Академический Документы
Профессиональный Документы
Культура Документы
S.P. Tsarev
Department of Mathematics
Krasnoyarsk State Pedagogical University
Lebedevoi,89
660049, Krasnoyarsk, Russia
tsarev@edk.krasnoyarsk .SU
usuaIly fix D-1 R = s~m R dz for arbitrary expression R The problem of checking identities involving also /in similar
and suppose that R is rapidly vanishing at -co. Practical expressions will be touched in the last section.
examples of such computations are numerous, especially in The case IV = O (no indefinite functions appear) was
the modern theory of integrable nonlinear partial differential studied in many papers, see [16, 6] and references therein.
equations (also called “soliton theory”), see e.g. [8, 9]. We introduce the notion of a “subnormal form” (which
Due to identities similar to is “close” in its properties to a normal form) for expressions
where formal inverses L-1 and solutions of (3) do not appear
p = (D-lp)(D-lg) = D-l(@-lg) + D-l(gD-lp) (1) in denominators.
310
K = ~ o L = ~ o M (the order of K is minimal). The so- i.e. L] o L2 = ~z o ~1 iff lI(z) = 12(z) – Dlog R – R,
lution space {~) of L CM(L, AI) (in a proper Picard-Vessiot R = 12(z) – ~1(z) E K. In particular if L1, L2 are similar,
extension of K) is spanned by sums (p + @) of solutions of 11(z) = 12(z) – D log Q, then they are interchangeable iff
Lq = O, &f@ = 0. We say that operator L is (right) trans- ~Qdz~K.
formed into L1 by an operator (not necessary reduced) B, The notion of symmett-tc product L @ M of two operators
andwrite L~Ll, if K= LCA4(L, B)= LIo B= BIo L. plays an important role in the differential Galois theory (cf.
In this case any solution of Ly = O is mapped by B into a [18, 19]). We define L @ M to be the lowest order operator
solution BY of Lly = O. L1 will also be denoted as LB. If annihilating all products (~ti) of solutions of Lp = O, MO =
GCD(L, B) = 1 one may find with rational algebraic opera- 0. The algorithm of computation of coefficients of L @ M
(which also lie in K) is straighforward. For example D @
tions an operator Bz such that LB ~ L, &.oB = 1(modL).
D = D, and if L = D2 – u(z) (the Schrodinger operator)
Operators L, LB in this case will be also called similar or oj
L @ L = D3 + 8uD + 2u’D — the well-known in solitonics
the same kind (in the given differential field K ). So for sim-
Magri operator.
ilar operators the problem of solution of the corresponding
Completely factorable operators (i.e. operators which are
equations Ly = O, LB z = O are equivalent.
decomposable into products of first order operators)
A (reduced) LODO is called prime or irreducible (in the
given differential field K) if it has no nontrivial factors aside L=(D–fl (z)) oo(D-jn (z)), f,~K, (4)
from itself and 1. Every LODO similar to a prime LODO is
also prime. Two (prime for simplicity) LODO P and Q are and solutions
called mterclmngeable m the product P o Q and this product
will be called mterchartgeable if P o Q = Q1 o PI, Q1 # P,
P’l # Q. In this case P is similar to PI, Q is similar to Ql
and P, 4 P.
p=pn D-l
(%’D-’(””’D-’:)
”””)“)
p~/p, = f,, of equations Lp = O were studied in [1, 4, 2, 3]
Theorem 1 Any two diflerent decompositions of a given in the more restricted case N = O (i.e. K = Q(z)). Obvi-
~(?DO L @to products of prime LODO L = PI o o Pk = ously (5) is a generalization of the studied in [7] expressions
PI o ..0 PP have the same number of factors (k = p) and t~D-lt”_l . .. D-ltl. t, 6 Q[ul(z), . ..)u~(z). Dul]. ..]
the factors are simdar in pairs (m some transposed order,), (i.e. polynomial t,)while p, in (5) may have the form
One decomposition may be obtained from the other through exp (~ fa dz). As one can show using the techniques of
interchanges of contiguous factors (in the pairs P, o P,+l ). [1, 4, 2, 3] in this general case sums and products of several
different functions of the form (5) may be again rendered
If P and Q are interchangeable then P o Q = Q1 o P1 =
into the same form (called d’.Alembert functions). This cor-
LCM (Q, PI) and vice versa. It is useful to study de-
responds to the fact that LCM(LI, Lz) and LI @ Lz are
compositions of operators which are LCk4 of prime oper-
also completely factorable. The former also follows from
ators Such operators will be called completely reducible.
the Loewy-Ore theory:
If L = LCM(PI, . . . PP) where P,, are prime, —oral(L) =
—
ord(Pl) + + ord(P~,), then L = P] o o PP-] o PP, Theorem 5 Let L = LCM(A)B) and A = A1 o ~~~o .4k,
F,< iirl, similar to P.,. B = Bl o o B~ be decompositions of A, B into prime
factors. Then m any prime decomposition L = L, o o L“
Theorem 2 If L = LCM(P,, . . . . PP) has an irreducible one may put all L, into correspondence to a subset of
r~ghtfactor L, which does not coincide with Pi, . . . . PP, then {.41) . ..4~}~} U {B], ... ,B~}, in such a way that the cor-
among P., one can find at least two operators similur to L,. responding operators will be similar.
corollary 1 If all PI, , Pk are of d~flerent kinds then L The complete factorizability of L] @Lz may be proved recur-
have only them as rtght factors (i.e. there are no parameters sively using (1) and LCM of operators annihilating the right
m the right factors). hand terms. This propert,v will be (partially) generalized in
section 4.
Theorem 3 The necessary and suficient conditions that a
L ODO be completely reducible is that two arbitrary factors 3 Combinations of d’Alembertian expressions (5)
in ang arb~trury decomposttton be interchangeable.
Such expressions are often used in solitonics. We give in this
Now let us suppose that some P, are similar, namely P1, section their separate treatment in order to give a ready to
. . . . P,,,, m ~ k are similar to some Lo, and Pk, k > m use technique for possible applications and explain typical
are not similar to Lo, and the subfield of constants of K is complexity problems arising in computations with integrod-
algebraically closed. ifferential expressions. Most of the results are easily gener-
alizable, see sections 4, 5.
Theorem 4 Let Lo %_Pl , . . . . Lo % P,,,. Then the As we have mentioned above any expression formed with
most general right factor P of L = LCM(P], . . . . Pk) similar (5) and operations +, –, *, D-1 again is reducible to the
to Lo is the result of transiorrnation of LO by ~ = >1Bl + same form. For example u = D–lp + gD– 1h (with nota-
tions p = u1(z), g = u2(z), h = us(z) for brevity) has the
AzB, ++ A,, Brn, A EQ.
(minimal) annihilator (i.e. the lowerst order LODO L such
For first-order operators the following simple criterions that Lu = O)
(( ))
p’+2fJ%+gl+AP+9h)
U=D–’ g’D–l (7) on the fact that actual checking of annihilation of ~k+l by
9’ ~k (and computation of ~k itself) is NOT necessary if at
least one of the first order factors of the minimal annihilator
The fist obvious disadvantage of (6) and (7) compared to Lk+l of ~k+l is not similar to any of factors of the minimal
the initial simple u consists in their complexity. Moreover annihilators of Pi, i ~ k or if the number of similar factors
due to non-unique factorization of L1: in Lk+ 1 of certain kind is greater than their number in ~,k.
This is obvious from Thgmems 1, 5. In the contrary case
“=(D-aO(D-($-@)O(D-~) ‘8)
the actual verification of Lk~k+l = O does not require com-
plete computation of ~k and may be greatly simplified. The
details will be covered elsewhere. Note that the minimal an-
} E ~ (which is its general form) we afso have
nihilators of individual p. of the form (5) are readily given
by (4), the algorithm of checking similarity of first order
g’ ~_l W(g + A)
71=(g+A)D-1— (9) operators is given in section 2; in the case of higher order
g+A 9’ factors (see [14, 10]) it requires finding rational (i.e. in K)
solutions of some LODES.
Another problem of (7), (9) concerns their initial data: the
usuaf D-l* = ~~m. dz fails in presence of denominators.
4 Expressions with non-d’Alembert ian elements
One shall properly choose the initial point and the integra-
tion constants in order to make (7), (9) correct. Let us consider expressions involving elements of K, solu-
The left hand side of (1) has as its minimal annihilator tions of e uations (3), operations +, –, * and (formal) in-
verses L-Y of mreducible operators of various orders. We
“=[(D-+)”D18[(D+
”D1= (10)
=( D-dl)o(D -dj)o(D-$)o D,
need some new results from the formal theory of LODO.
(D-$)
DM2=(D-$-$)O(Di$
)D=their
minimal annihilators, and their sum should be annihilated
ti(z) ~~(z), E = Ga, ~ = Eb, a(z), b(z) being solutions of
Aa=(Dn+ ..+ctn(z))a=O,
by (12)
Bb=(Dm+ ..+ fl~(z))b=O.
La = LCM(M1, M2) = (D–el)O(D–ez)O
(11) Let us fix
( ‘-$+) ”(D-fOOD
ao = a(th), al = Da(zo), . . . ,an_l = Dn-la(zo),
(13)
(with some very complicated el, ez) which is obviously not bo = b(zo), bl = Db(Zo),. . . . bm_l = Dm-lb(xo),
the minimal annihilator (10). As one can check La is di-
visible by Lz. So in order to make a one-to-one corre- then
spondence between d’Alembert expressions (5) and com-
pletely factorable LODO one shall tind minimal annihilat- a(n) (ssO)= – ~ cllai, a(n+l) = ~ K;(~:, a~)a~, . . . ~14)
ing LODO. This problem was solved (for the case IV = O, b(m)(~o) = – ~@ibi, btm+’) = ~S,(~i,@~)bi,. .
K = Q(z)) in [3, 2]. In section 5 we will give another solu-
tion for the more general case.
and (Ca)(Eb) 1===0= xi j Ci,jaibj where Ci,j are some ex-
Provided the minimal annihilator of some particular
d’Alembert expression (with hyperexponential pi, i.e. P: = pressions built from the coefficients of A, B, C, E. Since
exp(~– 1ji), ~i E K) is found we may think of this operator F(ab) = ~i,j Fi,j(. . .)aib, , Fi,j being expressions built
(with attached proper initial data for its solution in ques- from the coefficients of A, B, F which are linear alge-
tion) as a “normal form” of such expression encoding the braic w.r.t. the coefficients of F (no derivatives!) then
complete (independent of any particular representation) in- from F(ab) 1===0= (Ca)(E~) 1===0we have a system of mn
formation about this expression. The obvious complexity of
linem algebraic (not differential!) equations for rnn coef-
this approach urges for another solution. Partially imitat-
ficients of F. This system is non-degenerate since in the
ing [7] we will recursively use (1) for representation of any
contrary case we could find an operator F E K[D] of or-
polynomial combination of (5) to a linear combination of
der < mn annihilating any product ab which contradicts to
(5). Further reduction to a “subnormal” form is performed
the fact that A @ B (the minimal annihilator of such prod-
induct ively on the number of terms in thk linear combinw
ucts) has order mn. Thus we can find F ● K[D] such that
tion. One (5) is subnormal by definition. Two of them are
F(ab) = (Ca) (Eb) in any generic point zo for any a, b, which
subnormal if their minimal annihilators do not coincide oth- means that the solution spaces of (A@B)~ and (At)@ (BE)
erwise the sum is just one (5) with another initiaf data. If coincide.
we have p = ~i<~ Wi + $Ok+land the minimal annihilator
312
Remark. If orci(.4 R B) < ord(A)ord(f3) then the state- everywhere we will suppose that they are chosen in such a
ment of Proposition 1 may be wrong: take A = (D – 1) o way that that the formulas below hold. We will generalize
(D – 2), B = (D + 1) o (D + 2), then A @ B annihilates (1) fist.
(Llb, = err–’ = 1 = a2b2, alb2 = eZr-2T = e-’, a2b2 =
~,?.r(,-r —
_ P’, so .4@B = Do(D–l)o(D+I). But .4(D+Z)8B Proposition 3 For arbitrary functions f, 9 and opemtors
A, B E K[D]
has the solution space tilbl = ((D+ .r)ez) e-r = z + 1,
?i1b2 = (z+l)e-’, Z2fJ1= x+2, ?i2b2 = e2rc–r =ez, so has
(A-l~) (B-’g) = (A QB)-10
order 4.
rOrdL – I
Proposition 2 Ij L = L] o Lz then for operators M such
that the order ojLl @ Itf is rnazimal, LB M = (LI @ M)B o
[2,=0 (.4~t)-’(Djj) P,(g) +
(16)
(L2 @ Af) for some operator B E K[D]. ordM-I 1
313
The case LCM(A, B)- 1(f) for arbitrary ~ may be studied (similar to several of L; of the same kind, cf. Theorems 1, 4).
analogously. At this point it is useful to study decompo- Splitting each Pa off L*: L* = ~P; in turn we check if the
sitions of symmetric products A @ B of prime operators. given expression u is annihilated by ~. The same should
As the example (communicated to the author by M. Sosnin) be done for Qi (Al,. . . . A~i), in this case we get a set of equa-
A= B= D2–~D–z, A@ B= LCM(D, D2– ~D–2x) tions for Ai aa the criterion of annihilation.
shows, symmetric products of irreducible operators may be
reducible and moreover may decompose as L CM ‘s. One can Lemma 1 If L* = R; A = R; B and both R1, R2 annilti-
check that the following is true at least for symmetric prod- late u then u is annihilated by R3 obtained by splitting off
ucts of some second order prime operators. LCM(A, B): L“ = W3LCM(A, B).
Conjecture. If the symmetric product of two prime op-
Proof. LCM(A, B) =Bo A=/io B, Lu=KB”u =
emtors is not prime then their symmetric product decom-
poses as LCM of several prime operators, so in fact the so- B“~v for ~ = Rsu. We have B*v = ~“v = O. Since
lutions of the opemtors are Liouvillian. GCD(~”, B“ ) = 1 (otherwise LCM(A, B) should have
Solutions with algebraic logarithmic derivatives play a lower order) there exist operators F, G E K[D] such that
very important role in the modern theory of LODO (cf. [18, FB” + GA* = 1. Thus R3U = V = (~~” +G~”)v = O.
19]). Using this Lemma we see (cf. Theorem 4) that the set of
After construction of the “unratified sum” for a given solutions of the equations for the parameters Ai should form
.
expression and simplification using L CM decomposable op- a linear subspace. Let Ai,j = (~1, i,j, . , Aqi,,,j ~) be the basis
erators, one shall check inductively that each term of this of this subspace. Then we may split off the operator P =
sum is not annihilated by the minimal annihilator of the
LCM(Pil, . . . . P~~,Qjl (~jl,l), . . .): L* = R*P, where Pi-
sum of previous terms. This may be sometimes verified fast
are such that for splittings L* = R~i o Pji , I?ji u = O. Due
using the method indicated in the end of section 3. In the
contrary case one should “include” such term into the pre- to Lemma 1 Ru = O. As the obvious examples of products
vious ones changing appropriately their coefficients and/or of non-interchangeable operators show, the constructed h
initial data. The problem of construction of the minimal an- may still be not minimal. So the above process should be
nihilator is addressed in the next section. After all described continued. It obviously terminates producing the desired
transformations we get a particular form of the given initial minimal annihilator.
expression which bears many features of the normal form
treated in the simple denominator-free d’Alembertian case 6 Conclusion
in [7]. We will call such form “subnormal” due to its obvious
non-uniqueness. We discussed in the above sections only manipulation of
expressions whose (intermediate) denominators are only el-
5 Minimal annihilating operators ements of K, for example (D2 + D– 1 0
( *D-1 (i))
Here we solve (answering a question by B. Salvy) the prob- But it is obviously possible to prove or disprove identities of
lem of construction of the LODO which arminilates a given the more general form namely with arbitrary denominators
expression u and has the minimal possible order. Such min- (including solutions of (3) and operator inverses). Using the
imal annihilator is obviously unique. Using the techniques Ore identity (18) we may transform the given expression
of LCM and symmetric products we can always tlnd some into a sum of the form L:l(. . .) + . . . + L~l(. . .). Applying
annihilating operator. As the example (10), (11) shows this LCit4(Ll,.. ,, L~ ) (also checking beforehand that the initial
may result in an operator whose order is greater than nec- data vanish) and removing the obtained denominator we
essary. will decrease the “denominator embedding depth” (i.e. the
Remark. Although the LCM(L1, Lz) of order 5 in (11) maximal level of embedding of denominators cent aining op-
gives the minimal operator which annihilates all linear com- erator inverses and solutions of (3)) probably incre~ing the
binations u = AID-I (~D-19) + AzD-l(gD-l~) one can total comlexity of the expression. After removing all denom-
easily prove that any particular such combination is an- inat ors we may precede further as described in section 4.
nihilated by some 4-th order operator: u’ = Al f D– 1g + Many solitonic problems include also inverses of matrix
A~?D-l f , U“ = Al f’D-lg + Azg’D-lf + (h -t &)fg, LODO. A proper treatment of this case as well as the case
u = &f’’D-1g+A2g’’1f +( Al+&) (}g)g+Alflg+A2g2 j’j of partial linear differential operators deserves furthur study.
UIV _—.. . Complexity estimates of the above algorithms are still high
are obviously linearly dependent over K. On the
other hand for two solutions of two prime operators any par- and should be reduced.
ticular sum is annihilated only by the LCM of the operators
(cf. Theorem 5). 7 Acknowledgement
Let L = L1 0...0 Lk be some (not minimal) annihilating
operator decomposed into prime fact ors using the techniques Several ideaa described here appeared after fruitful discus-
of [5, 17, 20] (applicable over K). If all Li have order 1 then sions with Prof. S.A. Abramov, to whom the author wishes
the construction of the minimal annihilator (a divisor of L) to express his special gratitude. I like to thank Professor
ww given in [3] using the techniques of adjoint operators and B. Salvy for the correct reference to [16].
“accurate integration” of [2] (in the more general case of Ore
rings). In the case of arbitrary ord(Li) ~ 1 we shall apply References
the technique of section 2. Namely, all possible right factors
of the adjoint operator L* = L; 0 ...0 L; are either “is@ [1] ABRAMOV, S.A. On d’Alembert Substitution. Proc.
lated” factors PI, . . . . P, (each similar to some L;), s < k, ZSSAC’93, pp. 169-174.
or parametric “sheaves” Qi(Al, . . . . A*i) = QXIBl+.. +A~i Bgi
314
[2] ABRAMOV, S. A., AND VAN HOEIJ, M. Ore Rings: Ad- [19] SINGER, M. F., AND ULMER, F. Linear Differential
joint Polynomials, Operators, Equations. submitted to Equations and Products of Linear Forms. Preprint
ISSAC’97. (1996), 13p.
[3] ABRAMOV, S. A., AND ZIMA, E.V. Completely Fac- [20] TSAREV, S. P. Onsomeproblems in factorization of
torabe Annihilators. submitted to LSSAC’97. Iinearordinary differential operators. Programming 8
Gomputer Software 2’0, 1 (1994), pp. 27-29.
141
., ABRAMO\’, S. A., AND PETKOVSEK,M. D’Alembertian
Solutions of Linear Differential and Difference Equa- [21] TSAREV, S. P. An algorithm for complete enumeration
tions. PTOC.ISSAC’94, pp. 169-174. of afl factorization of a linear ordinary differential op-
erator. Proceedings of ISSA C ’96 (1996), ACM Press,
[5] BRONST~I~, Itf, An improved algorithm for factoring pp. 226-231.
linear ordinary differential operators. In Proceedings of
ISSA C ’94 (Ozford, U.K.) (1994), ACM Press, pp. 336-
340.
315