Fast multiplication for skew polynomials

TL;DR

This paper introduces a new, faster algorithm for multiplying skew polynomials that leverages evaluation and interpolation techniques, achieving optimal asymptotic complexity for large degrees and improving existing methods.

Contribution

The authors develop a novel algorithm for skew polynomial multiplication based on evaluation and interpolation, improving complexity bounds and extending efficiency to small degrees.

Findings

Achieves optimal asymptotic complexity for large degree skew polynomial multiplication

Provides an efficient algorithm for small degree skew polynomial multiplication

Improves complexity bounds for various arithmetic problems involving skew polynomials

Abstract

We describe an algorithm for fast multiplication of skew polynomials. It is based on fast modular multiplication of such skew polynomials, for which we give an algorithm relying on evaluation and interpolation on normal bases. Our algorithms improve the best known complexity for these problems, and reach the optimal asymptotic complexity bound for large degree. We also give an adaptation of our algorithm for polynomials of small degree. Finally, we use our methods to improve on the best known complexities for various arithmetics problems.

Equations

$$\forall \alpha \in L,\text{leavevmode}X\alpha =\sigma (\alpha )X.$$

$$b_0=1+\sqrt[r]{a}+\sqrt[r]{a^2}+\cdots +\sqrt[r]{a^{r-1}}=\frac{a-1}{\sqrt[r]{a}-1}.$$

$$\begin{array}[]{rcl}L[X,\sigma]&\rightarrow&\text{\rm End}_{K}(L)\\ A=\sum_{i\geq 0}a_{i}X^{i}&\mapsto&A(\sigma)=\sum_{i\geq 0}a_{i}\sigma^{i}\end{array}$$

$$\epsilon \text{leavevmode}:\text{leavevmode}L[X,\sigma ]/(X^r-1)\simeq {\text{End}}_K(L).$$

$$C(T)=\tilde{A}(T)B(T)(\mathrm{mod}{T}^r-1).$$

$$C(T)=\left(\begin{array}{cccc}{b}_0& b_1& \cdots & b_{r-1}\end{array}\right)M\left(\begin{array}{c}1\\ T\\ ⋮\\ T^{r-1}\end{array}\right),$$

$$\tilde{A}(T)B(T)\equiv C(T)(\mathrm{mod}{T}^r-1),$$

$$c_i={\gamma }_i+{\tilde{\gamma }}_i,$$

$$c_i=\sum _{j^{\prime }=0}^i{a}_{i-j^{\prime }}{b}_{j^{\prime }}+\sum _{j^{\prime }=r-i}^{r-1}{a}_{i-j^{\prime }+r}{b}_{j^{\prime }},$$

$$\tilde{A}(T)B_n(T)\equiv Q(T)(\mathrm{mod}{T}^r-1),$$

$$U_i(T)B_n(T)+V_i(T)(T^r-1)=Q_i(T),$$

$$\begin{array}[]{rcl}\varphi_{i}:\quad L[T]_{<i}\times L[T]_{<i-1}&\longrightarrow&L[T]_{<r+i-1}/L[T]_{<r-i}\\ (U,V)&\mapsto&UR_{1}+VR_{0}\end{array}$$

$$U(T^r-1)+VB(T)=H(T)+T^{r-n+1}({\alpha }_0+\cdots +{\alpha }_{n-1}{T}^{n-1}).$$

$$\begin{array}[]{rcl}L[X,\sigma]&\longrightarrow&L[X,\sigma]/(X^{r}{-}1)\\ A(X)=\sum a_{i}X^{i}&\mapsto&A(\lambda X)=\sum_{i}\lambda\sigma(\lambda)\cdots\sigma^{i-1}(\lambda)a_{i}X^{i}\end{array}$$

$${\tilde{C}}_a(T)=\tilde{A}(T)\tilde{B}(T)(\mathrm{mod}{T}^r-a).$$

$$\sigma_{a}^{i}(b_{j})=\left\{\begin{array}[]{lc}b_{j-i}&\text{if }j\geq i\\ ab_{r+j-i}&\text{if }i>j\end{array}\right..$$

$$T^{i}B(T)=\sum _{j=0}^{r-1}{b}_j{T}^{i+j}=\sum _{j=i}^{r-1}{b}_{j-i}{T}^i+\sum _{j=0}^{i-1}a{b}_{r+j-i}{T}^j.$$

$$L[X,\sigma ]/Z(X^r)\simeq L^{\prime }[X,{\sigma }^{\prime }]/(X^r-a).$$

$$L[X,\sigma ]/Z(X^r)\to L[X,\sigma ]/Z_1(X^r)\times \cdots \times L[X,\sigma ]/Z_m(X^r)$$

$$A_i=A_i^{(0)}(X^r)+A_i^{(1)}(X^r)X+\cdots +A_i^{(r-1)}(X^r)X^{r-1}$$

$$A^{(j)}(X^r)=A_i^{(j)}(X^r)+Z_i(X^r)\cdot Q_i^{(j)}(X^r)$$

$$q^n\ge \max (64n,8r).$$

$$N_{L^{\prime }/K^{\prime }}(x)=N_{M^{\prime }/K^{\prime }}(x_1)\cdots N_{M^{\prime }/K^{\prime }}(x_g).$$

$$\text{Prob}\big{[}N_{L^{\prime}/K^{\prime}}=a\big{]}=\left(1-\frac{1}{q^{nf}}\right)^{g}\cdot\frac{1}{q^{n}-1}$$

$$\left(1-\frac{1}{n{c}_n}\right)\cdot \left(1-\frac{2}{n{c}_n}\right)\cdots \left(1-\frac{t-1}{n{c}_n}\right).$$

$${\left(1-\frac{1}{q^{nf}}\right)}^g\cdot \frac{c_n}{q^n-1}\cdot \left(1-\frac{1}{c_n}\right)\cdots \left(1-\frac{t-1}{n{c}_n}\right)$$

$$\frac{c_n}{q^n-1}\left(1-\frac{g}{q^{nf}}-\frac{t(t-1)}{2c_n}\right)\ge \frac{c_n}{q^n}-\frac{r}{q^n}-\frac{t(t-1)}{2n{q}^n}.$$

$$q^n-c_n\le \sum _{m|n,m<n}{q}^m\le 2\sqrt{n}\cdot q^{n/2}$$

$$\begin{array}[]{r@{\hspace{0.5ex}}ll}\text{\rm SM}(d,r)\hskip 2.15277pt&=d^{(\omega+1)/2}r&\text{for }d\leq r^{(5-\omega)/2}\vskip 3.0pt plus 1.0pt minus 1.0pt\\ \hskip 2.15277pt&=d^{\omega-2}r^{2}&\text{for }r^{(5-\omega)/2}\leq d\leq r\vskip 3.0pt plus 1.0pt minus 1.0pt\\ \hskip 2.15277pt&=dr^{\omega-1}&\text{for }d\geq r.\end{array}$$

$${\text{SM}}^{\ge 1}(d,r)=\underset{d^{\prime }\le d}{\sup }\left(\text{SM}(d^{\prime },r)\cdot \frac{d}{d^{\prime }}\right).$$

Full text

Introduction

The present paper is dedicated to the description of algorithms for fast arithmetics in skew polynomial rings. Since they were first introduced by Ore, skew polynomials and their variants have been widely studied in several areas of mathematics. In particular, skew polynomials over finite fields have various applications in coding theory [15], cryptography see [3], for $p$-adic Galois representations [11]. Fast arithmetics for manipulating these objects is useful for such applications, and has been improved over time since the first breakthrough paper on computational skew polynomials over finite fields, due to Giesbrecht [9].

Let $K$ be a field and let $L$ be a finite extension of $K$, endowed with the endomorphism $\sigma$. We assume that $\sigma$ has order $r\geq 1$ and that $K=L^{\sigma}$. We consider the ring $L[X,\sigma]$ of skew polynomials with coefficients in $L$. This is a non commutative ring where the relation $Xa=\sigma(a)X$ holds for all $a\in L$ (for more detail about the definitions, see section 1.1). The main problem addressed in this paper is the fast multiplication of elements of $L[X,\sigma]$. The complexity of algorithms is described in terms of the number of elementary operations in $K$ with respect to the degree $d$ of the skew polynomials to be multiplied, and the degree $r$ of $L$ over $K$.

State of the art. The naïve method for multiplication of skew polynomials of degree $\leq d$ yields an algorithm that has complexity $O(d^{2}r^{2})$ operations in $K$. In [9], this complexity is improved to $O(dr^{2}+d^{2}r)$. Let $\omega$ denote the exponent of matrix multiplication. The authors of the present paper gave several algorithms for multiplication in [4], with best complexity $\tilde{O}(dr^{\omega-1})$ achieved for $d>r^{2}$. The most recent results by Puchinger and Wachter-Zeh [13] give a bound of $\tilde{O}(d^{\frac{\omega+1}{2}}r)$ operations in $K$ for multiplication in $L[X,\sigma]$, which improves on the previous results [4] when $d\in\Theta(r)$, which is the most relevant case for applications in coding theory (see [13], §4.2). In the context of differential operators (which share many similarities with skew polynomials), Benoit, Bostan and Van der Hoeven have obtained a complexity of $\tilde{O}(\min\{d,r\}^{\omega-2}dr)$ (see [2], Theorem 1) for multiplication in $L[x]\langle\partial\rangle$. We expect that this complexity should be doable in $L[X,\sigma]$ as well, but we have only achieved it for $d\geq r$.

Contributions of the paper. This paper’s main algorithm improves the complexity of the best known algorithms for multiplication in $L[X,\sigma]$ to $\tilde{O}(dr^{\omega-1})$ when $d\geq r$. For $d\in\Theta(r)$, this gives a complexity of $\tilde{O}(r^{\omega})$ operations in $K$. This is quasi-optimal in the sense that matrix multiplication can be reduced to skew polynomial multiplication (this is for example a consequence of Proposition 1.6 below), so that any improvement on the exponent of skew polynomial multiplication would lead to a similar improvement for matrix multiplication. We also design a new algorithm for multiplication of polynomials of small degree $d\ll r$ in $L[X,\sigma]$, whose complexity is $\tilde{O}(d^{\omega-2}r^{2})$.

We also show that our method can be used to improve the best known complexities of various related problems, such as multi-point evaluation, minimal subspace polynomial, and interpolation which are studied in [13]. We also improve the complexities for greatest common divisors and least common multiples.

Organization of the paper. The first section of the paper focuses on elementary operations for skew polynomials with normal bases: evaluation and interpolation. More precisely, if $P\in L[X,\sigma]$, then $P(\sigma)$ is an endomorphism of the $K$-algebra $L$, and the map $P\mapsto P(\sigma)$ is a morphism of $K$-algebras. In this section, we describe how we can compute efficiently $P(\sigma)$ using a normal basis and, conversely, how to recover $P$ (the reduction modulo $X^{r}{-}1$ of) $P$ from the datum of $P(\sigma)$ (see Proposition 1.6). We also look into more detail how the can solve the same evaluation/interpolation problems with $P$ of small degree $n$ at only the first $n$ elements of a normal basis.

In the second section, we present our algorithm for fast multiplication of skew polynomials. First, we study how the multiplication can be done efficiently modulo $X^{r}{-}1$ through evaluation/interpolation on a normal basis and matrix multiplication. We generalize this study to multiplication modulo $Z(X^{r})$ for any irreducible polynomial $Z\in K[T]$. This allows us to give an algorithm for multiplication of skew polynomials of degree $d$ that works in $O(dr^{\omega-1})$ operations in $K$ (where $r^{\omega}$ denotes the complexity of multiplication of square matrices of size $r$).

In the third section, we give several applications to fast arithmetics for skew polynomials. We show how we can perform general multi-point evaluation, minimal subspace polynomial, and interpolation, as well as usual operations on skew polynomials such as (extended) Euclidean division, greatest common divisor, least common multiple.

1 Fast evaluation

and interpolation

In this section, we present the notion of skew polynomials, and we study the problems of their evaluation and interpolation using normal bases.

1.1 Definitions and notations

Let $K$ be a field and let $L$ be an étale $K$-algebra (since $K$ is a field, this just means that $L$ is isomorphic to a product of field extensions of $K$). Let $\sigma$ be an automorphism of $L$. We assume that $\sigma$ has finite order $r$ and that $K=L^{\sigma}$. The ring $L[X,\sigma]$ of skew polynomials with coefficients in $K$ is the ring whose underlying group is $L[X]$ and whose multiplication is determined by the relation

[TABLE]

The ring $L[X,\sigma]$ is not commutative unless $r=1$.

Examples. The following situations are examples of the general setting that we are considering:

• •

$L=K^{r}$, and $\sigma$ is the shift operator $(x_{0},\dots,x_{r-1})\mapsto(x_{1},\dots,x_{r-1},x_{0})$,

• •

(Extensions of finite fields) $K=\mathbf{F}_{q}$, $L=\mathbf{F}_{q^{r}}$ and $\sigma:x\mapsto x^{q}$ is the Frobenius endomorphism of $L$,

• •

(Cyclotomic extensions) $K=\mathbf{Q}$ and $L=\mathbf{Q}(\zeta_{p^{n}})$ where $\zeta_{p^{n}}$ is a primitive $p^{n}$-th root of unity and $p$ is prime; $\sigma$ is a generator of the Galois group $\textrm{Gal}(L/K)$ (which is the cyclic group $(\mathbf{Z}/p^{n}\mathbf{Z})^{\times}$).

• •

(Kummer extensions) $K$ contains a primitive $r$-th root $\zeta_{r}$ of $1$, $L=K(\sqrt[r]{a})$ for some suitable $a\in K$ and $\sigma$ takes $\sqrt[r]{a}$ to $\zeta_{r}\sqrt[r]{a}$.

The two last examples are addressed in [14] and have applications to space-time codes.

Remark 1.1**.**

Usually, $L$ is assumed to be a field extension of $K$. We are considering the more general context of an étale $K$-algebra because it is stable under base change: if $L/K$ is étale and $K^{\prime}$ is an extension of $K$, then $L^{\prime}=L\otimes_{K}K^{\prime}$ is étale over $K^{\prime}$ (but it is not a field in general, even if $L$ is). This feature is used mostly in Section 2.1.2, and does not make the classical results any more difficult to prove.

Definition 1.2**.**

A normal basis of $L/K$ is a basis $(b_{0},\dots,b_{r-1})$ of $L$ over $K$ such that $\sigma(b_{i+1})=b_{i}$ (the indices being taken modulo $r$).

Proposition 1.3** ([6], Satz 1).**

Assuming $\sigma$ has order $r$ and $K=L^{\sigma}$, $L$ has a normal basis.

The problem of the construction of normal bases has been widely studied, see for example [8] for the case of finite fields, and [10] for the case of number fields. In both cases of cyclotomic extensions and Kummer extensions, it is easy to exhibit a normal basis: in the cyclotomic case, the basis starting with $b_{0}=\zeta_{p^{n}}$ does the job while in the Kummer case, one can take:

[TABLE]

From now on, we assume that we have fixed a normal basis $(b_{0},\dots,b_{r-1})$ of $L$ together with a working basis in which the elements of $L$ are represented. Let $\Omega$ be the matrix of change of basis from the working basis to the normal basis. We assume that the multiplication in $L$ and the application of $\sigma$ can be both performed in $\tilde{O}(r)$ operations in $K$ in the working basis.

1.2 Evaluation and interpolation

on a normal basis

We introduce a relation between polynomials that allows to evaluate the linear map associated to a skew polynomial at the elements of the normal basis $(b_{0},\dots,b_{r-1})$.

Lemma 1.4**.**

The map:

[TABLE]

is a homomorphism of $K$-algebras. It induces an isomorphism of $K$-algebras:

[TABLE]

Proof.

The first map is a homomorphism because for all $a\in L$, $Xa=\sigma(a)X$ in $L[X,\sigma]$. Since $\sigma$ has order $r$, $X^{r}{-}1$ lies in the kernel of this map, so $\varepsilon$ is well-defined. Both $L[X,\sigma]/(X^{r}{-}1)$ and $\text{\rm End}_{K}(L)$ are $K$-vector spaces of dimension $r^{2}$, hence it suffices to prove injectivity. By Artin’s Lemma on independence of characters, $\{id,\sigma,\ldots,\sigma^{r-1}\}$ is a linearly independent family over $L$, so that if $P(\sigma)=0$ for some $P\in L[X,\sigma]$ of degree $<r$, then $P=0$. ∎

Lemma 1.4 shows that multiplication of skew polynomials modulo $X^{r}{-}1$ is essentially the same as multiplication of $r\times r$ matrices over $K$, assuming that the isomorphism $\varepsilon$ can be computed efficiently (in both ways). We now address this question.

Notation 1.5**.**

Throughout this paper, we will denote $P(x)$ for $P(\sigma)(x)=\varepsilon(P)(x)$ if $P\in L[X,\sigma]$ and $x\in L$.

Let $T$ be a new (commutative) variable and consider the classical polynomial ring $L[T]$. Let $B=\sum_{i=0}^{r-1}b_{i}T^{i}\in L[T]$ be the polynomial whose coefficients are the elements of the normal basis.

Proposition 1.6**.**

Let $A=\sum_{i=0}^{r-1}a_{i}X^{i}\in L[X,\sigma]$ and let $\tilde{A}(T)=\sum a_{i}T^{i}\in L[T]$. Let $c_{j}=A(b_{j})$ and let $C(T)=\sum_{j=0}^{r-1}c_{j}T^{j}$. Then

[TABLE]

Proof.

By linearity, it is enough to check that the relation holds when $A=X^{i}$ for $0\leq i\leq r-1$. Let $0\leq i\leq r-1$. We have $X^{i}(b_{j})=\sigma^{i}(b_{j})=b_{j-i}$, where indices are taken modulo $r$.

On the other hand, doing the calculations modulo $T^{r}{-}1$, $T^{i}B(T)=\sum_{j=0}^{r-1}b_{j-i}T^{j}$. ∎

Proposition 1.6, although elementary, shows that the isomorphism $\varepsilon$ of Lemma 1.4 can be computed efficiently. Moreover, it also shows how the inverse isomorphism can be computed. More precisely:

Corollary 1.7**.**

Multiplication in $L[X,\sigma]/(X^{r}{-}1)$ can be performed in $O(r^{\omega})$ operations in $K$.

Proof.

Let $A_{1},A_{2}\in L[X,\sigma]/(X^{r}{-}1)$. Let $\tilde{A}_{1}(T),\tilde{A}_{2}(T)\in L[T]$ be the commutative polynomials with the same coefficients as $A_{1},A_{2}$ respectively. Let $C_{1}(T)=\tilde{A}_{1}(T)B(T)\in L[T]/(T^{r}{-}1)$ and $C_{2}(T)=\tilde{A}_{2}(T)B(T)\in L[T]/(T^{r}{-}1)$. Both $C_{1}$ and $C_{2}$ can be computed in $\tilde{O}(r^{2})$ operations in $K$. Now let $M_{1}$ (resp. $M_{2}$) be the matrix whose $j$-th column is the decomposition of the $j$-th coefficient of $C_{1}$ (resp. $C_{2}$) in the working basis. By Proposition 1.6, $M_{1}$ (resp $M_{2}$) is the matrix of $\varepsilon(A_{1})$ (resp. $\varepsilon(A_{2})$) where the codomain in endowed with the normal basis and the codomain is endowed with the working basis. Set $M=M_{1}\Omega M_{2}$; this product can be computed within $O(r^{\omega})$ operations in $K$. We know that $M$ is the matrix of $\varepsilon(A_{1}A_{2})$ where again the codomain in endowed with the normal basis and the codomain is endowed with the working basis. Let

[TABLE]

and compute $\tilde{A}(T)=C(T)B(T)^{-1}\pmod{T^{r}{-}1}=\sum_{i=0}^{r-1}a_{i}T^{i}$, which can also be computed in $\tilde{O}(r^{2})$ operations in $K$. Then, again by Proposition 1.6, $A_{1}A_{2}=\sum_{i=1}^{r-1}a_{i}X^{i}$. This shows that the global complexity of this computation is $O(r^{\omega})$. ∎

In Section 2, we will generalize this algorithm and show how it yields a fast multiplication algorithm for skew polynomials (not only in the modular case).

1.3 Evaluation and interpolation at

an incomplete normal basis

Evaluation. We shall see later how we can compute the product of two skew polynomials of small degree $d$ by determining how their product acts on $2d$ elements of a normal basis. With this motivation in mind, let us describe how we can compute efficiently the image of the first few elements of a normal basis under the action of the skew polynomial $A\in L[X,\sigma]$. Recall that, using Proposition 1.6 with $\lambda=1$, and writing $B(T)=\sum_{i=0}^{r-1}b_{i}T^{i}$, we know that

[TABLE]

where $C(T)=\sum_{i=0}^{r-1}A(b_{i})T^{i}$. Let $n<r$, and let $A\in L[X,\sigma]$ of degree $n$. We are interested in computing only $A(b_{i})$ for $0\leq i\leq n-1$.

Lemma 1.8**.**

Let $A\in L[X,\sigma]$ of degree $n$ and let $c_{i}=A(b_{i})$ for $0\leq i\leq n-1$. Let $U(T)=\sum_{i=0}^{n}b_{i}T^{i}$ and $V(T)=\sum_{i=0}^{n}b_{r-i}T^{r-i}$. Then, for $0\leq i\leq n$:

[TABLE]

where $\tilde{A}(T)U(T)=\sum_{i=0}^{r-1}\gamma_{i}T^{i}$ and $\tilde{A}(T)V(T)=\sum_{i=0}^{r-1}\tilde{\gamma}_{i}T^{i}$ (the products being taken modulo $T^{r}{-}1$).

Proof.

Since $c_{i}=\sum_{j+j^{\prime}=i\pmod{r}}a_{j}b_{j^{\prime}}$, and $a_{j}=0$ for $j>n$, we are left with the formula:

[TABLE]

and both sums correspond precisely to the coefficients of $\tilde{A}U$ and $\tilde{A}V$ respectively. ∎

Corollary 1.9**.**

Let $A\in L[X,\sigma]$ of degree $\leq n$, then the collection of $A(b_{0}),\ldots,A(b_{n-1})$ can be computed in $\tilde{O}(rn)$ operations in $K$.

Proof.

By Lemma 1.8, the evaluation of $A$ at $b_{0},\dots,b_{n-1}$ can be obtained by two multiplications of (classical) polynomials of degree $n$ with coefficients in $L$, hence with complexity $\tilde{O}(nr)$ operations in $K$. ∎

Interpolation. Still bearing in mind the aim of multiplying two skew polynomials by composing the corresponding linear maps, we are interested in the following question of interpolation: given $n$ values $\alpha_{0},\dots,\alpha_{n-1}\in L$, find $A\in L[X,\sigma]$ of degree $n$ such that $A(b_{i})=\alpha_{i}$ for all $0\leq i\leq n-1$.

Let us explain first how the solution to this problem can be computed when $\alpha_{0}=\cdots=\alpha_{n-1}=0$. In this case, the skew polynomial we are looking for is the so-called minimal subspace polynomial corresponding to the span $\left<b_{0},\dots,b_{n-1}\right>$. A generic fast algorithm for solving this problem has been proposed by Puchinger and Wachter-Zeh in [13], Theorem 26; it has complexity $\tilde{O}(n^{\max\{\log_{2}(3),\frac{\omega+1}{2}\}}r)$ operations in $K$. In the special case we are considering, we shall see that this can be improved to $\tilde{O}(nr)$.

Let $B_{n}(T)=\sum_{i=0}^{r-1}b_{i+n}T^{i}$, so that $B_{n}(T)\equiv T^{-n}B(T)\pmod{T^{r}{-}1}$. If $A$ is such that $A(b_{i})=0$ for $0\leq i\leq n-1$, then there exists $Q\in L[T]$ of degree $\leq r{-}n{-}1$ such that $\tilde{A}(T)B(T)\equiv T^{n}Q(T)\pmod{T^{r}{-}1}$. Of course, the converse is also true, and this equation is equivalent to:

[TABLE]

with $\deg\tilde{A}\leq n$ and $\deg Q\leq r-n-1$. The latter equation can be solved thanks to the extended Euclidean algorithm. Indeed, computing the gcd of $T^{r}{-}1$ and $B_{n}(T)$ and stopping after the first remainder of degree $<r-i$, we get a relation of the form:

[TABLE]

with $\deg U_{i}\leq i$ and $\deg Q_{i}\leq r-1-i$, which yields a solution to the problem when $i=n$. This computation can be done in $\tilde{O}(nr)$ operations in $K$ thanks to the half-gcd algorithm (see [7], Theorem 11.5).

In the general case, let $\alpha_{0},\ldots,\alpha_{n-1}\in L$, and let $C(T)=\sum_{i=0}^{n-1}\alpha_{i}T^{i}$. We are looking for $A\in L[X,\sigma]$ with degree $\leq n$ and $Q\in L[T]$ with degree $\leq r-n-1$ such that $A(T)B(T)\equiv C(T)+T^{n}Q(T)\pmod{T^{r}{-}1}$. This equation is equivalent to $A(T)B_{n}(T)\equiv\sum_{i=0}^{n-1}\alpha_{i}T^{r-n+i}\pmod{T^{r}{-}1}$.

Lemma 1.10**.**

Let $R_{0}(T)=T^{r}{-}1$, $R_{1}(T)=B(T)$ and for $i\geq 2$, let $R_{i}$ be the remainder of the Euclidean division of $R_{i-2}$ by $R_{i-1}$. Then for $0\leq i\leq r$, $\deg R_{i}=r-i$.

Proof.

Consider the map

[TABLE]

It is well-defined, linear, and both sides have the same dimension over $L$. Moreover, the determinant of this map is nonzero if and only if $\deg R_{i}=r-i$ (see [16], §4.1). Therefore, it is sufficient to prove that $\varphi_{i}$ is injective.

Let us consider $(U,V)$ in the kernel of $\varphi_{i}$. By definition, $\deg(UR_{1}+VR_{0})<r-i$, so that $U(T)B(T)\equiv W(T)\pmod{T^{r}{-}1}$, where $\deg W(T)<r-i$. By Proposition 1.6, the skew polynomial $\sum_{j=0}^{i-1}u_{j}X^{j}$ (whose coefficients are the coefficients of $U$) evaluates to [math] at $b_{r-i},\dots,b_{r-1}$. Hence, it is a left multiple of the minimal subspace polynomial $M$ of $\left<b_{r-i},\dots,b_{r-1}\right>$. Since $(b_{r-i},\dots,b_{r-1})$ is linearly independent over $K$, $M$ has degree $i$ (it is a generator of the kernel of the $K$-linear map $L[X,\sigma]\rightarrow L^{i}$ mapping $P$ to $(P(b_{r-i}),\dots,P(b_{r-1}))$). In particular, since $\deg U<i$, $U=0$, so $V=0$ and $\varphi_{i}$ is injective. Hence $\det\varphi_{i}\neq 0$ and $R_{i}$ has the required degree. ∎

Theorem 1.11**.**

Let $n\leq r$ and $\alpha_{0},\ldots,\alpha_{n-1}\in L$. Then there exists $U,V,H\in L[T]$, with $\deg U\leq n-1$, $\deg V\leq n$ and $\deg H\leq r-n$ such that

[TABLE]

Moreover, Algorithm SmallDegreeInterpolation outputs $U$ and $V$ for a cost of $\tilde{O}(rn)$ operations in $K$.

Sketch of the proof.

The result follows from the correctness of Algorithm 1, but is also a theoretical consequence of Lemma 1.10. Indeed, this lemma shows that there exists a linear combination of $R_{0}=T^{r}{-}1$, $R_{1}=B(T),\ldots,R_{n-1}$ whose higher degree terms have coefficients $c_{0},\ldots,c_{n-1}$, and the bounds on the degrees follow from the fact that for $i\leq n$, $R_{i}=U_{i}R_{0}+V_{i}R_{1}$ with $\deg U_{i}\leq i-1$, $\deg V_{i}\leq i$. Algorithm 1 is an adaptation of the half-gcd algorithm, which computes simultaneously the sequence of the remainders in the extended Euclidean division or $R_{0}$ and $R_{1}$, and the combination of $R_{1}$ and $R_{0}$ that has the given higher degree terms. ∎

Thanks to Corollary 1.9, Theorem 1.11 and Algorithm 1, we can solve the problem of evaluation and interpolation at the first $n$ elements of an incomplete normal basis in $\tilde{O}(nr)$ operations in $K$.

2 Fast multiplication

In this section, we study the problem of multiplying efficiently two elements $A_{1},A_{2}\in L[X,\sigma]$ both of degree $\leq d$. The complexity is the number of operations in $K$, given as a function of $d$ and $r=\dim_{K}L$.

2.1 Modular multiplication

2.1.1 Multiplication modulo $X^{r}{-}a$

We consider the ring $L[X,\sigma]$. Let $\lambda\in L^{\times}$, and let $a=N_{L/K}(\lambda)$. We are now going to describe an algorithm for multiplication in $L[X,\sigma]$ modulo $X^{r}{-}a$.

Proposition 2.1**.**

The map

[TABLE]

factors as an isomorphism $L[X,\sigma]/(X^{r}{-}a)\simeq L[X,\sigma]/(X^{r}{-}1)$.

Proof.

This maps $X^{r}$ to $\lambda\sigma(\lambda)\cdots\sigma^{r-1}(\lambda)X^{r}=aX^{r}$, thus mapping $X^{r}{-}a$ to $a(X^{r}{-}1)$. ∎

Corollary 2.2**.**

Multiplication in $L[X,\sigma]/(X^{r}{-}a)$ can be performed in $O(r^{\omega})$ operations in $K$.

Proof.

By Proposition 2.1 and Proposition 1.6, it is enough to show that for $A\in L[X,\sigma]/(X^{r}{-}a)$, $A(\lambda X)$ can be computed in $O(r^{\omega})$ operations in $K$. For this we write $\lambda_{i}=\lambda\sigma(\lambda)\cdots\sigma^{i-1}(\lambda)$ and remark that the $\lambda_{i}$’s ($0\leq i<r$) can be all computed within $\tilde{O}(r^{2})$ operations in $K$ thanks to the recurrence formula $\lambda_{i+1}=\lambda\cdot\sigma(\lambda_{i})$. Now evaluating the formula $A(\lambda X)=\sum_{i}\lambda_{i}a_{i}X^{i}$ allows us to compute $A(\lambda X)$ in $\tilde{O}(r^{2})$ operations in $K$. ∎

We could use the proof of Corollary 2.2 directly to design an algorithm for multiplication modulo $X^{r}{-}a$. Such an algorithm would require computing $A_{1}(\lambda X)$ and $A_{2}(\lambda X)$ each time we use it to compute $A_{1}A_{2}$. Alternatively, we can slightly modify the basis on which we are evaluating the corresponding maps, which can provide a gain if there are many multiplications to do modulo $X^{r}{-}a$.

Let $\lambda\in L^{\times}$, and let $\sigma_{a}=\lambda\sigma$. Let $\tilde{b}_{r-1}\in L$, and for $0\leq i\leq r-2$, $\tilde{b}_{i}=\sigma_{a}^{r-1-i}(\tilde{b}_{r-1})$, such that $\tilde{\mathcal{B}}=(\tilde{b}_{0},\dots,\tilde{b}_{r-1})$ is a basis of $L$ over $K$. By construction, we have for $1\leq i\leq r-1$, $\sigma_{a}(\tilde{b}_{i})=\tilde{b}_{i-1}$, and $\sigma_{a}(b_{0})=a\tilde{b}_{r-1}$. For example, if $\mathcal{B}=(b_{0},\ldots,b_{r-1})$ is a normal basis of $L$ over $K$, then $\tilde{b}_{r-1}=b_{r-1}$ and $\tilde{b}_{i}=\lambda\sigma(\lambda)\cdots\sigma^{i-1}(\lambda)b_{i}$ defines a suitable basis. Now, let $\tilde{B}=\sum_{i=0}^{r-1}\tilde{b}_{i}T^{i}\in L[T]$.

Proposition 2.3**.**

Let $A=\sum_{i=0}^{r-1}a_{i}X^{i}\in L[X,\sigma]$ and let $\tilde{c}_{j}=A(\sigma_{a})(\tilde{b}_{j})$. Let $\tilde{A}(T)=\sum a_{i}T^{i}\in L[T]$. Let $\tilde{C}_{a}=\sum_{j=0}^{r-1}\tilde{c}_{j}T^{j}$. Then

[TABLE]

Proof.

The proof is similar to that of Proposition 1.6. By linearity, it is enough to check that the relation holds for $A=X^{i}$ for $0\leq i\leq r-1$. Let $0\leq i\leq r-1$. We have :

[TABLE]

On the other hand, doing the calculations modulo $T^{r}{-}a$:

[TABLE]

Hence, $T^{i}B(T)=C_{X^{i}}(T)$ for all $0\leq i\leq r-1$, so $C_{a}(T)=\tilde{A}(T)B(T)$ for all $A\in L[X,\sigma]/(X^{r}{-}a)$. ∎

Algorithm ModMult below makes precise the algorithmical content of Proposition 2.3; it uses a primitive $\text{Mat}_{\text{work}}$ that takes as input a tuple $(x_{1},\ldots,x_{r})\in L^{r}$ and outputs the $r\times r$ matrix whose $j$-th column are the coordinates of $x_{j}$ is the working basis.

Proposition 2.4**.**

Algorithm ModMult computes the product $A_{1}A_{2}$ in $L[X,\sigma]/(X^{r}{-}a)$ in $O(r^{\omega})$ operations in $K$.

Proof.

Multiplication of polynomials in $L$ modulo $T^{r}{-}a$ requires $\tilde{O}(r^{2})$ operations in $K$. Multiplication of matrices of size $r$ in $K$ requires $O(r^{\omega})$ operations in $K$. Hence the global complexity is $O(r^{\omega})$ operations in $K$. ∎

2.1.2 Multiplication modulo $Z(X^{r})$

Let $K^{\prime}/K$ be a finite extension. Define $L^{\prime}=K^{\prime}\otimes L$; it is an étale $K^{\prime}$-algebra endowed with the endomorphism $\sigma^{\prime}=\text{id}\otimes\sigma$ that extends $\sigma$ and has order $r$.

Remark 2.5**.**

The algebra $L^{\prime}$ is not necessarily a field (for instance, when $K^{\prime}=L$, it splits as a product $L^{r}$). It is the reason why we needed to place this paper in the more general setting of étale algebras.

Let $\lambda\in(L^{\prime})^{\times}$. Set $a=N_{L^{\prime}/K^{\prime}}(\lambda)=\lambda\sigma(\lambda)\cdots\sigma^{r-1}(\lambda)\in K^{\prime}$. We assume that $K^{\prime}=K(a)$. Let $Z\in K[T]$ be the minimal polynomial of $a$. We want to generalize the results of §2.1.1 to multiplication modulo $Z(X^{r})$ (in §2.1.1, we have $K^{\prime}=K$, $L^{\prime}=L$ and $\lambda\in L^{\times}$). Note that if $(b_{0},\dots,b_{r-1})$ is a normal basis of $L/K$, then $(1\otimes b_{0},\dots,1\otimes b_{r-1})$ is a normal basis of $L^{\prime}/K^{\prime}$.

Lemma 2.6**.**

The canonical morphism $1\otimes\textrm{id}$ : $L[X,\sigma]\rightarrow L^{\prime}[X,\sigma]$ induces an isomorphism

[TABLE]

Proof.

First note that $(X^{r}{-}a)$ is a two-sided ideal of $L^{\prime}[X,\sigma]$, and that the canonical morphism $L[X,\sigma]\rightarrow L^{\prime}[X,\sigma]$ induces a morphism $L[X,\sigma]\rightarrow L^{\prime}[X,\sigma]/(X^{r}{-}a)$ which maps $X^{r}$ to $a$, hence the latter surjective. Moreover, by $K$-linearity, $Z(X^{r})$ lies in the kernel of this map. We then get a surjective morphism of $K$-algebras $L[X,\sigma]/Z(X^{r})\rightarrow L^{\prime}[X,\sigma]/(X^{r}{-}a)$. Since both sides have dimension $r^{2}\deg Z$ over $K$, this morphism is an isomorphism. ∎

We are now back exactly in the situation of Section 2.1.1, where $K$ has been replaced by $K^{\prime}$ and $L$ by $L^{\prime}$: all the computations can be carried out the same way, and passing back through the isomorphism of Lemma 2.6, we can perform fast multiplication modulo $Z(X^{r})$. The algorithm is as follows:

Proposition 2.7**.**

Algorithm 3 computes the product $A_{1}A_{2}$ in $L[X,\sigma]/(Z(X^{r}))$ with $O(r^{\omega}\deg Z)$ operations in $K$.

2.2 Reconstruction with CRT

Let $A_{1},A_{2}\in L[X,\sigma]$ be two skew polynomials. We recall that our aim is to design a fast algorithm for computing the product $P=A_{1}A_{2}$. We set $d=\deg P$.

Multiplication in large degree. We first assume that the polynomial $P=A_{1}A_{2}$ has degree larger than $r$. In this case, the idea is to evaluate the $P$ modulo various $Z_{i}(X^{r})$ using Algorithm ModMultZ and then to reconstruct the result using a non commutative version of the Chinese Remainder Theorem. The precise result we need is given by the following Proposition.

Proposition 2.8**.**

Let $Z_{1},\dots,Z_{m}\in K[T]$ be pairwise coprime polynomials, and let $Z=Z_{1}\cdots Z_{m}$. Then the natural map:

[TABLE]

is an isomorphism of $K$-algebras.

Proof.

Since the domain and the codomain have the same dimension over $K$, it is enough to prove the surjectivity. For $i$ between $1$ and $m$, consider $A_{i}\in L[X,\sigma]/Z_{i}(X^{r})$ and write it:

[TABLE]

where the $A_{i}^{(j)}$’s are polynomials with coefficients in $L$. For a fixed $j\in\{0,\ldots r{-}1\}$, let $A^{(j)}\in L[T]$ be a polynomial such that the congruence $A^{(j)}\equiv A_{i}^{(j)}\pmod{Z_{i}}$ holds in the commutative ring $L(T)$. We can therefore write $A^{(j)}=A_{i}^{(j)}+Z_{i}Q_{i}^{(j)}$ for some polynomials $Q_{i}^{(j)}\in L[T]$. Noting that the inclusion $L[T]\to L[X,\sigma]$, $T\mapsto X^{r}$ is a ring homomorphism (i.e. the multiplication on $L[T]$ agrees with that on $L[X,\sigma]$), we deduce that the equality

[TABLE]

holds in $L[X,\sigma]$. Multiplying it by $X^{j}$ on the right and summing up over $j$, we end up with $A\equiv A_{i}\pmod{Z_{i}}$ for all $i$. Surjectivity is proved. ∎

Remark 2.9**.**

The above proof is constructive. More precisely it shows that solving the Chinese Remainder problem of degree $d$ in $L[X,\sigma]$ with central moduli reduces to solving $r$ independant Chinese Remainder problems of degree $\frac{d}{r}$ in the commutative ring $L[X^{r}]$ and therefore can be achieved for a cost of $\tilde{O}(d)$ operations in $L$, corresponding to $\tilde{O}(dr)$ operations in $K$ (see [7], §10.3).

It remains now to explain how the moduli $Z_{i}(X^{r})$’s can be constructed. We will do it in two different concrete contexts: first, the case of finite fields and second, the case of number fields.

The case of finite fields. We assume that $K$ and $L$ are finite fields and write $q$ for the cardinality of $K$. We consider an auxiliary finite extension $K^{\prime}$ of $K$ of degree $n$ and build the compositum $L^{\prime}=K^{\prime}\otimes_{K}L$. We endow $L^{\prime}$ with the uniform measure. We assume that $n$ is chosen sufficiently large so that:

[TABLE]

Asymptotically the latest condition is fulfiled as soon as $n$ grows at least as fast as $\log r$.

Lemma 2.10**.**

Let $t$ be an integer such that $4t^{2}\leq nq^{n}$. Let $\lambda^{\prime}_{1},\ldots,\lambda^{\prime}_{t}$ be random independant elements of $L^{\prime}$. Then the $N_{L^{\prime}/K^{\prime}}(\lambda^{\prime}_{i})$’s all generate $K^{\prime}$ over $K$ and are pairwise non-conjugate over $K$ with probability at least $\frac{1}{2}$.

Proof.

The étale algebra $L^{\prime}$ splits as a product $(M^{\prime})^{g}$ where $M^{\prime}$ is a finite extension of $K^{\prime}$ of degree $f$ and $g$ is a positive integer. Moreover if $x\in L^{\prime}$ decomposes as $x=(x_{1},\ldots,x_{g})$, we have:

[TABLE]

Observe that the norm map $N_{M^{\prime}/K^{\prime}}$ takes the value [math] only at [math]. Hence the probability that $N_{M^{\prime}/K^{\prime}}$ vanishes is $q^{-nf}$. Therefore $N_{L^{\prime}/K^{\prime}}$ vanishes with probability $1-(1-q^{-nf})^{g}$. As for the nonzero values of $K^{\prime}$, they are reached by $N_{L^{\prime}/K^{\prime}}$ with uniform probability because $N_{L^{\prime}/K^{\prime}}$ is a surjective group homomorphism, i.e.

[TABLE]

for all $a\in K^{\prime}$, $a\neq 0$. Let $c_{n}$ be the number of elements of $K^{\prime}$ that generate $K^{\prime}$ over $K$. The probability that a fixed $\lambda^{\prime}_{i}$ satisfy the requirement K\big{(}N_{L^{\prime}/K^{\prime}}(\lambda^{\prime}_{i})\big{)}=K^{\prime} is then $(1-q^{-nf})^{g}\cdot\frac{c_{n}}{q^{n}-1}$. Assuming that this occurs, the probability that the $N_{L^{\prime}/K^{\prime}}(\lambda^{\prime}_{i})$’s are pairwise non-conjugate is:

[TABLE]

Putting all together, we find the probability of success:

[TABLE]

which is at least:

[TABLE]

Clearly $q^{n}{-}c_{n}$ is the cardinality of the union of all strict subextensions of $K^{\prime}$. Therefore:

[TABLE]

the latter inequality coming from the fact that $n$ has at most $2\sqrt{n}$ divisors. From (1), we derive $q^{n}{-}c_{n}\leq\frac{q^{n}}{4}$. On the other hand, it follows from our assumptions that $r\leq\frac{q^{n}}{8}$ and $\frac{t(t{-}1)}{2n}\leq\frac{t^{2}}{2n}\leq\frac{q^{n}}{8}$. Combining with (2), we find that the probability of success is at least $\frac{1}{2}$. ∎

Theorem 2.11**.**

Let $A_{1},A_{2}\in L[X,\sigma]$ of degree $d\geq r$. Then Algorithm Mult computes the product $A_{1}A_{2}$ within $O(dr^{\omega-1})$ operations in $K$ with probability of success at least $\frac{1}{2}$.

Proof.

Observe first that $n$ can be chosen such that $n=O(\log d+\log r)$. Computing the product in $L[X,\sigma]/Z_{i}(X^{r})$ requires $O(r^{\omega}n)=\tilde{O}(r^{\omega})$ operations in $K$. Moreover by Remark 2.9, the reconstruction (line 4) can be done for a cost of $\tilde{O}(rd)$ operations in $K$. The overall cost of Mult is then $\tilde{O}(dr^{\omega-1})$ as announced. The fact that the probability of success is at least $\frac{1}{2}$ follows from Lemma 2.10. ∎

The case of number fields. We assume that $K$ and $L$ are number fields. It is then known that the image of the norm map $N_{L/K}:L^{\star}\to K^{\star}$ has index $r$ in $K^{\star}$. More precisely, class field theory teaches us that $K^{\star}/N_{L/K}(L^{\star})$ is canonically isomorphic to the Galois group of the abelian extension $L/K$, i.e. to $\mathbf{Z}/r\mathbf{Z}$. In particular, the image of $N_{L/K}$ is infinite meaning that if we take a finite set of random elements $\lambda\in L$, it is likely that the norm of the $\lambda$’s will be pairwise distinct. We can then reapply the strategy used in the case of finite field without having to work with an auxiliary extension $K^{\prime}$. We end up this way with a probabilistic Las Vegas algorithm whose complexity is $\tilde{O}(dr^{\omega{-}1})$ operations in $K$ and whose probability of success is high.

Multiplication in small degree. The idea for fast multiplication in small degree is that if a skew polynomial has degree $d\ll r$, it is determined by its values on $d{+}1$ linearly independent elements of $L$. Hence, starting with two skew polynomials $A_{1},A_{2}$ whose degrees add up to $d$, we should be able to compute their product by composing of two $K$-linear maps over vector spaces of dimension $d{+}1$. However, we know some efficient algorithm for evaluating $A(\sigma)$ only on a subspace of $L$ which is spanned by the first vectors of a normal basis. For this reason, it order to compute $A_{1}A_{2}(b_{0}),\ldots,A_{1}A_{2}(b_{d-1})$, we shall need to know the whole of the linear map $A_{1}(\sigma)$ (because $A_{2}(b_{0}),\ldots,A_{2}(b_{d-1})$ are in general nothing to do with a truncated normal basis).

The complexity of the above algorithm is given by the next Theorem whose proof is straightforward after what we have already done (the bottleneck comes from the matrix multiplication step).

Theorem 2.12**.**

Let $A_{1},A_{2}$ such that $\deg A_{1}+\deg A_{2}\leq d<r$. Then Algorithm 5 computes the product $A_{1}A_{2}$ with $O(d^{\omega-2}r^{2})$ operations in $K$.

Conclusion. As a conclusion, several algorithms with different complexities are available for the multiplication of skew polynomials. Precisely, we have designed in this paper one algorithm of complexity $\tilde{O}(dr^{\omega-1})$ when $d\geq r$ and an another algorithm of complexity $\tilde{O}(d^{\omega-2}r^{2})$ when $d\leq r$. Apart from that, Wachter-Zeh’s algorithm [13] performs the same computation with complexity $\tilde{O}(d^{(\omega+1)/2}r)$ without any assumption on $d$. The corresponding complexity curves are represented on Figure 1.

Putting all together, we find that the product in $L[X,\sigma]$ can be performed within $\tilde{O}(\text{\rm SM}(d,r))$ operations in $K$ where:

[TABLE]

As already discussed in the introduction, we expect to lower the complexity to $\tilde{O}(d^{\omega-1}r)$ in the range $d\leq r$ and, until now, we have not succeeded in doing so.

3 Other operations

and applications

Classically, fast multiplication algorithms can be used to speed up many other computations. This general philosophy works for skew polynomials as well and was concretized in [4], §3.2. Below, we analyze briefly the impact of the algorithms designed above in this paper.

In order to state our complexity results more elegantly, we introduce the function $\text{\rm SM}^{\geq 1}$ defined by:

[TABLE]

A direct computation shows that:

[TABLE]

The function $\text{\rm SM}^{\geq 1}$ (viewed as a function of the variable $d$) is the smallest function above SM whose “log-log slope” is always at least $1$ (see Figure 1). The notation comes from this interpretation.

With $\omega=2.37$, we have $\text{\rm SM}^{\geq 1}(d,r)\approx d^{1.69}r$ for $d\leq r^{0.76}$ $\text{\rm SM}^{\geq 1}(d,r)\approx dr^{1.52}$ for larger $d$.

Euclidean division. An algorithm that performs (right) Euclidean divisions in $L[X,\sigma]$ and takes advantage of fast multiplication algorithm is depicted in [4], §3.2.1 (Algorithm REuclideanDivision). Proposition 3.2.3 of loc. cit. extends readily to the settings of this paper and shows that the aforementioned algorithm has a complexity cost of $\tilde{O}(\text{\rm SM}^{\geq 1}(d,r))$ operations in $K$.

gcd** and lcm computation.** The classical half-gcd algorithm that we already mentioned above (see §1.3 and [7], §11) works in the same way to compute left and right gcd’s of skew polynomials. The precision corresponding algorithm is written in [4], §3.2.2 (Algorithm FastExtendedRGCD).

Proposition 3.1**.**

The algorithm FastExtendedRGCD of [4], §3.2.2 (using fast multiplication algorithms described above in this paper as primitives) runs in $\tilde{O}(\text{\rm SM}^{\geq 1}(d,r))$ operations in $K$.

Proof.

A careful look at the algorithm FastExtendedRGCD shows that its complexity in operations in $K$ is bounded by $T(d,r)$ where $T(d,r)$ satisfies the recurrence relation:

[TABLE]

By induction, it follows that for $m\geq 0$,

[TABLE]

Taking $m=\lfloor\log_{2}d\rfloor$, we get $T(d,r)=\tilde{O}(\text{\rm SM}^{\geq 1}(d,r))$ as expected. ∎

Remark 3.2**.**

A similar complexity is available for the computation of lcm’s.

Minimal subspace polynomial. Let $(x_{1},\ldots,x_{d})$ be a family of elements of $L$ which is free over $K$. We are interesting in computing the unique monic polynomial $P\in L[X,\sigma]$ of degree $d$ such that $P(x_{i})=0$ for all $i\in\{1,\ldots,d\}$.

Lemma 3.3**.**

For $x\in L$, $x\neq 0$, the value $\frac{P(x)}{x}$ is the remainder in the right Euclidean division of $P$ by $X-\frac{\sigma(x_{i})}{x_{i}}$.

Proof.

It is a direct computation. ∎

Lemma 3.3 shows that the polynomial $P$ we are looking for is nothing but the left-lcm of the polynomials $X-\frac{\sigma(x_{i})}{x_{i}}$. As a consequence, $P$ can be computed for a cost of $\tilde{O}(\text{\rm SM}^{\geq 1}(d,r))$ operations in $K$ using fast algorithms for lcm computation together with a “tree division strategy” [7], §10.1.

General multievaluation. We consider again a free family $(x_{1},\ldots,x_{d})$ of elements of $K$. The general multievaluation problem consists in evaluating a given polynomial $P\in K[X,\sigma]$ of degree $d$ at the $x_{i}$’s. Thanks to Lemma 3.3, the value $P(x_{i})$ agrees with $x_{i}$ times the remainder of the right division of $P$ by We are then reduced to compute the reduction of a given polynomials modulo some given moduli. This can be done efficiently using the strategy of [7], §10.1 for a cost of $\tilde{O}(\text{\rm SM}^{\geq 1}(d,r))$ operations in $K$. If $d$ are $r$ have the same order of magnitude, one can preferably compute the matrix of $P(\sigma)$ using the formula of Proposition 1.6 and derive from it the values of the $P(x_{i})$’s thanks to a single matrix multiplication. The cost of the resulting algorithm is $O(r^{\omega})$.

Remark 3.4**.**

If the $x_{i}$’s are the first vectors of a normal basis of $L$ over $K$, one can use directly the algorithm of §1.3 which has a better complexity.

General interpolation. We keep the family $(x_{1},\ldots,x_{d})$ and consider in addition some values $y_{1},\ldots,y_{d}\in L$. We address the question of computing a polynomial $P$ of degree at most $d{-}1$ such that $P(x_{i})=y_{i}$ for all $i$. Thanks to Lemma 3.3, the above problem reduces to solve the following Chinese Remainder system:

[TABLE]

which again can be done for a cost of $\tilde{O}(\text{\rm SM}^{\geq 1}(d,r))$ operations in $K$.

Remark 3.5**.**

If the $x_{i}$’s are the first vectors of a normal basis of $L$ over $K$, one can use directly the Algorithm SmallDegreeInterpolation which has a better complexity.

Gabulin codes. The solution sketched above to the general multievaluation problem allows us to encode messages in the framework of (generalized) Gabidulin codes [14] in complexity $O(n^{\omega})$ where $n$ is the length of the code. (Better complexities are possible when the dimension of the code is much smaller than its length.) In the similar fashion, efficient decoding is also possible using the key equation together with the half-gcd algorithm. The resulting algorithms run in $\tilde{O}(SM^{\geq 1}(n,k))$ operations in $K$ where $n$ and $k$ denotes the length and the dimension of the Gabidulin code respectively.