FETI-DP

A dual–primal nonoverlapping domain-decomposition solver that assembles selected interface degrees of freedom globally while enforcing continuity of the remaining duplicated interface variables with Lagrange multipliers.

Version

v1 · 2026-08-30 · History

Domain-specific #

1830

Origin domain

numerical analysis

Subdomain

iterative substructuring methods

Aliases

Finite Element Tearing and Interconnecting Dual Primal, Dual Primal Feti

Core Idea

FETI-DP—Finite Element Tearing and Interconnecting, Dual–Primal—is a nonoverlapping domain-decomposition method for solving large linear systems produced by finite-element and related discretizations. It divides a mesh into subdomains, solves most subdomain work locally, and coordinates those solutions through a smaller interface problem. Its defining choice is to treat selected interface degrees of freedom as primal, assembled and continuous from the outset, while treating the remaining interface degrees of freedom as dual, duplicated by subdomain and made continuous with Lagrange multipliers.^[1][2]

The locked identity is discretized boundary-value problem + nonoverlapping subdomains + condensed interior variables + explicit split of interface unknowns into primal and dual sets + direct assembly of primal continuity + multiplier enforcement of dual continuity + local subdomain solves + a global coarse component induced by primal variables + preconditioned iterative solution of the dual interface system + recovery of the global field.

This dual–primal split is not a branding detail. Purely dual FETI formulations enforce all interface continuity with multipliers and must explicitly handle subdomain nullspaces. Selecting enough primal constraints can remove local singularity, create a coordination space, and transmit global information while preserving substantial parallelism. Farhat and colleagues introduced the unified FETI-DP formulation as a faster, scalable alternative to the two-level FETI construction.^[1]

FETI-DP survives as an autonomous domain-specific abstraction because the same role structure governs a family of solvers, analyses, preconditioners, coarse-space enrichments, and applications. It is not a single software product or historical experiment. Yet its essential vocabulary—finite-element subdomains, interface degrees of freedom, Lagrange multipliers, Schur complements, primal constraints, and condition-number bounds—does not transfer literally outside numerical PDE solution, so it is not a prime.

Structural Signature

• the discretized operator — usually a sparse system arising from a finite-element, virtual-element, or closely related PDE discretization;
• the nonoverlapping partition — the computational domain and its degrees of freedom are divided into subdomains with shared interfaces;
• the local operators — each subdomain retains an independently applicable stiffness or system matrix;
• the interior/interface split — interior variables communicate only through interface variables and can normally be eliminated locally;
• the primal set — selected corners, vertices, edge averages, face averages, or adaptive constraints that are assembled globally and continuous by construction;
• the dual set — remaining interface variables kept as local copies during subdomain solution;
• the jump operator — records disagreement between neighboring copies of dual variables;
• the Lagrange multipliers — enforce zero jump, hence dual-interface continuity;
• the static condensation — eliminates interior and other local variables to expose an interface Schur-complement problem;
• the coarse problem — couples primal variables across the full partition and propagates low-frequency or global information;
• the dual interface equation — reduced multiplier system solved iteratively, commonly by a preconditioned Krylov method;
• the preconditioner and scaling — approximate interface inverse and balance contributions from adjacent subdomains;
• the local/coarse alternation — parallel local actions combine with collective coarse coordination at each solve stage;
• the reconstruction — converged interface data determine each subdomain’s interior field and therefore the global solution;
• the scalability criterion — iteration growth and communication/coarse costs are assessed as mesh resolution, subdomain count, coefficient contrast, and processor count change.

Deleting either half of the dual–primal split changes the method family. Assembling all interface unknowns produces a primal substructuring route; constraining them all through multipliers produces a fully dual FETI route.

What It Is Not

• Not the Finite Element Method generally. Finite elements create a discrete system; FETI-DP is one architecture for solving a partitioned system.
• Not domain decomposition generally. Schwarz, multigrid, BDDC, FETI, and many other methods decompose domains with different coupling rules.
• Not original FETI. The defining primal subset and its global assembly distinguish FETI-DP.
• Not BDDC. BDDC is usually formulated as a primal substructuring preconditioner, though the two families have closely related spectra under standard assumptions.^[3]
• Not a mesh partitioner. Partitioning supplies the subdomains but does not solve the interface equations.
• Not “parallel finite elements” as a generic implementation label. FETI-DP requires the specific interface algebra.
• Not a direct solver. Local factorizations may be direct, but the global reduced problem is characteristically iterative.
• Not any use of Lagrange multipliers. The multipliers must enforce continuity of the designated dual interface variables within the full dual–primal structure.
• Not one corner-selection recipe. Classical, deluxe-scaled, adaptive, and multilevel variants alter the coarse space while retaining the family identity.
• Not guaranteed scalability without qualifications. Poor primal constraints, difficult coefficients, or a coarse-solve bottleneck can degrade performance.

Scope of Application

FETI-DP originated in structural mechanics and applies broadly to large discretized elliptic and elasticity problems. The literature extends it to plates and shells, heterogeneous and nearly incompressible materials, electromagnetic problems, virtual elements, Stokes-type systems, and nonlinear or multilevel formulations when the corresponding operator and interface treatment support the method.

The method is particularly attractive on distributed-memory machines because subdomain factorizations and backsolves are local and parallel. Communication concentrates at interfaces and in the coarse problem. That architecture has enabled very large simulations, while later multilevel work addresses the point at which a single global coarse solve itself becomes a scaling bottleneck.^[4]

Adaptive FETI-DP enriches the primal/coarse space from local generalized eigenproblems when coefficient jumps or heterogeneous media defeat standard corner and average constraints.^[5][6] Such extensions confirm the abstraction’s recurrence: they revise how primal constraints are selected, not the dual–primal division itself.

Clarity

“Primal” and “dual” identify how interface continuity is represented, not two different physical solutions. Primal degrees of freedom are single-valued through assembly. Dual degrees remain duplicated until Lagrange multipliers drive their jumps to zero. Both contribute to one reconstructed solution.

Corners are the classical minimal example of primal variables, not a universal invariant. Three-dimensional scalability often requires edge or face averages, and heterogeneous problems may require adaptive spectral constraints. The invariant is a selected primal coarse subset sufficient for solvability and coordination, not “corners only.”

Scalability also has several meanings. Numerical scalability concerns iteration or condition-number behavior as resolution and partitioning grow. Parallel scalability concerns wall time, communication, local work balance, and coarse bottlenecks as processors grow. A solver can satisfy one more strongly than the other.

Manages Complexity

The method compresses one enormous coupled solve into many reusable local solves plus an interface coordination problem. Interior variables disappear from global iteration through condensation. The primal set carries indispensable global modes, while multipliers reconcile the many remaining local interface copies. This preserves parallel locality without accepting discontinuous subdomain solutions.

It also makes solver design diagnosable. Slow convergence can be localized to primal-set insufficiency, bad coefficient scaling, a weak interface preconditioner, partition geometry, or a coarse bottleneck. Those are distinct repair surfaces rather than one opaque complaint that “the parallel solver scales poorly.”

Abstract Reasoning

1. If a subdomain interface variable is primal, neighboring subdomains share one assembled value and need no multiplier for its continuity.
2. If it is dual, independent local copies require a jump constraint enforced by a multiplier.
3. If the chosen primal set fails to remove relevant local null modes, local solves or the reduced operator can be singular or ill-conditioned.
4. If only local solves are accelerated while the coarse solve dominates, adding processors will not restore parallel scalability.
5. If coefficient contrast produces interface modes unseen by the classical coarse space, adaptive spectral constraints can move those modes into the primal space.^[5]
6. If all interface variables are made primal, parallel independence shrinks and the method approaches a fully assembled primal solve.
7. If all are made dual, the method loses its defining dual–primal architecture and inherits fully dual nullspace handling.
8. If a partition creates long or irregular interfaces, constraint selection and scaling may matter more than raw subdomain count.
9. If two implementations have the same partition but different primal constraints or preconditioners, they can have sharply different convergence while remaining FETI-DP variants.
10. If an application has no decomposable operator or enforceable interface continuity, invoking “FETI-DP” metaphorically adds no valid method.

Knowledge Transfer

Exact transfer occurs across PDEs, discretizations, and machines where nonoverlapping substructures, interface variables, primal constraints, dual continuity, local solves, and a coarse problem remain literal. Virtual-element and electromagnetic variants can therefore instantiate the family even though their local operators differ.

The portable residue is Decomposition + Parallelism + Interface + Constraint + Coordination. Those primes explain why the method is intelligible outside numerical analysis, but an organizational system with local teams and global coordinators is only an analogy. Without Schur-complement interface algebra and primal/dual continuity treatment, it is not FETI-DP.

Examples

• classical elasticity solve: a finite-element mesh is partitioned; subdomain corners are primal, remaining interface displacements dual, and multipliers enforce their continuity;
• three-dimensional extension: edge or face averages join the primal set to improve global propagation;
• heterogeneous diffusion: local eigenproblems identify difficult interface modes and promote corresponding constraints into an adaptive coarse space;^[5]
• multilevel FETI-DP: the coarse problem is recursively decomposed when a single second-level solve limits extreme-scale execution;^[4]
• BDDC comparison: similarly selected coarse constraints can yield closely related nonunit spectra, while the formulations remain distinct;^[3]
• non-example—mesh partition only: subdomains are distributed to processors, but a generic sparse direct solver handles the assembled matrix;
• failure—corners insufficient: strong material jumps along an interface cause iteration growth until scaling or adaptive constraints are added;
• failure—coarse bottleneck: local work scales down but collective factorization and communication dominate total time.

Structural Tensions

• local independence vs. global continuity — tearing enables parallel subdomain work while interconnecting must recover one conforming solution;
• small coarse space vs. robust convergence — fewer primal constraints reduce collective cost while richer constraints capture difficult global modes;
• dual flexibility vs. primal stability — multipliers preserve locality while primal assembly removes singular modes and coordinates the partition;
• setup cost vs. solve cost — adaptive constraints and strong factorizations cost more initially but can reduce iterations;
• numerical scalability vs. parallel scalability — bounded iteration growth does not eliminate communication or coarse-solve limits;
• generic recipe vs. problem adaptation — standard corners and averages are reusable while heterogeneous operators may demand local spectral evidence;
• local load balance vs. interface complexity — equal element counts can still produce unequal factorization, communication, or constraint costs.

Structural–Framed Character

FETI-DP is strongly structural. Its identity is given by algebraic spaces, constraint operators, assembled and duplicated variables, local eliminations, and convergence properties. Human choices select the partition, primal constraints, tolerances, preconditioner, and hardware mapping, but those choices are evaluated against mathematically defined solvability and performance rather than community convention alone.

Structural Core vs. Domain Accent

The structural core is decompose + solve locally + preserve selected global coordinates + enforce remaining interface agreement + iterate on the reduced coupling problem + reconstruct. The domain accent is finite-element degrees of freedom, stiffness matrices, Schur complements, Lagrange multipliers, primal coarse spaces, and PDE convergence analysis. Removing that accent yields Decomposition and Coordination, not FETI-DP.

Instantiates / Related Primes

• Decomposition — the global operator is split into nonoverlapping local subproblems and recombined.
• Interface — shared boundary variables form the explicit coupling surface.
• Constraint — zero jumps impose compatibility among local copies.
• Coordination — the primal coarse problem propagates information that independent local solves cannot.
• Parallelism — local factorizations and backsolves execute concurrently.
• Trade-off — coarse-space richness exchanges setup and communication cost for robustness.

The minimal prospective DAG uses strict subsumption under prime:decomposition. Interface, Constraint, Coordination, and Parallelism remain related analytical primes.

Relationships to Other Abstractions

Current abstraction FETI-DP Domain-specific

Parents (1) — more general patterns this builds on

• FETI-DP is a kind of Decomposition Prime

  the global operator is split into nonoverlapping local subproblems and recombined.

Hierarchy path (1) — routes to 1 parentless root

• FETI-DP → Decomposition

Neighborhood in Abstraction Space

FETI-DP sits in a sparse region of the domain-specific corpus (93^rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

• Spectral Element Method — 0.80
• Frontal Solver — 0.80
• Fictitious domain method — 0.78
• Balancing domain decomposition method — 0.77
• Exponential Integrator — 0.77

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With

• FETI, FETI-1, or FETI-2;
• BDDC or primal balancing methods;
• overlapping Schwarz methods;
• finite-element discretization itself;
• mesh generation or graph partitioning;
• generic Schur-complement methods;
• generic Lagrange-multiplier formulations;
• multigrid coarse spaces;
• adaptive FETI-DP as the entire family;
• a particular FETI-DP software package or benchmark result.

References

[1] Charbel Farhat, Michel Lesoinne, Patrick Le Tallec, Kendall Pierson, and Daniel Rixen, “FETI-DP: A Dual–Primal Unified FETI Method—Part I,” International Journal for Numerical Methods in Engineering 50 (2001), 1523–1544, https://doi.org/10.1002/nme.76. registry ↩a ↩b

[2] Jan Mandel and Radek Tezaur, “On the Convergence of a Dual-Primal Substructuring Method,” Numerische Mathematik 88 (2001), 543–558, https://doi.org/10.1007/s211-001-8014-1. registry ↩

[3] Jan Mandel and Bedřich Sousedík, “BDDC and FETI-DP under Minimalist Assumptions,” Computing 81 (2007), 269–280, preprint https://arxiv.org/abs/0708.4031. registry ↩a ↩b

[4] Jari Toivanen, Patrick Avery, and Charbel Farhat, “A Multilevel FETI-DP Method and Its Performance for Problems with Billions of Degrees of Freedom,” International Journal for Numerical Methods in Engineering 116 (2018), 661–682, https://doi.org/10.1002/nme.5938. registry ↩a ↩b

[5] Axel Klawonn, Patrick Radtke, and Oliver Rheinbach, “FETI-DP Methods with an Adaptive Coarse Space,” SIAM Journal on Numerical Analysis 53(1) (2015), 297–320, https://doi.org/10.1137/130939675. registry ↩a ↩b ↩c

[6] Axel Klawonn, Martin Kühn, and Oliver Rheinbach, “Adaptive Coarse Spaces for FETI-DP in Three Dimensions,” SIAM Journal on Scientific Computing 38(5) (2016), A2880–A2911, https://doi.org/10.1137/15M1049610. registry ↩

[7] “FETI-DP,” Wikipedia, frozen revision 1183761874, https://en.wikipedia.org/wiki/FETI-DP. registry