Deflation-Assisted Hybrid NR/Laguerre Algorithm

• The paper introduces an implicit deflation technique that integrates approximated (tame) roots into the Newton-Raphson iteration without explicit quotient formation.
• It leverages logarithmic derivative corrections to maintain the original polynomial structure, enhancing numerical stability and computational efficiency.
• The framework extends to Laguerre-type methods and variable transformations, suggesting a hybrid approach that accelerates batched evaluations in polynomial root-finding.

A deflation-assisted hybrid Newton-Raphson/Laguerre algorithm is a polynomial root-finding scheme in which already approximated roots are incorporated into subsequent iterations through implicit correction terms rather than through explicit formation of a quotient polynomial. In the framework developed for Newton’s method in "Root-finding with Implicit Deflation" (Imbach et al., 2016), the polynomial is partitioned into tame roots, which have already been approximated, and wild roots, which remain unresolved. The resulting mechanism preserves the original polynomial representation, redirects the iteration toward the unknown factor, and suggests a direct extension to Laguerre-type corrections because those corrections are also built from logarithmic derivatives and related rational terms.

1. Problem setting and deflation paradigm

The underlying polynomial is written as

$$p(x)=\sum_{i=0}^d p_i x^i = p_d \prod_{i=1}^d (x-x_i), \qquad p_d\neq 0.$$

The setting of interest is the practical phase in which some roots have already been approximated well and further computation is directed only to the remaining roots. The paper characterizes the known roots as tame and the unresolved roots as wild. This situation arises when Newton iterations started from many initial points converge to most roots but miss a small set, when simultaneous methods such as Weierstrass or Ehrlich tame most roots quickly while a few difficult ones remain, and when subdivision methods isolate some root clusters well before the rest (Imbach et al., 2016).

If $m<d$ roots have already been approximated, the standard continuation strategy is deflation. If the number of wild roots is $w$, the deflated polynomial is

$$q(x)=\sum_{i=0}^w q_i x^i = p_d \prod_{j=1}^w (x-x_j), \qquad q_w\neq 0.$$

Here $x_1,\dots,x_w$ are wild, while $x_{w+1},\dots,x_d$ are tame.

The distinction between explicit and implicit deflation is central. In explicit deflation, one computes the coefficients $q_i$ of the quotient polynomial and then applies a root-finder to $q$. The paper emphasizes two drawbacks of this approach: loss of sparsity, since $q(x)$ generally is not sparse even when $p(x)$ is, and numerical instability or coefficient growth, especially when approximate roots are factored out. The example $m<d$0 is cited as a typical case in which explicit deflation may be problematic for large $m<d$1. Implicit deflation avoids these difficulties by keeping the original polynomial representation intact and modifying the iteration formulas so that the known roots are accounted for analytically rather than through coefficient recomputation.

2. Newton-Raphson with implicit deflation

The Newton ratio is

$m<d$2

and Newton’s iteration is

$m<d$3

The key identity used for implicit deflation is the logarithmic derivative formula

$m<d$4

After splitting the roots into wild and tame components through

$m<d$5

taking logarithmic derivatives yields

$m<d$6

and hence

$m<d$7

This is the core implicit deflation identity: the Newton correction for the deflated polynomial $m<d$8 is obtained from the Newton correction for the original polynomial $m<d$9 by subtracting the rational contributions of the tame roots (Imbach et al., 2016).

Algorithm 2 in the paper implements this identity directly. Given the polynomial $w$0, sufficiently close approximations to the tame roots $w$1, an initial approximation $w$2 to a wild root, a stopping criterion, and a black-box evaluator $w$3 for $w$4, the iteration computes

$w$5

then

$w$6

forms

$w$7

and updates

$w$8

The previously found roots enter only through the correction sum $w$9, so the iteration behaves as Newton’s method on the deflated factor $$q(x)=\sum_{i=0}^w q_i x^i = p_d \prod_{j=1}^w (x-x_j), \qquad q_w\neq 0.$$0 without ever constructing $$q(x)=\sum_{i=0}^w q_i x^i = p_d \prod_{j=1}^w (x-x_j), \qquad q_w\neq 0.$$1 explicitly.

The paper also records a scaled variant, suggested by Dario Bini for improved numerical stability,

$$q(x)=\sum_{i=0}^w q_i x^i = p_d \prod_{j=1}^w (x-x_j), \qquad q_w\neq 0.$$2

This is algebraically identical to $$q(x)=\sum_{i=0}^w q_i x^i = p_d \prod_{j=1}^w (x-x_j), \qquad q_w\neq 0.$$3, but can be numerically preferable when $$q(x)=\sum_{i=0}^w q_i x^i = p_d \prod_{j=1}^w (x-x_j), \qquad q_w\neq 0.$$4 and $$q(x)=\sum_{i=0}^w q_i x^i = p_d \prod_{j=1}^w (x-x_j), \qquad q_w\neq 0.$$5 are close or vary in scale.

3. Extension to Laguerre-type corrections and hybridization

The paper establishes that implicit deflation extends readily to root-finders involving $$q(x)=\sum_{i=0}^w q_i x^i = p_d \prod_{j=1}^w (x-x_j), \qquad q_w\neq 0.$$6, especially Ehrlich’s method. Ehrlich’s iteration is

$$q(x)=\sum_{i=0}^w q_i x^i = p_d \prod_{j=1}^w (x-x_j), \qquad q_w\neq 0.$$7

with

$$q(x)=\sum_{i=0}^w q_i x^i = p_d \prod_{j=1}^w (x-x_j), \qquad q_w\neq 0.$$8

otherwise

$$q(x)=\sum_{i=0}^w q_i x^i = p_d \prod_{j=1}^w (x-x_j), \qquad q_w\neq 0.$$9

If the tame roots are treated as already exact, $x_1,\dots,x_w$0 for $x_1,\dots,x_w$1, then

$x_1,\dots,x_w$2

This demonstrates that implicit subtraction of known-root contributions is not restricted to Newton-Raphson but transfers to other rational correction formulas built from logarithmic derivatives (Imbach et al., 2016).

Laguerre’s method is not discussed in the paper. However, the text states that Laguerre-type methods typically use

$x_1,\dots,x_w$3

or equivalent combinations of logarithmic derivatives. The same framework then strongly suggests the corresponding deflation identities:

$x_1,\dots,x_w$4

and, by differentiation,

$x_1,\dots,x_w$5

This suggests that a hybrid Newton-Raphson/Laguerre scheme can share a common implicit-deflation layer: Newton uses the corrected first logarithmic derivative, while Laguerre would use corrected first- and second-order logarithmic derivative terms. A plausible implication is that method switching could occur without changing the underlying representation of the polynomial, because both branches can be expressed through analytically corrected rational quantities rather than through coefficients of a deflated quotient.

The paper also states that it explicitly advocates concurrent use of several iterative root-finders on various transformed polynomials. This suggests that a Newton/Laguerre hybrid is architecturally aligned with the paper’s general strategy even though no explicit Laguerre formula is derived there.

4. Variable transformations, reversion, and root-squaring

A major theme of the framework is that one can apply iterative methods not only to the original polynomial but also to transformed polynomials, while still avoiding coefficient formation. The principal map is

$x_1,\dots,x_w$6

for complex $x_1,\dots,x_w$7 and $x_1,\dots,x_w$8, with root correspondence

$x_1,\dots,x_w$9

The Newton ratio transforms as

$x_{w+1},\dots,x_d$0

or equivalently

$x_{w+1},\dots,x_d$1

The authors suggest applying Newton to several such transformed polynomials and then mapping the recovered roots back via the inverse relation. Because implicit deflation does not require recomputation of transformed coefficients, this strategy remains algebraically light and numerically cleaner than explicit quotient formation (Imbach et al., 2016).

Reversion appears as the special case $x_{w+1},\dots,x_d$2, $x_{w+1},\dots,x_d$3, so that $x_{w+1},\dots,x_d$4. Then

$x_{w+1},\dots,x_d$5

and

$x_{w+1},\dots,x_d$6

The reversed polynomial also satisfies

$x_{w+1},\dots,x_d$7

The paper notes that reversion can change which roots are hard or easy: large roots become small, interior and exterior geometry is exchanged, and Newton basins can change.

The Dandelin–Lobachevsky–Gräffe root-squaring map is likewise treated through correction formulas rather than coefficient manipulation:

$x_{w+1},\dots,x_d$8

Its Newton ratio satisfies

$x_{w+1},\dots,x_d$9

The broader significance is methodological: transformed Newton corrections are evaluated through the original polynomial, so transformation, reversion, and deflation are all expressed at the level of correction formulas rather than coefficient arrays. This suggests that a hybrid Newton-Raphson/Laguerre implementation could inherit the same transformation philosophy if analogous Laguerre correction identities are derived.

5. Computational efficiency and fast batched evaluation

For Newton’s method, the paper records the Ostrowski local efficiency parameters

$q_i$0

It then argues that this per-root local measure ignores opportunities created by fast batched evaluation. Simultaneous or many-start root-finding requires repeated evaluation of $q_i$1, $q_i$2, and Cauchy-type sums such as

$q_i$3

which is essentially Trummer’s problem. The paper notes that exact superfast algorithms exist in $q_i$4 arithmetic operations for multipoint polynomial evaluation and related Cauchy summation problems, that classical exact superfast versions are numerically unstable for $q_i$5, and that numerically stable alternatives based on the Fast Multipole Method are available (Imbach et al., 2016).

These observations are directly relevant to deflation-assisted iterations. In Newton with implicit deflation, the costly ingredients are evaluations of $q_i$6 at many points and evaluations of correction sums

$q_i$7

for many iterates $q_i$8. Those are precisely the types of batched rational evaluations accelerated by FMM. The paper’s efficiency claim is that with superfast evaluation, the effective number of function evaluations per iteration per root drops to

$q_i$9

so local efficiency of simultaneous methods grows with $q$0. It states this for Weierstrass and Ehrlich, and similarly for Newton when initialized and applied simultaneously at $q$1 points.

For one implicitly deflated Newton step, the paper gives a modest cost accounting. Stage 1 requires one call to $q$2 to obtain $q$3. Stage 2 uses $q$4 divisions, $q$5 subtractions, and $q$6 additions. Stages 3 and 4 require two subtractions and one division. Thus, if only a few roots remain wild, the additional cost over ordinary Newton is mainly the sum over tame roots; if many wild iterates are processed together, that sum becomes a natural target for FMM acceleration.

A plausible implication for a Newton-Raphson/Laguerre hybrid is that the Laguerre branch would add second-order rational sums such as $q$7, which the paper identifies as having an FMM-like structure. Under that interpretation, hybridization increases algebraic richness without necessarily requiring a coefficient-based implementation.

6. Stability, limitations, and algorithmic significance

The practical preference for implicit deflation is grounded in numerical stability. Explicit division by factors corresponding to approximated roots can introduce coefficient perturbations that are amplified when the polynomial is sparse or structured, when root moduli vary greatly, when the degree is large, or when the tame roots are only approximate. Implicit deflation avoids building quotient coefficients and instead works with the original evaluator for $q$8, the current approximations to tame roots, and sums of simple rational terms $q$9. This preserves the original representation of $q(x)$0 and avoids replacing it by a potentially ill-scaled quotient (Imbach et al., 2016).

The approach nonetheless has limitations. It still depends on sufficiently accurate tame roots. If a current iterate $q(x)$1 approaches one of the supposedly known roots, terms such as $q(x)$2 can become very large. The paper therefore treats the assumption that tame roots can be “very quickly refined” as practically standard but nontrivial. The stopping criterion is not fixed; instead, the text points to MPSolve-related sources such as [7] and [12]. The paper is exploratory rather than a fully worked implementation study: it does not provide a full convergence analysis for implicitly deflated Newton with inexact tame roots, an explicit extension to Laguerre, comparative experiments benchmarking implicit versus explicit deflation, or detailed pseudocode for transformed or hybrid workflows. The provided text also reports no numerical experiments.

Within those limits, the algorithmic significance is clear. The paper’s exact contribution is the identity

$q(x)$3

together with its stabilized form

$q(x)$4

This gives Newton on the unknown factor $q(x)$5 without computing $q(x)$6. The same logic, combined with the paper’s treatment of Ehrlich corrections, variable mappings, reversion, root-squaring, and fast batched evaluation, suggests an extensible architecture for hybrid Newton-Raphson/Laguerre solvers: keep the original polynomial evaluator as the computational object, subtract known-root contributions analytically from logarithmic-derivative-based corrections, and use transformations plus multipoint acceleration to target resistant roots.