Skip to content

Local convergence

A numerical method's guarantee that every sufficiently nearby initial approximation converges to a specified solution.

Version
v1 · 2026-10-25 · History
Domain-specific #
14410
Domain group
Formal Sciences
Origin domain
Mathematics
Subdomain
Numerical Analysis → Mathematics

Core Idea

An iterative method is locally convergent to a specified solution when there exists a neighborhood of that solution such that every admissible initial approximation in the neighborhood generates iterates tending to the solution.[1] The claim is about an update rule, a target, and a set of starts. One successful run is evidence about that run; it does not establish the neighborhood guarantee. A statement that the iterates converge from a particular start is also weaker than the local-method claim.

Local convergence does not specify the speed of approach. Newton's method, for example, converges locally and quadratically at a simple zero under the stated regularity conditions; the rate is an additional property.[1] Nor does “local” assert that starts outside one guaranteed neighborhood fail. It says what is guaranteed inside a suitable neighborhood and leaves other starts to separate analysis.

Structural Signature

Sig role-phrases: declared update; target solution; starting neighborhood; all-starts convergence guarantee; method conditions.

  • Declared update. Specify the state space and a rule that turns an admissible approximation into its successor. Newton's scalar step and Bairstow's coefficient-pair recurrence are different rules.[1]
  • Target solution. Identify the root, fixed point, or factor representation toward which the iterates are claimed to tend. Changing the target can change whether the claim holds.
  • Starting neighborhood. Require some neighborhood around that target, rather than a hand-picked successful start. Its existence is the local quantifier; a numerical table alone cannot establish it.[1]
  • All-starts convergence guarantee. Under the theorem's hypotheses, every permitted start in that neighborhood must produce iterates approaching the specified target. If arbitrarily close admissible starts fail to do so, the asserted local property fails.[1]
  • Method conditions. State the assumptions that keep the update defined and warrant the guarantee. NIST gives Newton's simple-zero condition and a different exception for Bairstow's local quadratic result.[1]

What It Is Not

  • Not convergence of one sequence alone. A single trace may approach a limit even when no all-starts neighborhood has been shown.
  • Not a convergence-rate label. Linear, quadratic, and cubic describe orders of approach; “local” specifies where a guarantee on starting points holds.[1] NIST states these separately.
  • Not a claim of failure everywhere else. A point outside one proved neighborhood may converge, reach another solution, or fail; the local theorem alone does not decide which.
  • Not a universal claim about numerical methods. Each update, target, and set of assumptions needs its own analysis. The local property makes no claim about how often numerical methods have a broader starting-domain guarantee.

Scope of Application

The reviewed evidence establishes this property in numerical iteration near a specified solution. NIST treats nonlinear scalar root iteration, including Newton's method, and polynomial-factor iteration, including Bairstow's method.[1] A fixed-point method can also be investigated through the same neighborhood question, but a named method's particular update and hypotheses must be checked before calling it locally convergent.[1]

The scope is method and target specific. The same method can face different roots, multiplicities, or representations. NIST notes that Newton's convergence at a multiple zero has a different rate from its simple-zero result; changing the representation of its worked transcendental equation changes the observed successful-start range.[1]

Clarity

For a purported local-convergence claim, write down the update, target, admissible starting domain, theorem conditions, and the exact conclusion. This prevents the frequent slide from “the displayed run worked” to “all nearby runs work.” NIST's Newton table begins at 4.65 and approaches the required root, while its separate simple-zero statement supplies the local theorem.[1]

NIST reports that some more distant Newton starts do not converge to the required root in that example.[1] Those reported failure thresholds are not themselves a proof that every start between them succeeds. The distinction matters because a numerical search for a useful initial guess and a mathematical neighborhood guarantee answer different questions.

Manages Complexity

An iterative solver can generate many trajectories as its starting state varies. The local-convergence property compresses that family into a conditional statement: some neighborhood of a fixed target is guaranteed to converge under a specified update and assumptions. One then examines whether a planned start is covered, rather than mistaking a few numerical experiments for a theorem.

The compression has a strict limit. A local theorem need not give a practical radius, and an observed failure far away need not locate the boundary of every convergent start. In NIST's Newton example, a different expression for the same root-finding task enlarges the reported successful-start range; the article does not thereby claim a universal optimal representation or a complete basin.[1]

Abstract Reasoning

The useful inference is conditional. If a theorem establishes local convergence of the chosen update to the desired solution and an initial approximation satisfies its neighborhood conditions, the generated iterates tend to that solution. If only one trajectory has been computed, the analyst may report that trajectory but cannot promote it to an all-starts assertion. If a proposed start lies outside the proved region, the correct conclusion is that this theorem does not cover it; failure does not follow.

When a case includes a rate result, keep that conclusion separate. A local quadratic theorem explains rapid error reduction after the iterates are in the relevant regime; it does not turn an arbitrary initial approximation into a safe one.[1]

Knowledge Transfer

Within numerical analysis, the same question can be asked of unlike updates: What is the target, what neighborhood of starts is covered, and which assumptions make convergence follow? Newton's scalar root approximation and Bairstow's factor-coefficient iteration require different formulas, yet both can be evaluated through that common local guarantee.[1]

Only the question transfers automatically. Newton's simple-zero theorem does not prove Bairstow's factor result, and Bairstow's exception does not amend Newton's conditions. Terms such as “local convergence” in network routing or other fields should not be treated as literal instances of this numerical-method property without the same update, target, and neighborhood quantifier.

Examples

Canonical: Newton iteration for a transcendental root

NIST studies Newton's method on f(x) = x − tan(x) near its first positive nonzero root.[1] Starting at x₀ = 4.65, the tabulated iterates approach approximately 4.4934094579.[1] It also reports starting-value thresholds outside which this update does not reach the required zero.[1] The local-convergence conclusion rests on the simple-zero result, not on extrapolating from the one table run or treating the space between failure thresholds as a proved success interval.[1]

Mapped back: declared update → Newton's scalar quotient step; target → the specified nonzero root; neighborhood → one whose existence follows from the simple-zero local result, without claiming its full boundary; all-starts guarantee → convergence to that root for starts in such a neighborhood; conditions → a simple zero and the needed smoothness and defined derivative.[1]

Applied: Bairstow iteration for a quadratic factor

NIST applies Bairstow's coefficient recurrences to p(z) = z⁴ − 2z² + 1.[1] From (s₀,t₀) = (7/4,−1/2), the tabulated pair approaches (2,−1), representing the factor (z−1)².[1] Its separate local-quadratic statement has an exception when the wanted quadratic factor is a multiple factor of the quotient.[1] The repeated root inside (z−1)² is not, by itself, that quotient-factor exception.[1]

Mapped back: declared update → Bairstow's recurrence on the coefficient pair; target → (2,−1); neighborhood → the local region warranted by the stated theorem under its factor condition; all-starts guarantee → convergence toward the target from sufficiently near admissible pairs; conditions → a nonexceptional wanted factor.[1] The table illustrates one trajectory and does not independently prove the neighborhood.

Structural Tensions

There is no general rule here that faster local convergence must require a smaller successful-start region. NIST's alternative representation of the Newton example reports both faster convergence and a larger successful-start range.[1] The practical failure mode is to choose a start on the strength of a fast local rate while leaving the start conditions unchecked. Diagnostic: Which neighborhood guarantee is actually proved for this update and target, and is the proposed start within it?

Structural–Framed Character

Placement: strongly structural within a domain-specific mathematical setting. Vocabulary travel: “neighborhood,” “iterate,” “target,” and “convergence” retain formal meanings across the reviewed numerical cases. Evaluative weight: the definition does not express a preference for one method; “fast” or “robust” is an additional judgment. Human-practice dependence: people choose solvers and initial guesses, but the theorem's truth is not constituted by that choice. Institutional origin: the property is mathematical, not an institutional rule. Import versus recognition: other fields may use the same words, but literal recognition requires an update, specified target, and all-starts neighborhood guarantee. Its character: strongly structural as a conditional mathematical guarantee, while its named method-and-target identity remains domain-specific to numerical iteration.

Structural Core vs. Domain Accent

The core is a conditional convergence guarantee for a family of trajectories indexed by initial state. The live Prime Iteration supplies the repeated-update component, and the live Prime Convergence supplies the limit-approach component. This entry adds an existential neighborhood whose admissible starts all converge under a declared numerical update toward one target. Whether that stronger neighborhood-indexed skeleton qualifies as a future Prime across domains remains an open classification question; these two numerical cases alone do not establish it.

The domain accent is a numerical update used to solve a mathematical problem, with conditions on its target and initial approximation. NIST's scalar root and polynomial-factor examples show the property across distinct numerical procedures.[1] A broader Prime claim would need evidence that the same identity and quantifier survive outside numerical iteration.

The live Prime Convergence supplies the broader idea that an indexed sequence approaches a limit; Iteration supplies repeated state updates. In the present property, those structures are necessary for the claim to make sense, but neither Prime alone specifies a method-and-target neighborhood guarantee. The live domain-specific Iterative method names the numerical procedure on which this property is evaluated. These are conceptual relations; the property and procedure have distinct identities.

Neighborhood in Abstraction Space

Local convergence sits in a sparse region of the domain-specific corpus (63rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Dynamical Systems & Differential Structures (38 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-25

Not to Be Confused With

  • A convergence order. Quadratic convergence says how error shrinks near the target; local convergence says where a guarantee over starting states holds.[1]
  • The complete basin of attraction. A local theorem supplies at least one guaranteed neighborhood; it need not map every start that succeeds.
  • Global convergence. A broader starting-domain claim needs its own specified domain and proof. “Local” does not mean “only local.”
  • Γ-convergence. The live domain-specific file named convergence concerns convergence of variational functionals, not this property of an iterative numerical update.

References

[1] National Institute of Standards and Technology. “Nonlinear Equations”. Digital Library of Mathematical Functions, §3.8, especially §§3.8(i), 3.8(ii), and 3.8(iv), with the cited equations and worked tables. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x ↩y ↩z ↩27 ↩28