Journal of Symbolic Computation 41 (2006) 1091–1104

www.elsevier.com/locate/jsc

Gröbner fan and universal characteristic sets of prime
differential idealsI
Oleg Golubitsky ∗
School of Computing, Queen’s University, Kingston, Ontario K7L 3N6, Canada

Received 15 October 2004; accepted 18 May 2006
Available online 22 August 2006

Abstract

The concepts of Gröbner cone, Gröbner fan, and universal Gröbner basis are generalized to the case of
characteristic sets of prime differential ideals. It is shown that for each cone there exists a set of polynomials
which is characteristic for every ranking from this cone; this set is called a strong characteristic set, and an
algorithm for its construction is given. Next, it is shown that the set of all differential Gröbner cones is
finite for any differential ideal. A subset of the ideal is called its universal characteristic set, if it contains
a characteristic set of the ideal w.r.t. any ranking. It is shown that every prime differential ideal has a finite
universal characteristic set, and an algorithm for its construction is given. The question of minimality of
this set is addressed in an example. The example also suggests that construction of a universal characteristic
set can help in solving a system of nonlinear PDE’s, as well as maybe providing a means for more efficient
parallel computation of characteristic sets.
c 2006 Elsevier Ltd. All rights reserved.

Keywords: Differential algebra; Prime differential ideal; Characteristic set

1. Introduction

Consider a differential ring of differential polynomials and a prime differential ideal in it. We
study the dependence of the characteristic set of the ideal on the ranking of partial derivatives
and give two invariants of the ideal—the differential Gröbner fan and the universal characteristic
set.

I This is a full version of the article presented as a poster at ISSAC 2004.
∗ Tel.: +1 613 449 1267; fax: +1 613 533 6513.
E-mail address: oleg.golubitsky@gmail.com.

c 2006 Elsevier Ltd. All rights reserved.
0747-7171/$ - see front matter
doi:10.1016/j.jsc.2006.05.003

referring the reader to Ritt (1950). but their ranks coincide. of which the most well-known are the FGLM (Faugère et al. given any characteristic set of a prime differential ideal. . 2. δm } which commute pairwise. Basic concepts of differential algebra Here we give a short summary of the basic concepts of differential algebra. A differential ring is a commutative ring endowed with a finite set of derivations ∆ = {δ1 . (2) The set of all differential Gröbner cones corresponding to any differential ideal is finite. . i m are nonnegative integer numbers. These algorithms have been carried over to the case of characteristic sets of prime differential ideals (Boulier. b ∈ R satisfies δ(a + b) = δ(a) + δ(b). Then we show that (1) For every differential Gröbner cone. for every ranking. We call this set a universal characteristic set and give an algorithm for its construction. This was followed by the development of several efficient algorithms for transformation of Gröbner bases from one admissible order to another. 2004). constructs a strong one of the same rank. we construct a minimal universal characteristic set. the problem of minimality in general is open. Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 In the algebraic case.. (3) There exists a finite set of differential polynomials whose characteristic set.. . One can pose the converse problem: given an autoreduced set of ranks (a rank is a power of a derivative). Golubitsky. Finally. We illustrate the algorithm for construction of a universal characteristic set on an example from Boulier et al. We call this set of rankings a differential Gröbner cone. (2001). where i 1 . 1993) algorithms.1092 O. no claims are made about the usefulness of the universal characteristic sets for the solution of nonlinear PDE systems in general). a similar problem has been solved in Mora and Robbiano (1988) by introducing the concepts of Gröbner fan and universal Gröbner basis and giving an algorithm for their construction. Now we carry over the theoretical basis for these algorithms to the differential case as well. . . A derivation over R is a mapping δ : R → R which for every a. The commutative monoid generated by the derivations is denoted by Θ. there exists a set which is characteristic for each ranking from this cone. For a fixed ranking there are usually infinitely many characteristic sets. we show how the system of nonlinear PDE’s from this example can be easily solved with the help of the universal characteristic set (again. 1993) and the Gröbner walk (Collart et al. δ(ab) = δ(a)b + aδ(b). . For this particular example. we call this set a differential Gröbner fan. (1999) for a more complete exposition. We call this set a strong characteristic set and describe an algorithm that. Kolchin (1973) and Kondratieva et al. 1999. the problem of efficient computation of the universal characteristic sets is discussed in relation to the problem of fast parallel computation of (ordinary) characteristic sets. . however. . Let R be a commutative ring. The set of all rankings is equal to the disjoint union of the cones in the Gröbner fan. Its elements are derivation operators θ = δ1i1 · · · δm im . is a strong characteristic set of the prime ideal. describe the sets of rankings with respect to which characteristic sets of the ideal have this set of ranks. Moreover. .

1. The specification of rankings described in the above theorem has been implemented in Maple by A.D. j). λn ∈ Zm . Theorem 1 (Rust and Reid. . A ranking is a total order ≤ of Nm × Nn such that for all a. Let K be a differential field of characteristic zero. j)) ∈ R. . O. (a. Below we briefly describe the corresponding algorithm. . .2 A ranking is called a Riquier ranking. i) ≥ (0. 1 From private communication with A. km . u n } be a finite set whose elements are called differential indeterminates. j)) ∈ R. (a. An ideal is called prime. namely by trying all permutations σ and numbers ti j such that matrices Mi and M j agree on the first ti j rows [from private communication with C. b.01. Let N = {0. . which uniquely specify a ranking ≤0 on Nm × Nn such that for all ((a. 30). 2. j) ≤ (k1 . and for this reason has not been implemented. . l).D. . (b. j). i. . [A] denotes the smallest differential ideal containing A. if ∀a. i) ≤ (b. Let R be a finite subset of (Nm × Nn )2 such that for all pairs ((a.1 The problem of finding a ranking containing a given finite relation on partial derivatives has been solved for Riquier rankings and implemented in the Maple 9. Wittkopf (Rust and Reid. . . . Derivation operators apply to differential indeterminates yielding derivatives θu. ∀ A ∈ I. n}. j) ≤ (b. This code was first included in the Rif package. • (a. Wittkopf. Let m be a nonnegative integer and n be a positive integer. . i) ⇐⇒ (a. i).}. Rankings on Nm × Nn correspond to rankings on the set of derivatives ΘU : δ1i1 · · · δm im u j ≤ δ1k1 · · · δm km u l ⇐⇒ (i 1 . To the author’s knowledge. j ∈ Nn . Let ≤ be a ranking on Nm × Nn . . Theorem 30). but then replaced by a simpler and more efficient code that accounts for Riquier rankings only (see below). i) ≤ (b. i. procedure differential ring. . Denote by ΘU the set of all derivatives. i). . For a subset A ⊂ R. i m . . Then there exist non-singular m × m integer matrices M1 . . Nn = {1. for arbitrary rankings this problem remains open. b ∈ R ab ∈ I ⇒ a ∈ I or b ∈ I. if for all a. (b. Mn . j). The differential ring of differential polynomials K{U } is the ring of polynomials of infinitely many variables K[ΘU ] endowed with the set of derivations ∆. . vectors λ1 . (a. i) ≤ (b. b ∈ Nm . this solution is inefficient. . and a permutation σ of Nn . . . However. 1997. . . j ∈ Nn . Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 1093 A differential ideal I of differential ring R is an ideal of R stable under derivations. Remark after Th. Let U = {u 1 . 2 The problem can be reduced to a linear programming problem using Theorem 1. Rust].J. j) • (a. c ∈ Nm . . 1997. i) ≤ (b + c. . i) ≤0 (b.e. . . i). i. δ ∈ ∆ δ A ∈ I. . package diffalg. j) ⇐⇒ (a + c.

. j) 7−→ (i 1 . . . . Characteristic sets Let ≤ be a ranking on the set of derivatives ΘU .t. m. i m . . 3. . . . . 0)T . A Riquier ranking is a ranking ≤ for which there exists a positive integer s and an s × (m + n) real matrix M such that • for k = 1. . d) and compare two such pairs lexicographically. . . if we consider a rank t d ∈ R as a pair (t. Let f. . Differential polynomial i f = gd is called the initial of f . i m . kth column ck of M satisfies ck ≥lex (0. 1997. gd are uniquely P defined polynomials free of u f . . for example. d > 0} the set of all ranks.1094 O.   1 1 Vice versa. Using this embedding. km . 0 ≤lex M k1 . l) if and only if (m+ j) (m+l) M i 1 . Let d = degu f f . . . where g0 .r. . or . in which matrix M is unknown and satisfies the requirements of the above theorem.r. . Now. . . by applying the simplex method (see Danzig (1998)). this is the case if and only if there exists a vector w ∈ Nm+n satisfying {w · (φ(u) − φ(v)) < 0 | (u.. The derivative θu j of the highest rank present in f is called the leader of f (denoted ld≤ f or u f when the ranking is j clear from the context). p ∈ K{U }. . Then f = dj=0 g j u f .. For a set A ⊂ K{U }. .e. Differential polynomial f is partially reduced w. then f is said to be (fully) reduced. if f is free of all proper derivatives θu p (i.. p 6∈ K. The above linear system of inequalities is a linear programming problem. Let hA = iA ∪ sA. v) ∈ R}. we can characterize Riquier rankings by matrices: Theorem 2 (Rust and Reid. If f is partially reduced w. . . has a solution. the set of initials and the set of terminals of A are denoted i A and s A respectively.t. .0 . 0.11). . Denote by R = {t d | t ∈ ΘU. . Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 Let us embed Nm × Nn into Nn+m by using the following mapping: (m+ j) φ : (i 1 . .. . . . . . . i m . . . 0. The rank rk f is differential polynomial s f = j=1 jg j u ≤ the monomial (u f ) . i m . . f 6∈ K. (1) defines a Riquier ranking ≤ M . v) ∈ R}. A ranking on the set of derivatives ΘU induces a linear order on the set of ranks R. . According to Sturmfels (1996. . the question whether a finite set R ⊂ (Nm × Nn )2 is contained in a Riquier ranking reduces to the question whether the system {M(φ(u) − φ(v)) ≤lex 0 | (u. Proposition 1. . . j) ≤ (k1 . . and Pd j−1 is called the separant of f . 1 . 0) (1) • (i 1 . . . . . . p. 0. . any s × (m + n) real matrix M of rank m + n satisfying Eq. . km . . and let f ∈ K{U }. the set of its ranks is rk≤ A = {rk≤ f | f ∈ A}. . . . θ 6= 1) of the leader of p. Theorem 6). which can be solved. . . p and degu p f < degu p p. . . . d Similarly.

t.r. It follows from the definition that all characteristic sets of X w. u p j for 1 ≤ i 6= j ≤ r are distinct. A nonempty subset A ⊂ K{U } is called (partially) autoreduced if every f ∈ A is (partially) reduced w. Proof. which. then we have rk≤ A = rk≤ B.t.t. if it is not (partially) reduced w. . Chapter IV.r. in 0 particular. C. Lemma 2. A. 1973. w. by induction i f 0 ∈ I . Chapter I.r.r.t. If k = l and rk≤ f i = rk≤ gi (1 ≤ i ≤ k). f 0 reduces to 0 in less steps than f . gl } be two autoreduced sets. C.t. we have i f ∈ I . A \ { f }. .t. Since rk≤ f is irreducible w.r. f 6= 0 be such that rk≤ f is irreducible w. . p . if either there exists j ∈ N such that rk≤ f i = rk≤ gi (1 ≤ i < j) and rk≤ f j < rk≤ g j . consider the corresponding reduction of f : f 0 = h · f − τ rk≤ f · θ p. and f − i f rk≤ f is reducible w. a set of differential polynomials A ⊂ K{U }. p. Then rk≤ f 0 = rk≤ f and i f 0 = q. B = {g1 . a ranking ≤. If A = { p1 . f k }. Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 1095 irreducible. . Since also h 6∈ I (by Lemma 1) and I is prime. hence.r. i f ∈ I . implies that f 6= 0. Then rk≤ f 0 = rk≤ f and i f 0 = h · i f . . .r.t. Let q be the result of this reduction: q = h · i f − τ θ p.r. f is reducible w. then any two leaders u pi . 1973. . . (2) i f is irreducible w. an autoreduced subset of X of the lowest rank is called a characteristic set of X . Chapter I.r. if it is (partially) reduced w. Then i f ∈ I . A polynomial f is called (partially) reducible w. Since h 6∈ I (by Lemma 1) and I is prime. C. We say that A has lower rank than B and write rk≤ A < rk≤ B. proof of Lemma 2).t. For a subset X ⊂ K{U }. This implies h · i f ∈ I . An autoreduced set A is a characteristic set of X if and only if all nonzero elements of X are reducible w. Lemma 1 (Kolchin. we assume that elements of any autoreduced set are arranged in order of increasing rank of their leaders u p1 < u p2 < · · · < u pk . ≤ have the same rank. C. C. . by induction we obtain that i f 0 ∈ I .r. Section 9.r. and let f ∈ I.t. and C is a characteristic set. . since f 0 reduces to 0 in less steps than f . two cases are possible: (1) i f is reducible w. p. Now. f 6= 0. Let f 0 be the result of a reduction step of f w. Let A = { f 1 . Now. every polynomial p ∈ A. or k > l and rk≤ f i = rk≤ gi (1 ≤ i ≤ l). Since f ∈ I .  .t. C. . Section 9). Every autoreduced set is finite (Kolchin.r.r. Section 10).t. we obtain that h · i f ∈ I .t. Let C be a characteristic set of a prime differential ideal I w. p. If q 6= 0.t.r.t. Any nonempty family of autoreduced sets contains an autoreduced set of the lowest rank (Kolchin. O. then ∞ hC ∩ I = ∅. h ∈ h∞ If q = 0. A differential polynomial f is called (partially) reduced w.t. where h ∈ hC∞ . pk } is an autoreduced set.r. a polynomial p ∈ C. 1973.r.t. . If C is a characteristic set of a prime differential ideal I .

if we take an arbitrary characteristic set w. . Definition 2.r. Indeed. it may not be characteristic for all rankings from ≤∈ Σ I (R). the set Σ I (R) is convex in the following sense: Definition 3.t. Hence.1096 O.r.. and let ≤ be a ranking. ≤. ≤2 ] is defined as the following set of rankings: [≤1 . Define the subcone corresponding to the marked characteristic set C ≤ as follows: 0 subcone(C ≤ ) = {≤0 | C ≤ = C ≤ }. a ranking ≤∈ Σ I (R). ≤0 ] ⊂ R. differential) case. unlike the algebraic case. Moreover. one can define the rank of a polynomial f ∈ K[x1 . called the Gröbner cone.r. For the purposes of construction of the universal characteristic set.r. the autoreduced subset of I of the smallest rank is the reduced Gröbner basis of I (up to multiplication by constants from K). is proposed in the following section.t. For two rankings ≤1 .e. Definition 5. ≤. a set A is autoreduced if and only if it is autoreduced w. Furthermore. . in which the reduced Gröbner basis of I w. given an arbitrary characteristic set.t. ≤ is unique. and let C be a characteristic set of ideal I w. and let ≤ be a ranking. Let X ⊂ K{U }. Let I be an ideal. and let R = rkchar≤ I for some ranking ≤. For this definition of rank (which is rather unusual). Let C be a set of differential polynomials. this set corresponds to an open convex cone in Rn .t. ≤0 ∈ R implies that [≤. for a polynomial ring K[x1 . . if ≤. Differential Gröbner cones Definition 1. v ∈ Nm × Nn (u ≤1 v ∧ u ≤2 v) ⇒ u ≤ v}. The above definition is a generalization of the concept of Gröbner cone for algebraic ideals. Note that. . Define the differential Gröbner cone corresponding to R as follows: Σ I (R) = {≤0 | rkchar≤0 I = R}.r. Let also cone I (≤) = Σ I (rkchar≤ I ) denote the differential Gröbner cone containing ranking ≤. Σ I (R) = {≤0 | rkchar≤0 I = R} is the set of all admissible orderings. xn ] to be equal to its leading monomial w.r. ≤2 ] = {≤ | ∀ u. ≤2 . we are interested in finding a set that is characteristic for all rankings from a differential Gröbner cone. ≤. . 5. Strong characteristic sets Definition 4. Define the characteristic rank of X w.t. The algorithm for constructing such a set. for which the set of leading monomials of the reduced Gröbner basis is equal to R. in the general (i. According to Mora and Robbiano (1988). .t. Let ≤ be a ranking. rk≤ f ) | f ∈ C} and call it a marked set of differential polynomials. to be equal to the rank of any characteristic set of X w. ≤ in the usual algebraic sense. A set of rankings R is called convex. Denote by C ≤ the following set: C ≤ = {( f.r. .t. ≤. rkchar≤ X . and rkchar≤ I is the set of its leading monomials. a characteristic set of a differential ideal I may not be unique. Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 4. the segment [≤1 . . xn ] and an admissible monomial ordering ≤. . As is shown below.

which implies that f is reducible by f 1 w. and none of these sets is a strong characteristic set of I . denote by allrk( f ) the set of its ranks w. since the inverse inclusion always holds. if for any ranking ≤0 ∈ cone I (≤). Proof.. .r. . . subcone({u x }≤ ) and subcone({uu x }≤ ). For a polynomial f . consider prime differential ideal I = [u x ].t. Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 1097 It follows from the above definition that for any ≤0 ∈ subcone(C ≤ ). we have cone I (≤) = R.t.r. Lemma 3. . {vu x }.t. Also. ≤0 .r.r.r. sets {u x } and {uu x } are strong characteristic sets of I w. then we obtain that the differential Gröbner cone cone I (≤) is convex. any ranking. Now. Thus. Since ≤0 6∈ subcone(C ≤ ). since rk≤0 f < rk≤ f and therefore rk≤0 f cannot be reducible by rk≤ f . Assume that C is not strong. Indeed.r.r. there exists a polynomial f 1 ∈ C such that rk≤ f 1 = rk≤0 g. for every t ∈ allrk( f ).r.  . Let C be a characteristic set of a prime differential ideal I w. all possible rankings.t. ≤2 ∈ subcone(C ≤ ). a proper subset of the set of all rankings R (e. ≤.r. Hence.r. a ranking ≤ is called strong. suppose that rk≤0 f is reducible by g ∈ C 0 . Then rk≤0 f is irreducible w. hence the corresponding subcones are proper subsets of R. ranking ≤ is strong. Then rk≤0 f is also reducible by rk≤0 g. {u yy u x }. Proof. ≤.r. .r. Then there exists a ranking ≤0 ∈ cone I (≤) \ subcone(C ≤ ).t. Therefore.t. . rk≤0 C = rk≤ C. set C is a characteristic set of I w. since rkchar≤ I = {u x } for any ranking ≤.r. This contradicts the condition ∀ f ∈ C ∀ t ∈ allrk( f ) it f 6∈ I. C 0 and ≤0 .t. . Since ≤0 ∈ Σ I (rk≤ C). {u y u x } is a characteristic set of I for all rankings ≤ satisfying u y < u x ). ≤0 and has the same rank as w.t.r. are equal to the set of all rankings R. In other words. ∈ subcone(C ≤ ). ≤ satisfying the following condition: ∀ f ∈ C ∀ t ∈ allrk( f ) it f 6∈ I. A characteristic set C of ideal I w.t. Let ≤1 . for any ranking ∈ [≤1 . Then C is strong. if subcone(C ≤ ) = cone I (≤). there exists polynomial f ∈ C such that ld≤0 f 6= ld≤ f . a ranking ≤ such that t = rk≤ f . it f denotes the initial of f w. for any ranking ≤. ≤. ≤.r. ≤2 ] we have C  = C ≤1 = C ≤2 . any ranking ≤. Each of these sets is a characteristic set of I w.t.t. Then C ≤1 = C ≤2 = C ≤ . Hence. Let C 0 be a characteristic set of I w. Theorem 3.t. Note also that the above equality of two sets can be replaced by inclusion subcone(C ≤ ) ⊇ cone I (≤). . The set subcone(C ≤ ) is convex. and subcone(C ≤ ) is convex. Lemma 2 applies and we obtain that i≤0 f ∈ I . Below we prove that a strong characteristic set exists for any prime differential ideal w. a characteristic set C w. Then.t. O.g. Hence. If there exists a strong characteristic set of I w. It remains to notice that f 1 6= f .t. consider sets {u y u x }. Contradiction.t. sets {u x } and {uu x } are characteristic sets of I of rank {u x } w.  Definition 6. For example. the corresponding subcones. by Definition 3.

we also include the proof of this result here. Lemma 4 (Kolchin. since initially C is a characteristic set w. we have allrk( f ) = {u x . Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 The above theorem provides an algorithm for construction of a strong characteristic set.e. ≤. We illustrate this algorithm on our example. the fan is finite for any differential ideal.t.r. Also.r. rk≤ it f ≤ rk≤ f . For any differential ideal I ⊂ K{U }. Ld(I ) is the family of sets of leaders of characteristic sets of I w. {u y u x }. Let t1 . family Ld(I ) is finite.t. Chapter 0. ≤}. if there exists θ ∈ Θ such that t2 = θt1 . Algorithm StrongCharSet(C. t2 ∈ ΘU be two derivatives. we have it f = u x ∈ I .. Lemma 5. Let t1 . . We say that t2 is a derivative of t1 .t. Next. Hence. Gröbner fan and universal characteristic sets It follows from Definition 2 that distinct differential Gröbner cones do not intersect each other and that every ranking belongs to a cone. we obtain that rk≤ C is an invariant of the above algorithm. ≤) repeat flag = false for f ∈ C do for t ∈ allrk( f ) do if it f ∈ I then C = C \ { f } ∪ {it f } flag = true end if end for end for until flag = false return C end The algorithm terminates. it is a strong characteristic set of I .g. .1098 O. hence the algorithm replaces f by u x and thereby obtains a strong characteristic set of I . Then there exist indices i < j such that t j is a derivative of ti . For f = u y u x . To make this paper self-contained. In other words. 1973. the result of the algorithm is a characteristic set of I w. The set of all differential Gröbner cones is called the differential Gröbner fan. . and for any polynomial f and derivative t . . for t = u y .r. e. According to the above Theorem 3. 6. i. ⊂ ΘU be an infinite sequence of derivatives. ≤.r. We will first show that the set Ld(I ) is finite for any ideal I . t2 . set of I w. given any characteristic set of a prime differential ideal. Let Ld(I ) = {ld≤ (A) | ≤ is a ranking and A is a char. Let a characteristic set of ideal I = [u x ] be given. Lemma 15). all possible rankings. because each replacement of f by it f reduces the total number of derivatives present in C. u y }.t. As follows from Golubitsky (2004. Theorem 3). the set of all rankings can be represented as a disjoined union of differential Gröbner cones.

any ranking ≤∈ Σ1 . .t. if f ν(i1 +1) remains in the sequence. For each polynomial f i . if f ν 2 (1) remains in the sequence. . let i 2 = ν(i 1 + 1). Proceeding in the same way. we obtain an infinite sequence of indices i 1 < i 2 < . . let i 2 = i 1 +1. therefore ld≤ f ik is not a derivative of ld≤ f i j (≤∈ Σik ). Then A1 is also a characteristic set of I w. .  . A2 and any ranking ≤∈ Σ2 . f 2 . an infinite sequence of autoreduced sets A1 . ld≤ f j > ld≤ f i (≤∈ Σ j ). Applying the above argument. For each L ∈ Ld(I ). let i 1 = ν(1). ld≤ f ν j (1) is not a derivative of ld≤ f ν i (1) .t. . Indeed. . . a ranking ≤1 ∈ Σ1 . According to the pigeonhole principle. In this case f i ∈ A j for all j ≥ i. the characteristic set of A2 ∪ { f 3 } is the same and the polynomials in it have the same leaders. Therefore. . . there exists a polynomial f 2 ∈ I reduced w.. Otherwise. ld≤ f 1 = ld≤0 f 1 . f ν(1) . where ν is applied k times. Since f 1 contains only a finite number of derivatives. there exists an infinite subset Σ1 ⊂ Σ such that for all ≤. The set A2 cannot be a characteristic set of I for some ≤∈ Σ2 (according to the definition of Σ ). Denote by ν k (i) the expression ν(ν(.. Otherwise. and an infinite sequence of sets of rankings Σ1 ⊃ Σ2 ⊃ . If f i1 +1 remains in the sequence. such that for all j. . Hence. . Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 1099 Proof. A2 . f 2 . let i 1 = 1. But the latter is not possible. let i 1 = ν 2 (1).r. ld≤ f ν j (1) < ld≤ f ν i (1) (≤∈ Σν j (1) ). Hence.r. hence there exists a polynomial f 3 ∈ I reduced w. Suppose A1 is a characteristic set of I w. we construct an infinite sequence of polynomials f 1 . . . Suppose that Ld(I ) is infinite. . and so on.r.t. We will either find an index i 1 such that f i1 remains in the sequence. call this characteristic set A2 . But the fact that both f i j and f ik (i j < i k ) remain in the sequence means that they both belong to the autoreduced set Aik . Thus we have constructed an infinite sequence of derivatives {ld≤ j f i j | ≤ j ∈ Σi j }. A1 and any ≤∈ Σ1 coincide. there exists an infinite subset Σ3 ⊂ Σ2 such that for all ≤∈ Σ3 . we show that the process will eventually stop and we will find an index i 2 such that f i2 remains in the sequence. Otherwise. and denote the smallest such j by ν(i). Then the set Σ = {≤ L | L ∈ Ld(I )} is infinite. Let f 1 ∈ I be a differential polynomial. . . Now we will construct a subsequence of f 1 . ≤0 ∈ Σ1 . and so on.t. call this characteristic set A3 . In this case we say that f i is followed by a smaller derivative. ν(i) . f i j remains in the sequence.. denote by ≤ L the corresponding ranking. O. . or will construct an infinite sequence f 1 .)). A1 is autoreduced but not a characteristic set of I . This contradicts Lemma 4. According to the pigeonhole principle. it follows from the definition of ν(i) that for all i < j. . this contradicts the definition of the set of rankings Σ . (2) There exists j > i such that ld≤ f j < ld≤ f i (≤∈ Σ j ). which contradicts Lemma 4. f ν 2 (1) . the characteristic set of A1 ∪ { f 2 } is the same and the polynomials in it have the same leaders. . if f ν(1) remains in the sequence. A1 and any ≤∈ Σ1 . because characteristic sets corresponding to different rankings in Σ have different sets of leaders. . . none of which is a derivative of another one. there exists an infinite subset Σ2 ⊂ Σ1 such that for all ≤∈ Σ2 . one of the following two options is possible: (1) For all j > i. Continuing in the same way. and we say that f i remains in the sequence. If f 1 remains in the sequence.r.t. contradicting Lemma 4. However. and let A1 = { f 1 }. according to the pigeonhole principle. since the reduction relations w.r.

≤) U =StrongCharSet(C. ≤} is finite.t.e. Termination is guaranteed by the above Theorem 4.t. from the efficiency point of view. rk≤i Ci = ri (i = 1. ≤. . which contradicts the fact that C2 is a characteristic set of I w. . Algorithm UniversalCharSet(C. lkik }. ≤) R = {rk≤ U } while ∃ ≤0 such that rkchar≤0 U 6∈ R do C 0 =convert(U . all possible rankings is finite. set R grows with each iteration of the while-loop. ≤0 ) can be performed using any of the following algorithms: PARDI (Boulier et al. set R in the above algorithm cannot grow infinitely. . .. The set of ranks of characteristic sets of any differential ideal I w. 1996) that there exist two distinct ranks r1 = j j {l1i1 .t.r. k}. Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 Theorem 4. lk k } ∈ R such that for all m ∈ {1. Below we give an algorithm which. given a characteristic set of a prime differential ideal. ≤. . Proof. Suppose that the set of ranks of characteristic sets Rk(I ) = {rk≤ (C) | ≤ is a ranking and C is a char. the set of derivatives present in r is the same. 2004). set of I w. U contains the corresponding characteristic set. A subset C of ideal I is called a universal characteristic set of I .r.t. .  The finiteness of the differential Gröbner fan implies the existence of a finite universal characteristic set in the following sense: Definition 7. ≤0 ) C 0 =StrongCharSet(C 0 . Take an index m such that i m < jm .≤0 ) U = U ∪ C0 R = R ∪ {rk≤0 C 0 } end while return U end The while-loop in the above algorithm has the following invariant: for each set of ranks stored in R.r. if for any ranking ≤. . the set Ld(I ) = {ld≤ (C) | ≤ is a ranking and C is a char. it is worthwhile to add intermediate characteristic sets to U immediately once they are constructed. Note also that the conversion step C 0 =convert(C.t. On the other hand. . ≤} is infinite.r. constructs a universal characteristic set. .r. According to Lemma 5. Take also the corresponding characteristic sets C1 . 2001). . set of I w. denote this set of derivatives {l1 . .1100 O. ≤2 .t.. 2). Then there exists an infinite subset R ⊂ Rk(I ) such that for all sets of ranks r ∈ R. . . This invariant implies correctness of the algorithm.r. the union of strong characteristic sets corresponding to the cones in the differential Gröbner fan is a universal characteristic set. since family {rkchar≤ I | ≤ is a ranking} is finite.t. Let f ∈ C1 be the polynomial with rk≤1 f = lm . ≤ is a characteristic set of I w. i m ≤ jm .. lk }. . C2 and rankings ≤1 .r. In fact. or differential Gröbner walk (Golubitsky. Then f is irreducible w. r2 = {l11 . C2 and ≤2 . It follows from Dickson’s Lemma (Cox et al. the characteristic set of C w. . ≤. 1999). In the latter case. ≤2 and f ∈ I . im i. . DFGLM (Boulier. In fact. .

We illustrate the performance of the algorithm on an example from Boulier et al.t. since U is finite). several different rankings. the algorithm proceeds by calling the Rosenfeld Gröbner function from the diffalg package for 37 different rankings. −2v 2y + 1 − 2v 2yy + v 4yy . which the marked derivatives are leaders. which is restricted to Riquier rankings only. However. A rather strange phenomenon has been observed during this computation. vx x − u x ]. −v yy u + v y vx y + v yy . In particular. them. all previously constructed polynomials are used as its input. when the Rosenfeld–Gröbner algorithm was applied directly to the initial set of generators of the ideal I for one of these rankings. we already knew some characteristic sets of this ideal w. we decided. Before we started our experiments. O.r. determines whether there exists a Riqiuier ranking w.t. computes the above universal characteristic set in 9 seconds. and if so. Thus. Consider the prime differential ideal I = [u 2x − 4u. v y vx y + v yy − v 3yy . given a finite set U . u x y v y − u + 1. except for the polynomial vx x x − 2.r. compute rkchar≤ U for each of them (the computation of the characteristic set U is trivial. u y u − u y − v y vx x . . (2001). u 2y − 2u . We have obtained a set of polynomials that contains a characteristic set of I for each ranking from R0 . Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 1101 The above algorithm works for general rankings. instead of implementing the generation of all possible rankings on the derivatives present in U at each iteration of the while-loop. vx x − uvx x + 2v y vx y . Now notice that all polynomials in this set. As mentioned above. and the union of the corresponding characteristic sets consists of the following 19 polynomials: −32v 2y + 16 − 8vx2x + vx4x . we can generate all Riquier rankings on the derivatives present in it. It turned out that there are 37 Riquier rankings on the above derivatives (call the set of these rankings R0 ). In an attempt to simplify our implementation.3 the computation took more than 5 h and 1 GB of memory. given a finite marked set of polynomials.t. and check whether at least one of the resulting characteristic ranks is not in R. −vx x + 2v yy . u x u − u x − 2v y u y . u y . Our algorithm. after which it was interrupted.01. 4vx x + 8v y vx y − vx3x . to generate first all possible rankings on the above 7 derivatives and compute the union of characteristic sets w. θ1 u1 > θ2 u2 when weight(θ1 u1 )>weight(θ2 u2 ). These characteristic sets contained 7 derivatives: u.01 on a Celeron 900 MHz under Red Hat Linux 9. the computation of the universal 3 The following ranking was used: the derivatives of u and v are ordered by weights u = 6. −vx x + uvx x − 2v y u y . vx x − u x . vx y . this subroutine allows us to generate all possible Riquier rankings on a finite set of derivatives. y = 1. we have obtained a universal characteristic set of I . −2v 2y + 1 − 2u + u 2 . Thus. u y u − u y − v y u x . or weights and indeterminates are equal and θ1 > θ2 for the lexicographic order x > y. vx x . 4u − vx2x . x = 4. v = 0. −v 2yy + u. implemented in Maple 9.r. depend only on the above 7 derivatives. or weights are equal and u1 = v.01 contains a subroutine (which is embedded in the procedure for specification of differential rings) that. u x . vx x x − 2. u 2x − 4u. v y . and the rank of vx x x − 2 is equal to vx x x for any ranking. The latter was done using the Rosenfeld– Gröbner algorithm implemented in the diffalg package of Maple 9. u2 = u. Thus. The diffalg package of Maple 9. v yy . except for the following part. each time the Rosenfeld–Gröbner algorithm is called. 4v y u + u x u y − u x u y u. constructs such a ranking (see also end of Section 2).

u. ≤. c2 = c1 c3 . vx2x } {u x . c4 . vx y . since some polynomials can be removed from it. vx4x . Again. u 2y } {vx x . each of which corresponds to a cone in the differential Gröbner fan: {vx x . To prove it. 2 Substitute this solution in the original system u 2x − 4u. −u x + vx x . −vx2x + 4u. v 2yy } {vx y . vx2x }. . vx2x } {u. −v 2yy + u. if one of these 3 polynomials contained at least one of the remaining 4 ranks. u y . v 2y . u. different rankings.r. we obtain that each of the presented 11 ranks requires a separate polynomial to produce it. u y . vx x − u x . 4 polynomials are necessary to produce the characteristic set w. vx x . u 2x . u x y v y − u + 1. u 2 . v 2y . we write down the ranks of all characteristic sets w. in this particular example. u 2y − 2u. Note also that. vx y . vx x . v 2y . v 4yy .t. v y vx y − v yy u + v yy . u 2x − 4u. vx x . v 2y . vx4x } {v y . This universal characteristic set turns out to be minimal. then the characteristic set containing this polynomial would not be autoreduced (this can be verified by considering all possibilities for these 3 polynomials to participate in the 9 characteristic sets).1102 O. Applying the same considerations to the remaining 4 polynomials. u 2y . and the second one implies 1 c2 − 1 c1 = ± √ . a ranking ≤ with ranks {u x . Indeed. v 2yy . we reduce the set to 11 polynomials: vx4x − 32v 2y − 8vx2x + 16. v yy . u 2x . vx x x − 2.r. they cannot contain any of the following ranks: v yy . u 2 . vx x x }. Since these 4 polynomials must form an autoreduced set w. v 4yy − 2v 2y − 2v 2yy + 1. Now we show that at least 11 polynomials are necessary to produce the above sets of ranks. c5 = 2 2 c1 (the solution depends on three parameters: c3 . Applying a greedy minimization algorithm. the universal characteristic set allows us to obtain the general solution of the original PDE system4 : vx x x − 2 = 0 ⇒ v = 13 x 3 + x 2 f 1 (y) + x f 2 (y) + f 3 (y) vx x − 2v yy = 0 ⇒ 2x + 2 f 1 (y) − 2x 2 f 100 (y) − 2x f 200 (y) − 2 f 300 (y) = 0 f 100 (y) = 0 ⇒ f 1 (y) = c1 y + c2 f 200 (y) = 1 ⇒ f 2 (y) = 12 y 2 + c3 y + c4 f 300 (y) = f 1 (y) ⇒ f 3 (y) = c61 y 3 + c22 y 2 + c5 y + c6 vx x − 4u = 0 ⇒ u = (x + c1 y + c2 )2 . u 2x } {vx y . Observe that the first and last equations are fulfilled automatically. in total. c6 ). v yy . there are 9 different sets of ranks.t. v 2y .t.r. Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 characteristic set turned out to be faster than the direct computation of a characteristic set for a particular ranking! The above universal characteristic set is not minimal. −2v 2y + u 2 − 2u + 1. v 4yy } {u x . vx x x . 4 This observation was pointed out to the author by Vladimir Gerdt. The first three of the above ranks require at least 3 different polynomials to produce them. −2u y v y + vx x u − vx x . u 2y } {u y .

which has played a key role for the solution of the system.r. Conclusion The concepts of differential Gröbner fan and universal characteristic set introduced in this paper raise several further questions about the algorithmic properties of differential ideals: (1) How efficiently can a universal characteristic set be constructed? It seems obvious that since the universal characteristic set includes characteristic sets for all rankings. Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 1103 7. compute the universal Gröbner basis of this ideal. while there exists a particular ranking. (4) Do arbitrary (radical) differential ideals always have strong characteristic sets? If this turns out to be true.. However. O. It may well be the case that if a polynomial participates in several characteristic sets corresponding to several different rankings.t. but there is another mechanism involved. especially when no monomial ordering is given a priori. the cost of its computation depends on the ranking. for which the direct application of the Rosenfeld–Gröbner algorithm takes much longer. the following general approach seems feasible: one can consider the sum of algebraic ideals corresponding to particular characteristic sets. the complexity of its construction cannot be less than the complexity of construction of a particular characteristic set. Oddly enough. as well as at proving the minimality. This suggests that not only the complexity of computation of the universal characteristic set is much less than the sum of complexities of computations of all particular characteristic sets (because many polynomials are shared by several characteristic sets). The same idea can be applied in the algebraic case to compute characteristic sets or Gröbner bases of algebraic ideals more efficiently. another ranking. to Vladimir Gerdt for productive discussions . the ideal that has a given characteristic set may not be unique. in general. (2) Can universal characteristic sets be used to solve PDE systems? It has been the case in our example that the universal characteristic set happened to contain a linear polynomial vx x x − 2. the differential Gröbner fan is finite for these ideals as well. an algorithm which computes several characteristic sets simultaneously (possibly in parallel) and uses intermediate polynomials obtained for one ranking to perform reductions w. and try to obtain a universal characteristic set from this basis. in order to look for the polynomials that facilitate the solution of the system. it seems reasonable to compute the universal characteristic set. (3) How to find the minimal universal characteristic set? It has been just a matter of chance that we succeeded at obtaining a minimal characteristic set for our example. for the above example we have observed just the opposite! Our program computes the universal characteristic in several seconds. which ranking it is. Thus. 1995)/characterizable (Hubert. since. Since a priori it is not clear whether there exists a ranking which would yield a linear polynomial in the characteristic set and. may perform better than the direct computation of a single characteristic set. the role of characteristic sets for arbitrary (as opposed to prime) ideals is not as important. if so. However. 2000) decomposition of a radical differential ideal? Acknowledgements The author is very grateful to Carlo Traverso for the invitation to work in his research group at the University of Pisa and for financial support. (5) Can one efficiently construct a universal regular (Boulier et al.

1999... Dover. 1973. E.1104 O. Academic Press. Gröbner Bases and Convex Polytopes. United States.. Lemaire. The author is also thankful to the anonymous reviewers for their comments and suggestions. Klüwer Publishers.. 1993. J. Kondratieva.. PARDI! Tech. G. 183–208. 641–662. Danzig. Lazard.. S. TU München. Little. pp. . M. Mora. Springer-Verlag. Sturmfels. Golubitsky / Journal of Symbolic Computation 41 (2006) 1091–1104 during the work on this paper. The Gröbner walk. 2001. 1993. Linear Programming and Extensions. V. Ideals.. Differential Algebra. Kolchin. T. In: Ganzha... M. J. Alg. 9–16. Ollivier.. Mikhalev. Rep. The author was supported by the Progetto di Interesse Nazionale “Algebra Commutativa e Computazionale” of the Italian “Ministero dell’Istruzione dell’Università della Ricerca Scientifica Tecnologica”. D. 29... E. E.. in Sc. C. Hubert. J. 1995. ISSAC’97. T. Representation for the radical of a finitely generated differential ideal. Symbolic Comput.. 1999.. F. Rankings of partial derivatives.. Boulier. and Algorithms.. Petitot. Robbiano.. ISSAC’95. 1996. 02-01-01033. Faugère.. In: Proc. A. Tech. of Tech. Rep. Kihei... O. Germany. LIFL. L. Cox. (Eds. 2004.. Comp. Vorozhtsov. Efficient computation of regular differential systems by change of rankings using Kähler differentials. Kalkbrener. 6.. T. and NSERC grant PDF-301108-2004. Maui. Efficient computation of zero-dimensional Gröbner bases by change of ordering.. Factorization-free decomposition algorithms in differential algebra.. P. Levin.. J. D. Differential Algebra and Algebraic Groups. Princeton University Press. F. Tech. 1996..). 329–344. G. M. Mora. Proc. pp.. Collart. Gröbner walk for characteristic sets of prime differential ideals. O’Shea. Université Lille. Golubitsky.. E. References Boulier.. Pankratiev. and to Alexey Ovchinnikov and Marina Kondratieva for reading the draft of the paper. Lazard. B. RFBR grant no. ACM Press.. Rep. Symbolic Comput. The Gröbner fan of an ideal. 158–166. D.. In: Proc. D. Gianni.. Ritt. American Mathematical Society. J.. Mall. F. Swiss Federal Inst. Varieties. Rust. J. Reid. 1997.. 207–221. A.. Boulier. 16 (4). M. 1988. pp.. E. Symbolic Comput. 1950. Hawaii. Differential and Difference Dimension Polynomials. 7th Workshop on Comp. 1998. F. Moreno Maza. F. 2000. Mayr..