---
title: 'IETI: Isogeometric Tearing & Interconnecting'
url: https://www.emergentmind.com/topics/isogeometric-tearing-and-interconnecting-ieti
type: topic
---

# IETI: Isogeometric Tearing & Interconnecting

Searching arXiv for recent and foundational IETI papers to ground the article.
arxiv_search:
  query: "IETI Isogeometric Tearing and Interconnecting"
  max_results: 10

Isogeometric Tearing and Interconnecting (IETI) is a non-overlapping domain decomposition framework for multi-patch Isogeometric Analysis (IgA) in which the patches of the geometric model are used as subdomains, patch-local problems are solved separately, and global coupling is restored by interface constraints. In the cited literature, the dominant realization is the dual-primal variant, IETI-DP, which is presented as the isogeometric analogue of FETI-DP and as closely related to BDDC. Its basic attraction is the replacement of one large global solve by local patch solves, a coarse/primal problem, and an interface system in Lagrange multipliers that is amenable to iterative treatment and scalable preconditioning [2103.04801].

## 1. Domain-decomposition role in multi-patch IgA

In multi-patch IgA, the computational domain is represented as a union of non-overlapping patches \(\Omega^{(k)} = G_k(\widehat\Omega)\), with \(\widehat\Omega=(0,1)^d\), and each patch carries its own tensor-product spline space. This patch structure is natural for CAD-based geometry representation, but it also creates the algebraic setting for non-overlapping substructuring: interfaces between patches must be coupled, while local patch matrices remain accessible and highly structured [1611.08122].

For conforming discretizations, neighboring patch spaces match across interfaces and the global discrete space is \(H^1\)-conforming. Even in that setting, direct solution becomes expensive for large patch decompositions and high-degree splines, especially in three dimensions, which motivates IETI-DP as a patch-wise solver architecture [2103.04801]. For discontinuous Galerkin (dG) discretizations, the global space is broken from the outset, and continuity is enforced weakly by SIPG-type interface terms; in that setting IETI-DP becomes a solver for the dG-coupled system rather than for a strongly assembled conforming system [2005.09546].

A recurring theme in the literature is that IETI is not itself a discretization method. It is a solver framework built on top of conforming or dG multi-patch IgA discretizations. This distinction matters because several extensions that appear under the IETI name actually modify the interface modeling assumptions—fully matching interfaces, non-matching meshes, partial-edge interfaces, T-junctions, nested traces under adaptive splitting, or \(C^1\) couplings for the biharmonic problem—while retaining the same tearing-and-interconnecting logic [2103.02536].

## 2. Algebraic architecture

The standard IETI-DP construction begins with patch-local matrices \(A^{(k)}\) obtained from the local bilinear forms. Interface degrees of freedom are then “torn,” meaning that shared interface unknowns are duplicated patchwise. Global continuity is restored by a jump operator \(B\), whose rows encode differences of corresponding interface traces; in the standard formulations, each row has one \(+1\) and one \(-1\) [1912.07909].

Dual-primal structure enters through the choice of primal constraints. If \(C^{(k)}\) denotes the local primal constraint matrix, the dual space consists of local interface vectors satisfying \(C^{(k)} \underline w^{(k)}=0\), while the primal space is defined as an energy-minimizing complement. In the energy-minimizing formulation used repeatedly in the literature, the partially assembled interface space splits as
\[
\widetilde W = W_\Delta \oplus W_\Pi,
\]
with \(W_\Pi\) the \(S\)-orthogonal complement of the homogeneous dual space. This makes the local and coarse components of the inverse particularly clean [1705.04531].

A canonical formulation is the interface Schur-complement equation
\[
F \,\underline\lambda = \underline d,
\]
where \(F\) is assembled from local inverses or local Schur complements and \(\underline\lambda\) is the vector of Lagrange multipliers. In conforming Poisson-type settings, one may write
\[
F := \sum_{k=1}^{K+1} \widetilde B^{(k)} (\widetilde A^{(k)})^{-1} (\widetilde B^{(k)})^\top,
\]
with the primal coarse problem treated as an additional “subdomain” indexed by \(K+1\) [2103.04801]. In a reduced interface formulation for conforming multipatch IgA, the same structure appears as
\[
F:=\widetilde B\,\widetilde S^{-1}\widetilde B^T, \qquad d:=\widetilde B\,\widetilde S^{-1}\widetilde g,
\]
after elimination of interior and primal/interface displacement variables [1705.04531].

For dG-coupled IETI-DP, the same architecture is retained, but local spaces are enlarged by artificial interface traces. One works with
\[
V_e^{(k)} = V^{(k)} \times \prod_{\ell\in\mathcal N_\Gamma(k)} V^{(k,\ell)},
\]
so that each patch carries its own field together with copies of neighboring traces needed by the local SIPG terms [2103.02536]. A different algebraic realization appears in the low-rank tensor-train literature, where the pure dual IETI equation is written as
\[
\mathbf{A}\mathbf{K}^{-1}\mathbf{A}^T\,\boldsymbol{\lambda} = \mathbf{A}\mathbf{K}^{-1}\mathbf{f},
\]
with \(\mathbf{A}\) the interface constraint operator on the broken patchwise space [2405.06458].

## 3. Primal variables and interface topology

The primal space is the decisive modeling choice in IETI-DP because it determines local invertibility, the size of the coarse problem, and much of the eventual conditioning. In the two-dimensional Poisson literature, the standard choices are vertex values, edge averages, or both. In three dimensions, one additionally has face averages, and the tested combinations are \(V\), \(E\), \(F\), \(V+E\), \(V+F\), \(E+F\), and \(V+E+F\) [2103.04801].

The following table summarizes the principal choices that recur across the cited works.

| Setting | Typical primal variables | Representative role |
|---|---|---|
| 2D conforming Poisson | vertex values, edge averages, both | local invertibility and coarse correction |
| 3D conforming Poisson | vertex values, edge averages, face averages, combinations | robustness across faces, edges, and vertices |
| dG with T-junctions | all basis functions non-zero at a junction (“fat vertices”) | handling partial-edge interfaces |
| Stokes | corner values, edge-wise normal-component averages, patch pressure averages | local Stokes solvability and incompressibility |
| Adaptive nested traces | vertex values and T-junction values | conforming continuity under recursive patch splitting |
| Biharmonic | corner function values and corner first derivatives (“fat vertices”) | \(C^1\) coupling in \(H^2\)-conforming spaces |

For conforming three-dimensional Poisson problems, the computational message is explicit: vertices alone are not enough for robust behavior, while edge averages are the most effective choice among the tested primal variables. On a twisted Fichera corner decomposed into \(448\) patches, edge-based choices produced much smaller condition numbers and iteration counts than vertex-only constraints; face averages improved over vertices but not as much as edge averages [2103.04801]. This mirrors standard finite-element FETI-DP/BDDC practice in three dimensions.

T-junctions force a different notion of “vertex” because a spline T-junction is not represented by a single coefficient. In the dG extension to non-whole-edge interfaces, the proposed rule is: choose as primal all degrees of freedom associated with basis functions that are non-zero on a junction. At ordinary corners this recovers the usual vertex choice; at T-junctions it yields several primal degrees of freedom, hence the term “fat vertices” [2103.02536]. For dG multi-patch IgA with T-junctions, this fat-vertex strategy is the critical interface-topology adaptation, and the corresponding paper states a condition number bound that coincides with the bound for the conforming case [2109.13147].

For Stokes, primal constraints must also control the saddle-point nullspaces. The conforming multi-patch Taylor–Hood solver uses corner values of each velocity component, edge-wise integrals of the normal component of the velocity, and patchwise pressure averages. The pressure averages are needed so that patchwise constant pressure modes are handled cleanly, and the edge normal-flux averages are needed for local solvability and the Stokes analysis [2202.13707]. A related earlier empirical study on Stokes compared \(\Pi_{\mathrm c}\), \(\Pi_{\mathrm{ce}}\), and \(\Pi_{\mathrm{cn}}\), showing that corner values alone are clearly inferior to enriched edge-based primal spaces [2112.12163].

## 4. Preconditioning and convergence theory

The standard preconditioner in the IETI literature is the scaled Dirichlet preconditioner. In conforming and dG Poisson settings, it is written in forms such as
\[
M_{\mathrm{sD}}^{-1}:=B_D S B_D^T
\]
or, at the matrix level,
\[
M_{\mathrm{sD}} := B D^{-1} S D^{-1} B^\top,
\]
with multiplicity scaling as the default choice in the matching-interface case [1705.04531]. The central theoretical result for conforming two-dimensional Poisson problems is the explicit estimate
\[
\kappa(M_{\mathrm{sD}}F) \le C\, p \left(1+\log p+\max_{k=1,\ldots,K}\log\frac{H_k}{h_k}\right)^2,
\]
proved for vertex values, edge averages, and their combination, with \(C\) independent of \(h\), \(H\), \(p\), and spline smoothness [1912.07909].

For dG coupling, the corresponding \(h\)- and \(p\)-explicit theory distinguishes the primal choice. If primal constraints include vertex values, or both vertex values and edge averages, the same qualitative bound as in the conforming case is obtained. If only edge averages are primal, an additional dependence on the ratio of neighboring mesh sizes appears:
\[
\kappa(M_{\mathrm{sD}}F) \le C\,\delta\, p \left( \max_{k}\max_{\ell\in\mathcal N_\Gamma(k)} \frac{h_k}{h_\ell} \right)
\left(1+\log p+\max_k\log\frac{H_k}{h_k}\right)^2.
\]
This is one of the clearest results in the dG-IETI-DP theory: vertex primal constraints are essential if one wants the same robustness as in the conforming case [2005.09546].

Three-dimensional theory is less complete. For conforming multi-patch IgA in three dimensions, the available result is numerical evidence rather than a full proof. The reported experiments indicate that the favorable two-dimensional behavior persists qualitatively, with robust dependence on mesh refinement and spline degree when suitable edge-based primal constraints are used, but the paper explicitly does not present a full three-dimensional condition-number proof [2103.04801].

Several later developments adapt the preconditioner to non-matching topology. For adaptive recursive patch splitting with nested interface trace spaces, the proposed scaled Dirichlet preconditioner uses “selection scaling” rather than multiplicity or coefficient scaling. Under the nested-trace assumption, the paper proves
\[
\kappa(M_{sD}F) \le C\, p\, \max\left\{ \max_{\ell\prec k}\frac{\nu^{(k)}}{\nu^{(\ell)}}, 1 \right\}
\left(1+\log p+\max_{k=1,\ldots,K}\log \frac{H_k}{h_k}\right)^2,
\]
which has the same asymptotic form as the fully matching case in the coefficient-independent regime [2510.08148].

Beyond scalar Poisson problems, analogous estimates exist but with problem-specific factors. For conforming multi-patch Stokes discretizations with isogeometric Taylor–Hood spaces, the main theorem is
\[
\kappa(M_{\mathrm{sD}}\bar F) \le C\, p\, \left(1+\log p+\max_{k=1,\dots,K}\log\frac{H_k}{h_k}\right)^2 \left(\max_{k=1,\dots,K}\frac{\delta}{\beta_k}\right),
\]
where \(\beta_k\) are local inf-sup constants. The salient point is that the bound depends on local patch properties rather than the global continuous inf-sup constant of the whole geometry [2202.13707]. For the biharmonic problem with \(H^2\)-conforming, \(C^1\)-coupled multi-patch IgA, the proven estimate becomes
\[
\kappa(MF)\le C\, p^3 \left( 1+\log p + \max_{k=1,\dots,K}\log\frac{H_k}{h_k} \right)^2,
\]
reflecting the greater difficulty of \(C^1\) continuity and \(H^2\) regularity [2511.05247].

A common misunderstanding is that IETI theory is synonymous with a single \((1+\log(H/h))^2\) statement. The literature shows a more nuanced picture: the asymptotic \(H/h\) behavior is often polylogarithmic, but the explicit degree dependence, the role of local inf-sup constants, the effect of neighboring mesh ratios, and the admissibility of T-junctions all depend on the particular discretization and on the chosen primal variables.

## 5. Exact, inexact, parallel, and low-rank realizations

The earliest large-scale implementations in the cited corpus rely on exact local sparse direct solves and energy-minimizing primal subspaces. In this setting, \(\widetilde S^{-1}\) splits into a global coarse solve and block-local dual solves, which is one of the reasons the method parallelizes well [1611.08122]. Distributed-memory studies for both conforming and dG-IETI-DP on \(1024\) patches reported assembly speedups around \(900\) on \(1024\) cores and solver speedups around \(500\), with weak and strong scaling described as excellent [1611.08122].

A major later development is the systematic replacement of exact local sparse direct solvers by inexact local solvers. One line uses multigrid. In the conforming Poisson setting, several variants were studied: exact local solves everywhere, multigrid only for local Dirichlet problems, multigrid for all local problems, and a saddle-point variant solved by Bramble–Pasciak conjugate gradients. The results showed that replacing direct solvers by multigrid in the scaled Dirichlet preconditioner did not increase the number of outer iterations, while the fully inexact method traded additional CPU time for a much smaller memory footprint and the ability to solve cases where direct variants ran out of memory [1705.04531].

A second line uses Fast Diagonalization (FD). For conforming multi-patch discretizations, sparse LU factorizations in the local patch solves were replaced by FD-based local inverse actions while retaining the same explicit condition-number bound as in the exact theory:
\[
C\, p \left(1+\log p+\max_{k=1,\ldots,K}\log\frac{H_k}{h_k}\right)^2.
\]
In the reported quarter-annulus and Yeti-footprint tests, the FD-based saddle-point solver was faster overall than sparse-LU-based variants despite requiring more outer iterations, and it solved cases that direct variants marked as out of memory [2110.06087].

For dG-IETI-DP, tensor-structured inexact local solvers require an additional step because the local extended dG spaces do not directly preserve tensor-product structure. The proposed remedy is an orthogonal splitting of the local extended space into a tensor-product part and a small residual interface part. This permits FD on the dominant block and direct treatment of the small remainder. The resulting inexact dG-IETI-DP method preserves a polylogarithmic \(H/h\) condition-number estimate, mildly reduces computational time, significantly reduces memory requirements relative to standard sparse-direct local solvers, and shows good scaling behavior on distributed-memory machines [2206.08416].

Although the overwhelming majority of the literature treats the dual-primal variant, low-rank work has also revived the pure IETI perspective. In a tensor-train method for PDE-constrained optimization on multi-patch IgA domains, continuity is enforced by a constraint matrix and Lagrange multipliers, and the dual problem is written directly as
\[
\mathbf{A}\mathbf{K}^{-1}\mathbf{A}^T\boldsymbol{\lambda}
=
\mathbf{A}\mathbf{K}^{-1}\mathbf{f}.
\]
That work illustrates that the tearing-and-interconnecting idea also adapts to low-rank tensor numerics and to reduced optimality systems, not only to classical sparse-matrix substructuring [2405.06458].

## 6. Generalizations, misconceptions, and current scope

The present literature portrays IETI not as a single fixed algorithm but as a family of interface solvers adapted to increasingly diverse patch-coupling scenarios. In conforming scalar elliptic problems, the benchmark theory is now explicit in \(h\) and \(p\) in two dimensions and numerically convincing in three dimensions [1912.07909]. In dG settings, IETI-DP accommodates non-matching patch discretizations, artificial interfaces, and T-junctions, with fat-vertex primal constraints as the main device for restoring stability at junctions [2103.02536].

One common misconception is that “non-matching” in this literature always means unrelated meshes in the finite-element sense. In fact, several papers use the term in a more geometric and topological way. In the partial-edge and T-junction setting, non-matching means that one patch edge may meet only part of another edge; in the adaptive nested-trace setting, the interface grids may be different but the traces are nested, so continuity can still be enforced exactly without dG, mortar, or Nitsche couplings [2510.08148].

Another misconception is that IETI-DP is only a Poisson solver. The cited corpus contains a conforming multi-patch Stokes solver with a rigorous convergence analysis, in which local Stokes saddle-point structure is handled by appropriate velocity and pressure primal variables and a Poisson-based scaled Dirichlet preconditioner [2202.13707]. It also contains a biharmonic solver for \(H^2\)-conforming multi-patch IgA, where the interface conditions match both function values and normal derivatives and the primal variables become fat vertices in a \(C^1\) setting [2511.05247]. These extensions show that the framework is compatible with both mixed saddle-point systems and higher-order smoothness requirements.

A further limitation that remains visible across the literature is the coarse problem. Even when local solves are made inexact, the coarse/primal system is usually still treated by a direct solver. This suggests that coarse-space design and scalable coarse solves remain open implementation bottlenecks, especially as patch counts grow. A second recurring limitation is the partial nature of \(p\)-robustness: explicit and favorable \(p\)-dependence is known for several two-dimensional scalar cases, but inexact dG local solvers and some saddle-point settings still lack fully satisfactory degree-explicit theory [2206.08416].

Taken together, these results support a broad characterization. IETI is the isogeometric realization of non-overlapping tearing-and-interconnecting ideas for multi-patch spline discretizations; IETI-DP is its dominant dual-primal form; and the modern literature treats it as a flexible solver platform for conforming, dG, non-matching, adaptive, mixed, and higher-order smoothness settings rather than as a solver tied to one model problem or one interface topology.

Source: https://www.emergentmind.com/topics/isogeometric-tearing-and-interconnecting-ieti