The Rayleigh–Ritz method is a technique in numerical analysis for approximating variational problems by restricting the admissible solutions to a finite-dimensional space. An unknown function or vector is expressed as a linear combination of chosen trial functions, and its coefficients are determined by minimizing an energy functional or making a Rayleigh quotient stationary. The resulting problem is a finite system of algebraic equations or a matrix eigenvalue problem. The method is central to structural mechanics, quantum mechanics, and the finite element method, and also serves as an eigenvalue extraction procedure in numerical linear algebra.
Principle and historical background
The method combines two closely related ideas:
- Approximate a function by an element of a manageable trial space.
- Determine the approximation through a variational condition rather than by enforcing a differential equation at every point.
Its name associates the work of Lord Rayleigh on vibration frequencies and energy quotients with Walther Ritz’s systematic development of finite-dimensional variational approximations in the early twentieth century. Ritz’s work extended the approach to broad classes of boundary-value and vibration problems.
Terminology varies by field. Ritz method often denotes direct minimization of an energy functional over a trial space. Rayleigh–Ritz method frequently emphasizes eigenvalue approximation through a projected operator or a stationary quotient. These are related constructions, but an energy-minimization problem need not be an eigenvalue problem.
The underlying framework belongs to the calculus of variations: a problem about functions is reduced to a problem about finitely many coefficients. Its effectiveness depends on choosing trial spaces that both satisfy the mathematical constraints and approximate the desired solution well.
Variational formulation for boundary-value problems
Consider a real Hilbert space , a continuous symmetric bilinear form , and a continuous linear functional . Define the energy functional
Suppose is coercive: there is a constant such that
Under these assumptions, has a unique minimizer . Its first variation vanishes, giving
This is commonly a weak formulation of a boundary-value problem. For differential equations, it may require fewer derivatives of the solution than the corresponding pointwise equation and thereby admit a weak solution.
Choose a finite-dimensional trial space
where the trial functions are linearly independent. The Ritz approximation is
with coefficients chosen to minimize over . Differentiating with respect to each coefficient gives
Thus the infinite-dimensional problem becomes the linear system
where
The matrix , often called the stiffness matrix, is symmetric positive definite under the stated assumptions. It can therefore be solved using methods such as Cholesky decomposition.
Boundary conditions
Essential boundary conditions, such as prescribed displacements or prescribed function values, must be built into the admissible trial functions or imposed through an appropriate constraint formulation.
For nonhomogeneous essential conditions, an admissible function is first chosen to satisfy the prescribed boundary data, and the approximation is written as
where the satisfy the corresponding homogeneous conditions. The admissible set is then affine rather than a vector space.
Natural boundary conditions, such as specified fluxes or tractions, usually enter through the variational formulation after integration by parts. Trial functions need not generally satisfy those conditions pointwise.
Rayleigh–Ritz approximation of eigenvalues
Let be a symmetric energy form and a symmetric positive-definite mass or overlap form. The variational eigenproblem is
The associated Rayleigh quotient is
Its stationary values correspond to eigenvalues under the appropriate domain and spectral assumptions. Minimizing it gives the lowest eigenvalue when that eigenvalue exists and the variational principle applies; higher eigenvalues are characterized through additional subspace or orthogonality conditions.
Restricting the problem to produces
where
The approximate eigenvalues are called Ritz values, and the corresponding functions
are Ritz vectors or Ritz functions. If the basis is orthonormal with respect to , then ; otherwise, the generalized eigenvalue problem must retain the mass matrix.
Matrix formulation
For a Hermitian matrix , let the columns of form an orthonormal basis for a trial subspace. The reduced problem is
where denotes the conjugate transpose. Here is a Ritz vector approximating an eigenvector of .
Its residual
satisfies
Thus the residual is orthogonal to the entire trial space. This makes the method a projection procedure closely related to orthogonal projection.
For a nonorthonormal basis , the corresponding reduced problem is
Rayleigh–Ritz extraction can also be performed for nonsymmetric matrices, but the real-valued minimization principles and upper-bound guarantees of the self-adjoint case generally do not extend to that setting.
Variational bounds and convergence
Eigenvalue upper bounds
For a suitable semibounded self-adjoint operator with discrete eigenvalues, the min–max principle characterizes the ordered eigenvalues by
Restricting the candidate subspaces to those contained in gives
Consequently, a conforming Rayleigh–Ritz approximation satisfies
If the trial spaces are nested,
then
The approximate eigenvalues therefore decrease toward the exact ones when the spaces become sufficiently rich and the relevant spectral approximation hypotheses hold. Without nested spaces, upper bounds may still hold, but monotonicity between successive approximations is not assured.
These statements require conforming trial spaces and the correct forms. Nonconforming discretization, approximate integration, altered operators, or numerical errors can remove the upper-bound property.
Best approximation in the energy norm
For the coercive source problem, subtracting the exact and discrete equations yields Galerkin orthogonality:
Because is symmetric and positive definite, it defines the energy norm
The Ritz solution is the best approximation in this norm:
It also satisfies the energy identity
If has continuity constant and coercivity constant , the corresponding bound in the original space norm is
This is a form of Céa’s lemma. Convergence follows when the trial spaces approximate the exact solution arbitrarily well in the relevant norm.
Example: the lowest eigenvalue of a fixed string
Consider
with
The exact eigenvalues are
The variational setting uses the Sobolev space , with
Choose the one-dimensional trial space generated by
This function satisfies both endpoint conditions. Its Rayleigh quotient is
The lowest Ritz value is therefore , an upper bound for
The single trial function captures the broad shape of the first vibration mode. Adding suitable independent trial functions can improve the bound. Because the quotient is unchanged by scaling , the coefficient of a one-dimensional trial function determines normalization but not the approximate eigenvalue.
Relationship to other numerical methods
Galerkin methods
A Galerkin method requires the residual of a weak equation to vanish against every test function in the chosen test space. For symmetric coercive problems, using the same trial and test space gives exactly the Ritz equations.
The distinction is structural: Galerkin projection applies to problems that need not arise from energy minimization, including many nonsymmetric equations. The Ritz method, in its minimization sense, relies on an underlying variational functional. A stationary functional also need not possess a minimum if the problem is indefinite.
Finite elements and spectral approximations
The finite element method typically uses piecewise-polynomial trial functions with local support. For symmetric coercive problems, conforming finite element approximations are Ritz approximations. Local support leads to a sparse matrix, allowing large discretizations to be assembled and solved efficiently.
Global polynomials, trigonometric functions, or other problem-adapted functions can instead be used as trial bases. Such choices connect Rayleigh–Ritz approximation with spectral methods and approximation theory. Smooth solutions may admit very rapid convergence, whereas singularities or limited regularity can substantially reduce the rate.
Iterative eigenvalue algorithms
In numerical linear algebra, Rayleigh–Ritz extraction is a component of many large-scale eigenvalue algorithms. A Krylov subspace or another adaptively constructed trial space is generated first, and a smaller projected eigenproblem is then solved within it.
The extraction procedure and the construction of the subspace are separate steps. Rayleigh–Ritz determines the approximate eigenpairs available in the current space; an iterative algorithm determines how that space is expanded, filtered, or restarted.
Applications and limitations
In structural mechanics, the method approximates displacement fields, equilibrium configurations, vibration modes, and, in suitable formulations, buckling loads. The matrices and represent stiffness and mass in vibration analysis.
In quantum mechanics, trial wavefunctions give matrix approximations to a Hamiltonian. For a self-adjoint Hamiltonian bounded below, admissible normalized trial states have energy expectations no lower than the ground-state energy. Excited-state calculations require additional variational structure, such as subspace minimization and orthogonality constraints; unrestricted minimization alone selects the lowest state available in the trial space.
The principal limitations concern approximation quality and the preservation of the variational structure:
- Trial-space quality: A large space may still be ineffective if it poorly represents boundary layers, singularities, oscillations, or localization.
- Conditioning: Nearly dependent basis functions can make the stiffness or mass matrix ill-conditioned. Basis scaling and orthogonalization improve numerical stability but do not correct an inadequate approximation space.
- Regularity requirements: Trial functions must belong to the energy space. A formulation involving second derivatives may require more continuity than a standard second-order elliptic formulation.
- Spectral gaps and clusters: Individual eigenvectors can be sensitive when eigenvalues are close. Approximation of the associated invariant subspace is often more informative than comparison of single vectors.
- Spectral pollution: In some infinite-dimensional problems, especially approximation inside gaps of an essential spectrum, projected eigenvalues may converge to values outside the true spectrum. The simple low-lying discrete-eigenvalue picture does not cover these cases.
- Computational cost: Global trial functions often yield dense matrices, while local bases generally provide more favorable sparsity.
Ritz values alone do not quantify every aspect of solution accuracy. Residual norms, approximation estimates, spectral separation, and checks that the discretization preserves the intended variational problem provide distinct information about the reliability of a computed result.