betaFeedback

WorksMathematics

Far-field prediction of algebraic curves via generalized asymptotes

Abstract

The asymptotic behavior of algebraic curves is classically described by means of linear asymptotes. However, many infinite branches exhibit a non-linear behavior that cannot be represented accurately by straight lines. In this paper we propose a computational framework for far-field prediction of algebraic curves based on generalized asymptotes. Starting from the asymptotic expansion of a real branch at infinity, we retain its non-decaying terms in order to construct an algebraic surrogate, which is then used as an adaptive far-field model outside a sufficiently large compact region. We introduce explicit discrepancy indicators and a switching radius that determine when the exact branch may be replaced by its generalized asymptotic predictor. The framework recovers the classical linear asymptote whenever the latter exists, while remaining effective for branches with polynomial or Puiseux-type dominant behavior. In the implicit setting, we prove algebraicity and compatibility properties of the predictor and derive asymptotic error estimates from the first neglected term of the expansion. A residual certificate distinguishes guaranteed finite-tail replacement from asymptotic radius estimates. Tangential tails, slow decay, and a smooth transition rule clarify the limitations of truncation. Reproducible finite-window tests compare the predictors with fitted polynomial baselines; these tests are not a benchmark of a complete plotting system.

Introduction

The study of the behavior of algebraic curves at infinity is a classical topic with direct impact on geometric modeling, qualitative visualization, and symbolic-numeric computation. In practice, the remote part of a curve plays a role in clipping and windowing procedures, adaptive plotting, branch classification, and the construction of simplified surrogate models. A faithful description of that far-field geometry is therefore not only of theoretical interest but also of computational relevance.

The traditional framework for asymptotic analysis is based on linear asymptotes. While this description is useful in many situations, it becomes too restrictive when the dominant remote behavior of a branch is nonlinear. In those cases, replacing a branch by a line may fail to capture the actual geometry at infinity, and therefore provides little information for downstream computational tasks.

This limitation has motivated a series of recent advances on the symbolic computation of asymptotic data for algebraic curves. New algorithms have been proposed for parametric and implicit curves, for asymptotic families, and for generalized asymptotes; see, among others, [1, 2, 3, 4, 5, 6, 7, 8][1, 2, 3, 4, 5, 6, 7, 8]. In particular, generalized asymptotes provide a natural extension of the classical notion by allowing the asymptotic model itself to be a low-degree algebraic curve rather than necessarily a straight line.

The computation of Puiseux expansions and the truncation to non-decaying terms belong to the existing theory cited above. Algebraicity and convergence of these truncations are recalled here as background, not claimed as new results. The contribution studied here is a tolerance-based replacement procedure: a residual test for finite-tail certification, an explicit distinction between predictor construction and separation of tangential tails, and a smooth replacement rule. We test its behaviour on specified algebraic examples, including a deliberately slow-converging family. This does not establish superiority over all asymptotic or numerical continuation methods.

The scope of this paper is intentionally focused. We do not further develop the general theory of generalized asymptotes. Instead, we exploit it as a practical device for far-field prediction. More precisely, for each real branch at infinity of a given algebraic curve, we construct a generalized asymptotic predictor and use it as an adaptive surrogate outside a prescribed compact region. The result is a simple and interpretable model-reduction strategy for the remote geometry of the curve.

The main contributions of the paper are the following:

  • (i) we formulate a far-field prediction framework based on generalized asymptotic surrogates of real infinite branches;

  • (ii) we introduce explicit discrepancy indicators and a switching radius that determine when the exact branch can be replaced by its surrogate;

  • (iii) we prove uniform far-field convergence and a basic invariance result under affine changes preserving the chosen asymptotic direction;

  • (iv) we recall algebraicity and asymptotic error orders, and give a residual-based sufficient certificate for finite-tail replacement;

  • (v) we show that the framework recovers the classical linear asymptote whenever the branch is linearly asymptotic, while extending naturally to genuinely nonlinear dominant behavior;

  • (vi) we analyze tangential tails, slow convergence, smooth transitions, and a coordinatewise space-curve extension, and report reproducible finite-window comparisons.

The paper is organized as follows. § 2 introduces the asymptotic setting, fixes the notation, and develops the notion of generalized asymptotic predictor together with the associated switching criterion. In the same section, Subsection 2.3 discusses the relation with classical linear asymptotes. § 3 extends the framework to implicit curves with Puiseux branches, and Subsection 3.1 illustrates this setting with a representative example. Subsection 3.2 collects several exact algebraic examples and explicit discrepancy computations. § 4 presents plotting-oriented examples and applications, including a real-world motivation based on the far-field continuation of partially observed algebraic contours. Finally, § 5 contains the concluding remarks.

Generalized asymptotic predictors

We begin by briefly recalling the asymptotic viewpoint needed in the sequel. Full algorithmic and theoretical details can be found in [1, 2, 3, 4, 5, 7][1, 2, 3, 4, 5, 7].

Let C⊂R2\mathcal{C} \subset\mathbb{R}^{2} be a real algebraic curve. A real branch at infinity of C\mathcal{C} is an unbounded connected part of C\mathcal{C} that, after a suitable affine change of coordinates if necessary, admits for large xx a representation

B={(x,ψ(x)):x>R0},(1)B = \{(x,\psi(x)) : x > R_0\}, \tag*{(1)}

where ψ\psi has an asymptotic expansion in powers or Puiseux powers of xx. The guiding principle is that the dominant non-decaying terms in that expansion determine a simpler curve with the same far-field behavior.

Definition 2.1. Let BB be a real branch at infinity represented as in (1). A far-field surrogate of BB is a selected real tail ABA_B of an algebraic curve, represented for sufficiently large xx by

AB={(x,ψ~(x)):x>R1},A_B = \{(x,\widetilde{\psi}(x)) : x > R_1\},

and satisfies

lim⁡x→∞(ψ(x)−ψ~(x))=0.\lim_{x\to\infty}\left(\psi(x)-\widetilde{\psi}(x)\right)=0.

If, moreover, ABA_B has minimal degree among the algebraic curves with this property, then ABA_B is called a generalized asymptote of $B.

We use the standard non-decaying truncation of the Puiseux expansion as the predictor. We distinguish its selected real tail from the full algebraic curve containing it; minimal-degree characterizations are part of the cited theory and are not reproved here. This operational viewpoint is sufficient for far-field prediction and is consistent with the general framework developed in [3, 4, 5].

Remark 2.2. If the retained terms are affine, namely ψ~(x)=mx+n\widetilde{\psi}(x)=mx+n, then the generalized asymptote is the classical linear asymptote. Hence the generalized setting strictly contains the classical one.

Now, let BB be a real branch at infinity of a real algebraic curve, represented by

B={(x,ψ(x)):x>R0}.B = \{(x,\psi(x)) : x > R_0\}.

Assume that ψ\psi admits an asymptotic expansion

ψ(x)=∑k=1makxαk+r(x),(2)\psi(x)=\sum_{k=1}^{m}a_kx^{\alpha_k}+r(x), \tag*{(2)}

where

α1>α2>⋯>αm≥0,\alpha_1>\alpha_2>\cdots>\alpha_m\ge0,

the coefficients aka_k are real, and r(x)→0r(x)\to0 as x→∞x\to\infty. The exponents may be integers or rational numbers.

Definition 2.3. The generalized asymptotic predictor associated with the branch BB is the function

ψ~(x)=∑k=1makxαk.(3)\widetilde{\psi}(x)=\sum_{k=1}^{m}a_kx^{\alpha_k}. \tag*{(3)}

The corresponding algebraic curve

AB={(x,ψ~(x)):x>R1}A_B=\{(x,\widetilde{\psi}(x)):x>R_1\}

is called the far-field predictor tail of the branch. Its degree means the degree of its algebraic closure, not the degree of a semialgebraic tail.

This predictor captures exactly the non-decaying part of the asymptotic behavior.

Then

lim⁡x→∞∣ψ(x)−ψ~(x)∣=0.\lim_{x\to\infty} \lvert\psi(x)-\widetilde{\psi}(x)\rvert=0.

Consequently, for every ε>0\varepsilon>0 there exists Rε>0R_{\varepsilon}>0 such that

∣ψ(x)−ψ~(x)∣<ε,x≥Rε.\lvert\psi(x)-\widetilde{\psi}(x)\rvert<\varepsilon,\qquad x\ge R_{\varepsilon}.

Proof. By (2) and (3),

ψ(x)−ψ~(x)=r(x),\psi(x)-\widetilde{\psi}(x)=r(x),

and therefore the result follows from the fact that r(x)→0r(x)\to0.

Discrepancy indicators

To use the predictor adaptively, we define the far-field discrepancy indicators

E∞(R):=sup⁡x≥R∣ψ(x)−ψ~(x)∣,Erel(R):=sup⁡x≥R∣ψ(x)−ψ~(x)∣1+∣ψ(x)∣.(4)E_{\infty}(R):=\sup_{x\ge R}\lvert\psi(x)-\widetilde{\psi}(x)\rvert,\qquad E_{\mathrm{rel}}(R):=\sup_{x\ge R}\frac{\lvert\psi(x)-\widetilde{\psi}(x)\rvert}{1+\lvert\psi(x)\rvert}. \tag*{(4)}

The first one is an absolute discrepancy, whereas the second provides a scale-aware version.

Proposition 2.5. Assume that the remainder rr in (2) is continuous for x≥R0x\ge R_0 and satisfies r(x)→0r(x)\to0 as x→∞x\to\infty. Then, for every ε>0\varepsilon>0, there exists Rε≥R0R_{\varepsilon}\ge R_0 such that

sup⁡x≥Rε∣ψ(x)−ψ~(x)∣≤ε.\sup_{x\ge R_{\varepsilon}}\lvert\psi(x)-\widetilde{\psi}(x)\rvert\le\varepsilon.

Equivalently,

lim⁡R→∞E∞(R)=0.\lim_{R\to\infty}E_{\infty}(R)=0.

Proof. Since r(x)→0r(x)\to0, for every ε>0\varepsilon>0 there exists RεR_{\varepsilon} such that ∣r(x)∣≤ε\lvert r(x)\rvert\le\varepsilon for all x≥Rεx\ge R_{\varepsilon}. Using again ψ(x)−ψ~(x)=r(x)\psi(x)-\widetilde{\psi}(x)=r(x), one gets

sup⁡x≥Rε∣ψ(x)−ψ~(x)∣=sup⁡x≥Rε∣r(x)∣≤ε.\sup_{x\ge R_{\varepsilon}}\lvert\psi(x)-\widetilde{\psi}(x)\rvert =\sup_{x\ge R_{\varepsilon}}\lvert r(x)\rvert\le\varepsilon.

This proves the claim.

Corollary 2.6. Under the assumptions of Proposition 2.4, if ∣ψ(x)−ψ~(x)∣\lvert\psi(x)-\widetilde{\psi}(x)\rvert is eventually monotone decreasing, then

E∞(R)=∣ψ(R)−ψ~(R)∣E_{\infty}(R)=\lvert\psi(R)-\widetilde{\psi}(R)\rvert

for all sufficiently large RR.

Proof. If the discrepancy is monotone decreasing for x≥Rx\ge R, its supremum on [R,∞)[R,\infty) is attained at x=Rx=R.

Proposition 2.7. Let

T(x,y)=(ax+b,cy+d),a>0, c≠0,T(x,y)=(ax+b,cy+d),\qquad a>0,\ c\ne0,

be an affine change of variables, with X=ax+bX=ax+b, preserving the direction x→+∞x\to+\infty. Let BB be a branch represented by y=ψ(x)y=\psi(x) and let ψ~\widetilde{\psi} be its generalized asymptotic predictor. Then the image branch T(B)T(B) admits the algebraic surrogate

ψ~T(X)=cψ~(X−ba)+d,\widetilde{\psi}_{T}(X)=c\widetilde{\psi}\left(\frac{X-b}{a}\right)+d,
lim⁡X→∞∣Ψ(X)−ψ~T(X)∣=0,\lim_{X\to\infty}\left|\Psi(X)-\widetilde{\psi}_{T}(X)\right|=0,

where Ψ\Psi denotes the exact branch function of T(B)T(\mathrm{B}). The transported surrogate need not equal the canonical non-decaying truncation after a translation in xx; the assertion concerns convergence of the transported tail.

Proof. Writing X=ax+bX=ax+b, one has x=(X−b)/ax=(X-b)/a and

Y=cψ(X−ba)+d.Y=c\psi\left(\frac{X-b}{a}\right)+d.

Hence the transformed exact branch is

Ψ(X)=cψ(X−ba)+d.\Psi(X)=c\psi\left(\frac{X-b}{a}\right)+d.

Defining

ψ~T(X)=cψ~(X−ba)+d,\widetilde{\psi}_{T}(X)=c\widetilde{\psi}\left(\frac{X-b}{a}\right)+d,

we obtain

Ψ(X)−ψ~T(X)=c(ψ(X−ba)−ψ~(X−ba)).\Psi(X)-\widetilde{\psi}_{T}(X) =c\left(\psi\left(\frac{X-b}{a}\right)-\widetilde{\psi}\left(\frac{X-b}{a}\right)\right).

Since (X−b)/a→∞(X-b)/a\to\infty as X→∞X\to\infty, Proposition 2.4 yields the result.

Switching radius

Given a tolerance ε>0\varepsilon>0, we define the switching radius

Rε:=inf⁡{R≥R0:E∞(R)≤ε}.(5)R_{\varepsilon}:=\inf\{R\ge R_{0}:E_{\infty}(R)\le\varepsilon\}. \tag*{(5)}

Choose R0>0R_{0}>0 on a closed regular tail on which the branch extends continuously. The continuous remainder implies that the infimum is attained whenever it is finite. In computation we use a certified sufficient radius R^ε≥Rε\widehat{R}_{\varepsilon}\ge R_{\varepsilon}, rather than assume this infimum can be evaluated exactly. This leads to a simple computational scheme.

Algorithm 1. Generalized far-field prediction of a real branch

Input: A real algebraic curve CC, a real branch at infinity BB of CC, and a prescribed tolerance ε>0\varepsilon>0.

Output: A generalized asymptotic predictor ABA_{B} for the branch BB, together with a switching radius RεR_{\varepsilon} beyond which the predictor approximates the branch within the required tolerance.

  1. Determine a real branch at infinity BB of CC and represent it, after a suitable affine change of coordinates if necessary, in the form

B={(x,ψ(x)):x>R0}.B=\{(x,\psi(x)):x>R_{0}\}.
  1. Compute an asymptotic expansion of ψ(x)\psi(x) as x→∞x\to\infty,

ψ(x)=∑k=1makxαk+r(x),α1>α2>⋯>αm≥0,\psi(x)=\sum_{k=1}^{m}a_{k}x^{\alpha_{k}}+r(x),\qquad\alpha_{1}>\alpha_{2}>\cdots>\alpha_{m}\ge0,

where r(x)→0r(x)\to0.

  1. Construct the generalized asymptotic predictor by retaining the non-decaying terms:

ψ~(x)=∑k=1makxαk.\widetilde{\psi}(x) = \sum_{k=1}^{m} a_k x^{\alpha_k}.

Let ABA_B denote the algebraic curve associated with ψ~\widetilde{\psi}.

  1. Obtain a proved tail bound ∣ψ(x)−ψ~(x)∣≤M(x)\lvert\psi(x)-\widetilde{\psi}(x)\rvert\le M(x), valid for every x≥Rcertx \ge R_{\mathrm{cert}}, using an explicit remainder estimate or Theorem 2.11. A sampled discrepancy is diagnostic only.

  1. Find a sufficient R^ε≥Rcert\widehat{R}_{\varepsilon} \ge R_{\mathrm{cert}} such that sup⁡x≥R^εM(x)≤ε\sup_{x \ge\widehat{R}_{\varepsilon}} M(x) \le\varepsilon. If verification fails, keep the exact branch and report that replacement has not been certified.

  1. Replace the exact branch BB by its generalized asymptotic predictor ABA_B for all x≥R^εx \ge\widehat{R}_{\varepsilon}. That is, use the exact branch in the near field and the predictor in the far field.

The symbolic core of the method is the computation of branches and asymptotic expansions, for which effective algorithms are already available in the literature [1, 2, 3, 4, 5, 7].

Example 2.8. Consider the implicit algebraic curve

C:y3−xy+1=0.C : y^3 - xy + 1 = 0.

For its positive real branch at infinity, a dominant-balance argument gives the asymptotic expansion

ψ(x)=x1/2−12x−1+O(x−5/2).\psi(x) = x^{1/2} - \frac{1}{2}x^{-1} + O(x^{-5/2}).

Hence the generalized asymptotic predictor is

ψ~(x)=x1/2,\widetilde{\psi}(x) = x^{1/2},

which corresponds to the branch y>0y > 0 of the parabola

y2=x.y^2 = x.

Moreover,

ψ(x)−ψ~(x)∼−12x−1,E∞(R)∼12R.\psi(x) - \widetilde{\psi}(x) \sim-\frac{1}{2}x^{-1}, \qquad E_{\infty}(R) \sim\frac{1}{2R}.

If ε=0.1\varepsilon= 0.1, then the switching radius is approximately

Rε≈5.R_{\varepsilon} \approx5.

This is only an initial estimate. Example 2.12 shows that 5 does not meet the tolerance and proves that 6 is a sufficient switching radius.

Figure 1 illustrates the output of Algorithm 1. The exact branch is not introduced through a closed-form expression y=f(x)y=f(x), but through its implicit equation and a parametric representation derived from it. The generalized asymptotic predictor still provides a simple algebraic surrogate, namely the branch of the parabola y2=xy^2=x, which accurately captures the far-field geometry.

This example also anticipates the more structural implicit Puiseux analysis developed in § 3.

Exact branch of $y^3-xy+1=0$, predictor $y=\sqrt{x}$, and certified switching line $x=6$

Figure 1. Illustration of Algorithm 1 for the positive branch of the implicit curve y3−xy+1=0y^3-xy+1=0. The exact branch is represented parametrically by (t2+t−1,t)(t^2+t^{-1},t), while the generalized asymptotic predictor is the branch of the parabola y2=xy^2=x given by (t2,t)(t^2,t). Beyond the switching region, both curves become visually close.

Relation with classical linear asymptotes

The generalized predictor contains the classical theory as a special case.

Proposition 2.9. Suppose that a branch at infinity has an asymptotic expansion

ψ(x)=mx+n+r(x),r(x)→0 as x→∞.\psi(x)=mx+n+r(x), \qquad r(x)\to0 \text{ as } x\to\infty.

Then its generalized asymptotic predictor is the line y=mx+ny=mx+n. In particular, the proposed framework recovers the classical linear asymptote whenever it exists.

Proof. The only non-decaying terms in the expansion are mx+nmx+n, hence the predictor coincides with that affine function.

The real gain of the generalized framework appears when the dominant behavior is nonlinear.

Proposition 2.10. Suppose that a branch at infinity admits an expansion

ψ(x)=p(x)+r(x),\psi(x)=p(x)+r(x),

where p(x)p(x) is a nonconstant polynomial of degree at least two and r(x)→0r(x)\to0 as x→∞x\to\infty. Then no affine function is a valid asymptotic surrogate of the branch, whereas the polynomial curve y=p(x)y=p(x) is a generalized asymptotic predictor.

Proof. Let ℓ(x)=mx+n\ell(x)=mx+n be an affine function. Since deg⁡p≥2\deg p\ge2, one has*

∣p(x)−ℓ(x)∣→∞as x→∞.\lvert p(x)-\ell(x)\rvert\to\infty\qquad\text{as } x\to\infty.

Therefore

∣ψ(x)−ℓ(x)∣≥∣p(x)−ℓ(x)∣−∣r(x)∣→∞,\lvert\psi(x)-\ell(x)\rvert\ge\lvert p(x)-\ell(x)\rvert-\lvert r(x)\rvert\to\infty,

so ℓ\ell cannot define a far-field surrogate. On the other hand,

ψ(x)−p(x)=r(x)→0,\psi(x)-p(x)=r(x)\to0,

hence y=p(x)y=p(x) is a valid predictor.

Propositions 2.9 and 2.10 show that generalized asymptotes are not simply a refinement of linear asymptotes. They enlarge the class of branches that admit a compact asymptotic geometric description, which is precisely what is required in model-reduction tasks at infinity.

Finite-tail certification

An asymptotic equivalence estimates a switching radius but does not certify a finite tolerance. The following sufficient test uses the implicit equation.

Theorem 2.11 (Residual certificate in an isolating tube). Let pp be a continuous algebraic function on [R,∞)[R,\infty) and h>0h>0. Suppose, for all x≥Rx\ge R, that

∣F(x,p(x))∣≤ρ(x),∣Fy(x,y)∣≥m(x)>0(∣y−p(x)∣≤h),ρ(x)<m(x)h.|F(x,p(x))|\le\rho(x),\qquad|F_y(x,y)|\ge m(x)>0\quad(|y-p(x)|\le h),\qquad\rho(x)<m(x)h.

Then the tube contains exactly one real root y=ψ(x)y=\psi(x) in each vertical fibre, and ∣ψ(x)−p(x)∣≤ρ(x)/m(x)|\psi(x)-p(x)|\le\rho(x)/m(x). If the selected asymptotic branch is in the tube for sufficiently large xx, it is this root branch throughout [R,∞)[R,\infty). Thus sup⁡x≥Rρ(x)/m(x)≤ε\sup_{x\ge R}\rho(x)/m(x)\le\varepsilon certifies tail replacement.

Proof. At fixed xx, continuity and nonvanishing of FyF_y give a constant derivative sign in the vertical interval. The endpoint values differ from F(x,p(x))F(x,p(x)) in opposite directions by at least m(x)hm(x)h, hence have opposite signs. The intermediate value theorem gives existence, strict monotonicity gives uniqueness, and the mean value theorem gives the error bound. The implicit function theorem supplies a continuous root branch; uniqueness continues its identification with the selected asymptotic branch across the tail.

The criterion is sufficient, not necessary. It can fail if the tube contains several roots or FyF_y vanishes. A small residual alone does not identify a branch. For polynomial FF with rational or real algebraic coefficients, algebraic pp, and semialgebraic ρ,m\rho,m, the inequalities can in principle be verified by real quantifier elimination, introducing auxiliary algebraic variables and isolating intervals to specify the real determinations. This finite verification route may be expensive. Checking a finite sample does not certify an infinite tail.

Example 2.12 (A certified radius for the cubic). For F=y3−xy+1F=y^3-xy+1, take p=xp=\sqrt{x}, R=6R=6, and h=1/10h=1/10. On the tube, F(x,p)=1F(x,p)=1 and

Fy≥2x−35x+3100.F_y\ge2x-\frac{3}{5}\sqrt{x}+\frac{3}{100}.

This lower bound increases for x≥6x\ge6 and at 66 exceeds 1010. Theorem 2.11 certifies error less than 0.10.1 on [6,∞)[6,\infty). The asymptotic estimate R0.1≈5R_{0.1}\approx5 is not itself sufficient: F(5,5−0.1)=0.035−0.001>0F(5,\sqrt{5}-0.1)=0.03\sqrt{5}-0.001>0 and the large positive root lies below 5−0.1\sqrt{5}-0.1. Its error at 55 therefore exceeds 0.10.1.

From an implicit equation to labelled tails

Branch computation uses established Newton–Puiseux methods, not a new expansion algorithm. The required tasks are as follows.

  1. Remove repeated factors and handle components separately. For a chart x→+∞x\to+\infty, set z=1/xz=1/x, y=z−kuy=z^{-k}u with integer kk large enough for the relevant branches, and multiply by a power of zz to form a polynomial H(z,u)H(z,u).

  1. Compute Newton–Puiseux expansions of H=0H=0 at z=0z=0. Each Newton edge gives candidate exponents and a characteristic polynomial for the leading coefficient. Substitute each root and repeat at multiple roots until distinct local series separate. Keep determinations real for z>0z>0; remove duplicate parametrizations of the same graph. Other directions require other charts or x↦−xx\mapsto-x.

  1. Return to y=ψi(x)y=\psi_i(x) and retain every exponent at least zero, giving pip_i; if no such term occurs, set pi=0p_i=0. Continue farther to separate branch labels or obtain a remainder bound. These stopping tasks need not have the same truncation depth.

  1. At a candidate RR, isolate the real roots of F(R,y)F(R,y) and match the selected tail to its isolating interval. Certify a tube as in Theorem 2.11, or use an independently proved remainder bound. Real-root isolation and semialgebraic tests are exact subroutines for rational or algebraic input.

  1. Increase RR, for example by doubling, until the error certificate passes. The output is a sufficient R^ε\widehat{R}_{\varepsilon}, not necessarily the infimum RεR_{\varepsilon}. If a prescribed search budget is exhausted, report failure to certify, rather than accept an asymptotic estimate.

This is a verifiable workflow, not a claim of an efficient implementation for every curve. The numerical tests below implement the stated scalar examples; they do not implement a general Newton–Puiseux or quantifier-elimination package.

Tangential branches and truncation depth

No fixed number of terms suffices for all input curves. If ψ1−ψ2∼cx−γ\psi_1-\psi_2\sim cx^{-\gamma}, γ>0\gamma>0, distinct tails have the same non-decaying predictor. Their labels require computation through the first differing coefficient, even though it is not retained in that predictor. The common predictor alone cannot preserve their separation.

Example 2.13 (The parametrization (t4+t,t2)(t^4+t,t^2)). Elimination gives the irreducible curve

(x−y2)2−y=0.(x-y^2)^2-y=0.

Its two tails over large positive xx satisfy x=y2+σyx=y^2+\sigma\sqrt{y}, σ=±1\sigma=\pm1, y>0y>0, and

ψσ(x)=x1/2−σ2x−1/4+O(x−7/4).\psi_\sigma(x)=x^{1/2}-\frac{\sigma}{2}x^{-1/4}+O(x^{-7/4}).

Indeed, substituting y=x1/2(1+v)y=x^{1/2}(1+v) balances 2v2v against σx−3/4\sigma x^{-3/4}; the next coefficient in vv vanishes. Both predictors are p=x1/2p=x^{1/2}, with curve y2=xy^2=x. In the projective chart X=1X=1, the point Y=Z=0Y=Z=0 has common tangent Z=0Z=0 for these tails. Computing the x−1/4x^{-1/4} term separates them; retaining that decaying term as a corrected surrogate improves the error order to x−7/4x^{-7/4}. The corrected algebraic surrogate is not asserted to be a minimal-degree generalized asymptote.

Slow decay and smooth transitions

For a certified error bound Cx−γCx^{-\gamma} valid beyond RcertR_{\mathrm{cert}}, a sufficient radius is max⁡(Rcert,(C/ε)1/γ)\max(R_{\mathrm{cert}},(C/\varepsilon)^{1/\gamma}). Small γ\gamma may make this impractical. For example,

F=x(y−1)20−1,ψ±=1±x−1/20,p±=1F=x(y-1)^{20}-1,\qquad\psi_{\pm}=1\pm x^{-1/20},\qquad p_{\pm}=1

give the exact radius Rε=ε−20R_{\varepsilon}=\varepsilon^{-20} on x>0x>0: R0.1=1020R_{0.1}=10^{20}. Possible responses are to retain additional decaying terms and recertify, use a relative tolerance when justified, or keep the exact branch within the window of interest. Here one extra term is exact. In general the cost and algebraic degree can increase; no universal acceleration is claimed.

A hard switch need not be continuous. Choose L>0L>0 and a certified RR. For w(u)=6u5−15u4+10u3w(u)=6u^5-15u^4+10u^3, put

q(x)={ψ(x),x≤R,(1−w((x−R)/L))ψ(x)+w((x−R)/L)p(x),R<x<R+L,p(x),x≥R+L.q(x)= \begin{cases} \psi(x), & x\le R,\\ (1-w((x-R)/L))\psi(x)+w((x-R)/L)p(x), & R<x<R+L,\\ p(x), & x\ge R+L. \end{cases}

Since 0≤w≤10\le w\le1 and its first two derivatives vanish at both endpoints, qq is C2C^2 if ψ,p\psi,p are C2C^2 and ∣q−ψ∣≤ε|q-\psi|\le\varepsilon on the transition and tail. Tangent and curvature accuracy require separate derivative bounds. Exact evaluation remains necessary inside the transition. The composite model is piecewise algebraic, not generally one algebraic curve.

Coordinatewise extension to space curves

For a given tail (x,ψ2(x),ψ3(x))(x,\psi_2(x),\psi_3(x)) with convergent Puiseux coordinates, truncate each coordinate to its non-decaying part pjp_j. A common denominator substitution x=sNx=s^N yields a polynomial parametrization and hence an algebraic space-curve surrogate. If ∣ψj−pj∣≤Mj(x)|\psi_j-p_j|\le M_j(x), its Euclidean error is at most M2(x)2+M3(x)2\sqrt{M_2(x)^2+M_3(x)^2}. For example (x,x1/2+x−1,x1/3+x−2)(x,x^{1/2}+x^{-1},x^{1/3}+x^{-2}) has predictor (s6,s3,s2)(s^6,s^3,s^2), s>0s>0, and tail error bounded by R−2+R−4\sqrt{R^{-2}+R^{-4}}. This observation does not solve the extraction of space-curve branches from a general implicit system.

Implicit Puiseux branches and algebraicity of the predictor

The previous sections formulate the far-field predictor assuming that an asymptotic expansion of the branch is already available. We now add a more structural result showing that the framework also applies naturally to implicit algebraic curves whose branches at infinity are described by Puiseux expansions.

Consider a real algebraic curve

C={(x,y)∈R2:F(x,y)=0},F∈R[x,y].\mathcal{C}=\{(x,y)\in\mathbb{R}^2:F(x,y)=0\},\qquad F\in\mathbb{R}[x,y].

Assume that one of its real branches at infinity admits a Puiseux expansion

ψ(x)=∑j=0∞cjxβj,β0>β1>β2>⋯ ,(6)\psi(x)=\sum_{j=0}^{\infty}c_jx^{\beta_j},\qquad\beta_0>\beta_1>\beta_2>\cdots, \tag*{(6)}

where the exponents βj\beta_j are rational numbers and only finitely many of them are nonnegative. Let

β0>⋯>βm≥0>βm+1>βm+2>⋯ .\beta_0>\cdots>\beta_m\ge0>\beta_{m+1}>\beta_{m+2}>\cdots.

We define the corresponding truncation

ψ~(x)=∑j=0mcjxβj.(7)\widetilde{\psi}(x)=\sum_{j=0}^{m}c_jx^{\beta_j}. \tag*{(7)}

The next theorem gives two properties that strengthen the mathematical content of the predictor construction: the predictor remains algebraic even in the Puiseux setting, and the first neglected exponent controls the asymptotic order of the error.

Theorem 3.1. Let BB be a real branch at infinity of an implicit algebraic curve F(x,y)=0F(x,y)=0 admitting a Puiseux expansion (6), and let ψ~\widetilde{\psi} be the truncation (7). Then:

(a) the graph of ψ~\widetilde{\psi} is contained in a real algebraic curve;

(b) one has

ψ(x)−ψ~(x)=O(xβm+1)as x→∞,\psi(x)-\widetilde{\psi}(x)=O(x^{\beta_{m+1}})\qquad\text{as }x\to\infty,

and therefore

E∞(R)=O(Rβm+1)as R→∞.E_{\infty}(R)=O(R^{\beta_{m+1}})\qquad\text{as }R\to\infty.

Proof. Let NN be a common positive denominator of the rational exponents β0,…,βm\beta_0,\ldots,\beta_m. Then there exist integers q0,…,qmq_0,\ldots,q_m such that βj=qj/N\beta_j=q_j/N. If we set

x=sN,x=s^N,

then

ψ~(x)=∑j=0mcjsqj=:P(s).\widetilde{\psi}(x)=\sum_{j=0}^{m}c_js^{q_j}=:P(s).

Hence the predictor is parametrized by

x=sN,y=P(s).x=s^N,\qquad y=P(s).

Eliminating ss from these equations, for instance by taking the resultant of X−sNX-s^N and Y−P(s)Y-P(s) with respect to ss, yields a nonzero polynomial relation

G(X,Y)=0G(X,Y)=0

with real coefficients. Therefore the graph of ψ~\widetilde{\psi} is contained in a real algebraic curve.

For the error estimate, write

ψ(x)−ψ~(x)=∑j=m+1∞cjxβj.\psi(x)-\widetilde{\psi}(x)=\sum_{j=m+1}^{\infty}c_jx^{\beta_j}.

Under the usual asymptotic interpretation of the Puiseux expansion, the first neglected term determines the order of the remainder. Therefore,

ψ(x)−ψ~(x)=O(xβm+1),\psi(x)-\widetilde{\psi}(x)=O(x^{\beta_{m+1}}),

which immediately implies

E∞(R)=sup⁡x≥R∣ψ(x)−ψ~(x)∣=O(Rβm+1)E_{\infty}(R)=\sup_{x\ge R}|\psi(x)-\widetilde{\psi}(x)|=O(R^{\beta_{m+1}})

as R→∞R\to\infty.

Theorem 3.1 is relevant for the intended computational use of the predictor. It shows that even when the original branch comes from an implicit curve and is only accessible through local Puiseux analysis, the retained non-decaying terms still define an algebraic far-field surrogate. Moreover, the first discarded exponent provides an immediate asymptotic estimate for the decay of the discrepancy.

Remark 3.2. If the nonnegative exponents in (6) are integers, then the predictor is polynomial. If fractional nonnegative exponents occur, the predictor remains algebraic but is no longer polynomial; this is precisely the situation in which generalized asymptotes go beyond the polynomial setting while preserving algebraicity.

The previous theorem can be sharpened in the generic case in which the first discarded coefficient is nonzero.

Corollary 3.3. Under the assumptions of Theorem 3.1, and assuming that the Puiseux expansion is understood asymptotically in the usual sense, suppose in addition that cm+1≠0c_{m+1}\ne0. Then

ψ(x)−ψ~(x)∼cm+1xβm+1as x→∞,\psi(x)-\widetilde{\psi}(x)\sim c_{m+1}x^{\beta_{m+1}}\qquad\text{as }x\to\infty,

and, if the remainder is eventually monotone in absolute value,

E∞(R)∼∣cm+1∣Rβm+1as R→∞.E_{\infty}(R)\sim|c_{m+1}|R^{\beta_{m+1}}\qquad\text{as }R\to\infty.

Proof. By separating the first neglected term from the rest, one gets

ψ(x)−ψ~(x)=cm+1xβm+1+∑j=m+2∞cjxβj.\psi(x)-\widetilde{\psi}(x)=c_{m+1}x^{\beta_{m+1}}+\sum_{j=m+2}^{\infty}c_jx^{\beta_j}.

Since every exponent βj\beta_j with j≥m+2j\ge m+2 is strictly smaller than βm+1\beta_{m+1}, the tail is o(xβm+1)o(x^{\beta_{m+1}}). Hence the first asymptotic equivalence follows. The second one is then a direct consequence of Corollary 2.6.

A further structural consequence is that the predictor is compatible with the algebraic parametrization naturally associated with the Puiseux branch.

Proposition 3.4. Let ψ\psi be as in (6), and let x=sNx=s^{N} be a common denominator substitution for the nonnegative exponents of ψ\psi. Then the far-field predictor is exactly the image of the polynomial truncation in the parameter ss, namely

x=sN,y=P(s),x=s^{N}, \qquad y=P(s),

with

P(s)=∑j=0mcjsNβj.P(s)=\sum_{j=0}^{m}c_j s^{N\beta_j}.

In particular, the predictor may be computed either by truncating the branch in xx or by truncating its corresponding polynomialized parametrization in ss.

Proof. This is immediate from the proof of Theorem 3.1. After the substitution x=sNx=s^{N}, all retained exponents become integers, so the predictor is represented by the polynomial P(s)P(s) in the parameter ss. Replacing ss by x1/Nx^{1/N} gives back exactly the same truncated function ψ~(x)\widetilde{\psi}(x). □\square

An implicit example of Puiseux type

Consider the irreducible affine curve

C5: y3−xy+1=0.(8)C_5:\ y^3-xy+1=0. \tag*{(8)}

This curve is naturally given in implicit form, and its far-field behavior is not transparently read off from a global explicit representation. We analyze one of its real branches at infinity.

To determine the dominant balance, we look for a branch of the form y∼cxαy\sim cx^\alpha as x→∞x\to\infty. Substituting into

y3−xy+1=0y^3-xy+1=0

shows that the dominant terms balance when 3α=α+13\alpha=\alpha+1, that is, when α=1/2\alpha=1/2. The corresponding coefficient must satisfy

c3−c=0,c^3-c=0,

hence c=±1c=\pm1 for the nontrivial branches. Therefore the curve has two nonzero real branches with leading behavior

y∼±x1/2.y\sim\pm x^{1/2}.

We now compute the next term for the positive branch. Write

y=x1/2+u(x),u(x)=o(x1/2).y=x^{1/2}+u(x), \qquad u(x)=o(x^{1/2}).

Substituting into (8), we obtain

(x1/2+u)3−x(x1/2+u)+1=2xu+3x1/2u2+u3+1=0.\left(x^{1/2}+u\right)^3-x\left(x^{1/2}+u\right)+1=2xu+3x^{1/2}u^2+u^3+1=0.

The cancellation of the constant term requires

2xu+1=02xu+1=0

at first order, so that

u(x)=−12x+O(x−5/2).u(x)=-\frac{1}{2x}+O\left(x^{-5/2}\right).

Hence the branch admits the Puiseux expansion

y=x1/2−12x−1+O(x−5/2).(9)y = x^{1/2} - \frac{1}{2}x^{-1} + O(x^{-5/2}). \tag*{(9)}

According to the general construction, the far-field predictor is obtained by retaining the nonnegative exponent:

ψ~5(x)=x1/2.\widetilde{\psi}_5(x) = x^{1/2}.

Therefore the associated generalized asymptote is the algebraic curve

A5:y2−x=0,y>0.A_5:\quad y^2-x=0,\qquad y>0.

By Theorem 3.1, the predictor is algebraic and

ψ5(x)−ψ~5(x)=O(x−1),E∞(R)=O(R−1).\psi_5(x)-\widetilde{\psi}_5(x)=O(x^{-1}),\qquad E_\infty(R)=O(R^{-1}).

In fact, Corollary 3.3 yields the sharper estimate

ψ5(x)−ψ~5(x)∼−12x−1,E∞(R)∼12R.\psi_5(x)-\widetilde{\psi}_5(x)\sim-\frac{1}{2}x^{-1},\qquad E_\infty(R)\sim\frac{1}{2R}.

This example is mathematically more substantial than the polynomial ones of Subsection 3.2. The original curve is implicit, the branch is obtained by a Puiseux-type asymptotic argument, and the predictor is an algebraic curve defined only after truncation of the local expansion. Nevertheless, the same far-field philosophy applies unchanged: the remote geometry is captured by the algebraic surrogate

y2=x.y^2=x.

There is also a third real tail with ψ0(x)=x−1+O(x−4)\psi_0(x)=x^{-1}+O(x^{-4}) and predictor y=0y=0. The negative outer branch is treated in the same way and leads to the predictor

y=−x1/2,y=-x^{1/2},

that is, to the lower branch of the same parabola y2=xy^2=x.

Figure 2 illustrates the two real nonzero branches of the implicit curve together with their generalized asymptotic predictors. The exact branches are drawn through the parametric representation

Far-field behavior of the implicit curve $y^3-xy+1=0$, showing exact positive and negative branches and their predictors

Figure 2. Far-field behavior of the implicit curve y3−xy+1=0y^3-xy+1=0. The outer branches are represented parametrically by (t2+t−1,t)(t^2+t^{-1},t) and (t2−t−1,−t)(t^2-t^{-1},-t), while their generalized asymptotic predictors are the two branches of the parabola y2=xy^2=x, namely (t2,±t)(t^2,\pm t). The coincidence becomes visually clear in the remote region.

(x,y)=(t2+t−1,t)or(t2−t−1,−t),(x,y)=(t^2+t^{-1},t)\quad\text{or}\quad(t^2-t^{-1},-t),

whereas the asymptotic surrogate is given by the parabola

x=t2,y=±t.x=t^2,\qquad y=\pm t.

This visualization makes the far-field approximation transparent and confirms that the algebraic predictor captures the remote geometry of the implicit curve.

Exact algebraic examples

We now present four exact examples. They are intentionally elementary, since the objective is to make the far-field mechanism transparent and to derive closed-form discrepancy expressions.

Example 3.5 (A linear asymptote recovered exactly). Consider

C1:xy−x2−1=0.C_1: \quad xy-x^2-1=0.

For x≠0x\ne0, the curve can be written as

y=x+1x.y=x+\frac{1}{x}.

Hence the branch for x>0x>0 has predictor

ψ~1(x)=x,\widetilde{\psi}_1(x)=x,

which is exactly the classical asymptote. The discrepancy is

∣ψ1(x)−ψ~1(x)∣=1x,\lvert\psi_1(x)-\widetilde{\psi}_1(x)\rvert=\frac{1}{x},

and thus

E∞(R)=1R,Rε=1ε.E_{\infty}(R)=\frac{1}{R},\qquad R_{\varepsilon}=\frac{1}{\varepsilon}.

This example confirms that the generalized framework is consistent with the classical one.

Example 3.6 (A quadratic far-field surrogate). Consider

C2: xy−x3−1=0.C_2: \ xy-x^3-1=0.

For x≠0x\ne0,

y=x2+1x.y=x^2+\frac{1}{x}.

The predictor is therefore

ψ~2(x)=x2.\widetilde{\psi}_2(x)=x^2.

Moreover,

∣ψ2(x)−ψ~2(x)∣=1x,E∞(R)=1R,Rε=1ε.\lvert\psi_2(x)-\widetilde{\psi}_2(x)\rvert=\frac{1}{x},\qquad E_{\infty}(R)=\frac{1}{R},\qquad R_{\varepsilon}=\frac{1}{\varepsilon}.

The relevant point is not the size of the remainder, but the geometry of the predictor: no line can model the branch at infinity, while the parabola y=x2y=x^2 captures its dominant behavior exactly.

Example 3.7 (Same predictor, different switching radius). Consider

C3: x2y−x4−1=0.C_{3}: \ x^{2}y-x^{4}-1=0.

For x≠0x\ne0,

y=x2+1x2.y=x^{2}+\frac{1}{x^{2}}.

Hence the predictor is again

ψ~3(x)=x2,\widetilde{\psi}_{3}(x)=x^{2},

but now the discrepancy satisfies

∣ψ3(x)−ψ~3(x)∣=1x2,E∞(R)=1R2,Rε=1ε.\lvert\psi_{3}(x)-\widetilde{\psi}_{3}(x)\rvert=\frac{1}{x^{2}},\qquad E_{\infty}(R)=\frac{1}{R^{2}},\qquad R_{\varepsilon}=\frac{1}{\sqrt{\varepsilon}}.

This shows that branches with the same generalized asymptote may have different adaptive behavior, and therefore the predictor and the switching radius play complementary roles.

Example 3.8 (A cubic predictor). Consider

C4: xy−x4−1=0.C_{4}: \ xy-x^{4}-1=0.

For x≠0x\ne0,

y=x3+1x.y=x^{3}+\frac{1}{x}.

Thus

ψ~4(x)=x3,E∞(R)=1R,Rε=1ε.\widetilde{\psi}_{4}(x)=x^{3},\qquad E_{\infty}(R)=\frac{1}{R},\qquad R_{\varepsilon}=\frac{1}{\varepsilon}.

Again, no affine asymptotic model is meaningful, whereas the cubic predictor captures the far-field geometry in a natural way.

Table 1 collects the exact predictors and discrepancies obtained above.

CurvePredictorE∞(R)E_{\infty}(R)RεR_{\varepsilon}
xy−x2−1=0xy-x^{2}-1=0y=xy=x1/R1/R1/ε1/\varepsilon
xy−x3−1=0xy-x^{3}-1=0y=x2y=x^{2}1/R1/R1/ε1/\varepsilon
x2y−x4−1=0x^{2}y-x^{4}-1=0y=x2y=x^{2}1/R21/R^{2}1/ε1/\sqrt{\varepsilon}
xy−x4−1=0xy-x^{4}-1=0y=x3y=x^{3}1/R1/R1/ε1/\varepsilon

Table 1. Exact generalized predictors and switching radii for the four examples.

The examples make two points clear. First, the method recovers the classical linear asymptote when that model is adequate. Second, in the genuinely nonlinear cases, the generalized predictor yields a compact far-field description that has no linear counterpart.

Some plotting-oriented examples and applications

We next include a simple visual example showing how the generalized asymptotic predictor can be used in adaptive plotting.

Consider again the curve

C: xy−x3−1=0,C:\ xy-x^{3}-1=0,

whose branch for x>0x>0 is

y=x2+1x.y=x^2+\frac{1}{x}.

Its generalized predictor is the parabola

y=x2.y=x^2.

For this branch, the discrepancy is exactly 1/x1/x, and hence the switching radius associated with a tolerance ε\varepsilon is

Rε=1ε.R_{\varepsilon}=\frac{1}{\varepsilon}.

For instance, if one prescribes an absolute far-field tolerance ε=0.2\varepsilon=0.2, then

R0.2=5.R_{0.2}=5.

Therefore, in a plotting procedure, one may evaluate the exact branch on [x0,5][x_0,5], with x0>0x_0>0 and replace it by the predictor y=x2y=x^2 for x≥5x\geq5.

This strategy has a direct computational interpretation: the exact branch may involve expensive algebraic manipulations or numerical evaluation steps, while the predictor is a low-degree polynomial that can be sampled essentially at no cost. Thus the surrogate preserves the shape at infinity and, at the same time, simplifies the remote part of the plotting process.

Figure 3 shows the exact branch and its generalized predictor. The two curves are visually almost indistinguishable in the remote region, whereas any affine model would completely miss the correct far-field geometry.

Far-field prediction for the branch $y=x^2+\frac{1}{x}$

Figure 3. Far-field prediction for the branch y=x2+1xy=x^2+\frac{1}{x}. Beyond the switching radius, the generalized asymptote y=x2y=x^2 provides a visually faithful and computationally simpler surrogate.

From the viewpoint of adaptive plotting, the rule is therefore clear:

exact branch on [x0,Rε], x0>0,predictor on [Rε,∞).\text{exact branch on }[x_0,R_{\varepsilon}],\ x_0>0,\qquad\text{predictor on }[R_{\varepsilon},\infty).

This is a minimal but concrete application of generalized asymptotes to geometric processing.

An implicit high-degree example with non-polynomial far-field predictor

We next consider an implicit algebraic curve for which the variable yy is not conveniently available as an explicit function of xx, but whose far-field behavior can still be captured by a generalized asymptotic predictor. Let

C6: y5+y−x2=0.C_6: \ y^5+y-x^2=0.

This is an implicit algebraic curve. Although, for each sufficiently large x>0x>0, the branch we are interested in is locally the graph of a real function, the equation is not naturally handled by solving explicitly for yy in terms of xx. Instead, we extract its far-field behavior directly from the implicit relation.

To determine the dominant balance, we look for a branch of the form

y∼cxαas x→∞.y\sim cx^\alpha\qquad\text{as } x\to\infty.

Substituting into the equation yields

c5x5α+cxα−x2∼0.c^5x^{5\alpha}+cx^\alpha-x^2\sim0.

The dominant terms balance when

5α=2,that is,α=25,5\alpha=2, \qquad\text{that is,} \qquad\alpha=\frac{2}{5},

and the leading coefficient must satisfy

c5=1.c^5=1.

Hence the positive real branch satisfies

y∼x2/5.y\sim x^{2/5}.

We refine this approximation by writing

y=x2/5(1+u(x)),u(x)→0.y=x^{2/5}(1+u(x)), \qquad u(x)\to0.

Substituting into y5+y−x2=0y^5+y-x^2=0, we get

x2((1+u)5−1)+x2/5(1+u)=0.x^2\left((1+u)^5-1\right)+x^{2/5}(1+u)=0.

Since uu is small, the first-order balance gives

5x2u+x2/5≈0,5x^2u+x^{2/5}\approx0,

so that

u(x)≈−15x−8/5.u(x)\approx-\frac{1}{5}x^{-8/5}.

Therefore the branch admits the asymptotic expansion

y=x2/5−15x−6/5+O(x−14/5).y=x^{2/5}-\frac{1}{5}x^{-6/5}+O(x^{-14/5}).

According to the general construction, the generalized asymptotic predictor is obtained by retaining the non-decaying term:

ψ~6(x)=x2/5.\widetilde{\psi}_6(x)=x^{2/5}.

Thus the associated generalized asymptote is the algebraic curve

A6: y5−x2=0,y>0.A_6:\ y^5-x^2=0,\qquad y>0.

Moreover,

ψ6(x)−ψ~6(x)=−15x−6/5+O(x−14/5),\psi_6(x)-\widetilde{\psi}_6(x)=-\frac{1}{5}x^{-6/5}+O(x^{-14/5}),

and hence

E∞(R)=O(R−6/5).E_{\infty}(R) = O(R^{-6/5}).

This example is important for two reasons. First, the original curve is implicit and the asymptotic model is extracted without solving it globally for yy. Second, the far-field predictor is not linear and not polynomial in xx, but it is still algebraic and therefore perfectly consistent with the generalized asymptotic framework.

For plotting purposes, it is convenient to represent the exact branch parametrically. Setting y=ty=t with t>0t>0, the exact branch is given by

x=t5+t,y=t,x = \sqrt{t^5+t}, \qquad y=t,

whereas the predictor is represented by

x=t5/2,y=t.x=t^{5/2}, \qquad y=t.

This allows a direct visual comparison between the exact branch and its far-field surrogate.

As a simple plotting rule, one may choose a tolerance and replace the exact branch by the predictor once the discrepancy becomes small enough. Since the first neglected term has size approximately (1/5)x−6/5(1/5)x^{-6/5}, the discrepancy decays faster than in the basic examples with 1/x1/x decay. The decay exponent alone does not order finite switching radii: constants and the onset of the remainder bound also matter.

Plot of the exact branch, generalized predictor, and reference switching line

Figure 4. Far-field prediction for the positive branch of the implicit curve y5+y−x2=0y^5+y-x^2=0. The exact branch is plotted parametrically as (t5+t,t)(\sqrt{t^5+t},t), while the generalized asymptotic predictor is (t5/2,t)(t^{5/2},t). In the remote region, both curves become visually indistinguishable.

Remark 4.1 (Computational remarks). The proposed framework is particularly suitable for lightweight computational tasks involving the remote geometry of a curve.

Adaptive plotting. Once the switching radius has been reached, the exact branch may be replaced by its predictor. This avoids unnecessary evaluation of the original expression while preserving the correct asymptotic shape.

Far-field clipping and windowing. The predictor provides a simple explicit description of the branch outside a compact set. This can be exploited in clipping procedures and in the construction of better viewing windows for unbounded curves.

Qualitative branch analysis. The degree and shape of the predictor immediately reveal whether the branch behaves linearly, quadratically, cubically, or according to a more general Puiseux law. This may be useful as a symbolic preprocessing stage before more expensive numerical tasks.

These observations support the use of generalized asymptotes as computational surrogates rather than as purely descriptive asymptotic objects.

A real-world motivation: far-field continuation of partially observed algebraic contours

A natural real-world motivation for generalized far-field prediction arises in geometric vision and contour processing. In many applications, a planar object boundary is modeled, either exactly or approximately, by an implicit algebraic curve

F(x,y)=0.F(x,y)=0.

Typical examples appear in computer-aided inspection, geometric reconstruction from images, and contour-based shape analysis.

In such settings, one often has reliable information only in a bounded observation window. Outside that region, the contour may be partially occluded, truncated by the image frame, or too expensive to track numerically with the full implicit representation. A practical question is then the following: how can one continue the contour in a geometrically meaningful way beyond the observed region?

The framework developed in this paper provides a simple answer. Once a real branch at infinity has been identified, its generalized asymptotic predictor supplies a low-complexity algebraic surrogate that preserves the correct remote geometry of the contour. Therefore, instead of extrapolating the contour by linear heuristics or by purely numerical continuation, one may switch to the generalized asymptotic model beyond a prescribed radius RεR_{\varepsilon}.

This has two concrete advantages. First, the continuation remains geometrically consistent with the algebraic structure of the original contour. Second, the predictor may be cheaper to evaluate after symbolic preprocessing. No general speed-up is asserted, and for some examples its algebraic degree equals that of the original curve.

For instance, in the implicit example

y3−xy+1=0,y^{3}-xy+1=0,

the visible branch may be represented exactly in the observation region, while its far-field continuation is replaced by the predictor

y2=x.y^{2}=x.

In a contour-tracking pipeline, this means that once the branch enters a remote region where only its large-scale behavior matters, the exact model can be replaced by the parabola without losing the essential geometric information.

These are potential uses under the assumption that the implicit model and the selected tail are already known. No recovery of an unknown contour from bounded or noisy observations is proved here, and no image-processing dataset is evaluated. Subject to this restriction, generalized asymptotes may serve as mathematically informed far-field surrogates in:

  • (i) contour continuation outside a finite field of view,

  • (ii) continuation of a specified algebraic model beyond a bounded window,

(iii) geometric preprocessing for clipping, plotting, and visualization of implicit shapes.

We emphasize that the present paper does not address the full image-processing pipeline. Rather, it isolates a geometric core task: the asymptotically faithful continuation of algebraic contours. In this respect, generalized asymptotes provide a mathematically grounded alternative to linear extrapolation in situations where the remote behavior is intrinsically nonlinear.

Reproducible finite-window comparisons

The accompanying dependency-free Node.js script validate_examples.cjs generates Table 2 and the CSV file. We use IEEE double arithmetic, 401 logarithmically spaced abscissae xi=4 16i/400x_i = 4\,16^{i/400} on [4,64][4,64], and bisection with 100 iterations on explicitly bracketed real roots. Here dd is the total degree of the input curve, b+b_{+} the number of real graph tails over x→+∞x \to+\infty (not the number of projective places), and dpd_p the degree of the predictor’s algebraic closure. The labels Tplus/Tminus correspond to σ=+1/−1\sigma= +1/-1 in Example 2.13; Splus/Sminus denote the slow family. For the cubic, plus/minus denote the two outer roots and small the root tending to zero.

Tailddb+b_{+}dpd_peve_vdH,Nd_{H,N}eline_{\mathrm{lin}}equade_{\mathrm{quad}}
C12112.50⋅10−12.50\cdot10^{-1}2.50⋅10−12.50\cdot10^{-1}6.95⋅1006.95\cdot10^{0}1.46⋅1021.46\cdot10^{2}
C23122.50⋅10−12.50\cdot10^{-1}2.50⋅10−12.50\cdot10^{-1}3.73⋅1033.73\cdot10^{3}1.46⋅1021.46\cdot10^{2}
C34126.25⋅10−26.25\cdot10^{-2}6.25⋅10−26.25\cdot10^{-2}3.73⋅1033.73\cdot10^{3}1.59⋅1021.59\cdot10^{2}
C44132.50⋅10−12.50\cdot10^{-1}2.50⋅10−12.50\cdot10^{-1}2.60⋅1052.60\cdot10^{5}2.27⋅1052.27\cdot10^{5}
C5plus3321.39⋅10−11.39\cdot10^{-1}1.39⋅10−11.39\cdot10^{-1}1.82⋅1011.82\cdot10^{1}3.31⋅1023.31\cdot10^{2}
C5minus3321.15⋅10−11.15\cdot10^{-1}1.15⋅10−11.15\cdot10^{-1}9.00⋅1009.00\cdot10^{0}3.90⋅1013.90\cdot10^{1}
C5small3312.54⋅10−12.54\cdot10^{-1}2.54⋅10−12.54\cdot10^{-1}9.22⋅1009.22\cdot10^{0}2.92⋅1022.92\cdot10^{2}
C65153.87⋅10−23.87\cdot10^{-2}3.87⋅10−23.87\cdot10^{-2}1.05⋅1011.05\cdot10^{1}1.08⋅1021.08\cdot10^{2}
Tplus4223.52⋅10−13.52\cdot10^{-1}3.52⋅10−13.52\cdot10^{-1}1.32⋅1011.32\cdot10^{1}1.01⋅1021.01\cdot10^{2}
Tminus4223.52⋅10−13.52\cdot10^{-1}3.52⋅10−13.52\cdot10^{-1}9.76⋅1009.76\cdot10^{0}5.83⋅1015.83\cdot10^{1}
Splus21219.33⋅10−19.33\cdot10^{-1}9.33⋅10−19.33\cdot10^{-1}8.52⋅10−18.52\cdot10^{-1}9.98⋅1009.98\cdot10^{0}
Sminus21219.33⋅10−19.33\cdot10^{-1}9.33⋅10−19.33\cdot10^{-1}8.52⋅10−18.52\cdot10^{-1}9.98⋅1009.98\cdot10^{0}

Table 2. Deterministic finite-window tests on [4,64][4,64]. Errors are sampled measurements, not certified tail errors.

The measured vertical error is ev=max⁡i∣ψ(xi)−p(xi)∣e_v = \max_i|\psi(x_i)-p(x_i)|. We also report the symmetric Hausdorff distance dH,Nd_{H,N} between the two finite point sets {(xi,ψ(xi))}\{(x_i,\psi(x_i))\} and {(xi,p(xi))}\{(x_i,p(x_i))\}, computed by exhaustive nearest-neighbour searches in both directions. It is a sampled-set distance, not a certified Hausdorff distance between continuous curves or infinite tails. In these tests it equals eve_v because of sampling and endpoint effects; it supplies no sharper geometric certificate. For continuous tails at matching abscissae the actual Hausdorff distance is at most E∞(R)E_{\infty}(R), by using the matched point in each direction.

The baselines are least-squares affine and quadratic polynomials fitted to 81 equally spaced exact branch values on [2,4][2,4], using the scaled coordinate u=x−3u=x-3. Their errors eline_{\mathrm{lin}} and equade_{\mathrm{quad}} are measured on the same test window. A fitted affine baseline is not a classical asymptote; for C1 the true classical asymptote is the predictor. These baselines measure extrapolation beyond the fitted interval and do not represent every polynomial fitting strategy. The script specifies all root brackets and verifies scaled polynomial residuals.

The non-decaying predictor is substantially more accurate than these two fitted baselines for C1–C6 and the tangential example. This reflects the use of exact far-field information and is not a fair claim of superiority for recovering unknown curves from observations alone. On the slow family, the fitted line has a smaller measured error than the uncorrected predictor. The predictor degrees also show that reduction of algebraic degree is not universal: C6 has input and predictor degree five. No runtime benefit or general robustness is inferred from this small deterministic experiment.

Conclusions

Generalized asymptotic truncations provide algebraic models for specified real tails. Their asymptotic convergence is classical; finite-tolerance replacement requires additional information. We have separated the asymptotic radius estimate from a sufficient residual certificate, distinguished branch separation from predictor truncation, and described smooth replacement with controlled positional error. Tangential and slow-converging examples show why these distinctions matter.

The tests concern exact algebraic input and finite windows. They demonstrate specific error behaviour, not a universal gain in runtime or a validated vision application. The slow example shows that the minimal truncation can be impractical and can lose to fitted baselines on a bounded window. Symbolic preprocessing, real-branch extraction and certification can be costly. Vertical discrepancies depend on the coordinate chart; they bound paired-tail Euclidean Hausdorff distance from above but need not approximate it sharply. A common predictor can merge distinct tangential tails. The smooth transition controls position, not tangent or curvature errors. The space-curve observation assumes that a compatible branch expansion is already available. General branch extraction, runtime benchmarks and data-driven contour recovery remain outside the scope of this paper.

References

  1. [1]A. Blasco, S. Pérez-Díaz, A new approach for computing the asymptotes of a parametric curve, J. Comput. Appl. Math. 364 (2020), 112350.DOI
  2. [2]E. Campo-Montalvo, M. Fernández de Sevilla, S. Pérez-Díaz, A simple formula for the computation of branches and asymptotes of curves and some applications, Comput. Aided Geom. Design 94 (2022), 102084.DOI
  3. [3]E. Campo-Montalvo, M. Fernández de Sevilla, S. Pérez-Díaz, Determining the asymptotic family of an implicit curve, Comput. Aided Geom. Design 98 (2022), 102146.DOI
  4. [4]E. Campo-Montalvo, M. Fernández de Sevilla, J.R. Magdalena-Benedicto, S. Pérez-Díaz, Some new symbolic algorithms for the computation of generalized asymptotes, Symmetry 15(1) (2023), 69.DOI
  5. [5]J.R. Magdalena-Benedicto, M. Fernández de Sevilla, S. Pérez-Díaz, Design and implementation of symbolic algorithms for the computation of generalized asymptote, Ann. Math. Artif. Intell. 91 (2023), 537–561.
  6. [6]M. Fernández de Sevilla, J.R. Magdalena-Benedicto, S. Pérez-Díaz, A survey on recent advances in the computation of asymptotes, São Paulo J. Math. Sci. 18 (2024), published online.DOI
  7. [7]M. Fernández de Sevilla, J.R. Magdalena-Benedicto, S. Pérez-Díaz, An effective algorithm for computing the asymptotes of an implicit curve, J. Comput. Appl. Math. 437 (2024), 115468.DOI
  8. [8]X.-Y. Wang, L.-Y. Shen, J.R. Magdalena-Benedicto, S. Pérez-Díaz, Infinity branches and asymptotic analysis of algebraic space curves: new techniques and applications, Comput. Aided Geom. Design 117 (2025), 102422.arxiv.org/abs/2503.03390

Paper details

vlnc:W_sjhrgu3k
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents