Iterated Conformal–Padé Methods

• Iterated Conformal–Padé approach is a family of iterative approximation techniques that use conformal mappings, Padé-type rational approximants, and singularity diagnostics to extend analytic functions beyond their initial convergence domain.
• It employs adaptive conformal maps and repeated rational reconstructions to effectively handle threshold logarithms and improve spectral approximations near singularities.
• The method’s iterative architecture allows for dynamic re-centering and updates as new expansion coefficients emerge, with applications ranging from QCD amplitude reconstruction to numerical solutions of ODEs.

Searching arXiv for recent and relevant papers on conformal–Padé, two-point Padé, and related approaches. “Iterated Conformal–Padé approach” is an Editor’s term for a family of approximation strategies in which conformal changes of variable, Padé or Padé-type rational approximants, and singularity diagnostics are used in repeated or adaptively refined stages to continue analytic or multivalued functions beyond an initial disk of convergence, reconstruct amplitudes with threshold structure, or improve spectral approximations near nearby singularities. In the literature summarized here, iteration appears through repeated updating of conformal maps or branch-cut models, rebuilding approximants as new large-mass or threshold coefficients become available, varying centers of expansion, or applying rational post-processing to coefficient data (Yattselev, 2021, Davies et al., 2019).

1. Conceptual scope and defining components

The approach combines three ingredients that recur across otherwise different problem classes. The first is a conformal map chosen to normalize the analytic structure. In heavy-mass amplitude reconstruction, the central map is

$z=\frac{4\omega}{(1+\omega)^2},$

which maps the complex $z$-plane cut along $[1,\infty)$ onto the unit disc $|\omega|\le 1$, sends $z=1$ to $\omega=1$, and sends $z=0$ to $\omega=0$ (Davies et al., 2019). In Buslaev–Yattselev theory, the Joukowski map

$\mathfrak j(z)=\frac{z+z^{-1}}{2}$

is used to define separating Buslaev compacts $F$ from Chebotarev sets in the $z$0-plane (Yattselev, 2021).

The second ingredient is a rational approximant matched to local or multipoint data. Depending on context, this can be a classical $z$1 Padé approximant in the conformal variable $z$2, a two-point Padé approximant $z$3 simultaneously interpolating germs at $z$4 and $z$5, a Frobenius–Padé approximant built from orthogonal expansions, or a Padé-type approximant with prescribed denominator structure (Davies et al., 2019, Yattselev, 2021, Matos et al., 2017, Brezinski et al., 2014). The heavy-mass reconstruction literature uses rational ansätze of the form

$z$6

typically with $z$7 (Davies et al., 2019).

The third ingredient is feedback from analytic structure. In Buslaev’s setting, the compact $z$8 and its $z$9-property determine the asymptotic pole distribution of two-point Padé approximants (Yattselev, 2021). In spectral filtering, poles of Frobenius–Padé approximants are used to estimate singularities of the exact solution (Matos et al., 2017). In barycentric and partial-fraction formulations, prescribed poles and zeros become free parameters that can be chosen by the user (Brezinski et al., 2014).

Taken together, these ingredients define a methodology rather than a single algorithm. A plausible implication is that “iteration” should be understood broadly: not only repeated application of Padé after a fixed map, but also repeated redesign of the map, the interpolation geometry, the pole ansatz, or the center of expansion as singularity information accumulates.

2. Analytic foundations: conformal normalization and rational continuation

The central justification for conformal preconditioning is that it relocates the dominant singularities to a geometry in which rational approximation is more effective. For gluon-fusion amplitudes, the threshold square root $[1,\infty)$0 becomes

$[1,\infty)$1

which is analytic in $[1,\infty)$2 near $[1,\infty)$3, while threshold logarithms remain non-analytic and therefore must be subtracted explicitly before Padé reconstruction (Davies et al., 2019). The method accordingly separates a known subtraction $[1,\infty)$4 carrying threshold logarithms and reconstructs the analytic remainder by Padé in $[1,\infty)$5.

Two-point Padé theory provides the corresponding framework for multivalued functions defined by different germs at $[1,\infty)$6 and $[1,\infty)$7. In that setting one starts from

$[1,\infty)$8

and seeks a rational function $[1,\infty)$9, with $|\omega|\le 1$0, that interpolates $|\omega|\le 1$1 at $|\omega|\le 1$2 and $|\omega|\le 1$3 at $|\omega|\le 1$4. For the symmetric choice $|\omega|\le 1$5, Buslaev’s theorem gives convergence in capacity outside a unique compact $|\omega|\le 1$6, while Yattselev’s work strengthens this to strong asymptotics and locally uniform convergence in specific separating geometries (Yattselev, 2021).

Several adjacent rational frameworks broaden the same picture. Frobenius–Padé approximants replace Taylor coefficients by orthogonal-series coefficients and are used as rational filters for Tau approximations of ODEs near singularities (Matos et al., 2017). Barycentric and partial-fraction representations rewrite Padé and Padé-type approximants in forms with explicit free poles and zeros, including

$|\omega|\le 1$7

and

$|\omega|\le 1$8

which are especially relevant when a conformal map has already exposed a useful singularity pattern (Brezinski et al., 2014).

A recurrent misconception is that Padé alone resolves all non-analytic structure. The cited work is more restrictive. Threshold logarithms in heavy-quark amplitudes must be separated out before Padé in $|\omega|\le 1$9 (Davies et al., 2019); multivalued branch structure in two-point approximation is controlled by a Buslaev compact $z=1$0 rather than by arbitrary pole placement (Yattselev, 2021); and standard Padé may diverge or converge to a wrong value for irrational functions unless a non-rational control factor is introduced (Gluzman et al., 2015).

3. Iterative architectures and workflow patterns

One explicit workflow is formulated for heavy-mass amplitudes. The sequence is: compute the large-mass expansion around $z=1$1; compute the non-analytic part of the threshold expansion around $z=1$2; subtract a known function $z=1$3 containing threshold logarithms; map to $z=1$4; build Padé ansätze in $z=1$5 for the remainder; match coefficients to the available expansions; reject approximants with poles near the positive real axis; and evaluate the retained family of approximants, using the sample mean as central value and the standard deviation as uncertainty estimate (Davies et al., 2019). Iteration appears because the approximants are rebuilt when additional large-mass or threshold coefficients become available, because multiple Padé orders are explored, and because rescaling parameters $z=1$6 are sampled repeatedly (Davies et al., 2019).

A more abstract iterative pattern is stated for analytic continuation of multivalued functions. A typical strategy is described as: start with a power series expansion around some point; use a conformal map to move dominant singularities or branch cuts into a normalized configuration; compute one or more Padé or multipoint Padé approximants in the mapped variable; observe the poles to infer approximate branch cuts; adjust the conformal map or add new interpolation points iteratively; and repeat (Yattselev, 2021). This scheme is not named in Yattselev’s paper itself, but the structural components are explicit.

Self-similarly corrected Padé introduces iteration in another sense. One first constructs an irrational control function $z=1$7 from self-similar root or factor approximants, defines $z=1$8, and then applies diagonal Padé to $z=1$9 rather than to $\omega=1$0 (Gluzman et al., 2015). The approximant therefore has the form

$\omega=1$1

and the control function can itself be refined by increasing the order of the self-similar approximant. This suggests an iterative hierarchy in which conformal normalization of the variable and self-similar normalization of the function are complementary rather than competing operations.

Spectral filtering supplies a fourth pattern. One first computes a polynomial Tau approximation $\omega=1$2 in an orthogonal basis, then constructs a Frobenius–Padé approximant $\omega=1$3 from the first $\omega=1$4 Tau coefficients, subject to the constraint

$\omega=1$5

A Froissart table is then used to select a filter with no near-canceling pole-zero pairs, often the largest feasible diagonal filter $\omega=1$6 with zero Froissart doublets (Matos et al., 2017). This is not a conformal algorithm, but it provides an explicit post-processing step that can be inserted into a larger iterative conformal–Padé pipeline.

A plausible synthesis is that the phrase “iterated conformal–Padé” should denote a design principle with several admissible outer loops: over conformal maps, over Padé index families, over subtraction models for non-analytic terms, over control functions, or over centers of expansion.

4. Canonical geometries, Buslaev compacts, and pole organization

The most detailed geometric theory in the cited material is Yattselev’s analysis of Buslaev compacts $\omega=1$7 for two-point Padé approximation (Yattselev, 2021). Buslaev’s theorem yields a compact $\omega=1$8 such that

$\omega=1$9

and on each analytic arc of $z=0$0 the $z=0$1-property holds: $z=0$2 In the geometries treated in the paper, $z=0$3 separates the sphere into two simply connected components, providing a two-phase decomposition that is directly compatible with two-point analytic continuation (Yattselev, 2021).

For real $z=0$4, the compact is

$z=0$5

while for non-real $z=0$6 it is a four-arc bouquet connecting $z=0$7, where $z=0$8 is the Chebotarev center and $z=0$9 (Yattselev, 2021). In the real case the function

$\omega=0$0

satisfies

$\omega=0$1

and the approximation error decays geometrically because $\omega=0$2 in $\omega=0$3 and $\omega=0$4 in $\omega=0$5 (Yattselev, 2021).

In the non-real case the geometry is encoded on the genus-1 Riemann surface

$\omega=0$6

A moving point $\omega=0$7 enters the asymptotics, producing a “floating zero” or switching point and allowing small neighborhoods of “parasitic” poles that move with $\omega=0$8 (Yattselev, 2021). The literature is explicit that these outliers do not invalidate global convergence on compact subsets away from their neighborhoods.

Pole distribution is therefore not a side effect but a central diagnostic. Zeros of $\omega=0$9, hence poles of the approximants, accumulate on $\mathfrak j(z)=\frac{z+z^{-1}}{2}$0 according to the equilibrium problem induced by the external field of charges at $\mathfrak j(z)=\frac{z+z^{-1}}{2}$1 and $\mathfrak j(z)=\frac{z+z^{-1}}{2}$2 (Yattselev, 2021). In the spectral setting, low-order Frobenius–Padé poles similarly converge to branch points such as $\mathfrak j(z)=\frac{z+z^{-1}}{2}$3 in the Chebyshev example and $\mathfrak j(z)=\frac{z+z^{-1}}{2}$4 in the Legendre example (Matos et al., 2017). This suggests that a conformal map chosen from observed pole patterns is not merely heuristic; it is aligned with the asymptotic organization of rational approximants in several independent settings.

5. Applications in amplitude reconstruction and spectral computation

In perturbative QCD, the conformal–Padé strategy is applied to single-Higgs production, double-Higgs production, and $\mathfrak j(z)=\frac{z+z^{-1}}{2}$5 interference (Davies et al., 2019). For the three-loop single-Higgs triangle form factor, the approximation is written as

$\mathfrak j(z)=\frac{z+z^{-1}}{2}$6

with separate Padé reconstructions for the non-logarithmic part and for the coefficient of $\mathfrak j(z)=\frac{z+z^{-1}}{2}$7 (Davies et al., 2019). For double-Higgs production, the method is applied form-factor by form-factor using large-mass expansion up to $\mathfrak j(z)=\frac{z+z^{-1}}{2}$8 at NLO, threshold expansion up to $\mathfrak j(z)=\frac{z+z^{-1}}{2}$9 at LO, and up to $F$0 at NLO (Davies et al., 2019). For $F$1 interference, LO and NLO form factors are reconstructed using large-mass expansion up to $F$2 and threshold expansion up to at least $F$3 (Davies et al., 2019).

The reported phenomenology has a clear pattern. LO reconstructions are described as almost perfect for the benchmarked processes; NLO reconstructions agree well with numerical results, though uncertainties grow toward large invariant masses and are larger for components multiplying $F$4 (Davies et al., 2019). The operational lesson is that explicit subtraction of threshold logarithms and separate treatment of global logarithmic coefficients are not optional refinements but core parts of the conformal–Padé architecture.

In numerical ODEs, Frobenius–Padé filtering is used to improve Tau solutions when exact solutions have singularities close to the interval of approximation (Matos et al., 2017). For

$F$5

Chebyshev Tau convergence is slow because the singularity is at the endpoint $F$6, whereas the filtered solution $F$7 substantially reduces the absolute error (Matos et al., 2017). For the Legendre family

$F$8

the filter $F$9 for $z$00 and $z$01 for $z$02 again yields smaller errors away from the nearest singularity (Matos et al., 2017). Although no conformal map is used there, the role of poles as singularity estimators makes the method structurally compatible with later conformal refinement.

Barycentric and partial-fraction representations strengthen the implementation side of the approach (Brezinski et al., 2014). Because poles and zeros appear as free parameters, they can encode known singularities after a conformal transformation. The paper also reports that for $z$03, under random perturbations of the series coefficients of magnitude $z$04, the worst-case error of barycentric Padé on $z$05 is roughly two orders of magnitude smaller than that of the classical Padé approximant (Brezinski et al., 2014). This suggests that node-controlled barycentric forms are natural candidates for numerically robust inner solvers within an iterative conformal–Padé loop.

6. Genericity, re-centering, and adaptive domain changes

Two papers on generic approximation show that Padé convergence can be formulated as a residual property in spaces of holomorphic functions (Fournodavlos, 2011, Fournodavlos et al., 2011). On simply connected domains containing the expansion point, if a set $z$06 contains a sequence $z$07 with $z$08, then the set of functions admitting a subsequence of their own Padé approximants converging uniformly on compact subsets is a dense $z$09 subset of $z$10 (Fournodavlos, 2011). On arbitrary open sets, the analogous result requires both $z$11 and $z$12 (Fournodavlos, 2011).

The stronger 2011 formulation adds simultaneous approximation with respect to all centers of expansion (Fournodavlos et al., 2011). For every compact $z$13, there exists $z$14 such that for all $z$15 and all centers $z$16, the Padé approximants $z$17 exist and satisfy

$z$18

and analogous statements hold in $z$19 with convergence of all derivatives (Fournodavlos et al., 2011). For iterated conformal–Padé design, this is the precise statement that adaptive re-centering need not destroy approximation, at least for generic functions and along suitable subsequences.

These genericity results are non-constructive. They show existence of convergent Padé subsequences but do not identify them algorithmically, do not quantify convergence rates, and do not describe global pole distributions beyond avoidance of the compact region where convergence is asserted (Fournodavlos, 2011, Fournodavlos et al., 2011). A common misconception is to read generic convergence as an unconditional numerical guarantee. The results are instead topological existence theorems in Baire-category language.

A plausible implication is that conformal mapping and re-centering should be viewed as structurally compatible with Padé approximation even on domains of complicated connectivity. The cited results do not provide a conformal–Padé algorithm, but they do supply a function-space backdrop in which repeated choice of centers, compacts, and rational types is mathematically natural.

7. Limitations, failure modes, and methodological boundaries

The approach is effective only under assumptions that are explicit in the cited literature. In two-point Buslaev–Yattselev theory, the multivalued function must have finitely many branch points, and the strong asymptotic results are derived for specific separating compacts $z$20 and weight classes $z$21 (Yattselev, 2021). Functions with natural boundaries, dense singular sets, or essential singularities are not covered there.

In heavy-mass amplitude reconstruction, Padé in the conformal variable cannot reproduce threshold logarithms unless they are separated out first (Davies et al., 2019). The method also uses rejection of approximants with poles close to real positive values of $z$22, balanced degree choices with $z$23, and ensembles of approximants with random rescaling parameters to estimate uncertainty (Davies et al., 2019). This means that the practical algorithm is not a single deterministic Padé table path but an informed selection procedure.

The self-similarly corrected Padé literature emphasizes a different failure mode: standard Padé may oscillate, diverge, or converge to the wrong limit for irrational functions, for indeterminate Stieltjes-type problems, or for certain strong-coupling asymptotics (Gluzman et al., 2015). The correction

$z$24

is introduced precisely because a purely rational approximant may be unable to encode the required irrational structure. This is an objective limitation of Padé, not an implementation defect.

Finally, barycentric and partial-fraction variants introduce free parameters whose influence on robustness and stability remains a central open issue (Brezinski et al., 2014). Node choice can improve stability, but poor choice can also degrade it. In spectral filtering, ill-conditioning of the Padé system and Froissart doublets limit feasible orders and motivate the use of Froissart tables for filter selection (Matos et al., 2017).

The overall picture is therefore technically consistent rather than universal. The literature supports repeated combination of conformal normalization, subtraction of known non-analytic structure, rational approximation, and pole-based structural inference. It does not support the stronger claim that one fixed Padé scheme, with no preprocessing and no geometry, will reliably reconstruct arbitrary analytic structure.