduchow, M. and Bloom, Me, “Critical study of integral mechode in conpeessible lamina’ boundacy layers,” NACA TN'3655 (1952). 168 Morduchow, M., On laminar flow chrough a channel or tube with injection: application of the meshed of averages, Quart Apple Math 14, 361-368 (1957). 165 Reckenstein, Eu, On mass transfer in the coatiauous phase fiom spherical bubbles or deopo, ‘Chem Euguan Soe Ios 11-146 960). Method of moments: 170 Beek, W. J. and Kramers, Hs, Mass. teansfet with change in interfacial atea, Chom Engng, Sct. Io, S0S561 38d. Collocation method: Ref. 4 and 171 Conway, H. Day The approximate analysis of ceria boundary-malue’ problems, ‘Trans. ASME 82, 275-277 acon AMRC13(1960), Rew. 139. 172 Jain, M. Ko, An extremal point coligcation metho for fluid tow problems, Appl Sei. Res. AML. 177188 (1360, AMR 1611965), Rev. 4991. 173 Jeltreys, ay The surface elevation in celtulae con ection, Quart. J. Mech Appl. Math 4, 285¢288 (1951) 747Galerkin method: Refs 10, 28, 43, 44, 56, $8 100, 130, and 174 Ashley, Ms and Zartarian, Gu, Piston theory: a gew seralyaamie tool for the aeroelastician, Aero. Sci 23, Nopeti18 1956). 175 Chenky HS. and Pan, C. My Toy Stability analysis of gatlutricaeed, selfncting, plata, cylindeieal journal bear ‘ge of tinice lengehy using Galertin's method, Trans. ASME, J Basic Engnge 87, 185-192 0963). 176 Cheng, He 5, and Trompler, P. Ray Stabiliey of the high-speed Journal bearing uader steady load, Trans, ASME, J. Boggs Indust 85, 2742280 (1963)- LIF DiPrima, 1 C., Application of che Galerkin method co probleme. in hydrodynamic, stability, Quart Apple Maths 13, 55-62 11955). Tra DiPeima, Ry Co, The stabiliey of viscous flow berween rotating coaceotrc eVlindere with 9 pressuwe gradiege acting found the cylinders, fo Plaid Mech, 6, 462-468 (1959), AME 15(19@), Rew. 283. 179 DiPrima, Re Co, The stability of viscous fluid beeween rocating cylinders with an axial flow, J. laid Mech. 9, 621~ G31 C960} AMR 14961), Rev. 4272. 180 DiPrima, Re Ca, Stability of noneceationally syrmetsic disturbances for viscous flow between rotating e¥lioiers, Dye: Fluide 4, 7512738 (1961); AMR 15(1962), Revs 1530 181 DiPsima, R. C. and Pan, C,H. T., The stability of flow between concentric cylindrical surlaces with ciealar tiagactic Held, ZAMP. 15, 3606867 (1963), AMR 18(1965), Revs 2426. Ts2 Duncan, W. Ju, Note on Galerkin's method for the treatment of problems conceming elastic bodies, Phil. Mag. 25, 628683 (1938). 165 Duncany B, Jy Applications of the Galeskia method co the torsion aid ficzure ol cylinders and prisms, Phil. Mag. (125, B4-69 1238) 184 Duncan, Ws Je and Lyon, He Moy “Torsional oscila~ tions of a santiever when the stiffness i= of composite Senin,” Ge, Brit. Aero. Res. Council Rept, aad Memos 1809 (ios7h, sepsloced in Ge. Brie) Air Maistey Aero. Res. Comm. ‘Tech, Repes 1, 471-485 (1937) 185 Hedgepethy Je My On che fluter of panels at high Mach numbers, J."Aera. Sei. 23, 09 (1956). 186 Hedgepetiy J. May Flatter of sectangular simply supe ported panels at high supersoaic velocities, [+ Aero, Set. 24 Seresr8 (1957). Ts? Racayyay Jo Vo Je Aero/Space’ Sci, 38, Row. 7392. Flutes analysis of circular panels, 340539. (1962); AMR 1511962), Lease-squares method: Refo, 4, 50, Gt. Collocation method: Refs Gh 748