Structural Condensation and Computational Solution Techniques

Introduction and Context

Advanced structural analysis frequently produces systems with many degrees of freedom (DOFs), including translational and rotational nodal displacements, internal member variables, contact variables, or parameters introduced by finite-element discretisation. Solving the complete system may be unnecessary when engineering interest is restricted to a smaller set of master variables. Structural condensation provides a mathematically exact or approximate means of eliminating selected variables while retaining their influence on the variables of interest.

Computational solution techniques are the numerical procedures used to assemble, factorise, solve, and interpret the resulting algebraic equations. They are central to the stiffness method, finite-element analysis, substructuring, static condensation, dynamic reduction, nonlinear analysis, and structural optimisation. At postgraduate level, it is important not only to know how to perform condensation, but also to understand its assumptions, numerical conditioning, physical meaning, and implications for modelling decisions.

This chapter develops the Schur-complement basis of condensation, distinguishes static and dynamic reduction, examines direct and iterative solution methods, and presents a practical substructuring application.

Learning Objectives

On completing this chapter, the reader should be able to:

1. Explain the physical and algebraic meaning of master and slave degrees of freedom.
2. Derive the static-condensation equations using block-partitioned matrices.
3. Interpret the Schur complement as a condensed stiffness or flexibility relationship.
4. Distinguish static condensation, Guyan reduction, component-mode synthesis, and related dynamic methods.
5. Select appropriate direct or iterative computational solution techniques.
6. Assess conditioning, sparsity, storage, convergence, and error issues.
7. Apply condensation to a small structural system and interpret the result.
8. Recognise professional, verification, and governance requirements in computational structural analysis.

1. Algebraic Foundation of Structural Condensation

For a linearly elastic structure discretised by the stiffness method, equilibrium is written as

where K is the global stiffness matrix, u is the vector of unknown nodal displacements, and f is the vector of externally applied nodal forces. For a three-dimensional model, translational and rotational DOFs may occur in u. The units of K are force per displacement for translational terms and the corresponding consistent force–rotation units for rotational terms. The equation assumes small displacements, linear material behaviour, prescribed boundary conditions already incorporated, and a stable system unless otherwise stated.

The DOFs are partitioned into master variables, denoted by subscript m, and slave or internal variables, denoted by subscript s:

Thus,

The partition is a modelling choice. Master DOFs are usually retained because they are externally loaded, measured, connected to another component, or required for design output. Slave DOFs are eliminated because they are internal or because retaining them would make the system unnecessarily large.

Assume K_ss is nonsingular. The second block equation gives

Substitution into the first equation yields

The condensed stiffness matrix and condensed load vector are therefore

These are Schur-complement equations. The inverse notation is conceptual: in actual computation, K_ss is factorised and systems involving K_ss are solved; an explicit inverse should generally not be formed.

Physical interpretation

The term K_mm represents the direct stiffness seen at master DOFs. The correction K_ms K_ss^(-1) K_sm represents the flexibility and load transfer through the eliminated DOFs. Consequently, condensation does not simply delete rows and columns. It accounts for the deformation of the eliminated portion and transfers its effect to the retained system.

If f_s = 0, which is common when internal slave nodes carry no directly applied load, then f_c = f_m. However, slave displacements are still recovered from

If nonzero loads act on slave DOFs, omitting the term involving f_s produces an incorrect condensed load vector.

2. Conditions for a Valid Condensation

2.1 Boundary conditions and rigid-body modes

The matrix K_ss must be nonsingular for the basic static-condensation formula. A singular K_ss may indicate that the slave subsystem contains a rigid-body mode, an unconstrained mechanism, duplicate constraints, or an unsuitable partition. A globally stable structure can still have a singular K_ss if the selected slave set is not independently stable.

Possible remedies include retaining an appropriate rigid-body DOF as a master, imposing physically correct constraints, or using a reduction method designed to handle component rigid-body modes.

2.2 Symmetry and positive definiteness

For a conservative, linearly elastic structure with compatible coordinates, K is normally symmetric. If K is symmetric positive definite (SPD), and K_ss is a principal nonsingular submatrix, K_c is also SPD. This is important because Cholesky factorisation can then be used efficiently and the condensed problem retains a unique solution.

Unsymmetric matrices may arise from follower forces, nonconservative effects, certain contact formulations, or unsymmetric discretisations. In those cases, symmetry-based solvers are not generally valid.

2.3 Units and coordinate consistency

All blocks must use compatible displacement and force coordinates. Mixing millimetres with metres, or radians with degrees in rotational DOFs, can make entries numerically inconsistent. Rotational coordinates are dimensionless in SI, but the associated stiffness terms have moment units and must be assembled consistently. Transformation matrices must be applied when local element coordinates are related to global coordinates.

3. Static Condensation and Finite-Element Implementation

Static condensation is especially useful inside an element. Internal element DOFs are eliminated before the element matrices are assembled into the global system. This reduces global bandwidth and the number of global unknowns while preserving the exact static response of the element at retained DOFs.

Suppose an element has boundary DOFs b and internal DOFs i:

If K_ii is nonsingular, the element-level condensed matrix is

and the condensed load is

The bar indicates a condensed quantity. This process is also called internal-node elimination or substructuring. It is exact for the static linear problem represented by the original element equations; it is not necessarily exact for dynamic response because inertia associated with eliminated DOFs has not been represented in a frequency-independent stiffness alone.

Element-level versus global condensation

Element-level condensation is usually advantageous when internal variables are local and not needed elsewhere. Global condensation is useful when a large assembled model contains a clearly identifiable internal region. The choice affects fill-in, memory, parallelism, and the possibility of reusing factorizations.

A crucial implementation principle is that condensation should be performed using stable factorisation. For example, to compute K_bi K_ii^(-1) K_ib, solve

then calculate K_bi X. This avoids explicitly calculating K_ii^(-1).

4. Worked Example: Condensing an Internal Degree of Freedom

Consider the dimensionally consistent scalar system

Let u_1 be the master displacement and u_2 the internal displacement. The stiffness coefficients may be interpreted as kN/m if forces are in kN and displacements in m.

Here,

Since K_ss = 8 ≠ 0,

The condensed load is

Therefore,

Recovery of the internal displacement gives

Verification in the original equations gives 12(0.8) − 4(0.4) = 8 and −4(0.8) + 8(0.4) = 0. Thus, condensation has preserved the exact static solution. The apparent stiffness at the master DOF is 10 kN/m rather than the direct term 12 kN/m because the internal DOF deforms and thereby reduces resistance.

5. Dynamic Condensation and Model Reduction

For undamped free vibration, the finite-element equations are

where M is the mass matrix, K is the stiffness matrix, u is displacement, and ü is acceleration. A harmonic solution u = φ sin(ωt) produces the generalised eigenproblem

where ω is circular frequency in rad/s and φ is a mode shape.

Static condensation applied only to K is exact for static equilibrium but does not preserve the full dynamic behaviour. Several reduction methods are used.

5.1 Guyan reduction

Guyan reduction, or static reduction, assumes that slave inertia is negligible in the kinematic relation. With f_s = 0, the static relation is

where T_sm is the slave-to-master transformation. The full transformation is

The reduced mass and stiffness matrices are

The reduced eigenproblem is

Guyan reduction is generally most accurate for low-frequency modes when the retained DOFs adequately describe the dominant deformation. It may be inaccurate for higher modes or when slave inertia is significant.

5.2 Craig–Bampton component-mode synthesis

Component-mode synthesis retains physical interface coordinates and represents internal deformation using fixed-interface normal modes. In a common formulation,

where u_b are boundary/interface displacements, Ψ_cb are constraint modes, Φ_i are selected fixed-interface normal modes, and q_i are modal coordinates. Constraint modes describe static deformation caused by unit interface displacement with internal inertia ignored. Fixed-interface modes satisfy zero interface displacement and capture dynamic internal behaviour.

Craig–Bampton reduction is more suitable than pure static condensation when a component participates in dynamic analysis over a meaningful frequency range. Accuracy depends on the number of retained fixed-interface modes and the frequency range of interest.

5.3 Other reduction approaches

Dynamic condensation may be frequency-dependent. Partitioning the dynamic stiffness matrix K − ω²M and eliminating slave coordinates gives a frequency-dependent Schur complement. This can be accurate near selected frequencies but is less convenient for broad-band time-domain analysis. Modal truncation, balanced model reduction, Krylov-subspace methods, and proper orthogonal decomposition are alternative approaches, especially in large multiphysics or parametric models.

6. Computational Solution Techniques

After boundary conditions and condensation, the central computational task is to solve a linear system

where A may be a stiffness, tangent stiffness, reduced stiffness, or dynamic matrix. The choice of solver depends on symmetry, definiteness, sparsity, conditioning, size, and whether multiple right-hand sides must be solved.

6.1 Direct methods

Gaussian elimination transforms A into triangular factors. In structural analysis, sparse LU factorisation is common for unsymmetric systems. For SPD matrices, Cholesky factorisation gives

where L is lower triangular. For symmetric indefinite systems, an LDLᵀ factorisation with pivoting is typically required:

Pivoting improves numerical stability when diagonal pivots are small or sign-indefinite.

Direct methods are robust and particularly efficient when many load cases use the same A: factorisation is performed once, followed by triangular solves for each right-hand side. Their disadvantages are memory use and fill-in, where zero entries become nonzero during factorisation.

6.2 Iterative methods

Iterative methods generate successive approximations rather than computing a complete factorisation. The conjugate-gradient (CG) method is appropriate for SPD matrices. Generalised minimum residual (GMRES) is designed for general nonsymmetric systems, while MINRES is appropriate for symmetric indefinite systems under suitable conditions.

The residual at an approximate solution x_k is

A relative residual measure is

where ||·||₂ is the Euclidean norm. A small residual is necessary but not always sufficient for an accurate displacement: if A is ill-conditioned, small data or arithmetic errors can produce relatively large solution errors.

Preconditioning replaces the original problem by an equivalent system that is easier to solve. If P approximates A and is inexpensive to solve with, left preconditioning uses

Incomplete LU, incomplete Cholesky, algebraic multigrid, domain decomposition, and block preconditioners are common choices.

6.3 Sparse storage and ordering

Finite-element matrices are sparse because each element connects only a limited number of DOFs. Sparse storage schemes record nonzero coefficients rather than the full matrix. Reordering methods such as minimum degree and nested dissection reduce fill-in and factorisation cost. Ordering is therefore not merely a programming detail; it can determine whether a large analysis is computationally feasible.

6.4 Parallel and domain-decomposition methods

Large structures are often solved by partitioning the model into subdomains. Interface variables may be retained while internal variables are eliminated, producing a Schur-complement problem on the interfaces. This is the computational foundation of substructuring and many parallel finite-element solvers.

In primal methods, interface displacements are the primary unknowns. In dual methods, interface forces or Lagrange multipliers enforce compatibility. FETI-type and balancing domain-decomposition methods are examples of dual or hybrid approaches. Their performance depends on subdomain quality, scaling, coarse spaces, and communication overhead.

7. Conditioning, Accuracy, and Verification

The condition number of a nonsingular matrix in a chosen norm is

For the 2-norm, κ₂(A) = σ_max/σ_min, where σ_max and σ_min are the largest and smallest singular values. A large condition number indicates sensitivity to perturbations. In approximate terms, relative solution error may be bounded by a factor proportional to κ(A) times relative data or computational error.

Poor conditioning can result from:

• severe differences in material or geometric stiffness;
• inconsistent units;
• nearly dependent constraints;
• very slender or distorted elements;
• mechanisms or near-mechanisms;
• inappropriate multipoint-constraint penalties;
• excessive differences between translational and rotational scaling.

Verification should occur at several levels:

1. Equation verification: check equilibrium residuals after recovery.
2. Symmetry verification: confirm expected symmetry of conservative stiffness matrices.
3. Energy verification: for a linear system, strain energy is

and external work under a proportional load is consistent with the computed response.

1. Mesh and reduction verification: compare results with a refined model or additional retained modes.
2. Limiting-case checks: test zero load, rigid support, very large stiffness, and disconnected components where relevant.
3. Independent implementation: compare selected results with a separate solver or hand calculation.

For condensation, recovered slave displacements must be substituted into the original equations. It is not enough to inspect only the reduced solution.

8. Applied Case: Substructuring a Floor-Frame Region

Background

The following is an illustrative engineering case. A finite-element model of a long-span composite floor contains a detailed local region around a column opening. The local region includes shell elements, stiffeners, and several internal nodes. The global building model requires only the interface displacements and forces at the boundary of that region.

Problem

Assembling every local shell DOF into the building-wide matrix produces a large system and makes repeated load-case analysis expensive. The engineering team needs a reduced component that preserves static stiffness at the interface and can also support vibration analysis.

Analysis and approach

The local component is partitioned into interface DOFs b and internal DOFs i. For static load cases, the team factorises K_ii once and forms the condensed interface stiffness

Internal loads are included through the corresponding condensed force vector. The interface matrix is then assembled into the global building model. After solving for interface displacements, internal displacements are recovered by

For vibration analysis, the team does not reuse the static condensed stiffness alone. Instead, it constructs a Craig–Bampton component: all interface DOFs are retained, and fixed-interface modes are calculated for the internal region. Modes are selected up to a frequency above the design frequency range, with additional modes retained for margin.

Outcome

The static model preserves the interface response of the detailed component, subject to numerical tolerances and consistent assembly. The dynamic model captures both interface motion and internal deformation modes. Repeated load cases benefit from reuse of the internal factorisation, while the reduced global dynamic problem requires substantially less memory.

Learning points

• Condensation is a transfer of structural influence, not deletion of internal physics.
• Static and dynamic objectives require different reduction strategies.
• Interface selection controls both accuracy and computational cost.
• Recovery and residual checks are essential before accepting reduced results.
• A reduction basis should be documented with its frequency range, retained modes, constraints, and validation evidence.

9. Practical Limitations and Professional Considerations

Condensation is exact only relative to the original mathematical model and the assumptions of the reduction. It cannot correct poor geometry, incorrect material properties, unsuitable boundary conditions, or inadequate contact modelling. Nonlinear problems require additional care: a condensed tangent matrix may change at every Newton iteration, and eliminating variables can complicate path-dependent constitutive updates, contact status, and internal history variables.

For nonlinear equilibrium, an incremental Newton method commonly solves

where K_t is the tangent stiffness at iteration k, ∆u_k is the correction, and R is the unbalanced residual. Condensation can be applied to the tangent system, but the internal state must be updated consistently and the full residual must be checked. Unsymmetric tangents may arise in frictional contact or nonconservative problems.

Penalty methods for constraints can cause ill-conditioning when the penalty is excessively large. Lagrange multipliers add unknowns and may create an indefinite system. Constraint elimination is often efficient, but it must preserve compatibility and not inadvertently remove physical rigid-body motion.

Professional computational practice requires traceability. A model report should identify software version, element formulations, units, coordinate conventions, constraints, solver settings, tolerances, reduction choices, retained modes, and validation tests. Engineers must not present a reduced-model output as exact physical truth outside its validated range. Design decisions remain subject to applicable structural codes, peer review, quality assurance, and engineering judgement.

10. Contemporary Developments

Modern structural analysis increasingly combines condensation with high-performance computing, model updating, digital twins, and uncertainty quantification. Automated substructuring can generate reusable component models for repeated design studies. Reduced-order models support real-time monitoring and control, provided that changing damage, boundary conditions, and nonlinear effects are represented.

Machine learning can assist in surrogate modelling, preconditioner selection, anomaly detection, and prediction of reduced bases. However, data-driven models should be constrained by equilibrium, compatibility, objectivity, and conservation principles where possible. AI-generated numerical results require independent verification, provenance, and a clear statement of the training domain. They should supplement, not replace, validated finite-element analysis and professional responsibility.

Best-Practice Workflow

A reliable workflow is:

1. Define the analysis objective and the required response quantities.
2. Establish units, coordinates, material laws, supports, and load conventions.
3. Assemble and inspect the unreduced model where feasible.
4. Select master/interface DOFs based on physics, connectivity, measurement, and output requirements.
5. Check that the slave block is appropriate and nonsingular.
6. Factorise rather than explicitly invert the slave block.
7. Form the condensed stiffness, mass, and load quantities required by the analysis.
8. Select a solver consistent with symmetry, definiteness, sparsity, and conditioning.
9. Solve, recover eliminated variables, and evaluate residuals and energy measures.
10. Validate against a detailed model, limiting case, refined discretisation, or independent calculation.
11. Document the reduction basis, tolerances, convergence evidence, and limitations.

Chapter Summary

Structural condensation eliminates selected degrees of freedom while retaining their mechanical influence through a Schur complement. Partitioning the equilibrium equations into master and slave variables produces a condensed stiffness and load vector, provided the slave block is nonsingular. Internal displacements are recovered after solving the reduced system, and residual checks confirm the validity of the computation.

Static condensation is exact for the represented linear static problem, whereas dynamic analysis requires mass-aware reduction. Guyan reduction is suitable when low-frequency behaviour dominates and slave inertia is modest. Craig–Bampton component-mode synthesis combines interface coordinates, constraint modes, and fixed-interface normal modes for more effective dynamic substructuring.

Direct sparse factorisation is robust and useful for repeated load cases, while iterative methods can be advantageous for very large sparse systems when effective preconditioning is available. Conditioning, fill-in, ordering, scaling, constraint treatment, and verification strongly influence reliability. Professional use requires documenting assumptions, validating the reduced model, and recognising that computational efficiency cannot substitute for physical correctness.

Key Terms

Master DOF: A retained degree of freedom in a reduced structural model.

Slave DOF: A degree of freedom eliminated during condensation.

Static condensation: Elimination of internal variables from a linear static system.

Schur complement: The matrix obtained after eliminating one block of a partitioned system.

Condensed stiffness: Effective stiffness relating retained forces to retained displacements.

Guyan reduction: Static reduction used as an approximation to dynamic behaviour.

Component-mode synthesis: Dynamic model-reduction approach based on component modes.

Constraint mode: Component deformation caused by unit interface displacement with internal coordinates constrained dynamically.

Fixed-interface mode: Natural mode of a component with its interface held fixed.

Sparsity: The property that most matrix entries are zero.

Fill-in: New nonzero entries created during matrix factorisation.

Condition number: A measure of sensitivity of a matrix problem to perturbations.

Preconditioner: An auxiliary operator that improves iterative convergence.

Residual: The imbalance b − Ax for an approximate solution.

Cholesky factorisation: Factorisation A = LLᵀ for an SPD matrix.

Substructuring: Decomposition of a structure into computational components.

Interface DOF: A retained variable connecting a component to other components.

Model reduction: Representation of a large model with fewer coordinates.

Tangent stiffness: The derivative of internal force with respect to displacement in nonlinear analysis.

Craig–Bampton method: A component-mode synthesis method retaining interface coordinates and internal fixed-interface modes.

Review Questions

Short Answer Questions

1. What is the physical meaning of the Schur-complement correction in a condensed stiffness matrix?
2. Why should an explicit inverse of K_ss generally not be formed in numerical implementation?
3. State one condition under which Cholesky factorisation is appropriate.
4. Why is static condensation alone insufficient for general dynamic analysis?
5. What is the purpose of recovering slave displacements after solving the reduced system?

Long Answer Questions

1. Derive the condensed stiffness and load equations from a block-partitioned structural equilibrium system.
2. Compare static condensation, Guyan reduction, and Craig–Bampton component-mode synthesis.
3. Explain how sparsity, fill-in, and matrix ordering affect computational structural analysis.
4. Discuss the causes and consequences of ill-conditioning in finite-element systems.
5. Develop a verification plan for a condensed finite-element model used in design.

Analytical / Application Questions

1. For a system with K_mm = 20, K_ms = −6, K_sm = −6, K_ss = 12, f_m = 14, and f_s = 3, calculate the condensed stiffness, condensed load, master displacement, and recovered slave displacement. Use consistent units.
2. A component is required to predict vibration accurately up to 80 Hz but has many internal DOFs. Recommend a reduction strategy and justify the selection of retained coordinates and internal modes.
3. A reduced model gives a small reduced residual but large errors in recovered internal stresses. Identify possible causes and propose a systematic diagnostic procedure.

Multiple Choice Questions with Answers

Question 1

For a partitioned system, which expression is the condensed stiffness matrix after eliminating slave DOFs?

A. K_mm + K_ms K_ss^(-1) K_sm

B. K_mm − K_ms K_ss^(-1) K_sm

C. K_ss − K_sm K_mm^(-1) K_ms

D. K_mm K_ss − K_ms K_sm

Correct Answer: B. K_mm − K_ms K_ss^(-1) K_sm

Question 2

What condition is required for the basic static-condensation formula to be directly applicable?

A. K_ss must be diagonal

B. K_ss must be nonsingular

C. K_mm must be zero

D. The global matrix must be unsymmetric

Correct Answer: B. K_ss must be nonsingular

Question 3

In the worked scalar example, the condensed stiffness was:

A. 8 kN/m

B. 10 kN/m

C. 12 kN/m

D. 14 kN/m

Correct Answer: B. 10 kN/m

Question 4

Which method is specifically designed to retain interface coordinates and internal fixed-interface modes?

A. Cholesky factorisation

B. Guyan reduction

C. Craig–Bampton component-mode synthesis

D. Gaussian scaling

Correct Answer: C. Craig–Bampton component-mode synthesis

Question 5

Cholesky factorisation is normally appropriate when the matrix is:

A. Symmetric positive definite

B. Nonsymmetric and singular

C. Rectangular and rank deficient

D. Symmetric with arbitrary indefiniteness

Correct Answer: A. Symmetric positive definite

Question 6

The residual for an approximate solution x_k of Ax = b is:

A. Ax_k + b

B. A^(-1)x_k − b

C. b − Ax_k

D. x_k − Ab

Correct Answer: C. b − Ax_k

Question 7

Which computational phenomenon creates additional nonzero matrix entries during factorisation?

A. Condensation

B. Fill-in

C. Aliasing

D. Modal orthogonality

Correct Answer: B. Fill-in

Question 8

For an SPD system, which iterative method is generally appropriate?

A. Conjugate gradient

B. GMRES only

C. Explicit inversion

D. Fourier elimination

Correct Answer: A. Conjugate gradient

Question 9

If direct loads act on slave DOFs, the condensed load vector includes:

A. Only f_m, regardless of f_s

B. The term −K_ms K_ss^(-1) f_s

C. The term +K_ss f_s

D. No load contribution

Correct Answer: B. The term −K_ms K_ss^(-1) f_s

Question 10

Why can a small residual fail to guarantee a highly accurate solution?

A. Residuals are unrelated to equilibrium

B. The matrix may be ill-conditioned

C. The matrix must be diagonal

D. Condensation always changes units

Correct Answer: B. The matrix may be ill-conditioned