Amortized bivariate multi-point evaluation

 Joris van der Hoevenab, Grégoire Lecerfac

 a. CNRS, École polytechnique, Institut Polytechnique de Paris Laboratoire d'informatique de l'École polytechnique (LIX, UMR 7161) 1, rue Honoré d'Estienne d'Orves Bâtiment Alan Turing, CS35003 91120 Palaiseau, France

 b. Email: vdhoeven@lix.polytechnique.fr

 c. Email: lecerf@lix.polytechnique.fr

 Preliminary version of June 27, 2022

. This paper is part of a project that has received funding from the French “Agence de l'innovation de défense”.

 The evaluation of a polynomial at several points is called the problem of multi-point evaluation. Sometimes, the set of evaluation points is fixed and several polynomials need to be evaluated at this set of points. Efficient algorithms for this kind of “amortized” multi-point evaluation were recently developed for the special case when the set of evaluation points is sufficiently generic. In this paper, we design a new algorithm for arbitrary sets of points, while restricting ourselves to bivariate polynomials.

## 1.Introduction

Let be an effective field, so that we have algorithms for the field operations. Given a polynomial and a tuple of points, the computation of is called the problem of multi-point evaluation. The converse problem is called interpolation and takes a candidate support of as input.

These problems naturally occur in several areas of applied algebra. When solving a polynomial system, multi-point evaluation can for instance be used to check whether all points in a given set are indeed solutions of the system. In [14], we have shown that fast algorithms for multi-point evaluation actually lead to efficient algorithms for polynomial system solving. The more specific problem of bivariate multi-point evaluation appears for example in the computation of generator matrices of algebraic geometry error correcting codes [18].

The general problem of multivariate multi-point evaluation is notoriously hard. If or is a field of finite characteristic, then theoretical algorithms of Kedlaya and Umans [17] achieve a complexity exponent , where represents a constant that can be taken arbitrarily close to zero. Unfortunately, to our best knowledge, these algorithms do not seem suitable for practical purposes [13, Conclusion].

The best known bound for over general fields is due to Nüsken and Ziegler [22]: the evaluation of at points can be done with operations in . This bound is based on the Paterson–Stockmeyer technique for modular composition [23]. Here, the constant is a real value such that the product of a matrix by a matrix takes operations; one may take ; see [16, Theorem 10.1]. We further cite [15] for an efficient algorithm in the case of special sets of points .

Last year, new softly linear algorithms have been proposed for multi-point evaluation and interpolation in the case when is a fixed generic tuple of points [12, 20]. These algorithms are amortized in the sense that potentially expensive precomputations as a function of are allowed. When the dimension is arbitrary but fixed, the algorithms from [12] take softly linear time: they generalize the classical univariate “divide and conquer” approach, as presented for instance in [6, chapter 10]. The results in [20] are restricted to the case . They take into account the partial degrees of and are based on changes of polynomial bases that are similar to the ones of [11, section 6.2].

In the present paper, we turn our attention to arbitrary (i.e. possibly non-generic) tuples of evaluation points , while restricting ourselves to the bivariate case and . Combining ideas from [12] and [20], we present a new softly linear algorithm for amortized multi-point evaluation. For the sake of simplicity, we have not optimized all constant factors involved in the cost analysis of our new algorithm, so our complexity bound is mostly of theoretical interest for the moment. The opposite task of interpolation is more subtle: since interpolants of total degree do not necessarily exist, the very problem needs to be stated with additional care. For this reason, we do not investigate interpolation in this paper.

## 2.Weighted bivariate polynomials

Our bivariate multi-point evaluation makes use of polynomial arithmetic with respect to weighted graded monomial orderings. This section is devoted to the costs of products and divisions in this context.

### 2.1.Complexity model

For complexity analyses, we will only consider algebraic complexity models like computation trees [3], for which elements in are freely at our disposal. The time complexity simply measures the number of arithmetic operations and zero-tests in .

We denote by the time that is needed to compute a product of two polynomials of degree . We make the usual assumptions that is non-decreasing as a function of . Using a variant of the Schönhage–Strassen algorithm [4], it is well known that we may take . If we restrict our attention to fields of positive characteristic, then we may even take [7].

### 2.2.Monomial orderings

General monomial orderings, that are suitable for Gröbner basis computations, have been classified in [24]. For the purpose of this paper, we focus on the following specific family of bivariate monomial orderings.

Definition 1. Let . We define the -degree of a monomial with by

We define the -ordering to be the monomial ordering such that

Let us mention that the idea of using such kinds of families of monomial orderings in the design of fast algorithms for bivariate polynomials previously appeared in [11].

### 2.3.Multiplication

Consider the product of two non-zero bivariate polynomials and the obvious bound

for the number of terms of . Then it is well known that Kronecker substitution allows for the computation of the product using operations in ; see [6, Corollary 8.27], for instance.

Writing for the valuation of in , the number of non-zero terms of is more accurately bounded by

Via the appropriate multiplications and divisions by powers of , we observe that can be computed using operations in .

Let us next show that a similar bound holds for slices of polynomials that are dense with respect to the -ordering. More precisely, let denote the minimum of over the monomials occurring in . By convention we set .

Lemma 2. Let . The product can be computed using operations in , where

Proof. From the monomial identity

we observe that the monomials of -degree are in one-to-one correspondence with the monomials of degree in and degree in in . It also follows that the number of terms of a non-zero polynomial is bounded by

and that

In addition, the number of non-zero terms in the product is bounded by . So it suffices to compute via the formula

in order to obtain the claimed complexity bound.

### 2.4.Division

Let be a polynomial in of -degree and of leading monomial written . Without loss of generality we may assume that the coefficient of this monomial is . We examine the cost of the division of of -degree by with respect to :

where and are in , and such that no monomial in is divisible by . Such a division does exist: this is a usual fact from the theory of Gröbner bases. In this context, a polynomial is said to be reduced with respect to when none of its terms is divisible by .

If for polynomials and such that is reduced with respect to , then , so and . In other words, and are unique, so we may write for the quotient of by with respect to .

In the remainder of this section, we assume that has been fixed once and for all. Given , we define

The naive division algorithm proceeds as follows: if has a term that is divisible by , then we pick a maximal such term for and compute

Then and the largest term of that is divisible by is strictly less than for . This division step is repeated for and for its successive reductions, until and are found.

During this naive division process, we note that only depends on and , for . When nothing needs to be computed. Let us now describe a more efficient way to handle the case , when we need to compute the quasi-homogeneous component of of maximal -degree :

Lemma 3. We may compute and using operations in .

Proof. We first decompose

and note that is reduced with respect to . In particular, the division of by yields the same quotient as the division of by , so

 (1)

for some quasi-homogeneous polynomial of -degree with . Dehomogenization of the relation (1) yields

with . Consequently, the computation of and takes operations in , using a fast algorithm for Euclidean division in ; see [6, chapter 9] or [8], for instance.

For higher values of , the following “divide and conquer” division algorithm is more efficient than the naive algorithm:

 Algorithm 1 Input and an integer , where and . Output . If or then return . If then compute and return using the method from Lemma 3. Let . Recursively compute . Let . Recursively compute . Return .

Proposition 4. Algorithm 1 is correct and takes operations in .

Proof. Let us prove the correctness by induction on . If , then and the result of the algorithm is correct. If , then and the result is also correct. The case has been treated in Lemma 3.

Now assume that and , so . The induction hypothesis implies that is reduced with respect to and that is reduced with respect to . After noting that

we verify that

 (since degk(Q0*B)⩽d-h)

Consequently, is reduced with respect to , whence

This completes the induction and our correctness proof.

Concerning the complexity, step 2 takes operations in , by Lemma 3. In step 5, the computation of takes operations in , by Lemma 2.

Let stand for the maximum of the costs of Algorithm 1 for and . We have shown that and that

Unrolling this inequality, we obtain the claimed complexity bound.

Corollary 5. Let be of -degree . The remainder in the division of by with respect to can be computed using operations in .

Proof. By Proposition 4, the computation of takes operations in . The computation of the corresponding remainder takes further operations in by Lemma 2.

Remark 6. The complexity bound from Proposition 4 is also a consequence of [9, Theorem 4] by taking for the cost of sparse polynomial products of size . This cost is warranted mutatis mutandis by the observation that all sparse bivariate polynomial products occurring within the algorithm underlying [9, Theorem 4] are either univariate products or products of slices of polynomials that are dense with respect to the -ordering. We have seen in section 2.3 how to compute such products efficiently.

## 3.General position

Let be a tuple of pairwise distinct points. We define the vanishing ideal for by

where we recall that .

A monic polynomial is said to be axial if its leading monomial with respect to is of the form . The goal of this section is to prove the existence of such a polynomial modulo a sufficiently generic change of variables of the form

 (2)

This change of variables transforms into a new tuple with

 (3)

and the ideal into

For any degree , we define

Given a polynomial such that we may decompose

where is homogeneous of degree and . The change of variables (2) transforms into

where

The coefficient of the monomial in is .

The -vector space is the solution set of a linear system consisting of equations and unknowns, that are the unknown coefficients of a polynomial in . Such a system admits a non-zero solution whenever . Now assume that is a non-zero element of of minimal total degree and let denote the set of roots of in . Since is minimal we have

that implies

 (4)

On the other hand, we have . And if , then is the leading monomial of for . Assuming that , this proves the existence of an axial polynomial of degree after a suitable change of variables of the form (2).

We say that is in general position if there exists a polynomial of minimal degree in that is axial.

## 4.Polynomial reduction

Let be a tuple of points in general position, as defined in the previous section. In this section, we describe a process for reducing a polynomial modulo . The reduction of is a polynomial whose support is controlled in a way that will be made precise.

### 4.1.Heterogeneous bases

Since is assumed to be in general position, thanks to (4), we first precompute an axial polynomial for of degree

For , let denote the number of monomials of -degree . We have

The -vector space of the polynomials in with -degree is the solution set of a linear system consisting of equations and unknowns. Consequently, if , then there exists a non-zero polynomial in of -degree . Let be a monic polynomial whose leading monomial is minimal for and set

We may precompute , e.g. by extracting it from a Gröbner basis for with respect to . By the minimality of the -degree of , we have

Now write with and . Then

Consequently, and

 (5)

We let be the smallest integer such that , hence

 (6)

and . There exists a monic non-zero polynomial in of minimal degree . Since , we may take for . We call a heterogeneous basis for . We further define

 (7)

Note that is the first integer such that the upper bound (7) is zero, although holds even for .

Remark. If the are pairwise distinct and if the cardinality of is sufficiently large, then, after a sufficiently generic linear change of coordinates, a Gröbner basis for with respect to the lexicographic ordering induced by can be computed in softly linear time: see [22, section 6], for instance. Then, a Gröbner basis for with respect to can be deduced with operations in thanks to the FGLM algorithm [5]. This bound has recently been lowered to in [21, Theorem 1.7, Example 1.3], where is a feasible exponent for matrix multiplication.

### 4.2.Elementary reductions

Given and , we may use the division procedure from section 2.4 to reduce with respect to . This yields a relation

 (8)

where and such that none of the monomials in the support of is divisible by the leading monomial of . We write and recall that is a -linear map.

We also define the projections and by

### 4.3.Compound reductions

For , we let denote the set of tuples of polynomials such that

• is the first integer such that ,

• , for .

Intuitively speaking, such a tuple will represent a sum modulo . Note that , so .

Given , we define three new sequences of polynomials by

Lemma 7. With the above notations. Let and let . If , then

for .

Proof. We first note that and

 (9)

Let us prove the degree bounds by induction on . For , we have

by using (7) and (9). Assume now that and that the bounds of the lemma hold for all smaller . Since , the induction hypothesis yields

Using (7) and (9) again, we deduce

Lemma 8. Under the assumptions of Lemma 7, we further have

 (10)

Assume that , let , and let be the first integer such that . If , then

otherwise, and

Proof. By construction, we have

whence

Since is axial, we have , which entails (10). From Lemma 7 we deduce

 (11)

Now contains no monomial that is divisible by the leading monomial of for and . Using (5), it follows that

 (12)

Consequently, for , inequalities (11) and (12), combined with , lead to

From and (6), we deduce that , whence . If follows that

From , we thus obtain . Since , the polynomial belongs to and has degree . It follows that .

Since and , we further deduce

Finally and imply , whence .

We call the tuple the reduction of with respect to .

Lemma 9. With the notation and assumptions of Lemma 7, the reduction of with respect to takes operations in .

Proof. First note that and . By Corollary 5, the computation of therefore takes operations in . For , Lemma 7 and (5) imply

By Corollary 5, we deduce that the computation of takes

operations in . Summing the costs to compute for , we obtain the claimed bound.

Proposition 10. Let . Then a sequence with can be computed using operations in .

Proof. We set , so that . We set and define recursively as the reduction of with respect to , for . The integer is the first integer such that

We finally take . Lemma 8 implies that belongs to . The complexity bound follows from Lemma 9.

## 5.Multi-point evaluation

Let be a tuple of pairwise distinct points and consider the problem of fast multi-point evaluation of a polynomial of total degree at . For simplicity of the exposition, it is convenient to first consider the case when is a power of two and is in general position. The core of our method is based on the usual “divide and conquer” paradigm.

We say that is in recursive general position if is in general position and if and are in recursive general position. An empty sequence is in recursive general position. With the notation of section 3 a recursive general position is ensured if the cardinality of is strictly larger that that is recursively defined by

for and with . Consequently . Whenever , we know from section 3 that we can reduce to the case where is in recursive general position. In this case, we can compute a recursive heterogeneous basis, that is made of a heterogeneous basis for and recursive heterogeneous bases for and .

 Algorithm 2 Input with and . Output . Precomputed A recursive heterogeneous basis for . Assumption is in recursive general position. If , then return . Compute the reduction of with respect to the heterogeneous basis for , via Proposition 10. Recursively apply the algorithm to and . Recursively apply the algorithm to and . Return the concatenations of the results of the recursive evaluations.

Theorem 11. Algorithm 2 is correct and runs in time .

Proof. The algorithm is clearly correct if . If , then we first observe that both and are in recursive general position by definition. Furthermore, Proposition 10 ensures that

The concatenation of the results of the recursive applications of the algorithm therefore yields the correct result.

As to the complexity bound, the cost of step 2 is bounded by according to Proposition 10. Hence, the total execution time satisfies

The desired complexity bound follows by unrolling this recurrence inequality.

Corollary 12. Consider an arbitrary effective field and , where is not necessarily a power of two. Modulo precomputations that only depend on and , we can evaluate any polynomial in at in time .

Proof. Modulo the repetition of at most more points, we may assume without loss of generality that is a power of two .

If is finite then we build an algebraic extension of of degree , so that . Multiplying two polynomials in takes

operations in . Consequently, up to introducing an extra factor in our complexity bounds, we may assume that .

Modulo a change of variables (2) from section 3, we may then assume that , defined in (3), is in recursive general position, and compute a recursive heterogeneous basis.

Given , we claim that we may compute using operations in . Indeed, we first decompose

where and each is zero or homogenous of degree . Computing then reduces to computing . This corresponds in turn to a univariate Taylor shift, which takes operations in ; see [1, Lemma 7], for instance. Finally, we apply Theorem 11 to and .

## 6.Conclusion

We have designed a softly linear time algorithm for bivariate multi-point evaluation, modulo precomputations that only depend on the evaluation points. This result raises several new questions. First, it would be useful to optimize the constant factors involved in the cost analysis, and study the efficiency of practical implementations. Second, one may investigate extensions of Corollary 12 that take into account the partial degrees of the input polynomial, e.g. by applying Theorem 11 to inputs of the form . A final challenge concerns the possibility to perform all precomputations in sub-quadratic time. For this, one might use techniques from [2, 19], as in [12]. The problem of achieving an almost linear cost appears to be even harder.

##### Acknowledgments.
We thank the anonymous referees for their useful comments.

## Bibliography

[1]

J. Berthomieu, G. Lecerf, and G. Quintin. Polynomial root finding over local rings and application to error correcting codes. Appl. Alg. Eng. Comm. Comp., 24(6):413–443, 2013.

[2]

A. Bostan, C.-P. Jeannerod, and É. Schost. Solving structured linear systems with large displacement rank. Theor. Comput. Sci., 407(1):155–181, 2008.

[3]

P. Bürgisser, M. Clausen, and M. A. Shokrollahi. Algebraic Complexity Theory, volume 315 of Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, 1997.

[4]

D. G. Cantor and E. Kaltofen. On fast multiplication of polynomials over arbitrary algebras. Acta Inform., 28:693–701, 1991.

[5]

J.-C. Faugère, P. Gianni, D. Lazard, and T. Mora. Efficient computation of zero-dimensional Gröbner bases by change of ordering. J. Symbolic Comput., 16(4):329–344, 1993.

[6]

J. von zur Gathen and J. Gerhard. Modern Computer Algebra. Cambridge University Press, 2nd edition, 2002.

[7]

D. Harvey and J. van der Hoeven. Faster polynomial multiplication over finite fields using cyclotomic coefficient rings. J. Complexity, 54:101404, 2019.

[8]

J. van der Hoeven. Newton's method and FFT trading. J. Symbolic Comput., 45(8):857–878, 2010.

[9]

J. van der Hoeven. On the complexity of multivariate polynomial division. In I. S. Kotsireas and E. Martínez-Moro, editors, Applications of Computer Algebra. Kalamata, Greece, July 20–23, 2015, volume 198 of Springer Proceedings in Mathematics & Statistics, pages 447–458. Cham, 2017. Springer International Publishing.

[10]

J. van der Hoeven. The Jolly Writer. Your Guide to GNU TeXmacs. Scypress, 2020.

[11]

J. van der Hoeven and R. Larrieu. Fast reduction of bivariate polynomials with respect to sufficiently regular Gröbner bases. In C. Arreche, editor, Proceedings of the 2018 ACM International Symposium on Symbolic and Algebraic Computation, ISSAC '18, pages 199–206. New York, NY, USA, 2018. ACM.

[12]

J. van der Hoeven and G. Lecerf. Fast amortized multi-point evaluation. Technical Report, HAL, 2020. https://hal.archives-ouvertes.fr/hal-02508529.

[13]

J. van der Hoeven and G. Lecerf. Fast multivariate multi-point evaluation revisited. J. Complexity, 56:101405, 2020.

[14]

J. van der Hoeven and G. Lecerf. On the complexity exponent of polynomial system solving. Found. Comput. Math., 21:1–57, 2021.

[15]

J. van der Hoeven and É. Schost. Multi-point evaluation in higher dimensions. Appl. Alg. Eng. Comm. Comp., 24(1):37–52, 2013.

[16]

X. Huang and V. Y. Pan. Fast rectangular matrix multiplication and applications. J. Complexity, 14(2):257–299, 1998.

[17]

K. S. Kedlaya and C. Umans. Fast polynomial factorization and modular composition. SIAM J. Comput., 40(6):1767–1802, 2011.

[18]

D. Le Brigand and J.-J. Risler. Algorithme de Brill–Noether et codes de Goppa. Bulletin de la société mathématique de France, 116(2):231–253, 1988.

[19]

V. Neiger. Fast computation of shifted Popov forms of polynomial matrices via systems of modular polynomial equations. In Proceedings of the ACM on International Symposium on Symbolic and Algebraic Computation, ISSAC '16, pages 365–372. New York, NY, USA, 2016. ACM.

[20]

V. Neiger, J. Rosenkilde, and G. Solomatov. Generic bivariate multi-point evaluation, interpolation and modular composition with precomputation. In A. Mantzaflaris, editor, Proceedings of the 45th International Symposium on Symbolic and Algebraic Computation, ISSAC '20, pages 388–395. New York, NY, USA, 2020. ACM.

[21]

V. Neiger and É. Schost. Computing syzygies in finite dimension using fast linear algebra. J. Complexity, 60:101502, 2020.

[22]

M. Nüsken and M. Ziegler. Fast multipoint evaluation of bivariate polynomials. In S. Albers and T. Radzik, editors, Algorithms – ESA 2004. 12th Annual European Symposium, Bergen, Norway, September 14-17, 2004, volume 3221 of Lect. Notes Comput. Sci., pages 544–555. Springer Berlin Heidelberg, 2004.

[23]

M. S. Paterson and L. J. Stockmeyer. On the number of nonscalar multiplications necessary to evaluate polynomials. SIAM J. Comput., 2(1):60–66, 1973.

[24]

L. Robbiano. Term orderings on the polynomial ring. In B. F. Caviness, editor, EUROCAL '85. European Conference on Computer Algebra. Linz, Austria, April 1-3, 1985. Proceedings. Volume 2: Research Contributions, volume 204 of Lect. Notes Comput. Sci., pages 513–517. Springer-Verlag Berlin Heidelberg, 1985.