On the complexity of skew arithmetic

Let

be an effective field of constants of characteristic zero, so that all field operations can be carried out by algorithms. Given an indeterminate

and the derivation

, where

, it is well known [26, 8, 12, 28, 29] that the skew polynomial ring

behaves very much like an ordinary polynomial ring: there are skew analogues for each of the classical operations of division with remainder, greatest common divisors, least common multiples, etc. In this paper, we will study the complexity of these operations. For this purpose, it will be more appropriate to work in the ring

instead of

. In analogy with the commutative case, we will give bounds for the computational complexities of the various operations in terms of the complexity of operator multiplication.

For our complexity measures, we make the customary assumption that all field operations in

can be carried out in constant time

. We will try to express the complexities of our algorithms in terms of the following standard complexities:

We will denote by

the subset of

of polynomials of degree

. Likewise, we denote by

the set of operators

of degree

and degree

Now consider two linear differential operators

. We start with studying the following complexities:

The special case

was first studied in [20], where it was shown that

, using evaluation-interpolation techniques. The inverse bound

has been proved in [5]; this paper also contains detailed information on the constant factors involved in these bounds. Recently (and after the writing of a first version of this paper), the quasi-optimal bound

was proved in [2].

For fixed constants

, one has

, etc., by splitting the multiplicands in a finite number of pieces. In this paper, we will freely use this remark without further mention. In order to simplify our complexity estimates, it will be convenient to make a few additional assumptions. First of all, we will assume that

, whence in particular

. We will also assume that the function

is increasing and that

is increasing in both

and

. This will indeed be the case for the complexity bounds for

that will be given in Section 3.

In Section 2, we will first prove (see Theorems 2 and 3) that the problems of multiplication and operator-vector application are essentially equivalent when

. We also recall the best existing bounds for operator multiplication.

In Section 3, we show that the problems of computing fundamental systems of solutions and its inverse can be reduced to operator multiplication modulo a logarithmic overhead (see Theorems 7 and 8). This provides a dual way to perform operations on differential operators by working on their fundamental systems of solutions. In Section 3 and all subsequent sections, we always assume that

. This is indeed required for the truncations of the fundamental systems of solutions at order

to be linearly independent.

In Section 4, we start with the operations of exact right division and right division with remainder. In Section 5, we consider greatest common right divisors (gcrds) and least common left multiples (lclms). Again, we will show how to express the complexities of these operations essentially in terms of the complexity

of multiplication (see Theorems 13, 15, 18 and 21).

For several of our algorithms, we need to work at a point where certain operators are non singular. If we only need the input operators to be non singular, then it is easy to find a point where this is the case. If we also need the output operators or certain auxiliary operators to be non singular (as in Section 5), then we resort to picking random points, which are non singular with probability 1. In Section 5.2 we present additional techniques for turning algorithms which rely on random point picking into randomized algorithms of Las Vegas type and into fully deterministic algorithms.

For technical reasons, we found it convenient to work with respect to the Euler derivation

instead of

. Nevertheless, operators

can be converted efficiently into operators in

and vice versa, modulo an increase of the degree

with the degree

(see Lemma 6). Using our assumption that

, such increases of the degree

only gives rise to constant overheads in the complexity bounds. Hence, the complexity bounds for our main algorithms from Sections 3, 4 and 5 still hold when replacing

. In addition, some of the algorithms can be adapted to directly use

instead of

, without the need for any conversions (see Remark 11).

To the best of our knowledge, the idea to perform operations on linear differential operators via power series solutions was first proposed (but only partially worked out) in [4, Chapter 10]. In this paper, we use a slightly different technique: instead of a single power series solution, we prefer to consider a fundamental system of solutions. This has the advantage of forcing a clean bijection between operators and solution spaces, thereby avoiding part of the randomness in the proposals from [4, Chapter 10].

It is also possible to mimic classical divide and conquer algorithms for right division, greatest common right divisors and least common left multiples, while using adjoints in order to perform the recursive operations on the appropriate side. Such algorithms were partially implemented inside Mathemagix [24] and we plan to analyze this technique in more details in a forthcoming paper.

Various complexity results for computations with linear differential operators and other skew polynomials were previously obtained [17, 14, 15, 25, 16, 4]. Especially the computation of greatest common right divisors and least common left multiples of two or more operators has received particular attention. After the publication of a first version of this paper [23], the complexities of several classical algorithms [19, 33, 25] for the computation of least common right multiples were studied in great detail in [6], and new improvements were proposed there.

The complexities of most of the algorithms in this paper are stated in terms of the input and output sizes. The uncertified randomized algorithms for gcrds and lclms are optimal up to logarithmic factors from this perspective, which yields an improvement with respect to the previously known complexity bounds. In the context of certified randomized algorithms (i.e. of Las Vegas type), the complexity bounds remain quasi-optimal in terms of the size of a suitable certificate. From the deterministic point of view, the new algorithms for gcrds and lclms are suboptimal.

Acknowledgment. We are grateful to the second referee whose questions and remarks have lead to several improvements with respect to the first version of this paper.

2.Evaluation and interpolation

The key argument behind the proof from [20] that

is the observation that an operator

is uniquely determined by its images on the vector

. This makes it possible to use a similar evaluation-interpolation strategy for the multiplication of differential operators as in the case of FFT-multiplication of commutative polynomials. More precisely, given

, let

be the matrix of the mapping

with respect to the bases

and

The evaluation and interpolation steps can be done efficiently using the following lemma, which is essentially contained in [5]:

of bandwidth

. The coefficients on the

-th subdiagonal of

are exactly the result of a multipoint evaluation of

. It is classical [27, 35, 3] that both multipoint evaluation and the inverse operation of interpolation can be performed in time

. Doing this for each of the polynomials

yields the result.

Proof. Let

and assume that we want to compute

. We may evaluate

in time

. We may also evaluate

in time

. Using Lemma 1, we may recover

from

in time

. This completes the proof.

Proof. Assume now that we are given

, as well as a vector

and that we want to evaluate

. This is equivalent to evaluating the operator

at the vector

. It is classical [1] that

can be computed in time

. Using Lemma 1, we may compute the unique operator

with

in time

. We may next compute the product

in time

. Lemma 1 finally allows us to evaluate

in time

, thereby yielding

The above results immediately imply the bound

from [20] by the computation of a product

to the computation of a matrix product

After the publication of a first version of this paper, the following quasi-optimal bound for

was established in [2, Theorems 3 and 5].

Proof. By the result from [5], the problem is equivalent to the computation of

operators

with a fixed operator

. Setting

, we may compute

in time

. We may directly read off the products

from the result.

In this paper, we have chosen to work with respect to the derivation

instead of

. The following result from [5, Section 3.3] can be used to efficiently convert between operators in

and

(in [20], we proved a somewhat weaker result which would also suffice for the purposes of this paper). We have written

for the set of operators of degree

and degree

3.Local solutions

From now on, we will assume that

. We recall that an operator

of order

is said to be non singular at

, if its leading coefficient

does not vanish at

. We will say that an operator

of order

is non singular (at the origin) if

and

is non singular as an operator in

Given a non singular differential operator

of order

, the equation

admits a canonical fundamental system

of solutions in

, with the property that

and

for all

with

. Conversely, given a

-linearly independent vector of power series

, there exists a unique monic operator

of order

with

. Let us show how to convert efficiently between these two representations.

Proof. Modulo multiplying

on the left by

, we may assume without loss of generality that

is monic. Since

is non singular at the origin, we have

. Rewritten in terms of

, this means that

is of the form

for certain

. Setting

, we observe that

maps

into

. We now compute

using the “recursive” formula

The equation (4) is a schoolbook example for applying the strategy of relaxed resolution of power series equations [21, 22]. Since

operates coefficientwise, it can be computed in linear time. The main cost of the computation therefore reduces to the relaxed evaluation of

. Using fast relaxed multiplication, this amounts to a cost

In what follows, given a non zero series

, we denote by

its valuation. Given a vector

of elements in a

-vector space, we will also denote by

the subvector space generated by the entries of

, and

Proof. Modulo a triangularization of

, we may assume without loss of generality that

. We define operators

Then

annihilates

and for any other operator

with

, we would have

, which is in contradiction with the fact that

. Moreover, by induction over

, we observe that the coefficient of

is given by

and the coefficients of

can be expressed in terms of the coefficients of

. In particular, the truncation of

at order

is uniquely determined by the truncation of

at order

In order to explicitly compute

up to a given order, it is more efficient to use a divide and conquer approach. More precisely, given

we compute

using the following method:

Remark 9. If

is increasing in

for some

, then the bound further simplifies to

Remark 10. We notice that the operator

in Theorem 8 is singular if and only if

, and if and only if

Remark 11. The algorithm from the proof can be adapted so as produce a vanishing operator in

instead of

. Indeed, for this, it suffices to take

Although a general operator

can be singular at the origin, many operations on operators (such as right division and greatest common right divisors) commute with translations

, and Lemma 6 may be used in conjunction with the following lemma in order to reduce to the case when

is non singular at the origin.

Proof. Let

be the leading coefficient of

. Since

, we have

for some

. Using fast multipoint evaluation [7], we may find such a point

in time

4.Right division

From the formula (3) it is clear that both the degrees in

and

are additive for the multiplication of operators

. In particular, if

and

is left or right divisible by

, then the quotient is again in

Proof. By Lemmas 12 and 12, and modulo a shift

, we may assume without loss of generality that

and

are non singular at the origin. We now use the following algorithm:

It is classical that euclidean division generalizes to the skew polynomial ring

. In other words, given operators

where

, there exist unique operators

and

with

and

. If

and

is the leading term of

with respect to

, then left multiplication of

allows us to remain in the domain

: there exist unique

and

with

and

. The operators

and

are usually called pseudo-quotients and pseudo-remainders. In some cases, a non trivial polynomial can be factored out in the relation (6). Let

be monic, of maximal degree, such that

. Then we call

and

the “simplified” pseudo-quotient and pseudo-remainder of

and

Proof. Let

for each

. Modulo a base change, we may assume without loss of generality that

. Let

be the operator with

In other words,

and its inverse

operate coefficientwise and

coefficients can be computed in time

and we observe that the coefficient of

in the righthand side of (8) only depends on earlier coefficients of

. In particular, we may solve the equation using a relaxed algorithm. Then the main cost is concentrated in the relaxed evaluation of

. As in the proof of Theorem 7, this evaluation can be done in time

Proof. Modulo a shift

, we may assume without loss of generality that

and

are non singular at the origin. We now use the following algorithm:

Remark 16. In the above proof, we have assumed that

is known beforehand. In general, we may still apply the above algorithm for a trial value

. Then the algorithm may either fail (for instance, if

), or return the triple

under the assumption that

. We may then check whether the triple is correct in time

. Applying this procedure for successive guesses

, the algorithm ultimately succeeds for an

with

. Using the monotonicity hypothesis, the total running time thus remains bounded by

5.Euclidean operations

5.1.Randomized algorithms

It is classical that greatest common right divisors and least common left multiples exist in the skew euclidean domain

: given two operators

, the greatest common right divisor

and the least common left multiple

are the unique monic operators with

Assume now that

and let

and

be monic polynomials of minimal degrees, such that

and

are in

. Then we call

and

the (simplified) pseudo-gcrd and pseudo-lclm of

and

Proof. Let

and

. If

is non singular, then it admits a canonical fundamental system of solutions

with

and

for all

with

. In particular,

. Since

is the least common left multiple of

and

, we also have

, which completes the proof of the first equality. If

is non singular, then we obtain the second equality in a similar way.

is non singular, then we also have

and

, since

and

are non singular. Now

, whence

. If

is non singular, then we obtain the first equality in a similar way.

Remark 19. In the above proof, we have again assumed that

is known beforehand. Below, we will discuss ways to check the correctness of the computed result for a trial value

, after which a similar strategy as in remark 16 can be applied. During the relaxed computation of

and

, we may also check whether

at each next coefficient. In the particular case when

, the running time of the algorithm will then be bounded by

, where

is the smallest order at which common solutions no longer exist. This kind of early termination only works for this very special case.

Remark 20. Notice that

might be singular at the origin, even if

and

are not. This happens for instance when

is the minimal annihilator of the vector

and

the minimal annihilator of the vector

, so that

5.2.Certifying correctness

The assumption that

should be non singular is still a bit unsatisfactory in Theorems 18 and 21, even though the probability that a randomly chosen point is singular is infinitesimal. If we drop this assumption, then we still have

in the proof of Theorem 18. Consequently, “candidate” pseudo-gcrds

found by the algorithm are genuine pseudo-gcrds whenever

pseudo-divides both

and

. Using the right division algorithms from the previous section, this can be checked in time

in the case of gcrds and

in the case of lclms.

Remark 22. Using the polynomial linear algebra techniques from [17, 6], it is likely that one may prove that

for some

and

. If this is indeed the case, then the trial divisions of

and

can actually be carried out in time

An alternative way to check whether candidate gcrds and lclms are correct is to compute Bezout and Ore relations. More precisely, given

with

, there exist operators

with

and

. The

matrix at the righthand side will be called the Euclidean matrix

and

. In a similar way as above, we may define a (simplified) pseudo-Euclidean matrix

with entries

, whenever

. We will say that

is non singular at

, if the denominators of

and

do not vanish at

We finally compute

from

and

using rational function reconstruction. The complexity analysis and the remainder of the proof is done in a similar way as in the proofs of Theorems 15 and 18.

With the above techniques, we may at least verify whether computed pseudo-gcrds or pseudo-lclms are correct. For a fully deterministic algorithm, we still need a way to find a point where

is non singular. This can be done by brute force. Let us state the result in the most general setting of pseudo-Euclidean matrices.

Proof. Let

, and assume first that we know

. Let

be pairwise distinct, randomly chosen points in

at which

and

are non singular. At each

, we compute canonical fundamental systems of solutions

and

for

and

at order

. We claim that this can be done in time

Indeed, it requires a time

to rewrite each operator with respect to

. We next perform a multipoint evaluation of the coefficients of these operators to obtain the shifted operators at

(this requires a time

). The truncations of these operators at order

are then converted back to the respresentation with respect to

. This can be done in time

. Using Theorem 7, we finally compute the required fundamental systems of solutions in time

From

, we get

. Since we assumed

to be sufficiently large, it follows that

is non singular at one of the points

. At such a point

, the canonical fundamental systems of solutions

and

generate a vector space

of maximal dimension

, and with a basis

such that

for all

. We finally apply Theorem 23 in order to obtain

. If

is unknown, then we use a sequence of guesses

, as in the previous proofs.

Remark 25. In the case of least common left multiples, we may directly compute

using Theorem 21 and certify the result using trial division by

and

. This allows us to use the weaker assumption

instead of

, whereas the complexity bound becomes

5.3.Summary of the complexity bounds for Euclidean operations

We have summarized our complexity bounds for Euclidean operations on two operators

in Table 1. We systematically write

for the degree in

of the result. We also write

for the degree of the Euclidean matrix in

The algorithms in the first line correspond to applying Theorems 18, 21 and 23 at a randomly chosen point, without checking the result. The second line corresponds to the Las Vegas randomized algorithm for which the answers are certified through trial division (the bound for gcrds might further drop to

in view of Remark 22; more generally, the bounds can be restated in terms of sizes of certificates). In the third line, we rather use Euclidean matrices for the certification. The fourth line shows complexity bounds for the deterministic versions of our algorithms.

In comparison, several randomized Las Vegas algorithms were given in [6] that achieve the complexity bound

for lclms. This is in particular the case for Heffter's algorithm [19], when using Theorem 4. The non determinism is due to the use of a fast Las Vegas randomized algorithm for the computation of kernels of matrices with polynomial entries [6, Theorem 2]. Grigoriev established complexity bounds for gcrds which rely on a similar reduction to polynomial linear algebra. Along the same lines as in [6], this should lead to a Las Vegas randomized algorithm of complexity

, although we did not check this in detail.

In summary, the new algorithms do not achieve any improvements in the worst case. Nevertheless, the uncertified versions of our algorithms admit optimal running times up to logarithmic factors in terms of the combined input and output size. The certified randomized versions satisfy similar complexity bounds in terms of the size of a suitable certificate; such bounds can sometimes be better than the previously known worst case bounds. When performing our expansions at a randomly chosen point in

, we also recall that the probability of failure is exponentially small as a function of the bitsize of this point.

5.4.Generalizations

The algorithms from Section 5.1 extend in a straightforward way to the computation of greatest common right divisors and least common left multiples of more than two operators. For instance, using obvious notations, we obtain the following generalizations of Theorems 21 and 18.

Remark 27. When taking

and using [2], the complexity bound simplifies to

. By [6, Theorem 6], we may always take

, after which the bound further reduces to

. In our randomized setting, this improves upon the bounds from [6, Figure 1].

Remark 28. If we also require a certification of the result, then we may use the trial division technique. This amounts to

exact divisions of operators in

. Using the division algorithm from Section 4, and taking

and

as above, this can be done in time

Proof. The proof is similar to the one of Theorem 26, except for the way how we compute a basis for

. Indeed, we first compute a basis

for

. This requires a time

for the computation of

modulo

and a time

for the remaining linear algebra. We next compute the unique constant matrix

such that

modulo

. Since

is non singular, we have

at any order, so it suffices to compute

up to order

in order to obtain

up to order

Remark 30. An interesting question is whether there exists a faster algorithm to compute the orders

and

, without computing

and

themselves. For this, it suffices to compute the dimensions of

and

. Assuming that we are at a “non singular point”, the answer is therefore yes: using the techniques from the proofs of Theorems 29 and 26, we may compute

in time

and

in time

Bibliography

[1]: A. V. Aho, K. Steiglitz, and J. D. Ullman. Evaluating polynomials on a fixed set of points. SIAM Journ. of Comp., 4:533–539, 1975.
[2]: A. Benoit, A. Bostan, and J. van der Hoeven. Quasi-optimal multiplication of linear differential operators. In Proc. FOCS '12, pages 524–530. New Brunswick, October 2012. IEEE.
[3]: A. Borodin and R.T. Moenck. Fast modular transforms. Journal of Computer and System Sciences, 8:366–386, 1974.
[4]: A. Bostan. Algorithmique efficace pour des opérations de base en calcul formel. PhD thesis, École polytechnique, 2003.
[5]: A. Bostan, F. Chyzak, and N. Le Roux. Products of ordinary differential operators by evaluation and interpolation. In J. Rafael Sendra and Laureano González-Vega, editors, ISSAC, pages 23–30. Linz/Hagenberg, Austria, July 2008. ACM Press.
[6]: A. Bostan, F. Chyzak, B. Salvy, and Z. Li. Fast computation of common left multiples of linear ordinary differential operators. In J. van der Hoeven and M. van Hoeij, editors, Proc. ISSAC '12, pages 99–106. Grenoble, France, July 2012.
[7]: A. Bostan and É. Schost. Polynomial evaluation and interpolation on special sets of points. Journal of Complexity, 21(4):420–446, August 2005. Festschrift for the 70th Birthday of Arnold Schönhage.
[8]: E. Brassine. Analogie des équations différentielles linéaires à coefficients variables, avec les équations algébriques, pages 331–347. Note III du Tome 2 du Cours d'analyse de Ch. Sturm. École polytechnique, 1864.
[9]: D.G. Cantor and E. Kaltofen. On fast multiplication of polynomials over arbitrary algebras. Acta Informatica, 28:693–701, 1991.
[10]: J.W. Cooley and J.W. Tukey. An algorithm for the machine calculation of complex Fourier series. Math. Computat., 19:297–301, 1965.
[11]: D. Coppersmith and S. Winograd. Matrix multiplication via arithmetic progressions. In Proc. of the Annual Symposium on Theory of Computing, pages 1–6. New York City, may 25–27 1987.
[12]: S.S. Demidov. On the history of the theory of linear differential equations. Arch. Hist. Exact. Sci., 28(4):369–387, 1983.
[13]: F. Le Gall. Powers of tensors and fast matrix multiplication. In Proc. ISSAC 2014, pages 296–303. Kobe, Japan, July 23–25 2014.
[14]: M. Giesbrecht. Factoring in skew polynomial rings over finite fields. In Proc. LATIN '92, volume 583 of LNCS, pages 191–203. 1992.
[15]: M. Giesbrecht. Factoring in skew polynomial rings over finite finite fields. JSC, 26:463–486, 1998.
[16]: M. Giesbrecht and Y. Zhang. Factoring and decomposing ore polynomials over . In Manuel Bronstein, editor, Proc. ISSAC '03, pages 127–134. Philadelphia, USA, August 2003.
[17]: D. Yu. Grigoriev. Complexity of factoring and calculating the gcd of linear ordinary differential operators. J. Symb. Comput., 10(1):7–37, 1990.
[18]: J. von zur Gathen and J. Gerhard. Modern Computer Algebra. Cambridge University Press, 2-nd edition, 2002.
[19]: L. Heffter. Über gemeinsame Vielfache linearer Differentialausdrücke und lineare Differentialgleichungen derselben Klasse. J. Reine Angew. Math., 116:157–166, 1896.
[20]: J. van der Hoeven. FFT-like multiplication of linear differential operators. JSC, 33(1):123–127, 2002.
[21]: J. van der Hoeven. Relax, but don't be too lazy. JSC, 34:479–542, 2002.
[22]: J. van der Hoeven. Relaxed multiplication using the middle product. In Manuel Bronstein, editor, Proc. ISSAC '03, pages 143–147. Philadelphia, USA, August 2003.
[23]: J. van der Hoeven. On the complexity of skew arithmetic. Technical Report, HAL, 2011. http://hal.archives-ouvertes.fr/hal-00557750.
[24]: J. van der Hoeven, G. Lecerf, B. Mourrain et al. Mathemagix. 2002. http://www.mathemagix.org.
[25]: Z. Li. A subresultant theory for ore polynomials with applications. In O. Gloor, editor, Proc. ISSAC '98, pages 132–139. Rostock, Germany, August 1998.
[26]: G. Libri. Mémoire sur la résolution des équations algébriques dont les racines ont entre elles un rapport donné, et sur l'intégration des équations différentielles linéaires dont les intégrales particulières peuvent s'exprimer les unes par les autres. J. Reine und Angew. Math., 10:167–194, 1833.
[27]: R.T. Moenck and A. Borodin. Fast modular transforms via division. In Thirteenth annual IEEE symposium on switching and automata theory, pages 90–96. Univ. Maryland, College Park, Md., 1972.
[28]: O. Ore. Theorie der linearen Differentialgleichungen. J. Reine und Angew. Math., 167:221–234, 1932.
[29]: O. Ore. Theory of non-commutative polynomials. Ann. of Math., 34(3):480–508, 1933.
[30]: V. Pan. How to multiply matrices faster, volume 179 of Lect. Notes in Math. Springer, 1984.
[31]: A. Schönhage. Schnelle Multiplikation von Polynomen über Körpern der Charakteristik 2. Acta Informatica, 7:395–398, 1977.
[32]: A. Schönhage and V. Strassen. Schnelle Multiplikation großer Zahlen. Computing, 7:281–292, 1971.
[33]: R.P. Stanley. Differentially finite power series. European J. Combin., 1:175–188, 1980. MR #81m:05012.
[34]: V. Strassen. Gaussian elimination is not optimal. Numer. Math., 13:352–356, 1969.
[35]: V. Strassen. Die Berechnungskomplexität von elementarsymmetrischen Funktionen und von Interpolationskoeffizienten. Numer. Math., 20:238–251, 1973.

1.Introduction

2.Evaluation and interpolation

3.Local solutions

4.Right division

5.Euclidean operations

5.1.Randomized algorithms

5.2.Certifying correctness

5.3.Summary of the complexity bounds for Euclidean operations

5.4.Generalizations

Bibliography