Superconvergence¶
A higher approximation order at specified points or for recovered quantities than the same method's general error order under stated conditions.
Core Idea¶
Superconvergence is a better approximation error order at a specified point or for a recovered quantity than the same numerical method's general error order, under stated assumptions. The comparison requires a method, target, general baseline, special output and hypotheses. It does not promise higher accuracy at every point.[ref-89df2daf34bb][ref-0eab5c7c0eaf]
Scope of Application¶
Levine's original abstract describes a recovered finite-element gradient that is superconvergent at triangle-edge midpoints for a second-order elliptic problem. De Boor and Swartz study Gaussian collocation for a nonlinear ODE boundary-value problem and report a stronger endpoint order under particular smoothness and parameter conditions. Recovery in a PDE method and endpoint values in an ODE method are unlike cases of the same comparison pattern.[ref-89df2daf34bb][ref-0eab5c7c0eaf]
The finite-element abstracts do not justify a precise rate for arbitrary raw gradients or meshes; the ODE abstract supplies explicit rates only with its stated assumptions.[ref-89df2daf34bb][ref-611449bdb6cf][^ref-0eab5c7c0eaf]
Clarity¶
“Super” means better than a matched general bound, not merely “small error.” The special point or output matters. In Gaussian collocation, the equation is enforced at interior Legendre-zero sites, while the stated stronger results are at interval endpoints.[^ref-0eab5c7c0eaf]
Manages Complexity¶
An approximation has errors across a whole field. A proved special-output result tells an analyst where to evaluate or recover the quantity needed for a task without pretending the whole field acquired the higher order. It also keeps different methods' proof conditions from being collapsed into one generic rule.[ref-89df2daf34bb][ref-0eab5c7c0eaf]
Abstract Reasoning¶
If a general error is O(h^p) and a compatible special-output error is O(h^q), the displayed bounds establish a higher order only when q>p under the same stated assumptions. For de Boor and Swartz, the original abstract gives global O(h^(m+k)) and endpoint O(h^(2k)), so the endpoint exponent is strictly larger when k>m. The abstract also specifies m side conditions, m+2k derivatives and Gaussian collocation sites.[^ref-0eab5c7c0eaf]
Knowledge Transfer¶
The test—target, baseline, special output, conditions—transfers from recovered finite-element gradients to collocation endpoint values. Their algorithms, locations and hypotheses do not transfer unchanged. The live Convergence Prime supplies approximation approaching a limit; this entry adds a numerical special-order requirement. Domain-specific Γ-convergence is a different technical identity.[ref-89df2daf34bb][ref-0eab5c7c0eaf]
Example¶
Recovered finite-element gradient: Levine reports recovery from triangular piecewise-linear approximations to an elliptic problem, with improved accuracy at edge midpoints. Mapped roles: approximation → refined triangular finite elements; target → exact gradient; baseline → general gradient error, kept qualitative here; special output → recovered midpoint gradient; conditions → the analyzed recovery/element setting.[^ref-89df2daf34bb]
Gaussian collocation endpoints: de Boor and Swartz enforce an ODE at k interior Legendre-zero sites per interval. Their abstract gives global order m+k and endpoint order 2k for the solution and its first m−1 derivatives, establishing a strict improvement when k>m. Mapped roles: approximation → piecewise-polynomial collocation; target → solution and first m−1 derivatives; baseline → global O(h^(m+k)); special output → endpoint O(h^(2k)) for those quantities; conditions → isolated solution, m side conditions, m+2k derivatives and specified sites.[^ref-0eab5c7c0eaf]
Relationships to Other Abstractions¶
Current abstraction Superconvergence Domain-specific
Parents (1) — more general patterns this builds on
-
Superconvergence is a kind of Convergence Prime
Superconvergence is convergence with stronger accuracy at a specified output under stated hypotheses.
Hierarchy path (1) — routes to 1 parentless root
- Superconvergence → Convergence
Neighborhood in Abstraction Space¶
Superconvergence sits in a sparse region of the domain-specific corpus (89th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Unclustered & Miscellaneous (2551 abstractions)
Nearest neighbors
- Finite Difference Method — 0.82
- Lanczos Approximation — 0.81
- Finite Difference Coefficient — 0.81
- PTAS Reduction — 0.81
- Finite Element Method — 0.80
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
Uniform convergence with no special improvement, raw gradients assigned a recovered estimate's property, or Gaussian collocation sites mistaken for the enhanced endpoints. The proposed strict parent is Convergence, not the unrelated domain-specific Γ-convergence entry. The named superconvergence test still requires a numerical approximation and a matched higher-order special output.[ref-89df2daf34bb][ref-0eab5c7c0eaf]
References¶
[^ref-89df2daf34bb]: Nick Levine, Superconvergent Recovery of the Gradient from Piecewise Linear Finite-element Approximations, IMA Journal of Numerical Analysis 5(4) (1985), 407–427, original publisher abstract. The full article was not consulted; the abstract identifies the recovered gradient, elliptic problem and edge-midpoint result. [^ref-611449bdb6cf]: Zhiming Zhang and Ahmed Naga, A Meshless Gradient Recovery Method Part I, Superconvergence Property, Wayne State University Mathematics Research Reports no. 2002.02 (2002), original institutional abstract. The linked title transcribes its colon as a comma; the report's title has a colon before “Superconvergence Property.” The full download returned 403 during review, so exact theorem rates are not imported. [^ref-0eab5c7c0eaf]: Carl de Boor and Blair Swartz, Collocation at Gaussian Points, SIAM Journal on Numerical Analysis 10(4) (1973), 582–606, original publisher abstract. The abstract states the ODE/side-condition/smoothness assumptions, Legendre-zero sites and global/endpoint orders; full proof was not consulted.