Resummation And Nonperturbative Extrapolation

When a useful local expansion stops behaving like an ordinary convergent approximation, diagnose its information content, transform it into a more revealing representation, and validate any continuation before trusting it.

Version

v1 · 2026-08-24 · History

Solution archetype #

885

Problem family

Correctness, Conformance & Formal Validity Failure

Problem subfamily

Computational Decidability & Bounded Approximation

Essence

Resummation and Nonperturbative Extrapolation is the disciplined recovery of information from a local approximation after ordinary term-by-term use has stopped being trustworthy. The source may be a perturbation series, an asymptotic expansion, a moment sequence, a high-temperature expansion, a hierarchy of corrections, or another structured local representation. Its partial sums may diverge, oscillate, converge too slowly, or break down before the regime that matters. The central move is not “calculate more terms.” It is to ask what the available terms still encode, change the representation used to read that information, and qualify every reconstructed result by structural constraints and independent evidence.

A divergent series can still be useful. Divergence may reveal coefficient growth, nearby singularities, competing scales, or the size of effects that ordinary powers cannot express. A rational approximant can encode a pole or branch-like obstruction more economically than a polynomial. A Borel transform can remove factorial growth before reconstruction. A conformal map can make known analytic structure computationally usable. Matched asymptotics can acknowledge that no single local representation is uniform across regions. These techniques differ, but they instantiate one transferable intervention: diagnose the failure, select a representation consistent with that diagnosis, compare admissible continuations, reject artifacts, and state what remains outside the recovered sector.

The archetype is deliberately more demanding than “use Padé” or “apply Borel summation.” A method by itself does not tell the analyst whether the coefficient sequence is trustworthy, which singularities are plausible, whether a branch choice is justified, or whether agreement is only the result of shared assumptions. The mature output is therefore a governed inference package: the reconstructed estimate, the continuation ensemble, the rejected alternatives, the benchmark results, a decomposed ambiguity budget, the validity envelope, and the nonperturbative completion boundary.

This pattern is useful because two opposite errors are common. One is premature abandonment: treating a nonconvergent expansion as meaningless even when its coefficients contain stable structural information. The other is mathematical overreach: treating a smooth transformed curve as an exact global answer. Resummation governance occupies the narrow, valuable space between those errors.

Compression statement

Begin with a sequence, local model, or perturbative expansion that contains real structural information but is slowly convergent, divergent, asymptotic, oscillatory, or invalid near the desired regime. Characterize coefficient growth and singularity clues; choose a transform whose assumptions fit those clues; impose known invariants, symmetries, limits, and analytic structure; compare multiple reconstruction routes; benchmark against exact or numerical cases; report transform dependence and ambiguity; and route irreducible discrepancies to a genuinely nonperturbative completion rather than hiding them inside a smooth extrapolated curve.

Canonical formula: local_expansion + failure_diagnosis + structural_constraints + representation_transform + continuation_ensemble + independent_benchmarks -> qualified_resummed_estimate + ambiguity_budget + nonperturbative_completion_route

When This Archetype Applies

Partial catalog groundingSome structural conditions are represented by existing abstractions, but no sufficient condition set is fully represented.

Diagnostic problem

A tractable local construction produces coefficients, correction terms, moments, or samples that are informative near a reference regime but cannot be used reliably by naive truncation or direct summation where an answer is needed. Adding terms may worsen the estimate; apparent convergence may be pre-asymptotic; a target may lie beyond a convergence disk or across a singular boundary; or effects invisible at every ordinary perturbative order may control the residual. The analyst must recover only the information the sequence supports without confusing a representation-dependent continuation with an exact global solution.

What this problem means

The structural problem is a mismatch between information and representation. A local expansion contains organized evidence about a target quantity, but the representation in which that evidence was produced is poorly suited to direct use where the answer is needed. Termwise addition assumes that later terms behave like smaller corrections. Near a singularity, at strong coupling, or in an asymptotic regime, that assumption can fail even though the coefficients remain highly informative.

Several distinct failures can look alike numerically. Slow convergence means the series is valid but inefficient. Asymptotic divergence means finite truncations can be useful even though the infinite sum does not exist in the ordinary sense. A finite convergence radius means a singularity blocks the local power-series representation. Oscillation may indicate alternating recoverable structure or unresolved competing scales. Factorial growth may make Borel transformation appropriate, while same-sign growth and positive-axis singularities may create a genuine reconstruction ambiguity. A boundary layer indicates that the dominant balance changes by region. A phase transition may indicate that one analytic branch cannot represent the target state at all.

Treating these as one generic “convergence problem” leads to method mismatch. Sequence acceleration cannot cure a missing phase. A rational approximation can invent pole-zero defects. Borel integration can be ambiguous when singularities lie on the integration path. A high-order polynomial can become smoother while becoming less trustworthy. The archetype therefore begins with classification of the breakdown, not with selection of a favorite transform.

There is also a governance problem. Analysts often explore many approximants, scale choices, maps, and orders. If only the successful-looking result survives into the report, transform selection becomes an unrecorded researcher degree of freedom. The result may be reproducible computationally yet unauditable epistemically. A continuation ensemble and rejection log are structural safeguards against this form of selection bias.

Finally, some information is not contained in the perturbative sector at all. Contributions that are exponentially small in the expansion parameter can vanish to every algebraic order and still matter in the target regime. Topological sectors, tunneling-like effects, new saddles, or new boundary data may be required. The nonperturbative completion boundary is not a confession of failure. It is the line between information recovered from the local sequence and information that must enter from elsewhere.

Applicability expression3 distinct conditions

Usable local seriesandPartial sums misbehaveandLocal representation fails

Algebraic1∧2∧3

groundedpartly groundedopen

3 conditions, all required.

3Required in every casenumbered 1–3

These hold no matter which pattern applies.

1

Usable local series · grounded · any one of 2

A baseline-plus-correction or local-series construction exists and has produced several usable terms. more

a

primePerturbation Theory— A technique for handling an intractable problem by splitting it into an exactly solvable baseline plus a small correction, then expanding the quantities of interest as a power series in that small parameter.

or — any one of these establishes this case

b

primeProgressive Refinement from Core Model— Incremental refinement.

and — also required

2

Partial sums misbehave · open

Direct partial sums converge too slowly, oscillate, diverge, or show an optimal truncation point. more

and — also required

3

Local representation fails · open

The target lies near a strong-coupling, low-temperature, critical, boundary-layer, or other regime where the local representation degrades. more

Other requirements and context (2)

Why these sit outside the expression

Deployment constraint — it constrains how the intervention must be deployed, not the situation that calls for it.

Solution feasibility — it describes whether the intervention can work, not whether the diagnostic problem exists.

• Deployment constraintKnown symmetries, conservation laws, exact limits, positivity conditions, singularity locations, or benchmark cases can constrain continuation. more

• Solution feasibilityA useful qualified estimate is possible before an exact or fully nonperturbative solution is available. more

Coverage

1 of 3 conditions grounded · 2 open.

Read the methodology · Download the trigger-logic data

When to Use This Archetype

Use this archetype when a local construction has yielded enough structured information to support diagnosis, yet its naive numerical use is unreliable in the regime that matters. Several terms are usually needed. With only one or two noisy coefficients, a transform can express more analyst choice than source information. The case becomes stronger when coefficient signs, ratios, growth, or order dependence display a recognizable pattern and when exact limits, symmetry, dimensional scaling, positivity, or independent numerical cases can constrain the admissible reconstructions.

The most characteristic trigger is an asymptotic sequence: early orders improve the estimate, but later orders eventually grow and degrade it. The useful question is then not whether the infinite series converges, but how much information is recoverable at optimal truncation and whether a transform can organize the late-order behavior. Another trigger is a finite radius of convergence: the local series is valid near a reference point, while the target lies nearer a singularity or outside the disk. Rational or conformal continuation may help, but only if the inferred analytic structure is tested rather than assumed.

Use the pattern near strong-coupling, critical, low-temperature, boundary-layer, or long-time regimes when a weak- or local-regime description remains the best available analytical source. It also applies when a surrogate or reduced model is trustworthy locally but high-fidelity evaluation in the target regime is expensive. The terminology changes across fields, yet the intervention remains stable: retain provenance, diagnose the breakdown, impose structure, build multiple admissible continuations, and compare them against evidence that was not simply fitted away.

Do not use the archetype merely because a calculation is difficult. If the expansion converges normally and an ordinary remainder estimate is available, the additional machinery can add avoidable model dependence. If the target regime has entirely new degrees of freedom, a different phase, or a discontinuity with no bridge from the local sector, resummation may be unable to recover the missing information. In such a case the correct output is a completion boundary and a handoff, not an extrapolated answer.

The intended use also matters. A qualified estimate may be adequate for hypothesis generation, design-space narrowing, or prioritizing a direct simulation. It may be inadequate for certification, legal compliance, or safety margins. The action–confidence boundary must be decided before numerical elegance turns an exploratory reconstruction into an unsupported operational commitment.

Structural Problem

The structural problem is a mismatch between information and representation. A local expansion contains organized evidence about a target quantity, but the representation in which that evidence was produced is poorly suited to direct use where the answer is needed. Termwise addition assumes that later terms behave like smaller corrections. Near a singularity, at strong coupling, or in an asymptotic regime, that assumption can fail even though the coefficients remain highly informative.

Several distinct failures can look alike numerically. Slow convergence means the series is valid but inefficient. Asymptotic divergence means finite truncations can be useful even though the infinite sum does not exist in the ordinary sense. A finite convergence radius means a singularity blocks the local power-series representation. Oscillation may indicate alternating recoverable structure or unresolved competing scales. Factorial growth may make Borel transformation appropriate, while same-sign growth and positive-axis singularities may create a genuine reconstruction ambiguity. A boundary layer indicates that the dominant balance changes by region. A phase transition may indicate that one analytic branch cannot represent the target state at all.

Treating these as one generic “convergence problem” leads to method mismatch. Sequence acceleration cannot cure a missing phase. A rational approximation can invent pole-zero defects. Borel integration can be ambiguous when singularities lie on the integration path. A high-order polynomial can become smoother while becoming less trustworthy. The archetype therefore begins with classification of the breakdown, not with selection of a favorite transform.

There is also a governance problem. Analysts often explore many approximants, scale choices, maps, and orders. If only the successful-looking result survives into the report, transform selection becomes an unrecorded researcher degree of freedom. The result may be reproducible computationally yet unauditable epistemically. A continuation ensemble and rejection log are structural safeguards against this form of selection bias.

Finally, some information is not contained in the perturbative sector at all. Contributions that are exponentially small in the expansion parameter can vanish to every algebraic order and still matter in the target regime. Topological sectors, tunneling-like effects, new saddles, or new boundary data may be required. The nonperturbative completion boundary is not a confession of failure. It is the line between information recovered from the local sequence and information that must enter from elsewhere.

Intervention Logic

The intervention changes the question from “What does the next partial sum say?” to “What function classes and continuations remain compatible with the sequence and everything else we know?” That shift makes coefficient behavior, analytic structure, invariants, and benchmarks active design inputs. The process is constructive, but it is also eliminative: much of its reliability comes from ruling out reconstructions that violate known structure or behave unstably under legitimate variation.

First, freeze the local information object. Record coefficient provenance, order, scheme, normalization, numerical precision, and any preprocessing. A coefficient computed in one convention cannot be mixed casually with another. Then define the target: observable, parameter range, required precision, and permitted decision. A method selected for an intermediate regime may not support a claim at a singular boundary.

Second, establish the naive baseline. Partial sums, optimal truncations, and simple extrapolations show what the transform must improve. Without that baseline, sophistication is rewarded for existing rather than for adding verified value. Examine coefficient ratios, sign patterns, apparent factorial growth, order-to-order drift, and sensitivity to removing the last term. Where theory supplies likely singularities or scaling, record those inputs separately from facts inferred from the coefficients.

Third, define admissibility before looking for the preferred answer. An admissible reconstruction may need to preserve positivity, monotonicity, symmetry, dimensional scaling, known weak- and strong-regime limits, or branch structure. These constraints should be classified by strength: exact, theoretically expected, empirically supported, or convenient. Convenience constraints deserve sensitivity analysis, not the authority of invariants.

Fourth, choose a transform family that targets the diagnosed failure. Use rational continuation when pole-like or meromorphic structure is plausible and defects can be tested. Use Borel-style methods when factorial growth is supported. Use conformal mapping when singularity geometry is sufficiently known. Use matched asymptotics when distinct regions have distinct dominant balances. Use renormalization-group improvement when large logarithms or scale dependence are the primary obstruction. No method earns priority merely through prestige.

Fifth, build an ensemble. Vary orders, rational numerator and denominator degrees, conformal parameters, scales, Borel prescriptions, and other legitimate choices. Reject members only under declared rules: unstable poles, gross coefficient sensitivity, violation of exact constraints, failure on held-out benchmarks, or inconsistency with the intended branch. Agreement among survivors is informative, but the ensemble spread is not automatically a calibrated probability interval.

Sixth, test and decompose uncertainty. Compare with exact or high-fidelity cases, preferably outside the tuning set. Distinguish coefficient error from truncation uncertainty, transform-family spread, parameter dependence, analytic-continuation ambiguity, and missing-sector uncertainty. When the residual has a scale characteristic of nonperturbative effects, state that explicitly and route to an independent completion rather than inflating or shrinking an undifferentiated error bar.

The loop ends with a validity envelope and reopening rule. A useful result states where it is stable, which constraints it satisfies, which decisions it can support, what remains unknown, and what new coefficient, benchmark, singularity result, or direct calculation would require revision.

Key Components

Component	Description
Local Expansion Record ↗	The Local Expansion Record is the evidence ledger. It contains coefficients or correction terms, the expansion variable, normalization, computational scheme, numerical precision, and provenance. If coefficients were inferred, fitted, or combined across sources, those transformations must be visible. A reconstruction built from undocumented coefficients cannot be distinguished from a curve fit after the fact.
Target Regime Statement ↗	The target is not simply “larger parameter values.” It identifies a quantity, interval or limit, desired precision, and use. The same continuation may be acceptable for locating a qualitative crossover and unacceptable for estimating a safety margin. Declaring the target early prevents a method from being selected because it looks stable where no decision is actually needed.
Large-Order or Singularity Diagnostic ↗	This diagnostic is one of the archetype's distinctive components. It records coefficient ratios, signs, growth, optimal truncation, suspected singularities, and the evidence for each interpretation. Its purpose is not to prove an analytic structure from a short sequence. It narrows the set of defensible transforms and marks where assumptions enter.
Structural Constraint Register ↗	The register holds exact limits, symmetries, conservation rules, dimensional scaling, positivity, monotonicity, branch conditions, and other admissibility constraints. Each item should identify its source and strength. A theoretically motivated singularity prior is not the same as an exact symmetry. Keeping those categories separate makes sensitivity analysis possible.
Resummation Transform Choice ↗	Transform choice records what family is used, which failure it addresses, and which assumptions it introduces. It should name rejected alternatives and explain why they are weaker fits. This record makes it possible to distinguish evidence-based method selection from transform shopping.
Continuation Ensemble ↗	An ensemble preserves legitimate variation across order, approximant shape, scale, mapping, and prescription. Its purpose is not to average every curve. It exposes stable features, fragile features, clusters, and method-dependent branches. The ensemble should include failures long enough for the rejection logic to be audited.
Artifact Rejection Log ↗	The log captures spurious poles, near-canceling pole-zero defects, wrong-branch behavior, constraint violations, sensitivity to one coefficient, and benchmark failures. It prevents an attractive failed approximant from returning later under a new filename or parameterization.
Benchmark Case Set ↗	Benchmarks may be exact values, solvable limits, numerical simulations, experimental measurements, dual descriptions, or higher-fidelity models. At least some should remain independent of tuning. Benchmarks should test the approach toward the target, not only the easy local regime where every method reproduces the supplied coefficients.
Ambiguity and Uncertainty Budget ↗	This budget separates uncertainties with different meanings. Coefficient noise can sometimes be propagated statistically. Transform spread is conditional on an admissible method set. Prescription dependence may represent a structural ambiguity. Missing nonperturbative sectors are not ordinary sampling error. Combining them into one precise-looking standard deviation destroys useful information.
Nonperturbative Completion Boundary ↗	The boundary states what the perturbative or local sector cannot determine and identifies possible additional inputs: direct numerics, exact constraints, transseries sectors, strong-regime models, boundary data, or experiments. It is the main safeguard against claiming that representation change can recover information never present in the source.
Validity and Reopening Rule ↗	The final component defines where the reconstruction may be used and what reopens it. New coefficients, a newly located singularity, a high-fidelity disagreement, a scheme change, or evidence of a phase boundary are typical triggers. A continuation without a reopening rule tends to harden into an orphaned canonical curve.

Common Mechanisms

10 documented mechanisms across 2 implementation forms.

The grouping reflects forms represented among the mechanisms currently documented for this archetype; an absent form is not necessarily an impossible implementation.

Analysis, Modeling & Optimization · 9 mechanisms

• Borel Resummation — Removes factorial coefficient growth in a Borel transform and reconstructs a value through a justified inverse integral, treating its contour ambiguity as a nonperturbative signal.
• Borel–Padé Resummation — Continues a Borel-transformed series by rational (Padé) approximation from finitely many coefficients, reading its Borel-plane poles to reconstruct a divergent expansion.
• Conformal Borel Mapping — Maps a cut Borel domain to a disk to improve transformed-series convergence, using an assumed singularity geometry as an explicit, testable input.
• High-Temperature Series Resummation — Continues a high-temperature expansion toward the critical point by biasing the continuation with known critical scaling, and stops where new critical structure takes over.
• Matched Asymptotic Expansion — Builds separate approximations in regions with different dominant balances and joins them through a consistent overlap into one composite.
• Padé Approximant — Replaces a truncated power series with a rational function that reproduces its coefficients, extending usefulness toward poles while flagging spurious ones across an ensemble of shapes.
• Renormalization-Group Improvement — Reorganizes scale-dependent and logarithmic terms using a flow equation, resumming large logarithms instead of computing more fixed orders.
• Sequence Acceleration Transform — Recombines a sequence's own partial sums to cancel a diagnosed leading remainder or damp oscillation, without changing representation.
• Strong-Coupling Extrapolation Check — Tests a reconstruction's strong-regime behavior against a single known exact limit of the same system, passing or failing its extension into that regime.

Experiment, Test & Rehearsal · 1 mechanism

• Exact-or-Numerical Benchmark — Supplies independent exact values and high-fidelity numerical results, held out from tuning, to test a reconstruction and cross-check competing methods near the target.

Parameter / Tuning Dimensions

The first tuning dimension is information depth: how many coefficients are available, how accurately they are known, and whether the observed order reaches genuine large-order behavior. Short sequences require stronger priors and wider uncertainty. Long sequences are not automatically safe if high-order coefficients are noisy or generated under inconsistent schemes.

The second dimension is transform family. Rational, Borel, conformal, acceleration, flow-based, and matched-regime methods assume different failure structures. Family selection should be treated as a model choice. When several families are defensible, maintain them as separate ensemble strata rather than collapsing all outputs into one average.

Within a family, order and shape matter. Padé numerator and denominator degrees, Borel–Padé orders, conformal-map parameters, sequence windows, matching locations, and truncation orders can all change results. Legitimate variation defines a stability surface. A narrow optimum chosen after seeing a target value should be flagged as tuning, not validation.

Scale and scheme form another dimension. Renormalization scale, normalization convention, coordinate choice, discretization, and coefficient preprocessing can alter apparent convergence. The goal is not to find the scheme with the smallest displayed error. It is to show that the substantive conclusion is stable under the range of schemes the problem permits, or to include the dependence explicitly.

Constraint strength must also be tuned openly. Exact symmetries and limits may be enforced. Theoretical expectations, estimated singularities, monotonicity assumptions, or strong-regime scaling should be varied or removed to measure their influence. An anchor that determines the answer more strongly than the coefficients should be described as a model input.

Benchmark allocation is a governance parameter. Decide which cases tune transforms, which reject them, and which remain held out. With very few benchmarks, rotate leave-one-out checks and avoid calling the residual an out-of-sample error. The benchmark set should include difficult approach-to-target cases, not only easy points near the expansion origin.

Finally, set a precision stop rule. Continue adding transform variants only while they improve constraint satisfaction, held-out performance, or stability relevant to the decision. More methods can widen awareness without narrowing uncertainty. That is sometimes the correct result, and the process should stop rather than optimize the uncertainty away.

Invariants to Preserve

Coefficient provenance is non-negotiable. Every term needs a definition, convention, scheme, and uncertainty statement sufficient for another analyst to reconstruct the input. Mixing coefficients across normalizations or silently correcting signs destroys the evidential chain.

Exact structure outranks cosmetic fit. A reconstruction that violates a conservation law, symmetry, positivity condition, boundary condition, or exact limit is not rescued by visual smoothness. If an apparent violation is theoretically possible, the reason must be explicit and reviewed as a change to the constraint register.

Local evidence must remain local in its claims. Continuation expands the usable inference range; it does not prove that the original function class or phase persists globally. Every output retains a validity envelope and branch statement.

Method dependence stays visible. Analysts must be able to see which features are common across orders, transforms, scales, and prescriptions and which appear only under one choice. Selection rules and rejection reasons are retained even when the final report emphasizes one estimate.

Tuning evidence and validation evidence remain distinguishable. When separation is impossible, disclose the reuse and treat validation as internal. This invariant prevents benchmarks from becoming both the source of a constraint and the proof that the constraint worked.

Nonperturbative ambiguity cannot be renamed as statistical noise. If different admissible prescriptions differ by a scale suggestive of a missing sector, that difference is a structural clue. It should be connected to completion logic rather than divided by the number of approximants.

The action supported by the estimate remains proportionate to evidence. A useful exploratory continuation can guide where to simulate next without being adequate for certification. The output should preserve this boundary even if downstream users prefer a single number.

Target Outcomes

The primary outcome is not maximum extrapolation distance. It is a qualified increase in usable information relative to naive truncation, demonstrated by stability, constraint satisfaction, and independent benchmarks. Sometimes the recovered range is modest. A smaller defensible interval is better than a dramatic unsupported continuation.

A second outcome is diagnostic clarity. The process should reveal whether failure is dominated by slow convergence, factorial growth, a nearby singularity, a scale mismatch, a regional dominant-balance change, or a missing nonperturbative sector. That classification helps choose the next calculation or experiment even when no final estimate is accepted.

A third outcome is an auditable method-selection record. Another analyst should be able to understand why a transform was chosen, which alternatives were tried, what was rejected, and what assumptions control the result. This reduces repeated rediscovery of failed approximants and makes expert disagreement more specific.

The ambiguity budget is itself a target outcome. Instead of one opaque interval, users receive a map of coefficient error, order dependence, transform spread, scale and scheme dependence, benchmark discrepancy, continuation prescription, and completion uncertainty. Different uncertainties invite different next actions.

Finally, the process should produce a rational handoff. It may endorse a resummed estimate, request more coefficients, prioritize a benchmark, introduce a nonperturbative sector, switch to direct numerical solution, or declare the target unsupported. A negative handoff can be a high-quality outcome when it prevents false precision.

Tradeoffs

Structural constraints reduce variance but increase dependence on theoretical commitments. An exact symmetry is usually beneficial; a guessed singularity map can dominate a short sequence. The tradeoff is managed by labeling constraint strength and rerunning the ensemble under plausible alternatives.

Transform diversity improves robustness assessment but consumes computation and interpretation time. Ten closely related Padé shapes do not provide the same independence as two genuinely different representation families. Diversity should be measured by assumption structure, not method count.

Aggressive continuation reaches farther but magnifies model dependence. Conservative continuation may appear less useful yet better support decisions. The target statement should determine how far to continue and which uncertainty level is tolerable.

Nonperturbative completion improves fidelity while adding new sectors, parameters, boundary conditions, or simulations. The new information may be expensive and less analytically transparent. The archetype makes that cost visible so resummation is not pursued indefinitely merely to avoid changing models.

Strong anchors can increase precision but reduce discovery. If a known strong-regime form is imposed, the reconstruction may interpolate between assumptions rather than predict. That can still be valuable, but the output should call it constrained interpolation or completion, not independent extrapolative confirmation.

Rich reporting aids audit but burdens downstream users. A concise decision summary can sit above the ensemble, rejection log, and ambiguity decomposition, but the underlying records must remain available. Compression should remove repetition, not uncertainty.

Failure Modes

Transform shopping occurs when many admissible-looking methods are tried and only the preferred answer is reported. The cause is an unconstrained researcher degree of freedom. The mitigation is to define admissibility and rejection criteria early, retain the ensemble, and show sensitivity to reasonable alternatives.

False convergence arises when several early orders agree before the large-order regime begins. It is especially dangerous because conventional stability plots look reassuring. Leave-one-order-out reconstruction, coefficient-ratio analysis, synthetic recovery tests, and a pre-asymptotic uncertainty term reduce the risk.

Spurious poles and pole-zero defects occur in rational approximants. A pole can look like a physical singularity or block a useful interval even when it is numerically unstable. Track pole locations across order and shape, inspect near cancellations, and require independent support before interpreting singularities substantively.

Branch and contour suppression occurs when one analytic continuation path or inverse-Borel prescription is treated as canonical without explanation. Report admissible paths and their differences. If a prescription choice is fixed by a physical boundary condition, cite that condition rather than presenting the numerical result as path-independent.

Phase-boundary overreach happens when an analytic continuation smooths across a regime where the relevant state changes. A local branch may continue mathematically while ceasing to describe the intended object. Require new-regime evidence, branch identification, and direct checks; if these are absent, stop at the boundary.

Constraint laundering encodes a desired answer as an “exact” limit, exponent, monotonicity rule, or singularity prior. Classify each constraint by evidential status and remove non-exact constraints in sensitivity runs. A reconstruction that fails without a convenience prior is prior-driven.

Benchmark leakage uses the same evidence to tune and certify a method. Reserve held-out cases when possible, or disclose that reported error is in-sample. With scarce data, rotate benchmarks and widen rather than narrow the uncertainty.

Nonperturbative omission presents a resummed perturbative sector as complete even when residual scales or theory indicate invisible contributions. Preserve the completion boundary, compare prescription ambiguity with expected missing-sector scales, and seek independent data.

Precision theater reports many digits because a computation is numerically stable. Numerical precision is not structural certainty. Round results to the uncertainty justified by coefficient quality, method dependence, and benchmarks, and describe qualitative conclusions separately.

Neighbor Distinctions

Solvable Baseline Decomposition is the closest parent and handoff. It chooses a tractable reference case, defines departures, computes ordered corrections, and tests their validity. Resummation and Nonperturbative Extrapolation becomes primary only when the correction representation itself must be reorganized or continued. If ordinary low-order correction remains adequate, stay with the parent.

Bounded Approximation is broader and often simpler. It constructs an approximation with a useful error bound. This archetype is needed when ordinary error behavior is precisely what has failed and the analyst must infer from divergent, asymptotic, or analytically obstructed structure. If a certified bound already governs the target interval, resummation may be unnecessary.

Perturbation Testing changes a system or input to learn sensitivity, robustness, or causal response. The present archetype may consume perturbative coefficients, but it transforms a formal representation to estimate a quantity. An experimental stress probe and a Borel transform solve different problems.

Progressive Fidelity Increase adds model detail, resolution, or computational expense in stages. Resummation can extract more from the same local information without adding model fidelity. Conversely, when missing degrees of freedom dominate, progressive fidelity or a new model is the correct handoff.

Generic analytic continuation is a mechanism-level operation. It does not specify how to diagnose the source, constrain branches, compare methods, reject artifacts, validate results, or bound nonperturbative omissions. Those governance stages define the archetype.

Convergence and Validity Range Assessment remains a possible neighbor. Its focus would be deciding whether a sequence or approximation is trustworthy, independent of constructing a rescued estimate. In the present archetype, assessment is an early and recurring stage serving a representation-changing intervention.

Cross-Domain Examples

Quantum field theory

A weak-coupling calculation yields coefficients that grow factorially. The team diagnoses sign and ratio behavior, constructs Borel–Padé and conformally mapped alternatives, records positive-axis singularity assumptions, and compares prescriptions. Lattice or exact-limit information is reserved for checking. The final estimate includes prescription dependence and identifies a possible missing nonperturbative sector rather than presenting one resummed value as exact.

Statistical mechanics

A high-temperature series is informative far from a critical point but fails as criticality approaches. Known scaling or exponent information constrains biased approximants. Several rational and transformed continuations are compared, with stability tested under exponent uncertainty. The validity envelope stops before a regime where the assumed critical structure no longer applies.

Fluid mechanics

A regular expansion fails in a thin boundary layer. Separate outer and inner solutions are constructed using their appropriate dominant balances and matched in an overlap. The overlap and boundary conditions serve as structural constraints. If no consistent overlap exists, the process reports a failed completion rather than forcing a composite solution.

Engineering approximation

A reduced model calibrated at low load must inform operation nearer a nonlinear regime change. The coefficient sequence and physical limits constrain rational and sequence-based continuations. High-fidelity simulations at selected intermediate points are held out. The reconstructed model is permitted for design-space screening but not for certification outside the validated envelope.

Numerical surrogate extrapolation

A spectral or local surrogate has accurate coefficients near a reference state but loses stability toward the edge of a costly simulation domain. Multiple transformed reconstructions are compared with sparse direct solves. Method spread becomes an acquisition signal: the next high-fidelity run is placed where admissible continuations disagree most, turning ambiguity into a targeted learning plan.

Across these domains, the mathematical mechanisms differ. The recurring structure is the same: local information, diagnosed failure, representation change, structural constraint, ensemble comparison, artifact rejection, independent benchmark, uncertainty decomposition, and completion boundary.

Non-Examples

A high-degree polynomial fit through sparse observations is not this archetype. It may extrapolate, but it lacks a structured local expansion, failure diagnosis, transform rationale, invariant register, and independent validation loop.

Adding a second- or third-order correction inside a reliable convergence range is ordinary solvable-baseline refinement. No representation-changing rescue is needed.

Choosing the Padé approximant that best matches a known target and then reporting that match as validation is benchmark leakage and transform shopping, not governed resummation.

Continuing one analytic branch across a first-order phase transition without a model of the new phase is not nonperturbative extrapolation. It is unsupported branch persistence.

Running a direct certified numerical solver in the target regime is not an instance of the archetype, although its results may be ideal benchmarks or completion inputs.

Using “nonperturbative” as a synonym for “nonlinear,” “complex,” or “hard to compute” is a naming error. The relevant boundary is whether essential effects are invisible to the local perturbative sector or require information beyond its ordinary algebraic expansion.

Related Abstractions

Abstractions this archetype builds on — directly (a source ingredient) or as a related pattern. Links follow the typed catalog namespace.

Built directly on (1)

• Perturbation Theory: A technique for handling an intractable problem by splitting it into an exactly solvable baseline plus a small correction, then expanding the quantities of interest as a power series in that small parameter.

Also references 13 related abstractions

• Approximation: Good-enough representation.
• Boundedness: Values remain within limits.
• Convergence: Movement toward stable state.
• Invariance: Properties unchanged under transformation.
• Nonlinearity: Disproportionate output.
• Representation: Model complex ideas.
• Robustness: Maintain functionality under stress.
• Scale: Properties change with size.
• Symmetry: Invariance under transformation.
• Tipping Points (or Phase Transitions): Abrupt state change.

▸ Show 3 more

Variants

Narrower or domain-specific specializations that share this archetype's core structure. Recognized variants are established; candidate variants are provisional.

Borel-Plane Rescue · mechanism family variant · recognized

Uses Borel transformation and controlled continuation to recover factorially divergent perturbative information.

• Distinct from parent: It specializes the rescue loop to factorial divergence and inverse-Borel reconstruction.
• Use when: Coefficients show factorial growth compatible with a Borel treatment; Borel-plane singularities and integration choices can be analyzed.
• Typical domains: quantum field theory, statistical mechanics, spectral problems
• Common mechanisms: borel resummation, borel pade resummation, conformal borel mapping

Critical-Regime Series Continuation · domain variant · recognized

Constrains series continuation near critical behavior using known singularity or scaling information.

• Distinct from parent: It emphasizes singular critical behavior rather than generic strong-regime continuation.
• Use when: A high- or low-temperature expansion approaches a critical point; Exponents, amplitudes, or singularity form provide defensible anchors.
• Typical domains: statistical mechanics, lattice models, critical dynamics
• Common mechanisms: high temperature series resummation, pade approximant

Matched-Regime Completion · scale variant · recognized

Builds separate local approximations for distinct dominant-balance regimes and joins them through an overlap.

• Distinct from parent: It changes spatial or scale representation and requires overlap consistency.
• Use when: A regular expansion fails in a thin, fast, or boundary region; Inner and outer approximations have a defensible overlap.
• Typical domains: fluid mechanics, reaction diffusion, wave propagation
• Common mechanisms: matched asymptotic expansion

Near names: Divergent Series Resummation, Padé Resummation, Strong-Coupling Extrapolation.

Editorial Notes

Problem Classification

Classification: Correctness, Conformance & Formal Validity Failure → Computational Decidability & Bounded Approximation

Problem kernel: local series information does not support naive global extrapolation

Rationale: Earliest causal condition: A tractable local construction produces coefficients, correction terms, moments, or samples that are informative near a reference regime but cannot be used reliably by naive truncation or direct summation where an answer is needed. Adding terms may worsen the estimate; apparent convergence may be pre-asymptotic; a target may lie beyond a convergence disk or across a singular boundary; or effects invisible at every or

Independent corroboration: The earliest necessary condition in the frozen evidence is: A tractable local construction produces coefficients, correction terms, moments, or samples that are informative near a reference regime but cannot be used reliably by naive truncation or direct summation where an answer is needed. That is a computational decidability and bounded approximation problem because Exact solution or total decision is impossible or infeasible, yet the system lacks an honest decision boundary, separated verification path, or bounded approximation.

Review outcome: Independent reviewer agreement; high confidence.