Existence and Uniqueness Theorems for Formal Power Series Solutions of Analytic Di erential Systems

C.J. Rust, G.J. Reid and A.D. Wittkopf Centre for Experimental and Constructive Mathematics Simon Fraser University Burnaby, B.C. Canada. V5A 1S6
Abstract
We present Existence and Uniqueness Theorems for formal power series solutions of analytic systems of pde in a certain form. This form can be obtained by a nite number of di erentiations and eliminations of the original system, and allows its formal power series solutions to be computed in an algorithmic fashion. The resulting reduced involutive form (rif0 form) produced by our rif0 algorithm is a generalization of the classical form of Riquier and Janet, and that of Cauchy{ Kovalevskaya. We weaken the assumption of linearity in the highest derivatives in those approaches to allow for systems which are nonlinear in their highest derivatives. A new formal development of Riquier's theory is given, with proofs, modeled after those in Grobner Basis Theory. For the nonlinear theory, the concept of relative Riquier Bases is introduced. This allows for the easy extension of ideas from the linear to the nonlinear theory. The essential idea is that an arbitrary nonlinear system can be written (after di erentiation if necessary), as a system which is linear in its highest derivatives, and a constraint system, which is nonlinear in its highest derivatives. Our theorems are applied to several examples.

1 Introduction
The Cauchy{Kovalevskaya Theorem 11] gives an Existence and Uniqueness Theorem for analytic solutions to systems of analytic pde in a certain form. For example, it applies to a pde of form uyy = F (u; ux ; uy ; uxx ; uxy ) subject to initial conditions u(x; y0 ) = f (x); uy (x; y0 ) = g(x) where f; g; F are analytic functions of their arguments. Riquier loosened the restriction of Cauchy{Kovalevskaya's form, which required that it has the same number of unknown functions, as equations. For example, the classical Riquier Existence and Uniqueness Theorem 16] can, under suitable conditions, be applied to the system uyy = G(u; ux ; uy ), uxy = H (u; ux ; uy ), for analytic G; H , while the Cauchy{ Kovalevskaya Theorem 11] can not. The Riquier-Janet Theory retained the restriction that each pde is written in solved form for one of its highest derivatives. In this paper we weaken the solved form requirements to obtain Existence and Uniqueness Theorems for formal power series solutions of analytic di erential systems that are nonlinear in their highest derivatives. The theoretical development of this paper parallels that of Grobner Basis Theory. In Section 2 we introduce nota-

tion, rankings and give our de nition of (di erential) reduction. In Section 3 we give a new development of essential aspects of the Riquier{Janet Theory. Our approach avoids a notoriously complicated process for identifying integrability conditions in the classical Riquier{Janet Theory. That process has been an obstacle to its widespread application. In Section 4, we give a proof of a Formal Riquier Existence Theorem for formal power series solutions. In Section 5, we describe our theoretical development of the nonlinear theory which hinges on our concept of Relative Riquier Bases. This allows techniques from the linear theory to be easily extended to the nonlinear case. Applications are discussed in Section 6. For polynomially nonlinear di erential systems, alternative methods for constructing formal power series solutions have been developed based on the work of Kolchin 10] and Seidenberg 21]. For such systems, using a result of Rosenfeld 17], Boulier et al 3] obtained an algorithmic solution for determining radical di erential ideal membership. The di erential elimination algorithms of these and related approaches are based on extending Ritt's concept of a (di erential) characteristic set 3, 12]. Applications of di erential elimination algorithms are given in 20, 7, 13, 5, 4, 9, 8, 3]. Our paper is based on results in Rust's Ph.D thesis 18] which is available on the web.1

In theoretical arguments we represent di erential equations as functions, much as in algebra polynomials have long since largely replaced polynomial equations. Thus an equation that is customarily written in the form lhs = rhs is instead represented by the function lhs ? rhs. Let n be a positive integer, representing the number of dependent variables u = (u1 ; :::; un ) and let m be a nonnegative integer, representing the number of independent variables x = (x1 ; :::; xm ) in a system of pde. Let N := f0; 1; 2; :::g denote the natural numbers including 0 and let Nn := f1; :::; ng. Let = f i ji 2 Nn ; 2 Nm g be a set of indeterminates, with = (a1 ; a2 ; :::; am ) corresponding to the various derivatives of the dependent variables with
1 C. Rust's PhD thesis, on which this paper is based, is available at http://www.cecm.sfu.ca/~rust or http://www.cecm.sfu.ca/~reid

2 Preliminaries 2.1 Functions and Derivatives

respect to the independent variables: @ a1 @ a2 ::: @ am ui (x) ! i @x1 @x2 @xm Let ~ := fx1 ; :::; xm g . Let F = R or C . Let f be an F-analytic function of ~ .2 One should think of f as representing an F-analytic di erential equation. We always insist that functions depend on only nitely many of the indicated variables. We also insist that the domain is an open subset of the cartesian product F ~ 0 , where ~ 0 is the nite subset of ~ on which the function depends. For i 2 Nn , the total derivative of f with respect to xi is de ned as: @f + X @f D ; Di f := @x i i 2 @

2.3 Reductions

where Di (ja1 ;:::;am) := (ja1 ;:::;aj +1;:::;am) . For = a a m m 1 f . (a1 ; :::; am ) 2 N , de ne D f := (D1 ) (Dm ) Let x0 be a point in Fm and let u(x) = (u10(xn); :::; un (x)) be a vector of formal power series in F x ? x ]] . We write D ui (x) to denote the usual (term by term) partial derivative of ui (x) with respect to = (a1 ; :::; am ) 2 Nm : D ui (x) = (D1 )a1 (Dm )am ui (x): For v(x) 2 F x ? x0 ]], v(x0 ) denotes the constant term of v(x). If f is analytic at the point x0 ; (D ui (x0 )) i 2 , let f u](x) denote the formal power series at x0 given by f u](x) := f x; (D ui (x)) i 2 : In the above equation, the subscript \ i 2 " indicates that i D u (x) is to be substituted in the argument of f corresponding to i , for each i 2 . A positive ranking of is a total order of which is compatible with di erentiation i < j =) i < j + + and which is a well-ordering i i; 0 for ; ; 2 Nm , i; j 2 Nn . For the rest of this paper, let a positive ranking be xed. For f a function of ~ , we write hd f for the highest derivative (highest with respect to ) which actually occurs in f . Note that hd f is unde ned if f depends only on fx1 ; :::; xm g and does not depend on . In this case, we insist as a matter of notational convenience that hd f hd g for arbitrary g. For the rest of this paragraph, let f be an analytic function of ~ . We say that f is -monic if f has the form f = hd f + F , with hd F < hd f . Following the terminology of 15], we say that f is -leading linear if there exists an analytic function g of ~ with hd g < hd f such that f=g is -monic. We say that g (which is easily seen to be unique) is the leading coe cient of f , denoted hc f . We say that f is -leading nonlinear if hd f is well-de ned, but f is not -leading linear.
2

For this subsection, let a nite set M of -monic analytic functions of ~ be xed. Let g; h be functions of ~ . We say that hm is a one step reduction of g if there exist f 2 M and 2 N such that, with 0 := hd D f , h can be given by substituting 0 ? D f for 0 in g: h = g(x; ( ) 6= 0 ; ( 0 ? D f ) = 0 ): This is denoted g 7?! h or, if we want to be more speci c, ;f ) g(7?! M h. See Example 1 for an explicit calculation. We say that u(x) 2 F x ? x0 ]]n (for some x0 2 Fm ) is a solution to a system of analytic pde if f u](x) is well-de ned and f u](x) = 0 for all f in the system. Throughout this paper, all solutions to di erential equations are assumed to be formal power series solutions unless otherwise speci ed. Suppose that u(x) 2 F x ? x0m ]]n is a solution to M. Clearly D f u](x) = 0 for all 2 N and f 2 M. Therefore for g; h analytic, if h is a one step reduction of g then h u](x) is well-de ned if and only if g u](x) is well-de ned and in this case g u](x) = h u](x). We say that g reduces to h if h can be obtained from g by a nite chain of one step reductions. That is, g reduces to h if there exists a positive integer k and h1 ; :::; hk such that g = h1 7?! h2 7?! ::: 7?! hk = h: We write g ?! h or g?!M h where is of the form = (( 1 ; f1 ); :::; ( k?1 ; fk?1 )) (1)
;fi ) i! with hi ( 7? M hi+1 for i 2 Nk?1 . We also write h = red(g; ). For u(x) a solution to M, we again have for g; h analytic that h u](x) is well-de ned if and only if g u](x) is well-de ned and in this case g u](x) = h u](x). Therefore, u is a solution to M fgg i u is a solution to M fhg. We say that g completely reduces to h if g reduces to h and h reduces to h0 implies that h = h0 . By a standard Dickson's Lemma argument (see Section 4.4 of 18]) any in nite chain h0 7?! h1 7?! ::: is eventually constant.

2.2 Rankings and Highest Derivatives

Example 1 (Non-uniqueness of complete reductions) Let m = 1, n = 2. Let M = fux ? u; uy ? u + y2 g and @ (ux ? u) and then by g = uxy . Reducing g rst by @y 2 uy ? u + y we obtain: g 7?! uy 7?! u ? y2 ; which cannot be @ (uy ? u + y 2 ) reduced any further. If instead we reduce g by @x and then by ux ? u we obtain: g 7?! ux 7?! u; which again cannot be reduced further, illustrating the non-uniqueness of complete reductions. Note that the above computations are independent of the ranking , since the only comparisons used were ux > u and uy > u which are true for any positive ranking. Following traditional terminology, we de ne the principal derivatives of M to be those derivatives that may be obtained by di erentiating the highest derivative of an element of M:
Prin M := f 2 j there exist f 2 M and 2 Nm with = D hd f g:

It is also possible to give a treatment where the pde are assumed to be formal power series, but certain restrictions are necessary, which we will discuss in future work

The parametric derivatives of M, which we denote Par M, are those derivatives that are not principal: Par M := r Prin M: Note that a reduction h of g is a complete reduction i h depends on fxg Par M only.

If 1 = 2 and D f ? D f 0 ?! 0, then: red(h; ) = g (x; (red( ; )) 6= 1 ; (red( ? D f; )) = 1 ) = g( x; (red( ; )) 6= 1 ; (red( ; ) ? red(D f; )) = 1 ) ? = g x; (red( ; )) 6= 1 ; (red( ; ) ? red(D f 0 ; )) = 1 = red(k; ):

3 Leading Linear Theory

In this section, we give a new development of the essential aspects of the Riquier{Janet Theory, motivated by algebraic Grobner basis theory. For this paper, x a non-empty open subset U of F ~ and a nite set M of -monic analytic functions on U which are polynomial in Prin M. The condition that M be polynomial in Prin M is not an onerous restriction. In particular, orthonomic systems { and therefore the usual standard forms 16, 14, 20] { satisfy this condition. Note that the class of analytic functions on U that are polynomial in Prin M has the useful properties that it is closed under algebraic operations, di erentiation and reductions by M. This property is used in the proof of the Formal Riquier Existence Theorem to show that the Initial Data map is well-de ned.

Lemma 2 Fix 2 . Suppose that0 for all ; 0 2 Nm and 0 , we have D f ? f; f 2 M with hd D f = hd D f D f 0 ?! 0. Let g be an analytic function on U that is polynomial in Prin M with hd g . Suppose g ?! h and g ?! k. Then there exists l with h ?! l and k ?! l. In
0 0

particular, g has a unique complete reduction.

3.1 Riquier Bases

In this section, we introduce Riquier bases (cf. Reid's \standard forms" 14] and Schwarz's \Janet bases" 20]), which are the di erential analogs of Grobner bases. We prove the uniqueness of complete reductions with respect to a Riquier basis, which is the analog of the uniqueness of normal forms with respect to a Grobner basis. This uniqueness is the key to the proof of our version of the Riquier Existence Theorem for formal power series solutions to analytic systems of pde in Riquier basis form (Theorem 2). Lemma 1 Let f; f 0 be elements of M and g an analytic function on Um that is polynomial in Prin M. Let ; be elements of N and let be as in Section 2.3. Let h and k denote the one step reductions: h = red(g; ( ; f )) and k = red(g; ( ; f 0 )): Then: 1. If hd D f < hd D f 0 then red(h; (( ; f 0 ); ( ; f ))) = red(k; ( ; f )): In particular, there exists an analytic function l of U that is polynomial in Prin M, such that h ?! l and k ?! l. 2. If D f ? D f 0 ?! 0, then red(h; ) = red(k; ): In particular, there exists l such that h ?! l and k ?! l. Proof: Let 1 = hd D f and 2 = hd D f 0 . If 1 < 2 , we have: red(k; D f ) = g(x; ( ) 6= 1 ; 2 ; ( 1 ? D f ) = 1 ; ( 2 ? D f 0 (x; ( ) 6= 1 ; ( 1 ? D f ) = 1 ) = 2 ) = red(h; (( ; f 0 ); ( ; f )): 3

This lemma follows from the previous lemma. The proof 18] is the virtually the same as in the algebraic case (for example see Theorem 4.75, pp. 176-7 of 1]). The analog of the Grobner basis in the di erential context is given by the following De 0 nition 1 We say that M is a Riquier basis if for all ; 2 Nm and f; f 0 2 M with hd D f = hd D f 0 , the integrability condition D f ? D f 0 can be reduced to 0. A \standard form" in the sense of Reid 14] is shown in 18] to be a Riquier basis S such that each f 2 S is completely reduced with respect to S r ff g. This more restrictive notion corresponds to a reduced Grobner basis in the algebraic theory. We therefore use the term reduced Riquier basis for this. An immediate consequence of Lemma 2 is: Theorem 1 Suppose that M is a Riquier basis and g is an analytic function on U that is polynomial in Prin M. Then g has a unique complete reduction. We denote the complete reduction of g by red(g; M). The uniqueness of complete reductions is the key idea underlying the formal Riquier Existence Theorem treated in the next section.
0 0

4 The Formal Riquier Existence Theorem

In this section, we give the proof of the formal Riquier Existence Theorem (Theorem 2), which gives an Existence and Uniqueness result for formal power series solutions to systems of analytic pde in Riquier basis form. A speci cation of initial data for M is a map id : fxg Par M ?! F: Note that fxg Par M may be in nite. For x0 2 Fm , we say that id is a speci cation at x0 if id(x) = x0 . (Note that we are abusing notation slightly. In reality, the domain of id is fx1 ; :::; x0m g Par M and we should write (id(x1 ); :::; id(xm )) = x not id(x) = x0 .) For g a function of ~ , let id(g) be the function of the principal derivatives obtained from g by evaluating x and the parametric derivatives using id: id(g )(( ) 2 = Par M ) := g (id(x); (id( )) 2Par M ; ( ) 2 = Par M ) :

Example 2 Let m = 2 and n = 1. Let M = fuxx ; uyy ? xeug. Then Par M = fu; ux ; uy ; uxy g. Let id be given by
id(x) = id(y ) = id(u ?) = 0, id(ux ) = 1, id(uy ) = 2 and id(uxy ) = 3. Then id uxyy + xux + u2 y = uxyy + 4.

Lemma 3 Let G be a nite set of functions of ~ . Then there exists such that red(g; ) is a complete reduction of g for all g 2 M. The easy proof of this lemma is given in 18]. In particular, if M is a Riquier basis and G is a nite set of analytic functions of ~ that are polynomial in Prin M, then there exists such that red(g; ) = red(g; M) for all g 2 G.
Suppose that M us a Riquier basis. Fix 0x0 2 Fm . Let id be a speci cation of initial data for M at x such that id(f ) is well-de ned for0 all f 2 M. Then there is a unique solution u(x) 2 F x ? x ]]n to M at x0 such that D ui (x0 ) = id( i ) for all i 2 Par M. Furthermore, every solution to M at x0 may be obtained in this way for some id. Proof: Let id be de ned as in the statement of the theorem. De ne u(x) 2 F x ? x0 ]]n by D ui (x0 ) := id red( i ; M) (2)

Note that red(D f (x; ( )); ) depends only on the parametric derivatives (and x). Therefore red(D f (x; ( )); ) is a complete reduction of D f (x; ( )) and we have red(D f (x; ( )); ) = red(D f (x; ( )); M) = red(D f (x; ( )); ( ; f )) = 0: 0 Thus (D f u])(x ) = 0, as required. This completes the proof of the existence part of the theorem. To prove the uniqueness part, we need only observe that u must satisfy Equation 2 for all i 2 Nn and 2 Nm and that this condition uniquely de nes u.

Theorem 2 (Formal Riquier Existence Theorem)

for i 2 Nn and 2 Nm . We must verify that u(x) is a well-de ned solution to M. Note that red( i ; M) depends only on the parametric derivatives and so id(red( i ; M)) is an element of F, so long as it is well-de ned. We must rst check that id(red( i ; M)) is well-de ned. Let g be an analytic function of ~ such that id(g) is wellde ned. Recall that any partial derivative of g is de ned wherever g is. It follows that for h any partial or total derivative of g, id(h) is also well-de ned. Suppose that g is polynomial in0 the principal derivatives. Take 2 Nm and f 2 M. Let be the highest derivative of D f . We have: red(g; ( ; f )) =
deg X g @kg k 0 k (?D f ) @ k=0
0

Our de nition of a Riquier basis requires that all the innitely many integrability conditions reduce to 0. As in the algebraic case, it turns out that it su ces to check a nite set of integrability conditions (see Theorem 3 and Corollary 1, below). The sequence of lemmas that follow is directed towards proving this. Lemma 4 Suppose 0h ?! 0 and k ?! 0. Suppose further that0 for all ; 2 Nm and f; f 0 2 M0 with hd D f = hd k, we have D f ? D f ?! 0. Then hd D f h + k ?! 0. Proof: Say h?! 0. Let l = red(k; ). Since k ?! l, by Lemma 2 there exists j with 0 ?! j and l ?! j . Since 0 ?! j , we have j = 0 and hence l ?! 0, say 0 = red(l; ). Then we have: red(h + k; ( ; )) = red(h; ( ; )) + red(k; ( ; )) = red(0; ) + red(l; ) = 0: Therefore, h + k ?! 0, as required.
0 0

4.1 Su cient Finite Sets of Integrability Conditions

and so id(red(g; ( ; f ))) is also well-de ned. This result, coupled with our assumption that G is polynomial in Prin M, shows that id(red( i ; M)) is well-de ned as required. Clearly, u(x) satis es D ui (x0 ) = id( i ) for all i 2 Par M. To verify that u0 (x) is a solution to M, it su ces to verify that (D f u])(x ) = 0 for all f 2 M and 2 Nm . Fix f; . We have (D f u])(x0 ) = (D f ) x0 ; (D ui (x0 )) ? = (D f ) x0 ; (id(red( ; M))) = (D f ) (id(x); (id(red( ; M)))) = id(D f (x; (red( ; M)))) where the permutation of id follows from its properties as an evaluation map. Let 0 be the nite subset of on which D f depends. By the lemma above there exists such that red( ; M)=red( ; ) for all 2 0 . Therefore, (D f u])(x0 ) = id(D f (x; (red( ; )))) = id(red(D f (x; ( )); )): 4

that the analogous implication always holds in the algebraic case: if a polynomial reduces to zero then any multiple of the polynomial reduces to zero. In Lemma 6 we will give a condition on M that guarantees g ?! 0 =) Di g ?! 0. Lemma 5 Take 2 Nm , f 2 M, i 2 Nn and g an analytic function of ~ . Let 1 be the highest derivative of D f . Let 0 be given by 1 = Di 0 , if this is well-de ned. Let g be an analytic function on U that is polynomial in Prin M. Then: red(Di g; ( ; f )) = Di red(g; ( ; f )) @g ; ( ; f ) Di D f + red @

Example 3 Let m = n = 1 and let M = fu ? x; ug. Let @u u? x xu 7? g = (u ? x)(ux ? 1) + uux . Observe that g7? ! ! 0 and x @x so g ?! 0. However, one computes that the only complete @g = (ux ? 1)2 + (u ? x)uxx + u2 + uuxx is 1 reduction of @x x @g 6?! 0. and so @x Thus in general g ?! 0 does not imply Di g ?! 0. Note

@g ; ( ; f ) D f: ? red @ 0

If 0 is not well-de ned, then the last term is omitted in the above formula. The proof 18] is by direct computation.

Lemma 6 Let g be an analytic function on U that is polynomial in Prin M such that g ?! 0. Fix i 2 Nm . Suppose that0 for all ; 0 2 Nm and 0f; f 0 2 M with hd D f = hd D f < hd Di g, D f ? D f ?! 0. Then Di g ?! 0. Proof: By induction on the length of the minimal chain required to reduce g to 0, we may assume that there exists an analytic function h of ~ with g 7?! h ?! 0 and Di h ?! 0, say h = red(g; ( h ; fh )). If g = 0 the result is clear. Thus we may assume h 6= g. By the previous lemma, we have an expression of the form
0 0

red(Di g; ( h ; fh )) = Di h + kDi D h fh + lD h fh ; (3) with k and l analytic functions of ~ satisfying hd k hd g and hd l hd g. Furthermore, either hd(Di h) < hd(Di g) or hd(Di D f ) < hd(Di g). Thus at least two of the three summands in (3) have highest derivative strictly less than hd(Di g ). Therefore by (two applications of) Lemma 4, we have red(Di g; ( h ; fh )) ?! 0 and so Di g ?! 0. The least common multiple of = (a1 ; :::; am ) and = (b1 ; :::; bm ) 2 Nm is de ned by lcm( ; ) := (max(a1 ; b1 ); :::; max(am ; bm )): We say that divides if ai bi for all i 2 Nm ; in this case obviously lcm( ; ) = . j For j ; j 2 we de ne lcm( j ; j ) = lcm ( ; ) . The differential analog of an S-polynomial is given in the following: De nition 2 Take f; f 0 2 M. Let their highest derivatives be i and i , respectively. Let be the least common multiple of and 0 . Then if i 0= i0 , de ne the minimal integrability condition of f and f to be ic(f; f 0 ) := D ? f ? D ? f 0 : If i 6= i0 , then ic(f; f 0 ) is said to be unde ned. Note that, in the notation of the de nition above, if ic(f; f 0 ) is well-de ned then hd D ? f = hd D ? f 0 and is minimal such that this holds, in that at least one of the coordinates of ? or of ? 0 is 0. In particular, the highest derivatives cancel in the expression for ic(f; f 0 ), that is hd(ic(f; f 0 )) < hd D ? f . The proof 18] of the following theorem uses a straightforward trans nite induction. Theorem 3 Suppose that for each pair (f; f 0 ) 2 M2 0with 0 ic(f; f ) well-de ned there exists an expansion of ic(f; f ) of the form
0 0 0 0

well-de ned and using the trivial expansion for ic(f; f ) as ic(f; f 0 ) = ic(f; f 0 ). Although the result of Corollary 1 is widely believed to be true, it appears that it has never been proved before for m > 2, as noted by Reid et al. 15], although for m 2 and M reduced, it is known 14]. This result gives a very natural nite set of integrability conditions that it is su cient to check reduce to 0. However, this set of integrability conditions is in general far from minimal. As shown in 18], a well known criterion due to Buchberger in the algebraic case (see e.g. 1], p. 223) goes through in the di erential case. Also see the preprint 2] for a related result in the case of di erential polynomials.

Proof: This follows 0immediately0 from Theorem 3, by f 0) setting Bf;f = f(f; f )g for f; f 2 M with ic(f; 0
0

5 Leading Nonlinear Theory

The Janet algorithm, and consequently the Riquier Existence Theorem, can only be successfully applied to systems that are linear or almost linear in some sense. The rif algorithm of Reid et al. 15] manipulates nonlinear analytic systems into variations of what Reid et al. call rif form (reduced involutive form), which is a generalization of Riquier Basis form discussed earlier. We state an Existence and Uniqueness Theorem for systems in a related form, called rif0 form, output by Rust's rif0 algorithm. A sketch of aspects of the proof by Rust 18] is also given. For polynomially nonlinear systems, it also applies to the output of Reid's rif algorithm. A ranking is said to be sequential if for all 2 , the set f 0 2 j 0 g is nite. For = (a1 ; :::; am ) 2 Nm , let j j denote a1 + ::: + am . For the rest of this paper, we insist that the xed ranking be sequential and we let N be a nite set of analytic functions on U that depend only on fxg Par M. For S a nite set of analytic functions of ~ , let L(S ) denote the leading linear elements of S . Let N(S ) denote the set of leading nonlinear elements of S . (The terms \leading linear" and \leading non-linear" were de ned in Subsection 2.2.)

5.1 Relative Riquier Bases and rif 0 Form Roughly speaking, we say that M is a Riquier basis relative to N if M \looks like" a Riquier basis so far as the algebraic solutions to N are concerned. We de ne what it means for M to be a Riquier basis relative to N and mimic the
development of the case where the system is linear in its highest derivatives. De nition 3 For an analytic function on U , we say that is a special consequence of N if lies in the ideal generated by N in the ring of analytic functions on U that depend on fxg Par M only. That is, is a special consequence of N if it admits an expansion of the form =
k X i=1

ic(f; f 0 ) =
0

(b;b )2Bf;f
0

D f;f ;b;b ic(b; b0 );

0 0 0 0 0

a subset of M2 such that for each (b; b0 ) 2 where Bf;f is Bf;f , ic(b; b0 ) ?! 0 and D f;f ;b;b lcm(hd b; hd b0 ) lcm(hd f; hd f 0 ). Then M is a Riquier basis.
0

No doubt the most natural special case is the following: Corollary 1 Suppose that ic(f; f 0 ) ?! 0 for all f; f 0 2 M with ic(f; f 0 ) well-de ned. Then M is a Riquier basis.

hi g i

(4)

with g1 ; :::; gk 2 N and h1 ; :::; hk analytic functions on U that depend on fxg Par M only. We say that M is a 0 2 Nm and Riquier basis relative to N on U if for all ; f; f 0 2 M with hd D f = hd D f 0 , the integrability condition D f ? D f 0 can be reduced to a special consequence of N.
0 0

De nition 4 We say that (M; N ) (or, by abuse, M N ) is in rif0 form on U if M is a Riquier basis relative to N on U and for all i 2 Nm and g 2 N , Di g can be reduced to a special consequence of N on U .
The following lemma generalizes Lemma 1. Its proof, given in 18], is similar to that of Lemma 1. Lemma 7 Let g be an analytic functions on U that is polynomial in Prin M. Fix f; f 0 2 M, ; 2 Nm and as in Section 2.3. Let h and k denote the0 one step reductions: h = red(g; ( ; f )) and k = red(g; ( ; f )): Then: 1. If hd D f < hd D f 0 then red(h; (( ; f 0 ); ( ; f ))) = red(k; ( ; f )):
2. Suppose that D f ? D f 0 ?! for some special consequence of N and chain of one step reductions 1 Then red(h; ) ? red(k; ) = h0 for some analytic function h0 on U with hd h0 hd g.

Proof: 0 The uniqueness property is clear. De ne u(x) 2 x ? x ]]n by D ui (x0 ) = id(redmod( i ; M)) for 2 Nm and i 2 Nn . We must verify that0 u(x) is a solution to M. It su mces to show that D f u](x ) = 0 for all f 2 M and 2 N . Fix f; . We have

The following lemma generalizes Lemma 2. It is proved in much the same way as in the algebraic case. Lemma 8 Let g be an analytic function on U that is polynomial in Prin M. Let 2 be the highest principal derivative0 that occurs in g. Suppose that 0 for all ; 0 2 Nm 0 and , D f ? D f ref; f 2 M with hd D f = hd D f duces to a special consequence of N . Suppose g ?! h and g ?! k. 0Then there exist l; l0 with h ?! l and k ?! l0 such that l ? l is a special consequence of N . An immediate consequence of Lemma 8 is: Theorem 4 Suppose that M is a Riquier basis relative to N and g is an analytic function on U , polynomial in Prin M. Then the di erence of any two complete reductions of g is a special consequence of N . In particular, any two complete reductions of g agree on the algebraic solutions to N . For g an analytic function on U polynomial in Prin M, let redmod(g; M) denote a particular choice of complete reduction of g with respect to M. Let id be a speci cation of initial data for M that satis es id( ) = 0 for all 2 N . Theorem 4 implies that redmod(g; M) is uniquely de ned up to a special consequence of N . In particular, id(redmod(g; M)) is independent of the choice of complete reduction. Note that if 1 lies in the ideal generated by N in an appropriate ring of analytic functions, then no such id exists. Theorem 5 Suppose that M is a Riquier basis relative to m . Let id be a speci cation of initial data for N . Fix 0 x0 2 F M at x 2 Fm with id 2 U and id( ) = 0 for all 0 2n N . Then there exists a unique solution u(x) 2 F x ? x ]] to M such that D ui (x0 ) = id(redmod( i ; M)) for all 2 Nm and i 2 Nn .
0 0

D f u](x0 ) = (D f ) x0 ; (Dui (x0 )) ? = (D f ) x0 ; (id(redmod( ; M))) = id(D f (x; (redmod( ; M)))): Let 0 be the nite subset of on which D f depends. By Lemma 3 there exists such that redmod( ; M) = red( ; ) for all 2 0 . Therefore, (D f u])(x0 ) = id(D f (x; (red( ; )))) = id(red(D f (x; ( )); )): Note that red(D f (x; ( )); ) depends only on the parametric derivatives (and x). Therefore red(D f (x; ( )); ) is a complete reduction of D f (x; ( )) and we have id(red(D f (x; ( )); )) = id(redmod(D f (x; ( )); M)) = id(red(D f (x; ( )); ( ; f ))) = id(0) = 0: Thus (D f u])(x0 ) = 0, as required. Just as in the leading linear case, the theory becomes most technical when it is shown that only a nite subset of integrability conditions have to be satis ed. We omit the full discussion which is given in 18]. Its result is the analog of Corollary 1: Corollary 2 Suppose that for each (f; f 0 ) 2 M2 with ic(f; f 0 ) well-de ned, ic(f; f 0 ) can be reduced to a special consequence of N . Suppose also that 0 for all (f; f 0 ) 2 2 0 m M with ic(f; f ) well-de ned and ; 2 N such that hd D f = hd D f 0 maxi2Nm; 2N (Di hd ), D f ? D f 0 can be reduced to a special consequence of N . Finally, suppose that Di can be reduced to a special consequence of N for all 2 N . Then (M; N ) is in rif 0 form.
0 0

We state an Existence and Uniqueness Theorem for systems of analytic pde in rif0 form. Let S be a nite set of analytic functions of ~ in rif0 form on an open subset U of fxg Par(L(S )). Let id be a speci cation of initial data for L(S ) with id 2 U . Clearly, for id to correspond to a formal solution u(x) of S it is necessary that id(g) = 0 for all g 2 N(S ). This turns out to be su cient also: Theorem 6 Let S be a nite set of analytic functions of ~ . Let L denote L(S ) and N denote N(S ). Let U be an open subset of fxg Par L. Suppose that S is in rif 0 form on U and that id is a speci cation of initial data for L that lies in U . Suppose further that id(g) = 0 for all g 2 N . Let u(x) 2 F x ? x0 ]]n be the solution to L that corresponds to id via Theorem 5. Then u(x) is a solution to S . 6

5.2 A Nonlinear Existence and Uniqueness Theorem

The proof, which is given in m 18], uses induction to show that for all g 2 N and 2 N , there exists an expansion P redmod(D g; L) = gi 2N hi gi with id(hi ) well-de ned. The other main step in the proof is that since u(x) is a solution to L, we have that D g u](x) = redmod(D g; L) u](x). Thus, D g u](x0 ) = redmod(D g; L) u](x0 ) X id(hi ) id(gi ) = 0: = id(redmod(D g; L)) =
gi 2N

6 Applications 6.1 Existence Uniqueness Theorems

Consider a system of real analytic or complex analytic pde of the form uy = f (ux ; v); vx = g(vy ; u): (5) With respect to a ranking that satis es uy > ux ; v and vx > vy ; u, and only with respect to such a ranking, this system is a Riquier basis. There do exist rankings that satisfy these conditions 19]. Here, the parametric derivatives are fu; ux ; uxx; :::; v; vy ; vyy ; :::g. Fix a point (x0 ; y0 ) in the domains of f and g. Ascribing values to the parametric derivatives at the point (x0 ; y0 ) uniquely determines a formal power series solution to the system at (x0 ; y0 ), and every such solution may 0 be obtained in this way. Equivalently, one may specify u ( x; y ), v(x0 ; y). A discussion of Existence Uniqueness Theory for analytic solutions will appear elsewhere. This example, although arti cial, also illustrates that the theorems apply to systems which are not necessarily polynomially nonlinear. Such non-polynomial systems arise in applications. For example in 15], an example of a frictionless bead sliding under gravity on wire with shape (x; y) = 0 is given. The shape of a physical wire is not necessary polynomial, although possibly it might be approximated by some function in a su ciently complicated di erential-algebraic extension.

6.2 Application to Non-classical Symmetries

In the previous subsection, the system studied was 0already in the form of a Riquier basis (or more generally rif -form). However, in many applications, such as the one described in this section, the system is not in such a form, and must be manipulated to such a form. The algorithms for this purpose (see 15, 18]), which are not described in detail here, have been implemented by Wittkopf in the symbolic language Maple. They essentially compute a sequence of approximations to an Existence and Uniqueness theorem for a system of pde. At each step integrability conditions are calculated, any which are not satis ed are appended to the system. Simpli cation is performed, and the process repeated until no new conditions arise. During the course of a calculation, the leading derivatives of leading linear equations must be solved for. When the coe cients of these leading derivatives contain as yet unknown functions, a decision must be made as to whether the coe cient (or pivot or separant) is identically zero, or non-zero. This results in a binary tree, with each branch corresponding to a possible rif0 -form for the system. This 7

binary tree is automatically generated by Wittkopf's Maple package for the example problem below. This example is concerned with the application of Bluman and Cole's so-called non-classical method which is a generalization of Lie's symmetry method. In particular this non-classical method is applied to the following system of coupled nonlinear Schrodinger (cnls) equations, which describes transverse e ects in nonlinear optical systems ? i t + r2 + j j2 + j j2 = 0; ? i t + r2 + j j2 + j j2 = 0: Nonclassical symmetry vector elds of form 1@x + 2@y + 3@z + @t + 1@p + 2@q + 3@r + 4@s of this system have been studied in detail by Mans eld, Reid and Clarkson 6]. Even though such vector elds do not generate a classical Lie group they lead to reductions in the number of variables of the cnls equations. The application of the non-classical method yields a polynomially nonlinear system of 856 equations, in 7 dependent ( 1; 2; 3; 4; 1; 2; 3) and 8 independent variables (x; y; z; t; p; q; r; s). In a prior calculation, certain dependencies were discovered allowing the variables to be expressed in terms of some addtional variables, g1; g2; g3; g4 which depend upon (x; y; z; t) only: k = gk1 p + gk2 q + gk3 r + gk4 s + gk ; k = 1; 2; 3; 4: Application of rif to this system, under the ranking described below, on a PII 333MHz machine required 120 CPU sec., consuming 4M of memory, and resulted in 3 consistent cases, of which only the following case corresponded to a genuine non-classical symmetry: case1 : Solved = 1 = ?g41 s + g44 p ? g43 q + g42 r; 2 = ?g42 s + g43 p + g44 q ? g41 r; 3 = ?g43 s ? g42 p + g41 q + g44 r; 4 = g44 s + g41 p + g42 q + g43 r; 1t; t = 4 g44 1t ; 2t; t = 4 g44 2t ; 3t; t = 4 g44 3t ; 1x = ?g44 ; 2x = 0; 3x = 0; g41 x = ? 1 g42 ; g42 x = 1 g41 ; 1 1 ? g44 1; g44 = 0; 1 = 0; 2 = ?g44 ; g43 x = 2 t y y x 3y = 0; g41 y = ? 2 g42 ; g42 y = 2 g41 ; 1 2 ; g44 = 0; 1 = 0; 2 = 0; g43 y = ?g44 2 + 2 t z z y 3z = ?g44 ; g41 z = ? 3 g42 ; g42 z = 3 g41 ; 1 3 ? g44 3; g44 = 0; g43 z = 2 t z 2 g41 t = 3 g41 g44 ? (? 1 + 2 g43 ? 32 ? 22 ) g42 ; g42 t = ?( 12 + 22 + 32 ? 2 g43 ) g41 + 3 g42 g44 ; g43 t = 2 g44 g43 ; g44 t = 2 g44 2 ; 1p = 0; 2p = 0; 3p = 0; 1q = 0; 2q = 0; 3q = 0; 1r = 0; 2r = 0; 3r = 0; 1s = 0; 2s = 0; 3s = 0; g11 = g44 ; g12 = ?g43 ; g13 = g42 ; g14 = ?g41 ; g21 = g43 ; g22 = g44 ; g23 = ?g41 ; g24 = ?g42 ; g31 = ?g42 ; g32 = g41 ; g33 = g44 ; g34 = ?g43 ; g1 = 0; g2 = 0; g3 = 0; g4 = 0] PolyRed = g41 2 + g42 2 = 0] Pivots = g41 + g32 6= 0]

The ranking used in the reduction was to rst eliminate all dependent variables in favor of g's, then by total order of derivative, then lexicographically by derivative (@x @y @z @t @p @q @r @s ) and nally lexicographically on dependent variable ( 1 2 3 g11 g12 g13 g14 g21 : : : g24 g31 : : : g34 g41 : : : g44 g1 g2 g3 g4). Here the Solved equations are leading linear in, and solved for their highest ranked derivatives, the PolyRed are leading nonlinear in their highest ranked derivatives, and the Pivots represent any non-trivial divisions (case splittings) that the algorithm performed to obtain the given system. Any equations in PolyRed are only algebraic constraints as all di erential consequences of these equations have been accounted for. Calculation of the initial data gives case1 initdata : In nite = ] Finite = g41 = a1 ; g42 = a2 ; g43 = a3 ; g44 = a4 ; 1 = a5 ; 1t = a6 ; 2 = a7 ; 2t = a8 ; 3 = a9 ; 3t = a10 ] where the a constants represent the values of the correspomding derivatives at a point (x0 ; y0 ; z0 ; t0 ; p0 ; q0 ; r0 ; s0 ). This yields a 10 dimensional variety, but the nonlinear constraint must also be considered. When we enforce a12 + a22 = 0 we are left with a 9 dimensional variety. This variety describes the data required to determine a unique (formal power series) solution for the input system. The tree of cases which was automatically generated by Wittkopf's package is
RIF Case Tree

<>

1 <> =

2 <> =

References
1] T. Becker, V. Weispfenning, and H. Kredel, Grobner Bases: A Computational Approach to Commutative Algebra, vol. 141 of Graduate Texts in Mathematics, Springer, 1993. 2] F. Boulier, Some improvements of a lemma of rosenfeld. Preprint, 1998. 3] F. Boulier, D. Lazard, F. Ollivier, and M. Petitot, Representation for the radical of a nitely generated di erential ideal, in Proc. ISSAC 1995, ACM Press, 1995, pp. 158{166. 8

4] K. Brenan, S. Campbell, and L. Petzold, Numerical Solutions of Initial-Value Problems in Di erentialAlgebraic Equations, Elsevier, 1989. 5] P. Clarkson and E. Mansfield, On a shallow water wave equation, Nonlinearity, 7 (1994), pp. 975{1000. 6] G. R. E. Mansfield and P. Clarkson, Nonclassical reductions of a 3+1-cubic nonlinear schrodinger system, Comp. Phys. Comm., 115 (1998), p. in press. 7] W. Hereman, Review of symbolic software for the computation of Lie symmetries of di erential equations, Euromath Bull., 1 (1994), pp. 45{79. 8] E. Hubert, The general solution of an ordinary di erential equation, in Proc. ISSAC '96, ACM Press, 1996, pp. 189{195. 9] , Etude Algebraique et Algorithmique des Singularites des Equations Di erentielles Implicites, PhD thesis, Institut National Polytechnique Grenoble, 1997. 10] E. Kolchin, Di erential Algebra and Algebraic Groups, Academic Press, 1973. 11] S. Kovalevskaya, Zur theorie der partiellen di erentialgleichungen, J. Reine Agnew. Math., 80 (1875), pp. 1{32. 12] E. Mansfield, Di erential Grobner Bases, PhD thesis, Univ. of Sydney, 1991. 13] T. Oaku, Algorithms for nding the structure of solutions of linear partial di erential equations, in Proc. ISSAC '94, ACM Press, 1994. 14] G. Reid, Algorithms for reducing a system of PDEs to standard form, determining the dimension of its solution space and calculating its Taylor series solution, European J. of Appl. Math., 2 (1991), pp. 293{318. 15] G. Reid, A. Wittkopf, and A. Boulton, Reduction of systems of nonlinear partial di erential equations to simpli ed involutive forms, European J. of Appl. Math., 7 (1996), pp. 635{666. 16] C. Riquier, Les systemes d'equations aux derivees partielles, Gauthier-Villars, 1910. 17] A. Rosenfeld, Specializations in di erential algebra, Trans. of the Amer. Math. Soc., 90 (1959), pp. 394{ 407. 18] C. Rust, Rankings of Derivatives for Elimination Algorithms and Formal Solvability of Analytic Partial Di erential Equations, PhD thesis, Univ. Chicago, 1998. www-address: www.cecm.sfu.ca/~rust. 19] C. Rust and G. Reid, Rankings of partial derivatives, in Proc. ISSAC '97, Maui, W. Kuchlin, ed., acm Press, 1997, pp. 9{16. 20] F. Schwarz, An algorithm for determining the size of symmetry groups, Computing, 49 (1992), pp. 95{115. 21] A. Seidenberg, An elimination theory for di erential algebra, Univ. California Publ. Math. (N.S.), (1956), pp. 31{38.