
. This paper is part of a project that has received funding from the French “Agence de l'Innovation de Défense”.
Let be a fixed effective field. The most straightforward approach to compute with an element in the algebraic closure of is to compute modulo its minimal polynomial. The determination of a minimal polynomial from an arbitrary annihilator requires an algorithm for polynomial factorization over . Unfortunately, such algorithms do not exist over generic effective fields. They do exist over fields that are explicitly generated over their prime subfield, but they are often expensive. The dynamic evaluation paradigm, introduced by Duval and collaborators in the eighties, offers an alternative algorithmic solution for computations in the algebraic closure of . This approach does not require an algorithm for polynomial factorization, but it still suffers from a nontrivial overhead due to suboptimal recomputations. For the first time, we design another paradigm, called directed evaluation, which combines the conceptual advantages of dynamic evaluation with a good worst case complexity bound.

Let be an effective field. Here, effective means that elements of can be represented by concrete data structures and that algorithms for the field operations are at our disposal, including a zerotest. Effective rings are defined in a similar way.
The aim of the present paper is the fast computation with algebraic numbers. We start by recalling various known strategies, including the most prominent one: dynamic evaluation. Our first main problem will be to develop a more efficient variant of dynamic evaluation. We next recall how to represent algebraic numbers using towers of algebraic extensions. Our second main problem is to develop efficient algorithms for computations in such towers.
A natural first approach begins with the computation of the defining polynomial of over : this essentially corresponds to factoring into irreducible polynomials. Unfortunately this task is not feasible over a generic field [19, 20]. Of course, when is explicitly finitely generated over its prime subfield, such factorizations can be computed. However, no general efficient algorithms (i.e. running in softly linear time) are currently known for this task; see [21].
Another approach consists in computing in while regarding as a parameter constrained by . So any element in is represented by a polynomial in of degree , such that . With this representation, additions, subtractions and products can easily be performed in , with seen as the class of . On the other hand, testing whether is zero or not requires us to distinguish the following cases:
If is the zero polynomial, then is zero for any root of , so the result of the test is “”;
If the resultant is nonzero, then is nonzero for any root of , so the result of the test is “”;
Otherwise, is not a constant polynomial and has degree . A proper decomposition of is thus discovered: . We split the current computation, into the two following ones:
In the first branch, is subjected to the constraint ; in this branch, the result of the zerotest becomes “”.
In the second branch, is subjected to the constraint and the result of the zerotest becomes “”.
If the inverse of is requested, then we first test if is zero or not. If is identically zero (in the resulting branch), then an exception is raised; if is nonzero, then the inverse can be computed from the Bézout relation for and , which can itself be computed using the extended gcd algorithm.
This manner of evaluating a program is called dynamic evaluation. It finishes after a necessarily finite number of splittings. At the end of all the computations, we obtain a factorization of , not necessarily into irreducible polynomials, along with one output value for each of the cases when for . In other words, instead of working with a specific root of , the program is evaluated for a generic root of . For this reason, is sometimes called an algebraic parameter: the program is executed as if were a field, and cases are distinguished only when inconsistencies actually occur during the execution. After a zero test or an inversion that gives rise to a splitting, the element is enforced to be identically zero or invertible in each branch. Notice that the separability assumption on is important, since it ensures that and are coprime in the above decomposition .
Example
Using dynamic evaluation, this gives rise to three branches, and the result
Dynamic evaluation is very attractive from a conceptual point of view, but rather tricky to implement: splittings correspond to nondeterministic control structures that are typically implemented using continuations [41]. Dynamic evaluation is also difficult to analyze from the complexity perspective. In particular, it is hard to determine the amount of computation that can be shared between several branches. At any rate, the worst case complexity of dynamic evaluation is much higher than the number of steps of the program multiplied by the cost of operations in .
The first aim of this paper is to design an alternative, socalled directed evaluation strategy, with the same abstract specification, but with a quasilinear computational complexity in . The key idea is to run the entire program, while systematically opting for the “principal” branch in case of splittings. This principal branch is the one for which the defining polynomial of is maximal. At every splitting, we also store the defining polynomial for the other “residual” branch. Once the computation for the principal branch has been completed, we recursively apply the same algorithm for each residual branch, on the reduced input data modulo the corresponding defining polynomial. The required reductions are computed simultaneously for all residual branches, using a fast divideandconquer algorithm for multimodular reduction.
The parameters naturally induce a tower of effective algebraic extensions
where
The parameter corresponds to the class of in for . A tower of this type is written and we call its height. Throughout this paper, we write and . The are called the defining polynomials of the tower, and its degree. Elements in are naturally represented by univariate polynomials in over of degree .
In order to reduce computational costs, it will be convenient to assume that the tower is explicitly separable in the sense that and that we have explicit Bézout cofactors with for . In particular, each is reduced. This means that contains no nonzero nilpotent elements or, equivalently, that is a product of separable field extensions of .
Towers of algebraic extensions are related to “triangular sets”. In fact, we may see the preimage of over as a multivariate polynomial in such that has degree in , degrees in for , and . The sequence then forms a special regular case of a triangular set. The separability of translates into the requirement that is separable for all roots of for .
The second main aim of this paper is to generalize the directed evaluation strategy from the univariate case, while ensuring that the complexity remains bounded by the number of steps in our algorithm multiplied by (which corresponds roughly speaking to the cost of arithmetic operations in the tower [29]). Intuitively speaking, we use the univariate strategy in a recursive fashion, but the analysis becomes far more technical due to the fact that splittings can occur at any level of the tower. This difficulty will be countered through the development of suitable data structures.
Given an effective ring , we use as a notation for the cost of multiplying two polynomials of degree , in terms of the number of operations in . It is known [6] that . We will also denote
In particular, given for some monic polynomial of degree , elements in may be represented as classes of polynomials in modulo , additions in require additions in , whereas multiplications in can be done using ring operations in .
We let denote an exponent such that two matrices over a commutative ring can be multiplied with operations in . The constant denotes another exponent such that a rectangular matrix over a commutative ring can be multiplied with a square matrix with operations in . Le Gall has shown in [34] that one may take ; Huang and Pan have shown in [31] that one may take . Throughout this paper, in order to simplify complexity analyses, we assume that and .
Recall that an element is said to be primitive if . If is a reduced algebra, then it is well known that a sufficiently generic linear combination of is a primitive element of . It may thus seem odd, at first sight, to manipulate towers of height whenever an isomorphic algebra exists. The problem is that both the computations of primitive elements and conversions between representations are usually expensive.
Over a general field , no efficient algorithm is currently known for finding a primitive element together with its minimal polynomial , and polynomials such that for . It is known that this problem can efficiently be reduced to a multivariate version of the problem of modular composition using a randomized Las Vegas algorithm. The corresponding conversions between tower and primitive element representations can also be done using modular composition. All this is a consequence of the transposition principle and Le Verrier's method; we refer to [29, 30, 33, 38, 39] for more details. Unfortunately, no softly linear algorithm is known for modular composition over a generic field, although efficient algorithms have recently been designed for specific cases [9, 27, 28]. If is a finite field, then Kedlaya–Umans and followers have established almost optimal theoretical bit complexity bounds [30, 33, 38].
Without primitive elements, using a naive induction on , the multiplication in can be performed in time for some constant . It is well known that, for a sufficiently large constant , such products can actually be carried out in time using Kronecker substitution; see for instance [21, chapter 8]. If many of the individual degrees are small, then can become as large as , which leads to an overall complexity bound of the form . Unfortunately, this bound is far from linear if is much larger than . Lebreton has shown that one may take ; see [35] and [29, Proposition 2.4].
Substantial improvements for the value of seem difficult to achieve, and it might not even be possible to reach values that are arbitrarily close to . An alternative approach for efficient computations in a given tower is to produce an equivalent, socalled accelerated tower of small height, such that conversions between both towers are reasonably cheap. This approach was first proposed in [29]: when the are all irreducible, we have shown how to multiply elements in with operations in , for any fixed real .
Dynamic evaluation has been developed by Della Dora, Dicrescenzo and Duval [10, 1315] as a way to compute with algebraic parameters without irreducible factorizations. The approach is sometimes called the “D5 principle”, after the initials of the authors of [10]. One may regard dynamic evaluation as a computer algebra counterpart of the concept of nondeterministic evaluation from theoretical computer science. The strategy of dynamic evaluation has been extended to transcendental parameters [17, 22], real algebraic numbers in [16], and more general algebraic structures [25, chapter 8].
Several implementations have been proposed for parameters that are constrained by triangular sets. Early implementations simply handled splittings by redoing the entire computation under stricter constraints [13]. Unnecessary recomputations can be avoided through the use of highlevel control structures such as continuations [4]. Unfortunately, efficient implementations of such control structures are rarely available for common programming languages. In this paper, we mostly leave implementation issues aside and focus on the complexity of computations with parameters.
In his PhD thesis, Dellière has investigated the relationship between dynamic evaluation and decompositions of constructible sets into triangular sets [11, 12]: the central operation is the computation of gcds with coefficients in products of fields, such as . Let us further mention that dynamic evaluation has influenced several polynomial system solvers relying on triangular sets [3], and is now involved in various other algorithms in computer algebra; see for instance [7, 26, 36].
Dedicated algorithms over dynamic fields such as have been proposed by Dahan et al. [8]. Without appealing to the general dynamic evaluation paradigm, they designed efficient algorithms for quasiinversion in , as well as gcd computations and coprime factorization in . Recomputations induced by splittings are handled using fast ad hoc techniques.
In the worst case, it is known that dynamic evaluation over suffers from an overhead of the order : see Examples 4.4 and 4.6. To our knowledge, this paper is first to propose a general strategy for removing this overhead.
The main contribution of the present paper is a new evaluation paradigm to compute with algebraic parameters in an asymptotically fast manner. In section 2, we recall the concept of computation trees as a formalization for the cost of algebraic algorithms. In section 3, we introduce the concepts of unpermissive, panoramic, and directed evaluation. This provides us with precise terminology for analyzing subtle differences between the complexities of various evaluation strategies.
Informally speaking, the unpermissive evaluation of a computation tree simply aborts the usual evaluation whenever a zerodivisor needs to be inverted or tested to zero. If a zerodivisor is discovered in this way, then a proper factorization of one of the defining polynomials can be deduced. This leads to a decomposition of into the direct sum of two proper subalgebras, defined by the towers and .
A panoramic evaluation of a computation tree returns all possible results when considering as algebraic parameters, and thereby constitutes our main objective. The unpermissive evaluation model gives rise to the following naive panoramic evaluation strategy: whenever the unpermissive evaluation produces a proper splitting of , we restart the evaluation over each of the two subalgebras of . The problem with this rather naive approach is that the evaluation may need to be restarted as many as times, which turns out to be suboptimal from a complexity point of view: see Example 4.4.
The directed evaluation is meant to remedy this situation by ensuring a sharp decrease of every time that we need to restart the evaluation. More precisely, we will show that the reevaluation depth remains bounded by .
Section 4.2 contains our main complexity result in the case of a single parameter, i.e. . We prove a softly linear complexity bound for panoramic evaluation and provide a detailed comparison with various implementations of dynamic evaluation.
In section 5, we turn to the case of an arbitrary number of parameters . With respect to the univariate case, the main difference is that is not necessarily a field, although we still conduct our computation as if it were (e.g. during gcds computations in ). Consequently, zerotests and inversions of elements in may lead to additional case distinctions according to the possible values of . Despite these technical complications, we again prove a quasilinear complexity bound for the cost of panoramic evaluation, under the assumption that is fixed (this means that is not allowed to grow with ).
For arbitrary, not necessarily bounded heights , the complexity analysis becomes tedious. We need to revisit the construction of accelerated towers of fields, introduced in [29], in the context of directed evaluation. The key idea is as follows: before the directed evaluation of a given computation tree, we perform the directed computation of an accelerated version of the current tower, so the computation tree can subsequently benefit from accelerated arithmetic. Overall, in section 7, we achieve an overhead of the form for the panoramic evaluation, where represents any fixed positive real number.
Our techniques are more general than those of [8]: they turn out to be applicable in all contexts that involve dynamic evaluation with algebraic numbers. In addition, thanks to our accelerated tower arithmetic, we improve the complexity bounds for the specific tasks studied in [8], whenever is allowed to be arbitrarily large.
Section 7.4 contains two important additional examples. We first show that products in a separable tower can be done in time , which extends [29]. We next present a less straightforward application of fast panoramic evaluation to matrix multiplication with entries in .
This section gathers known definitions and results about elementary operations in computer algebra, to be used in this paper.
Let us recall the notion of a computation tree; we essentially follow the presentation from [5, chapter 4, section 4]. In the present paper, all computation trees manipulate data in algebras over an effective field , and only the following operations are allowed:
The binary arithmetic operations in the algebra.
The unary operation of inversion, which is partially defined in the algebra.
The unary zerotest, written .
For each constant , the nullary constant function and the unary function of scalar multiplication by . We denote the corresponding sets of constant functions and scalar multiplications by and .
We write for the set of all arithmetic operations.
assigns an instruction of the form
to every computation node , where has arity , and are nodes in that are predecessors of (i.e. nodes in the path ascending from to the root of the tree);
assigns an instruction of form
to every branching node , where is a predecessor of in ;
assigns a return value to every output node , where are predecessors of in , and may depend on .
The value is a new symbol that is returned whenever an arithmetic operation cannot be executed. In the present framework this can only happen for inversions.
Given an input value , the evaluation of at proceeds by constructing a path from the root of the tree, along with the function
that associates a value to each node of the path, as follows:
The path begins with the input nodes of , and we set for . The next node of the path is set to the successor of .
If the current node of the path is a computation node of the form then we set . If this calculation is well defined, then the next node of the path is set to the successor of . Otherwise, the evaluation path ends at , we say that the computation tree is undefined at , and we set .
If the current node of the path is a branching node of the form , then we set . If is then the next node of the path is set to the left successor of , otherwise the next node is the right successor of .
If the current node of the path is an output node with return value , then we set . The path ends at and we set .
Since additions, multiplications and inversions typically admit different costs, it is convenient to introduce the following more detailed quantities:
stands for the number of input nodes, i.e. .
stands for the maximal arity of a return value, i.e. .
stands for the maximal number of operations in for a path in .
stands for the maximal number of multiplications for a path in .
stands for the maximal number of inversions or zerotests for a path in .
We will call the detailed cost of and notice that . Taking in Example 2.1, we have , , , , .
Usually, we are really interested in the maximal number of scalar operations in that are needed to evaluate over . If and are clear from the context, this is simply called the cost of or the time needed to evaluate . It will be useful to introduce the following additional quantities:
stands for the maximal number of elements in that are needed to represent an element in . Throughout this paper, we assume that operations in can all be computed using operations in .
stands for the number of operations in that are needed to multiply two elements in .
stands for the maximal number of operations in that are needed to perform a zerotest or inversion in .
We clearly have
Remark
negations are implemented as ;
divisions are implemented as ;
equality tests are implemented as .
These restrictions are harmless, in the sense that they only affect the constants hidden in the “” of complexity estimates.
Remark
Remark
Remark
A polynomial in of degree is represented by the vector of its coefficients from degree to . Polynomial additions (resp. subtractions) in take at most additions (resp. subtractions) in . Determining the degree of a polynomial in requires at most zerotests.
is a nondecreasing function in —this assumption is customary;
is sufficiently close to linear, in the sense that
(2.1) 
holds whenever . This assumption is less common, but it is only used within Propositions 2.6 and 7.7. Notice that (2.1) is equivalent to .
For general , it has been shown in [6] that one may take
If has positive characteristic, then it was shown in [23, 24] that one may even take
where .
, , , ;
, , , where is the quotient of by .
Consequently we have , that is the remainder in the division of by , and for all . The sequence ends after division steps with for , and . The last nonzero polynomial is and the Bézout relation is .
The fast extended Euclidean algorithm applied to and does not compute the complete Euclidean sequence, but only the sequence of the quotients in a divide and conquer fashion. This is enough to permit the efficient recovery of the Bézout relation . For , let . We claim that the number of zerotests in the fast extended Euclidean algorithm is exactly . This claim is a consequence of the two following observations about the fast extended Euclidean algorithm (see for instance [37, Algorithm 18]):
Computing the quotients does not involve any zerotests or inversions. On the other hand the determination of requires testing the coefficients of from degree to , which involves zerotests. This totalizes zerotests.
The degrees of the cofactors and , but also of all the entries of the “transition matrices” are explicitly determined from the ; see [37, Lemma 10]. So the computation of the transition matrices also does not involve any zerotests or inversions.
In order to perform polynomial divisions, the inverses of the leading coefficients of the are needed. This requires at most inversions.
To summarize the discussion, the extended gcd of and of degree takes additions, subtractions and products in , plus zerotests and inversions. In addition, every inversion of an element is preceded by an unsuccessful zerotest for that element.
Elements of a tower are in fine represented by vectors of coordinates in . So additions, subtractions, and products by scalars in can be done by straightline programs with linear costs. We recall the following result, where represents the cost of one product in .
In addition, the product of two polynomials in can be computed by a straightline program in time
Proof. This result is essentially due to Lebreton [35]; see also [29, Proposition 2.4].
In the introduction, we have informally discussed computations with parameters and various strategies for automating case distinctions. In this section, we formalize these ideas in the specific situation of algebraic parameters.
Let be an effective field and let be an ideal of such that:
is a algebra of finite dimension ,
is absolutely radical, which means radical over the algebraic closure of .
In the context of the present paper, it will be convenient to call a parametric algebra over with parameters, and is said to be its defining ideal, written . The respective images of in are denoted by ; they are regarded as parameters over , subject to the relations in . Geometrically speaking, the tuple represents a generic point in the set of the zeros of in . This setting is more general than the one of the introduction, but it is intended to be applied later to ideals generated by triangular sets.
Let be a second parametric algebra over with parameters. As a shorthand, we write whenever is a quotient algebra of . This is the case if, and only if, , or, equivalently, if . If , then we write for the canonical projection , and notice that represents a generic point in the set of the zeros of in . With a slight abuse of notation, is extended in a coefficientwise manner to .
Two parametric algebras and are said to be disjoint if . In that case, we have
More generally, given pairwise disjoint , we have
We will call such a decomposition a splitting of .
Now consider the prime decomposition
of the defining ideal of , and define for . Then we have and there exists a natural isomorphism
(3.1) 
It is convenient to call this relation the total splitting of . The corresponding zeroset is a disjoint union:
Given a zerodivisor , we have
Indeed, we clearly have . Conversely, given with , we have , whence since is radical, and . Setting
we obtain a natural isomorphism
Such a decomposition is called an atomic splitting of . We have
From an effective point of view, the computation of total splittings requires an algorithm for polynomial factorization, whereas atomic splittings can be computed using standard techniques from effective polynomial algebra, such as Gröbner bases, triangular sets, or geometric resolutions.
Let us now describe an abstract way to formalize computations in parametric algebras modulo splittings. Consider a set of parametric algebras over . We say that is a parametric framework if the following conditions are satisfied:
Each is an effective algebra.
Given , we can test whether is invertible and compute if so.
Given with , we have an algorithm for the projection .
Given , we can effectively compute and .
For any with , we have algorithms for both directions of this isomorphism.
Using Gröbner basis techniques, the above conditions are in particular verified if is the set of all parametric algebras that are explicitly given by a finite set of generators of . Indeed, a Gröbner basis of induces a basis of over along with the multiplication matrices by , so the above operations boil down to linear algebra. This observation is convenient from a conceptual point of view, but, of course, not very efficient from the asymptotic complexity perspective. Sections 4, 5, and 7 are devoted to more specific parametric frameworks that allow for asymptotically faster implementations.
From now on, denotes a parametric algebra, and we will use the notations from section 2.1 for computation trees. The unpermissive evaluation of a computation tree with input nodes as above at is defined as follows. Informally speaking, the tree is evaluated in the usual way over concerning ring operations, but the evaluation aborts whenever a zerodivisor needs to be inverted or tested to zero.
More precisely, we write for the unpermissive path, and for the unpermissive evaluation function. We introduce a new output value, written , in order to indicate that the evaluation process detects that the current algebra is not a field. We define by induction:
If the current node of the path is a computation node of the form , then we distinguish the following cases:
If is zero in , then the path ends at with .
If is invertible, then and the next node of the path becomes the successor of .
Otherwise, is a zerodivisor in , and the evaluation aborts; the path ends at , we set , and record the zerodivisor for future use.
If the current node in the path is a branching node , then we distinguish the following cases:
If is zero in , then , and the next node of the path becomes the right successor of .
If is invertible in , then we set , and the next node of the path becomes the left successor of .
Otherwise, is a zerodivisor in , and the evaluation aborts; the path ends at , we set , and record the zerodivisor for future use.
For the other types of nodes, the evaluation rules remain the same as for the standard evaluation described in section 2.1.
Proof. The proof is straightforward from the definitions.
is a panoramic splitting of for the pair if for . A panoramic value of at is a set of pairs such that is a panoramic splitting of and for .
where is the unique integer such that .
Proof. Since is a field, any nonzero element in is invertible. Consequently,
and Lemma 3.1 implies that . On the other hand,
and Lemma 3.1 implies that
Example
, and , with . If represents the image of in , then a possible panoramic splitting of at is
where
,
and , for ,
and .
Notice that the associated primes of are for , and the total splitting of is
The following naive algorithm for the computation of panoramic values was already sketched in the introduction:
Algorithm
Compute the unpermissive evaluation .
If , then return .
Otherwise a nontrivial zerodivisor has been recorded, which allows us to compute proper subalgebras and such that and .
Recursively apply the algorithm to for and return the union the panoramic values obtained in this way.
Proof. The algorithm finishes since is finite dimensional. As for the correctness it suffices to verify that the output in step 2 is correct, which is straightforward from the definitions.
From now we focus on the development of a faster alternative for Algorithm 3.1. The idea is to impose a finer control over the dimension of the components of splittings. Informally speaking, computation trees over will be evaluated entirely, while restricting operations to suitable subalgebras of : when a zero test or an inversion occurs for a value , the current subalgebra is further restricted to its largest subalgebra where the projection of is either invertible or zero.
In order to formalize this idea, we need some definitions and the underlying directed arithmetic operations.
is directed if
and for ,
for .
A directed evaluation of takes as input:
a directed splitting of ,
a point ,
and returns
a refined directed splitting of ,
such that .
We introduce a directed variant of the inversion in , that takes a directed splitting as input, the value in to be inverted, and that returns a directed splitting such that
is either zero or invertible;
The result of the directed inversion of is respectively or .
We also introduce a directed variant of the zerotest in , that takes a directed splitting and the value in to be tested as input, and that returns a directed splitting such that
is either zero or invertible;
If is invertible, then has been computed and recorded;
The result is .
Remark
Remark
Let us now define the directed evaluation of a tree . As input, we take a directed splitting and . Informally speaking, this evaluation simply applies the above directed operations in sequence, so we evaluate the entire computation tree while restricting the current algebra when encountering zerotests and inversions. The list of the residual subalgebras is maintained throughout the evaluation process. For consistency, the evaluation takes a directed splitting as input, but only the summand will be decomposed during the evaluation process.
Formally speaking, the directed evaluation path is written . To each node , we both associate a directed splitting
of , so
and a value
The directed evaluation of is defined inductively as follows:
The path starts with the input nodes of . We set and , for . The next node of the path is set to the successor of .
If the current node of the path is the root of the path, and is a computation node , then we necessarily have and , we set and . The next node of the path is set to the successor of .
If the current node of the path is a computation node with , whose direct predecessor is , then we set and
The next node of the path is set to the successor of .
If the current node of the path is a computation node , whose direct predecessor is , then we apply the directed inversion to for the splitting . This yields a new splitting that we assign to .
If , then we set , and the path ends at .
Otherwise , and the next node of the path becomes the successor of .
If the current node of the path is a branching node , whose direct predecessor is , then we apply the directed zerotest to for the splitting . This yields a new splitting , that we assign to .
If , then , and the current node of the path becomes the right successor of .
Otherwise, , and the current node of the path becomes the left successor of . We also record the inverse .
If the current node of the path is an output node , whose predecessor is , then we set and
Whenever are such that is a successor of , the splittings of at and satisfy . It can be checked by induction over the path that and are well defined.
Proof. We verify by induction on paths that and coincide, and that holds for all that is not a branching node.
Example
, and , with . Consider the directed evaluation of at , where represents the image of in . The first zerotest that we encounter is , which yields the directed splitting
After the zerotests , we end up with the directed decomposition
and the value . This is due to the requirement that directed evaluation always privileges the “branch of highest degree”. If , then the last zerotest potentially leads to another directed decomposition
and the value . In general, we recall that directed evaluation is a nondeterministic process: the output may depend on internal splitting choices that are made during the directed zerotests and inversions.
We are now in a position to state a central algorithm of this paper, which performs panoramic evaluations using directed ones. The presentation is abstract, in terms of general parametric frameworks. In the remainder of the paper we will study various concrete instantiations of this algorithm.
Algorithm
Perform the directed evaluation of at for the trivial directed splitting of . Let be the directed splitting obtained in return.
Compute the projections for .
Recursively call the algorithm in order to evaluate at for .
Return the union of and of the panoramic values obtained in step 3.
Proof. Algorithm 3.2 finishes because the dimension of the input algebra strictly decreases throughout the recursive calls in step 3. At the end of step 1, we have
by Proposition 3.9. Therefore the singleton is a panoramic value of at . We conclude by observing that the union of the panoramic values in step 4 gives a panoramic value of at .
Example
In the next sections, Algorithm 3.2 will lead to nearly linear asymptotic complexity bounds. So it turns out to be much faster than the naive Algorithm 3.1.
From the complexity point of view, there is still a problem with naive implementations of Algorithm 3.2: the directed evaluation approach involves many potentially expensive conversions of the form . A general solution to this issue is tedious, but for the concrete parametric frameworks from this paper, a natural and efficient approach is to delay these conversions. This can be done at the price of performing arithmetic operations in the input algebra instead of its subalgebras.
Consider for instance the simplest case when , where is monic of degree . Let with and . Then one projection corresponds to a division by of cost , when using the standard representation for elements in . This means that additions during the directed evaluation of a tree typically admit a cost instead of , which is suboptimal.
A natural remedy to this problem is to delay reductions modulo . This strategy naturally fits in our setting of parametric frameworks through the use of redundant representations. More precisely, consider some within a parametric framework . Then we may define a new parametric framework whose objects are algebras in , with this modification that elements in are represented by elements in . Given , this means that projections are free of charge in . The arithmetic operations on elements in are also performed in . The inversion of is done by computing the projection in and then performing the inversion of in ; zerotests are done similarly.
Of course, at the end of a directed evaluation that uses this kind of redundant representations, we may wish to convert the result to the standard representation in . This can be done by inserting a “finalization” step at the end of step 1 in Algorithm 3.2.
This section is devoted to parametric algebras with one algebraic parameter, that is of the form with monic, separable, and of degree . For our complexity bounds it will be convenient to assume that .
Elements in are represented as remainder classes with . The univariate situation is notably simple but already useful enough to be treated separately. We first specify the parametric framework and the directed operations, then we analyze the cost of our fast panoramic evaluation, and finally we compare the directed evaluation strategy with dynamic evaluation.
For efficiency reasons, we further assume that is explicitly separable in the sense that we are given the cofactors and in in the Bézout relation
(4.1) 
Computing this relation from the outset only requires operations in .
Let us first specify the required operations of the univariate parametric framework, and explicit their complexity.
Additions and subtractions in can clearly be performed by straightline programs in linear time , whereas multiplications take time .
An element with is invertible if and only if and, in that case, can be computed using the extended Euclidean algorithm. Both the test and the computation of can be done by a computation tree in time .
We have if and only if . In that case, we have
for all ; the projection can thus be computed in time by a straightline program (below, we will rather use the technique of delayed reductions from section 3.7 in order to avoid projections altogether).
Given with , we have and . This shows that we can compute and in time using computation trees.
Given univariate parametric algebras , we have if, and only if, (which in particular implies that the are pairwise coprime). In this case, the isomorphism in both directions can be computed in time by using the usual subproduct tree of ; see [21, chapter 10].
The directed zerotests and inversions of an element with preimage can be computed in time , as follows:
The input is a directed splitting of ,
the Bézout relation , and an element with preimage .
If , then the result of the zerotest is and the result of the inversion is . The input directed splitting is left unchanged in return.
We compute along with the Bézout relation .
If , then the result of the zerotest is and the result of the inversion is . The input directed splitting is left unchanged in return.
We compute , and deduce the Bézout relations of and , and of and by Lemma 4.1 below.
If then we return the directed splitting
and the image of is zero in . The value of the zerotest is , and the result of the inversion is .
Otherwise, we return the splitting
and the image of in is invertible. The latter splitting is actually directed since implies
The inverse of modulo is obtained as follows:
From the Bézout relation , we deduce and then
From the Bézout relation , we get , and deduce
The value of the zerotest is , and the result of the inversion is .
Proof. From the given Bézout relation we have
Since and are strictly less than , reduction of this relation modulo yields
This is the desired Bézout relation for and . By symmetry, the Bézout relation for and can be computed in a similar way.
We are now ready to present the main complexity bound of this section, in terms of the detailed cost functions defined in section 2.1.
Proof. We use Algorithm 3.2 in combination with the technique of delayed reductions from section 3.7. This means that we avoid the cost of conversions, at the price of performing arithmetic operations in subalgebras of in the algebra itself. More precisely, the cost of one directed evaluation of over is the sum of:
operations in for the nodes of type in ,
operations in for the nodes of type ,
operations in for the directed zerotests and inversions,
operations in for the reductions in the output nodes of the path (i.e. the “finalization step” in the terminology of section 3.7).
Let denote the directed splitting obtained at the end of step 1 of Algorithm 3.2. By construction we have
(4.2) 
Fast multiremaindering yields for in time .
We recursively perform the panoramic evaluation of over for all . Let be the cost of one panoramic evaluation of over an algebra of dimension . We have shown that there exists a universal constant such that
In view of (4.2), we conclude by unrolling this inequality at most times.
Example
which is better than the bound of Theorem 4.2 because the recursive depth is constant. If is a field, then we notice that the usual evaluation of at takes time .
Let us now compare the complexity of directed evaluation with the
complexities of other known strategies for a single algebraic parameter.
One problem with dynamic evaluation is that it is hard to implement in
common programming languages without support for parallelism or high
level control structures such as continuations [41]. Early
implementations of dynamic evaluation were therefore naive [13],
the computation tree essentially being reevaluated for every separate
case: see subsection 4.3.1. In theory, using
When using common sequential programming languages, dynamic evaluation is usually implemented as in Algorithm 3.2, with the following two differences:
Splittings are no longer required to be directed. In other words, at each evaluation step, the splitting at node of is not required to satisfy the conditions for . Consequently, whenever a zerotest or an inversion triggers a basic splitting for , the branch where we continue the evaluation is chosen nondeterministically. Such evaluations are said to be undirected in the sequel.
Once the undirected evaluation is finished, the projections in step 2 of Algorithm 3.2 are done naively, without fast multiremaindering.
Adapting the cost analysis of Algorithm 3.2 leads to the complexity bound
for naive dynamic evaluation, in terms of the detailed cost functions from section 2.1.
The following example shows that the latter bound is tight in the worst case. So naive dynamic evaluation is in general less efficient than fast panoramic evaluation.
Example
with , , , and let represent the image of in . Then a possible undirected evaluation of at performs a basic splitting and may select the splitting
So a possible undirected evaluation ends with the latter splitting and output value .
A possible undirected evaluation of over may end with the splitting
and value . Doing so for the rest of the evaluation, we may end with the splitting
and the value of over is for , and is over .
The total execution time therefore grows at least with
In comparison to Example 4.3, it turns out that dynamic evaluation may be significantly more expensive than our fast panoramic evaluation based on directed evaluation.
The naive implementation of dynamic evaluation using reevaluations of
the entire computation tree leads to unnecessary recomputations.
Highlevel control structures such as continuations can be used to avoid
this kind of recomputations, by resuming the
“recomputations” directly at the points where the splittings
occurred.
Example
with , , , and let represent the image of in . The dynamic evaluation of at encounters an inconsistency when testing whether is zero modulo . One computation continues modulo , and so returns . Another computation continues modulo and a new splitting occurs at the zerotest of , etc. Counting common computations in branches only once, the total cost of the dynamic evaluation is bounded by
which is essentially the same as the cost of our fast panoramic evaluation: see Example 4.3.
Example
for from to do
for from to do
if then
Evaluated over a field of degree over , the total cost of this program is
Let us now take with as in the previous example. The dynamic evaluation executes the first loop sequentially with cost . Then it splits at the zerotest of modulo into two branches:
The first branch works modulo and reduces modulo , which amounts to operations in . The costs of the zerotests totalize .
The second branch works modulo . The zerotest of costs , and causes the present branch to split up into two:
The first branch works modulo and reduces , which amounts to operations in . The costs of the zerotests totalize .
The second branch works modulo . The zerotest of costs , and causes the present branch to split up into two:
...
The total execution time when using dynamic evaluation is therefore
By Theorem 4.2, the execution time using fast panoramic evaluation is , which is significantly better.
The bottleneck of dynamic evaluation in Example 4.6 lies in the projections of data from parametric algebras into smaller ones. It is important to stress that this cost cannot be reduced by executing the branches in a judicious order or by cleverly factoring out common computations in branches using continuations. In our fast panoramic evaluation strategy, this means that fast multiremaindering plays a crucial role for handling these projections efficiently.
The above examples show that determining the best strategy to compute a panoramic value for a given computation tree is not an obvious problem: in fact, computation trees and more general programs may often be modified in order to accelerate panoramic evaluations. Let us now consider some specific computation trees for which efficient ad hoc strategies have been developed.
For panoramic quasiinversions in , gcds, and coprime factorizations in , Dahan et al. have designed fast algorithms in [8]. For instance, they showed that a panoramic gcd in could be obtained with cost
(4.3) 
see [8, Propositions 2.4 and 4.1]. For this problem, Theorem 4.2 leads to the cost
which can be further improved to
(4.4) 
by using Kronecker substitution to multiply polynomials in in time . If , then this cost is slightly higher than (4.3). If , then (4.4) simplifies to , which is slightly better than (4.3).
Another interesting application of fast panoramic evaluation is the computation of the determinant of an matrix with entries in . By means of Berkowitz' algorithm [1], such a determinant can be computed in time by a straightline program over , and therefore in time by a straightline program over . This cost decreases to by using [32], and even to thanks to [40], whenever inverses of are given in . We can do better with fast panoramic evaluation: we first compute the determinant of over as if it were a field by means of a computation tree that performs ring operations and zerotests and inversions. By Theorem 4.2, this can be done in time
The resulting panoramic value corresponds to the residues of modulo several pairwise coprime factors of . Chinese remaindering finally allows to recover the actual determinant using additional operations in .
From now on, we focus on the complexity of panoramic evaluation for algebraic parameters. Such a tower is determined by algebraic parameters , where is constrained by an algebraic equation over , and is constrained by an algebraic equation over for . It is natural to apply the method of the previous sections for a single algebraic parameter in a recursive manner. But we have to face a new difficulty: directed zerotests and inversions in will involve occasional zerotests and inversions in , that will themselves involve occasional zerotests and inversions in , etc. The first goal of this section is to design data structures that allow us to create new branches for free during directed evaluation. At the end of a directed evaluation we need to project input data into all the residual branches. Performing this task efficiently is the second goal of this section.
Data structures play an important role in the efficiency of operations required by parametric frameworks. From now on, we will only manipulate parametric algebras whose defining ideals are generated by triangular sets. In other words, all parametric algebras will be represented by towers over , as in section 1.1. We recall that is also assumed to be absolutely reduced, i.e. is a product of fields. A tower is said to be explicitly separable if we are given the Bézout cofactors and in such that for .
In order to simplify complexity analyses, it is convenient to assume that for . In particular, . This restriction is harmless, since extensions of degree can be discarded without cost in our model of computation trees.
Consider two towers and of the same height and over the same base field . Let and denote the respective defining polynomials of and . We say that is a factor tower of if . This is the case if, and only if, there exist natural projections (that naturally extend to projections in a coefficientwise manner) such that:
divides for .
sends an element represented by to
for .
Given such a factor tower of , let us now study the cost of computing one projection of an element with .
Proof. Let denote the preimage of . We recursively apply to the coefficients of . Then, the reduction of modulo can be done by a straightline program with cost . Consequently, there exists a universal constant such that the cost function satisfies
Unrolling this inequality yields
Taking from Proposition 2.6 concludes the proof.
In the sequel, “()” stands for the empty tuple. Let be an index set of tuples of integers with and
for integers and . Consider a family of factor towers of with the property that only depends on for all , and write . We say that such a family of factors forms a tree factorization of if the variety is partitioned into . If and , then we also write for the defining polynomial of over .
It is convenient to represent such a factorization by a labeled tree (see Figure 5.1): the nodes are identified with the index set made of the disjoint union
and each node is labeled with the algebraic extension . The parent of the node with is simply the node . Each individual factor tower corresponds to a path from the root to a leaf.

Given and , let
Projecting the equality
on the first coordinates, we observe that the projection of , for given , is the same for all , and equals , whence . Consequently, forms a tree factorization of the subtower . From an algebraic point of view, this means that we have a natural isomorphism
Consider an explicitly separable tower and assume that we wish to compute the panoramic evaluation of a computation tree over . In order to apply Algorithm 3.2 in this multivariate context, it would be natural to specify a parametric framework, as we did in the univariate case. Since we will exclusively focus on the directed approach from now on, it turns out to be simpler and more straightforward to describe how to perform the directed evaluation and the projections in steps 1 and 2.
Informally speaking, we want to work with respect to a “factor of highest possible degree” at each level. This leads to special types of factorizations of that are illustrated in Figure 5.2. The actual computations are done in the algebras from the highlighted “principal branch”. The other “residual branches” are dealt with at a next iteration, when the computations for the principal branch have been completed.
Technically speaking, a directed splitting of will be represented by:
Vectors of polynomials in for , such that:
are the defining polynomials of a factor tower of , written , and called the principal branch.
for .
for .
The Bézout relations for , so is explicitly separable.
Each factor with and gives rise to a factor tower of , called a residual branch, as follows:
For , we set .
.
For , the defining polynomial of over is the projection of in .
The tree representation of the resulting tower factorization is summarized in Figure 5.2

Notice that a directed splitting of naturally induces a directed splitting of for .
Now the main idea is to perform the directed evaluation of a computation tree over the principal branch, while postponing evaluations over residual branches. With our representation of directed splittings, creating residual branches is essentially free of charge; the corresponding tower representations will only be computed after completion of the entire directed evaluation, using a dedicated “finalization” procedure that will be detailed in subsection 5.4 below. After the finalization, the projections from step 2 of Algorithm 3.2 can be carried out efficiently using fast multiremaindering; this will also be detailed in subsection 5.4.
As in the univariate case, it will be convenient to use the technique of delayed reductions from section 3.7. During directed evaluations, this means that we will actually represent elements in by elements in , for . In particular, operations in on elements in are really performed in . We recall that the main benefit of this representation is that projections are free of charge, whenever is a quotient algebra of whose elements are also redundantly represented by elements in . On the other hand, we will resort to reduced representations for directed zerotests and inversions. We will also systematically reduce all output values. Notice that elements in can be reduced in time , by Lemma 5.1. In the next subsection, we detail how to perform directed zerotests and inversions.
In the case of a single algebraic parameter, we reduced directed inversions and zerotests to extended gcd computations. With several parameters the problem is more difficult. An element in can be tested to be invertible by means of a straightline program over with polynomial cost and a single zerotest in : it essentially suffices to compute the determinant of the multiplication endomorphism using Berkowitz' algorithm [1]. But the cost of this approach is more than quadratic in , which is not satisfactory for our purposes. We now develop a more efficient strategy based on recursive calls of the fast extended gcd algorithm in the directed evaluation model. Note that a similar recursive approach was previously used in the contexts of dynamic evaluation and triangular sets.
Algorithm
If then return if or and otherwise. The directed splitting is left unchanged.
Let denote the preimage of , and recursively compute the extended monic gcd of and over , using directed evaluation of a computation tree for extended gcds, with the directed splitting as extra input.
Let denote the directed splitting obtained in return and let be its principal branch. This extended gcd computation also returns a monic polynomial , and in such that the following Bézout relation holds:
where and .
Let , , so that the Bézout relation holds. Compute the reduced representations of , , , , and (recall that we delayed these reductions), and then compute .
Deduce the Bézout relations and by means of Lemma 4.1.
If , then return true, along with the directed splitting , and the Bézout relation of and .
If , then return , along with the directed splitting , the Bézout relation of and , and the inverse of .
If , then return true, along with the directed splitting , and the Bézout relation of and .
Otherwise return , along with the directed splitting , the Bézout relation of and , and the inverse of .
Until the end of the paper, represents a cost function for products in where can be any principal branch encountered in the directed evaluation model over . Similarly, is a cost function for reduced projections from onto .
Proof. In step 8 we have
Consequently, the splittings in the output are directed.
In step 8, the fact that taken modulo is the inverse of in can be verified as follows, in the same way as in section 4.1: the Bézout relation of and gives
which simplifies to thanks to the Bézout relation of and . The correctness of the algorithm follows from these facts.
Let us write for the cost of the algorithm and recall that we are using the technique of delayed reductions. If then the cost is . Otherwise, when computing extended gcds using the fast algorithm from section 2.2, step 2 takes
operations in . Step 3 takes
further operations, and step 4 takes time , by Lemma 4.1. Steps 5 to 8 require operations in . Therefore, there exists a universal constant such that
(5.1) 
Proposition 2.6 and Lemma 5.1 lead to
whence
for some universal constant . Using , unrolling the latter inequality leads to the claimed bound.
As before, let be a separable tower with defining polynomials of degree , and denote . At the end of a directed evaluation, we must recover the tower factorization of induced by the directed splitting. We also need efficient routines for projecting onto the residual branches. In this subsection (and contrary to the previous subsection), we use the traditional, reduced representations for elements in towers. We begin with a simple lemma for the computation of Bézout relations induced by factorizations.
Proof. We first construct the subproduct tree of : it is defined as the set of the subproducts defined recursively as follows:
the set if ,
the union of and the subproduct trees of and if .
A straightforward divide and conquer algorithm computes all these products with arithmetic operations in . Then, we obtain the Bézout relations for and , as well as for and , in time by Lemma 4.1. We conclude using a standard induction argument on .
(5.2) 
can be computed by straightline programs in time .
Proof. Until the end of the proof, it is convenient to keep in mind the notation for the tree factorization in Figure 5.2, associated to the directed splitting. Let us write:
for the cost of the construction of the explicitly separable factor towers ;
for the cost of the projections from onto ;
for the cost of the Chinese remaindering from to .
The proof is done by induction on . If , then the costs are zero. Assume now that and that explicitly separable factor towers have been constructed. We thus have the following effective isomorphism:
(5.3) 
We first project and the Bézout relation onto and for and with cost . This already gives the defining polynomials of over for and , along with the corresponding Bézout relations.
We next project the polynomials of onto , namely
for , using operations in (it is better to use Lemma 5.1 in practice, but the present bound in terms of simplifies the analysis). From the already computed , , and , we deduce the Bézout relations of and for , by means of Lemma 5.3, in time
Overall, we have shown that there exists a universal constant such that
(5.4) 
Let us now consider both directions of the isomorphism (5.2). Given , represented by , we compute the projections and for and in time ; this already gives the images of into for and . Then we compute the remainders of modulo for in time . So there exists a universal constant such that
(5.5) 
For the inverse Chinese remaindering problem, we are given for , and for and . We have at our disposal , its factors for and the Bézout relation of and . We compute whose projections are for :
where (resp. ) is the canonical preimage of (resp. of ). Computing costs ; see [21, Algorithm 10.20]. The isomorphism (5.3) gives rise to the isomorphism
whose righthand side contains . We finally recover the image of this element in using recursive calls of the Chinese remaindering procedure. Altogether, this shows that there exists a universal constant with
(5.6) 
By using , inequality (5.5) and Proposition 2.6 give
Similarly, the relation (5.6) yields
Finally, using our assumption that for , the relation (5.4) implies that
We are now ready the state the main complexity result of this section, in terms of the detailed cost functions from section 2.1.
operations in . Let denote the panoramic splitting in the output. By means of the auxiliary data, conversions in both directions of this isomorphism can be done in time
Proof. The proof is very similar to the one of Theorem 4.2. It is straightforward whenever , so we may freely assume from now. Using redundant representations, the directed evaluation of over requires the following operations in :
operations for the nodes of type in ;
operations for the nodes of type by Proposition 2.6;
operations for the zerotests and inversions by Lemma 5.1 and Proposition 5.2;
operations to reduce the output values, by Lemma 5.1.
At the end, the directed splitting
(5.7) 
satisfies
(5.8) 
The finalization of the residual branches takes time by Lemma 5.4. In step 2 of Algorithm 3.2 we next project the input values of the computation tree to the residual algebras . Again by Lemma 5.4, this can be done in time .
Let denote the cost function of one panoramic evaluation of over a algebra of dimension . The above analysis shows the existence of a universal constant such that
Unrolling this inequality at most times, thanks to (5.8), the claimed complexity bound follows.
Let denote the cost function of the conversions between and a panoramic splitting. Using Lemma 5.4, we conclude that
again thanks to (5.8).
Before we turn to the more technical topic of speeding up computations in algebraic towers, we will have a short intermezzo to study primitive tower representations. The results in this section are essentially from [29, sections 3 and 4.1], although we slightly generalized some of them.
Let be an effective algebra and let be an effective algebra with an explicit basis . We start by recalling the main result from [29, section 3] about the evaluation of multivariate polynomials in at points in . This can in particular be used in order to compute socalled multivariate modular compositions. We recall that (resp. ) stands for the number of operations in that are required in order to multiply two elements in (resp. ). We assume that and that additions and subtractions in and are cheaper than multiplications.
Proof. The case corresponds to the classical babystep giantstep algorithm over ; see for instance [29, section 3.1]. The extension to has been designed in [29, section 3.2]. In [29, Proposition 3.3], the assumption for was used for convenience. But if for some , then does not actually appear in , so discarding is free of charge in the straightline program model.
The main difference with [29] is that we will consider primitive tower over arbitrary algebras instead of fields . In the same spirit as section 3.7, it will also be convenient to also integrate the idea of redundant representations in the definition.
From an abstract point of view, given two effective algebras and , we may use elements in as a redundant representation of elements in if we have an effective monomorphism and an effective epimorphism such that is the identity map of . We say that is an effective retract of . We will write (resp. ) for the cost of one evaluation of (resp. ), and we denote .
From now on, assume that is a product of separable field extensions of . We consider a tower of ring extensions of of the form , for , where is monic and explicitly separable of degree , which means that there exists a Bézout relation with and .
A algebra , a monomorphism , and an epimorphism such that is the identity map of .
Linear forms , with , for .
Monic polynomials , for .
, for and .
Writing for the class of in , we assume that these data satisfy the following properties for :
The following map is a monomorphism that extends :
(j=1,…,i). 
The following map with is an epimorphism that extends :
The polynomials are called parametrizations of the in terms of .
In terms of the triangular set defined in the introduction ( is the canonical preimage of in ), an element is represented by a polynomial with defined modulo . So the conversion of into an element of boils down to evaluating modulo . The backward conversion consists in evaluating the image of a univariate polynomial at . Both conversions are instances of multivariate modular composition problems that can be handled using Proposition 6.1. More precisely:
in terms of the defining polynomials , and
in terms of , and ,
such that is a primitive tower representation of . Then there exist straightline programs for which:
One evaluation of takes operations in .
The images for can be computed in time .
Given , one evaluation of can be done in time .
Proof. If for some , then does not actually appear in polynomial representations of elements in . In the computation tree model, we may suppress such indices without any cost. Without loss of generality we may therefore assume that for .
The complexity bound for one evaluation of with is a direct consequence of Proposition 6.1. Indeed, any element in can be represented as , where admits partial degrees in for . Now we can compute
in time
since . We finally notice that .
For , the map extends coefficientwise into a map . We may thus convert the minimal polynomial of over into a polynomial in time
The computation of then takes time , since for all .
In order to evaluate at an element in , we write for the preimage of and observe that
We first evaluate at in , where is the class of in . This takes time by Proposition 6.1 and yields a polynomial
Since , we next notice that
so the preimage in of is
which represents . Applying to the coefficients of costs . Altogether, this proves the existence of a universal constant such that
We conclude that can be computed in time
using our assumptions that and for all .
If is a field, then our definition of primitive tower representations of coincides with the one from [29, Definition 2.2] whenever the are isomorphisms. If has sufficiently many elements, then primitive tower representations can for instance be computed using linear algebra techniques. For efficiency reasons, we will rely on the following result that will be used later with a parametric algebra in the role of .
Remark
Remark
It is more subtle to regard such randomized variants as computation trees. Although our model does not provide us with an instruction to compute random elements of , nothing prevents us from adding a “pool” of randomly chosen elements to the input. Since randomized Las Vegas algorithms may fail for certain random choices, they typically loop until suitable random numbers are drawn. As explained in Remark 2.4, we may unroll such loops to make things fit in our computation tree model, provided that we can provide an absolute bound on the number of times that we have to run the loop.
In our case, we claim that such an absolute bound indeed exists. Following [29, Proposition 4.3], the construction is incremental in : assuming that already found, one has to pick up a suitable value of in a subset of of cardinality . The crucial observation is that a suitable value always exists in this finite set , which bounds the number of random values that have to be tried. With ordered randomly, can actually be found with a bounded number of expected trials.
The suboptimality of the complexity bounds of Proposition 2.6 and of Theorem 5.5 appears when the height of the tower gets as large as . In fact, if indeed gets large, then many of the are necessarily small. This section is devoted to an efficient solution to this issue.
Roughly speaking, as long as a product of the form is reasonably small, it turns out to be favorable to change the given representation of into a more efficient primitive element representation . Repeating this operation as many times as necessary, the original tower is replaced by an equivalent tower of much smaller height and for which all defining polynomials have sufficiently large degree. Such accelerated towers were first introduced in [29, section 4.2]. In this section, we will further refine the concept and show how to use such towers in combination with directed evaluation.
Throughout this section, we carry on with the notations from section 5: the original tower is explicitly defined by , and we set for . As before, we will assume that and that for . In particular, .
In this subsection, we extend the notion of accelerated towers from [29, Definition 4.6] and study accelerated tower arithmetic. Let be an explicitly separable tower over , and let . By [29, Lemma 4.7], there exists a sequence of integers of length
(7.1) 
such that, for , we have
(7.2) 
where . In this subsection, we assume that the sequence has been fixed once and for all.
A tower over with defining polynomials , where and , and such that . The class of in is written .
For :
Linear forms , over for ,
for ,
for and .
Let represent the identity map of . For and , these data induce the following monomorphism that extends :
Let represent the identity map of . For and , we also have the following epimorphism that extends :
and . For and , we finally have
In particular, for , we have the following monomorphism that extends :
(l=i_{j1}+1,…,i_{j})  
When , taking
(7.3) 
yields
(7.4) 
whence products in can be computed using operations in , by Proposition 2.6. The worst case complexity of products in is therefore asymptotically better than in , which explains the terminology of “accelerated towers”. Let us now detail how to make elements in benefit from this acceleration.
Proof. The proof is adapted from [29, Lemma 4.8]. Let us first assume that the polynomials have been precomputed for . By Proposition 2.6,
and by Proposition 6.3, we have
If then such conversions simply cost . Otherwise we have by (7.2). In both cases it follows that
for some universal constant . Unrolling the latter inequality, while taking into account that and , we deduce that
Finally, given , we notice that can be precomputed in time
by Proposition 6.3, again by distinguishing the cases when and . Consequently, the total cost of these precomputations is
Two polynomials of degree over can be multiplied by a straightline program in time
Proof. One conversion between and can be done in time
by Lemma 7.2. On the other hand, Proposition 2.6 yields
Combining the latter costs, we deduce that
and the claimed cost for corresponds to setting .
One motivation behind Definitions 6.2 and 7.1 with respect to the less general definitions from [29] is to allow for redundant representations. More precisely, we have:
Proof. A tower factor of is a tower factor of . Let denote the canonical monomorphism, then is the identity of . Setting we first verify by induction that
(l=i_{j1}+1,…,k) 
is a monomorphism that extends , for , since .
Let represent the identity map of . For , we define the epimorphism , for , that extends :
Since , we deduce that is the identity map of .
Notice that the tower is only used for the acceleration of arithmetic operations; we do not require this tower to be separable. Writing for the image of in , our use of redundant representations also makes it unnecessary to determine the defining polynomials of . Although the computation of such defining polynomials along with Bézout relations for and would be convenient for speeding up multiplications in , it seems tricky to integrate these computations without deteriorating the overall complexity.
When reduced, nonredundant representations are explicitly required, we will resort to the following counterpart of Lemma 5.1:
Proof. The proof is the same as in Lemma 5.1, except that we appeal to accelerated products of Proposition 7.3. The cost of the projection thus becomes:
Now (7.2) yields for . Since , the claimed bound follows.
Proof. We adapt the proof of Proposition 5.2 so as to benefit from accelerated products. From Proposition 7.3 we have
and from Lemma 7.5 we have
Consequently, the relation (5.1) implies the existence of a universal constant such that
Unrolling the latter inequality yields
while using the facts that and for , by (7.2).
Now that we have seen how to benefit from accelerated arithmetic, let us show how to construct accelerated towers. We use induction on the height, in combination with Proposition 6.4. As usual, all computations in the directed model are done using the redundant representation.
Algorithm
If , then return for the directed splitting along with an empty accelerated tower.
Otherwise ; recursively apply the algorithm to and . Let represent the directed splitting of obtained in return, let be the principal branch, and let be the accelerated tower for .
Let be the canonical preimages of in ; compute ,...,.
By directed evaluation over , starting with the directed splitting , compute a primitive tower representation for
seen as a tower over . Let be the directed splitting obtained in return and let be the corresponding principal branch.
The latter primitive tower representation consists of
Linear forms , over for ;
for ;
for and .
With the notation of Definition 7.1, we set for and for and .
Compute and define, for :
as regarded as in ;
.
Complete the tower with the projections of the Bézout relations of and over ; return this tower, , and .
for a sequence of integers that satisfies
Proof. When entering step 5, is an accelerated tower for ; so we have the monomorphism and the epimorphism such that is the identity map of . We also have the following isomorphisms for :
(l=i_{s1}+1,…,k). 
So the following morphisms are monomorphisms that extend , for :
(l=i_{s1}+1,…,k)  
Let
We observe that extends and that
This completes the correctness proof.
Step 1 takes constant time. In step 3, the projections require
operations in , by Lemma 7.5 and the assumption that for . Notice that the same bound holds for step 6.
Step 4 is free of charge whenever . Otherwise holds by (7.2). During the directed evaluation of step 4, one product in the principal branch takes operations in by Proposition 7.3, and the cost of one directed inversion is by Proposition 7.6. Consequently, step 4 runs in time
In step 5 we perform
conversions from to , which amount to
operations in by Lemma 7.2.
Remark
which still dominates the cost of Algorithm 7.1 in this model.
In order to make the directed evaluation benefit from accelerated arithmetic, we first need to compute an accelerated tower. Then, the complexity of accelerated divisions is simply adjusted according to Proposition 7.6. At the end of a computation, we use the following lemma in order to recover the extended tower factorization of , as in section 5.4.
can be computed by straightline programs in time .
Proof. The algorithm is the same as in the proof of Lemma 5.4. We only revisit the cost analysis when using accelerated arithmetic. Given , Proposition 7.3 yields
whence
In a similar way, we obtain
We deduce that
(7.5) 
operations in , where
Let denote the panoramic splitting in the output. Using the auxiliary data, conversions in both directions of this isomorphism can be done in time
(7.6) 
Proof. The directed construction of the accelerated tower takes
by Proposition 7.7. Then, we project the input values onto the principal branch, written , in time
by Lemma 7.5.
Let us summarize the number of operations in needed for the directed evaluation of over , when using redundant representations in :
operations for the nodes of type in ;
for the nodes of type by Proposition 7.3;
for the zerotests and inversions by Lemma 7.5 and Proposition 7.6,
for the reduction of the output node by Lemma 7.5.
At the end, the directed splitting satisfies
(7.7) 
Then we finalize the construction of the residual branches in time
by Proposition 7.9. We next project the input values of the tree to , in time
again by Proposition 7.9. We finally perform the recursive panoramic evaluations of over . The conclusion follows as in the proof of Theorem 5.5.
Remark
The following corollary simplifies the above complexity bound by tuning the parameters and .
and the bound
Proof. It suffices to take and as in (7.3) and (7.4).
The main result of [29] is an efficient algorithm to compute products in separable towers of field extensions. We are now in a position to generalize this result to arbitrary explicitly separable towers.
Proof. Let . Applying Theorem 7.10 to the straightline program that simply computes a product, it requires operations to compute a parametric splitting , auxiliary data, and the projections for . Using the auxiliary data, the theorem also allows us to recover from these projections using further operations in . We conclude by taking and as in (7.3) and (7.4).
Another interesting application that has already been discussed in the univariate case at the end of section 4.3.3 is the computation of determinants.
Proof. The proof is similar to the one of Theorem 7.13.
As a final application, we revisit the matrix product. We need the following generalization of Proposition 2.6:
Proof. We proceed along the same lines as Lebreton in [35, section 3.2]; see also [29, Proposition 2.4]. Consider two matrices , we first lift the matrices to multivariate polynomial matrices in of partial degrees in for . We next convert these polynomial matrices to univariate polynomial matrices in using Kronecker substitution. We multiply these matrices using the evaluationinterpolation technique. Since admits at least distinct evaluation points, this can be done in time
We finally convert the result back to a polynomial matrix in , and we reduce each of the entries with respect to using the algorithm from [35, section 3.2]. This reduction step requires operations in . Now our assumption that for implies and . Using (2.1), we conclude that the total running time is bounded by
The next proposition is an example of a less straightforward use of Theorem 7.10, where we adapt the “acceleration factor” to a specific situation, and where we wish to decrease the overhead of the accelerated towers. Precisely, after a directed construction of an accelerated tower, we obtain a directed splitting that satisfies
As a byproduct, we have an accelerated tower of of degree at our disposal. The recursive calls to panoramic evaluations over for then lead to accelerated towers for each of the factor towers of . Let denote a bound for the sum of the degrees of all these accelerated towers. Then we have
Unrolling this inequality leads to , which is slightly suboptimal.
We may modify the construction of accelerated towers, described in Algorithm 7.1, in order to ensure that the degree of the accelerated tower is never more than twice the degree of the principal branch before entering the evaluation of the given computation tree. This is done as follows. At the end of Algorithm 7.1, we compare the degree of the accelerated tower to the degree of the principal branch . If then we are done. Otherwise, the principal branch is much smaller than in the sense that , and we use again Algorithm 7.1 in order to compute an accelerated tower for . We carry on computing a new accelerated tower of the current principal branch in the directed model over until the condition is met. By Proposition 7.7, the total cost of this process is roughly only twice as much as a single run of Algorithm 7.1 since
With these modifications in hand, we ensure that .
Proof. Let be a fixed positive constant. Take and fix a sequence such that (7.1) and (7.2) hold. Since , we observe that .
We first compute a single product in with the sole purpose of retrieving a parametric splitting , together with the auxiliary data for efficient conversions in both directions, and an accelerated tower of height for each of degree , where ; as discussed above. This precomputation takes time
One conversion between and also take time
Now assume that we wish to multiply two matrices . This problem is reduced to multiplication problems of matrices in that we further transform into matrix multiplications within the corresponding accelerated towers, so Proposition 7.15, yields the cost
using the assumptions that and .
Remark
Theorems 4.2, 5.5, and 7.10 show that the approach of directed evaluation allows us to compute panoramic evaluations in a time that is arbitrarily close to linear. We chose the setting of computation trees as our complexity model and it is natural to ask whether similar results hold in the RAM and Turing models. This seems plausible, although technical difficulties are likely to arise in the efficient implementation of data structures on Turing machines; see [30] for a detailed analysis of a similar kind, but for another problem.
So far, we have only addressed the sequential complexity of dynamic evaluation and its variants. It would be interesting to conduct a similar study for parallel computation models. Dynamic evaluation lends itself particularly well to parallel execution: whenever we encounter a splitting, it is natural to continue the execution of each of the branches in parallel. In the directed setting, we a priori have to complete the execution of the principal branch in order to benefit from the fastest possible multiremaindering. Nevertheless, this is really a tradeoff, and we may very well start to execute some “sufficiently thick” residual branches before completion of the principal branch. It is also important to notice that these considerations have only minor impact from the worst case complexity perspective: the opportunities for this kind of parallelism only arise when splittings actually occur. In the absence of splittings, we have to evaluate the computation tree over the full original algebra , so the parallel complexity is really a function of the amount of parallelism that is available in and in the algebra operations of .
For our presentation, we restricted our towers to be explicitly separable from the outset. In practice, towers are usually constructed by adding new parameters that satisfy polynomial constraints one by one. These polynomials may contain multiple factors, in which case one is usually only interested in the radical ideal generated by the constraints. This more general setting can actually also be covered efficiently using our techniques. Indeed, consider an explicitly separable tower and a new algebraic parameter that is subjected to , where is a monic polynomial that is not necessarily separable. Then we simply compute the separable factorization as if were a field (here assuming that the characteristic is zero or sufficiently large), using panoramic evaluation, together with the requested Bézout relations for the nontrivial factors . Let be the resulting splitting. Then any pair with gives rise to a new explicitly separable tower of height , which allows us to recurse.
In this paper, we only considered parameters that satisfy polynomial constraints, but the technique of directed evaluation naturally generalizes to various other settings. For instance, for an integer parameter , one may compute in as if it were a field, so is partially factored during directed evaluations. This situation is commonly encountered in number theory and computer algebra, where the deterministic construction of “large” prime numbers is rather expensive. Another possible extension of our work concerns real algebraic numbers in the continuation of [16].
Another direction for future work concerns the particular kind of tower arithmetic that we build on. In this paper, we essentially combined the best previously known generic algorithms from Theorem 2.6 with the idea of accelerated towers. For specific base fields , such as finite fields or subfields of , other efficient algorithms have been proposed for tower arithmetic [9, 27, 28]. It should not be hard to combine directed evaluation with these algorithms in order to improve on Theorem 7.10 and Corollary 7.12 for such more specific base fields .
Finally, the recent new theoretical ideas from [2729] and this paper should also allow for more efficient software implementations, although it remains a major challenge to make this happen. In particular, this requires to implement all available approaches with care and determine the thresholds for which they are most efficient. For particularly long parametric evaluations, one may also wish to switch to towers of smaller height, and to spend more time on factoring the defining polynomials.
S. J. Berkowitz. On computing the determinant in small parallel time using a small number of processors. Inform. Process. Lett., 18:147–150, 1984.
L. Blum, F. Cucker, M. Shub, and S. Smale. Complexity and real computation. Springer Science & Business Media, 2012.
F. Boulier, F. Lemaire, and M. Moreno Maza. Well known theorems on triangular systems and the D5 principle. In J.G. Dumas, editor, Proceedings of Transgressive Computing 2006: a conference in honor of Jean Della Dora, pages 79–92. U. J. Fourier, Grenoble, France, 2006.
P. A. Broadbery, T. GómezDíaz, and S. M. Watt. On the implementation of dynamic evaluation. In A. H. M. Levelt, editor, Proceedings of the 1995 International Symposium on Symbolic and Algebraic Computation, ISSAC '95, pages 77–84, New York, NY, USA, 1995. ACM.
P. Bürgisser, M. Clausen, and M. A. Shokrollahi. Algebraic complexity theory, volume 315 of Grundlehren der Mathematischen Wissenschaften. SpringerVerlag, 1997.
D. G. Cantor and E. Kaltofen. On fast multiplication of polynomials over arbitrary algebras. Acta Inform., 28:693–701, 1991.
M. Coste, H. Lombardi, and M.F. Roy. Dynamical method in algebra: Effective nullstellensätze. Ann. Pure Appl. Logic, 111(3):203–256, 2001.
X. Dahan, M. Moreno Maza, É. Schost, and Yuzhen Xie. On the complexity of the D5 principle. In J.G. Dumas, editor, Proceedings of Transgressive Computing 2006: a conference in honor of Jean Della Dora, pages 149–168. U. J. Fourier, Grenoble, France, 2006.
L. De Feo, J. Doliskani, and É. Schost. Fast algorithms for adic towers over finite fields. In Proceedings of the 38th International Symposium on Symbolic and Algebraic Computation, ISSAC '13, pages 165–172, New York, NY, USA, 2013. ACM.
J. Della Dora, C. Dicrescenzo, and D. Duval. About a new method for computing in algebraic number fields. In B. F. Caviness, editor, EUROCAL '85: European Conference on Computer Algebra Linz, Austria, April 1–3 1985 Proceedings Vol. 2: Research Contributions, pages 289–290. Springer Berlin Heidelberg, 1985.
S. Dellière. Triangularisation de systèmes constructibles : application à l'évaluation dynamique. PhD thesis, Université de Limoges. Faculté des sciences et techniques, 1999.
S. Dellière. On the links between triangular sets and dynamic constructible closure. J. Pure Appl. Algebra, 163(1):49–68, 2001.
C. Dicrescenzo and D. Duval. Algebraic extensions and algebraic closure in Scratchpad II. In P. Gianni, editor, Symbolic and Algebraic Computation. International Symposium ISSAC '88. Rome, Italy, July 4–8, 1988. Proceedings, number 358 in Lect. Notes Comput. Sci., pages 440–446. SpringerVerlag, 1989.
D. Duval. Rational Puiseux expansions. Compos. Math., 70(2):119–154, 1989.
D. Duval. Algebraic numbers: An example of dynamic evaluation. J. Symbolic Comput., 18(5):429–445, 1994.
D. Duval and L. GonzálezVega. Dynamic evaluation and real closure. Math. Comput. Simul., 42(46):551–560, 1996.
D. Duval and J.C. Reynaud. Sketches and computation – II: dynamic evaluation and applications. Math. Structures Comput. Sci., 4(2):239–271, 1994.
H. Friedman. Algorithmic procedures, generalized Turing algorithms, and elementary recursion theory. In R. O. Gandy and C. M. E. Yates, editors, Logic Colloquium '69, volume 61 of Studies in Logic and the Foundations of Mathematics, pages 361–389. Elsevier, 1971.
A. Fröhlich and J. C. Shepherdson. On the factorisation of polynomials in a finite number of steps. Math. Z., 62:331–334, 1955.
A. Fröhlich and J. C. Shepherdson. Effective procedures in field theory. Philos. Trans. Roy. Soc. London. Ser. A., 248:407–432, 1956.
J. von zur Gathen and J. Gerhard. Modern computer algebra. Cambridge University Press, New York, 3rd edition, 2013.
T. GómezDaz. Examples of using dynamic constructible closure. Math. Comput. Simul., 42(46):375–383, 1996.
D. Harvey and J. van der Hoeven. Faster polynomial multiplication over finite fields using cyclotomic coefficient rings. J. Complexity, 54:101404, 2019.
D. Harvey, J. van der Hoeven, and G. Lecerf. Faster polynomial multiplication over finite fields. J. ACM, 63(6), 2017. Article 52.
J. van der Hoeven. Automatic asymptotics. PhD thesis, École polytechnique, Palaiseau, France, 1997.
J. van der Hoeven and G. Lecerf. Composition modulo powers of polynomials. In M. Blurr, editor, Proceedings of the 2017 ACM on International Symposium on Symbolic and Algebraic Computation, ISSAC '17, pages 445–452, New York, NY, USA, 2017. ACM.
J. van der Hoeven and G. Lecerf. Modular composition via complex roots. Technical report, CNRS & École polytechnique, 2017. http://hal.archivesouvertes.fr/hal01455731.
J. van der Hoeven and G. Lecerf. Modular composition via factorization. J. Complexity, 48:36–68, 2018.
J. van der Hoeven and G. Lecerf. Accelerated tower arithmetic. J. Complexity, 55:101402, 2019.
J. van der Hoeven and G. Lecerf. Fast multivariate multipoint evaluation revisited. J. Complexity, 56:101405, 2020.
Xiaohan Huang and V. Y. Pan. Fast rectangular matrix multiplication and applications. J. Complexity, 14(2):257–299, 1998.
E. Kaltofen. On computing determinants of matrices without divisions. In P. S. Wang, editor, ISSAC 92 International Symposium on Symbolic Algebraic Computation, ISSAC '92, pages 342–349, New York, NY, USA, 1992. ACM.
K. S. Kedlaya and C. Umans. Fast polynomial factorization and modular composition. SIAM J.Comput., 40(6):1767–1802, 2011.
F. Le Gall. Powers of tensors and fast matrix multiplication. In K. Nabeshima, editor, ISSAC'14: International Symposium on Symbolic and Algebraic Computation, pages 296–303, New York, NY, USA, 2014. ACM.
R. Lebreton. Relaxed Hensel lifting of triangular sets. J. Symbolic Comput., 68(2):230–258, 2015.
G. Lecerf. Computing the equidimensional decomposition of an algebraic closed set by means of lifting fibers. J. Complexity, 19(4):564–596, 2003.
G. Lecerf. On the complexity of the Lickteig–Roy subresultant algorithm. J. Symbolic Comput., 92:243–268, 2019.
A. Poteaux and É. Schost. Modular composition modulo triangular sets and applications. Comput. Complex., 22(3):463–516, 2013.
A. Poteaux and É. Schost. On the complexity of computing with zerodimensional triangular sets. J. Symbolic Comput., 50:110–138, 2013.
F. P. Preparata and D. V. Sarwate. An improved parallel processor bound in fast matrix inversion. Inform. Process. Lett., 7(3):148–150, 1978.
J. C. Reynolds. The discoveries of continuations. LISP and Symbolic Computation, 6:233–247, 1993.