1 Statement and Intuition

1.1 Variational viewpoint of eigenvalues

For a self-adjoint operator on a Hilbert space, spectral values can be interpreted through optimization. Instead of viewing eigenvalues as abstract solutions of an operator equation, the min–max principle treats them as extremal values of a quadratic “energy” associated with the operator. In this perspective, one searches over subspaces and asks how small or large the operator’s energy can be made uniformly for vectors in those subspaces, depending on whether one uses minimization or maximization.

This variational viewpoint is especially effective for operators whose spectrum contains isolated eigenvalues of finite multiplicity (for example, compact self-adjoint operators). The optimization procedure then produces an ordered sequence of eigenvalues without requiring explicit diagonalization.

1.2 “Min over subspaces / max over vectors” interpretation

A common form of the principle can be summarized as follows: for the first eigenvalue, one considers all one-dimensional subspaces and, within each, maximizes the energy over unit vectors; then one takes the minimum across all such subspaces. The second eigenvalue arises by repeating the same idea using two-dimensional subspaces, again minimizing the worst-case energy over the chosen subspace. Higher eigenvalues follow by increasing the dimension of the subspace.

This “min over subspaces / max over vectors” structure captures the idea that lower eigenvalues are those for which every sufficiently large subspace must contain some direction with energy no greater than that value, while higher eigenvalues reflect the presence of directions with larger energy in a controlled way.

1.3 Rayleigh quotient and energy functionals

The optimization is typically performed using the Rayleigh quotient \[ R(x)=\frac{\langle Tx,x\rangle}{\langle x,x\rangle}, \] defined for vectors \(x\) in the domain of a self-adjoint operator \(T\). The numerator \(\langle Tx,x\rangle\) is the energy associated with \(x\), and the Rayleigh quotient measures the energy per unit norm. Eigenvectors make the Rayleigh quotient constant at the corresponding eigenvalue, since if \(Tx=\lambda x\) then \(R(x)=\lambda\).

In many formulations, the min–max principle is expressed entirely in terms of extremal values of \(R(x)\) on subspaces. This connects spectral information to a computable scalar functional.

1.4 Nested subspaces and ordering of eigenvalues

The ordering of eigenvalues is enforced by using nested chains of subspaces, typically \[ \{0\}\subset V_1\subset V_2\subset \cdots \subset V_k. \] The dimension of these spaces is what determines the index \(k\) in the min–max characterization. As \(k\) increases, the optimization becomes more restrictive in one direction and more permissive in another, producing an ordered sequence of eigenvalues.

The nesting structure also explains why the principle yields a monotone sequence: the constraints implied by “dimension \(k\)” change in a systematic way, which transfers to the extremal values of the energy functional.

2 Classical Finite-Dimensional Form

2.1 Self-adjoint operators on Hilbert spaces

In finite dimensions, a self-adjoint operator corresponds to a Hermitian (or real symmetric) matrix. The Hilbert space is \(\mathbb{C}^n\) (or \(\mathbb{R}^n\)) equipped with the usual inner product, and the operator admits an orthonormal eigenbasis. The min–max principle becomes a precise algebraic statement relating eigenvalues to extremal values of the Rayleigh quotient over subspaces.

Because the setting is compact and the spectrum consists solely of eigenvalues, the variational procedure recovers all eigenvalues with multiplicities.

2.2 Characterization of ordered eigenvalues

Let \(T\) be self-adjoint on an \(n\)-dimensional Hilbert space, with eigenvalues ordered decreasingly: \[ \lambda_1\ge \lambda_2\ge \cdots \ge \lambda_n. \] A typical Courant–Fischer min–max formulation states that for each \(k=1,\dots,n\), \[ \lambda_k=\min_{\substack{W\subset H\\ \dim W = k}} \ \max_{\substack{x\in W\\ x\ne 0}} R(x), \] and equivalently, \[ \lambda_k=\max_{\substack{U\subset H\\ \dim U = n-k+1}} \ \min_{\substack{x\in U\\ x\ne 0}} R(x). \] The first expression identifies \(\lambda_k\) as the minimum, over all \(k\)-dimensional subspaces, of the largest Rayleigh quotient achievable inside that subspace. The second expression uses the complementary perspective on \((n-k+1)\)-dimensional subspaces, maximizing the smallest Rayleigh quotient attainable there.

2.3 Relation to orthogonal complements

The equivalence between the two min–max forms relies on orthogonal complements. For a subspace \(W\), its orthogonal complement \(W^\perp\) has dimension \(n-\dim W\), and vectors constrained to one side correspond to constraints on the other side. Under self-adjointness, the Rayleigh quotient interacts well with this decomposition, allowing the “min over \(W\)” formulation to be transformed into a “max over \(U=W^\perp\)” formulation.

Conceptually, the first formula looks at the worst energy inside chosen directions, while the second looks at the best lower guarantee you can enforce by choosing complementary directions.

2.4 Examples with symmetric matrices

Consider a real symmetric matrix \(A\). For a fixed unit vector \(x\), the Rayleigh quotient \(R(x)=x^\top Ax\) is a scalar. The min–max principle then says that \(\lambda_1\) is the minimum over all one-dimensional subspaces of the maximum value of \(x^\top Ax\) on that line—equivalently, it is the maximum of \(x^\top Ax\) over all unit vectors. Likewise, \(\lambda_2\) can be obtained by taking all two-dimensional subspaces, looking at the maximum Rayleigh quotient inside each, and then minimizing over those choices.

In practice, these formulas illustrate how eigenvalues are determined by collective behavior across families of subspaces rather than by any single vector.

3 Infinite-Dimensional and Spectral-Approximation Form

3.1 Compact self-adjoint operators

In infinite-dimensional Hilbert spaces, a compact self-adjoint operator \(T\) has a spectrum consisting of real eigenvalues that accumulate only at \(0\). Each nonzero eigenvalue has finite multiplicity, and there is an orthonormal basis of eigenvectors spanning the closure of the range.

For such operators, one can index eigenvalues \(\lambda_1\ge \lambda_2\ge \cdots \to 0\) (including multiplicities). The min–max characterization extends by restricting to finite-dimensional subspaces and applying the same optimization structure. The result provides a way to recover the discrete eigenvalues through variational limits over increasing subspaces.

3.2 Weak convergence and continuity requirements

When passing from finite-dimensional approximations to infinite-dimensional limits, one must ensure the variational quantities behave well under convergence. Because \(\langle Tx,x\rangle\) is not automatically continuous with respect to weak convergence of \(x\), assumptions such as compactness (or stronger topological properties of the operator or the domain) are used to guarantee that extremal values converge appropriately.

For compact self-adjoint operators, compactness implies that bounded sequences can have subsequences on which \(Tx_n\) converges strongly, enabling control of Rayleigh quotient values and ensuring that the min–max procedure remains meaningful in the limit.

3.3 Min–max for essential spectral behavior (conditions-based)

The essential spectrum describes spectral values that are stable under compact perturbations and is not captured purely by isolated eigenvalues. Min–max principles can still be formulated to detect parts of the spectrum under suitable hypotheses, but the exact characterization depends on the operator class and the type of “spectrum” under consideration.

A typical approach uses conditions such as form-boundedness, semi-boundedness, or other structural assumptions, and replaces eigenvalue indexing with variational characterization of spectral thresholds. One formulation involves generalized Rayleigh quotients or limiting subspaces, where sequences may escape to infinity in a way that reflects essential spectrum rather than point spectrum.

3.4 Practical implications for operator limits

In spectral approximation, one often studies a sequence of operators \(T_n\) intended to approximate \(T\). Variational principles provide bounds on eigenvalues of \(T_n\) in terms of optimization over subspaces related to the approximation scheme. If \(T_n\) converges to \(T\) in an appropriate operator sense, the corresponding extremal values can converge as well, enabling rigorous justification of numerical eigenvalue methods.

Thus, the min–max principle functions not only as a theoretical characterization but also as a stability and convergence tool.

4 Connection to the Rayleigh–Ritz Method

4.1 Choosing trial subspaces

The Rayleigh–Ritz method approximates eigenvalues by restricting the optimization to a finite-dimensional “trial” subspace \(V_m\). One computes the Ritz values, which are eigenvalues of the compressed operator \(T_{V_m}\) (more precisely, of the operator associated with the restriction of the quadratic form to \(V_m\)). The min–max principle ensures that these Ritz values bound the true eigenvalues of \(T\).

The choice of trial subspace—often spanned by basis functions, finite element shape functions, or other ansatz functions—directly affects the accuracy. Enlarging \(V_m\) typically improves approximations in a manner consistent with the variational ordering.

4.2 Upper and lower bounds from truncation

For many settings with discrete eigenvalues, the Rayleigh–Ritz approximations produce monotone bounds: depending on how the subspaces are nested and how one indexes the Ritz values, the computed quantities can serve as upper or lower bounds for corresponding exact eigenvalues. The min–max principle explains these inequalities by relating optimization over \(V_m\) to optimization over all relevant subspaces.

This bounding property is valuable because it allows one to certify not only approximate eigenvalues but also their direction of error.

4.3 Convergence of approximate eigenvalues

If the union of trial spaces becomes dense in the relevant form domain (or if the approximation satisfies suitable convergence conditions), the Ritz values converge to the exact eigenvalues. The convergence can be understood through variational characterization: the restricted optimization problem approaches the full one as trial subspaces grow.

In the compact self-adjoint case, this often yields straightforward convergence for eigenvalues associated with eigenvectors that can be well approximated by vectors in the trial spaces.

4.4 Error monitoring using variational estimates

Beyond computing approximations, one may monitor error using the variational inequalities. For instance, one can compute approximate Rayleigh quotients for candidate eigenvectors and compare them with min–max bounds derived from complementary subspaces. Such comparisons can provide practical stopping criteria in iterative eigenvalue solvers.

Even when exact eigenfunctions are unknown, the variational framework offers computable quantities that bracket the true eigenvalues, improving the reliability of numerical results.

5 Applications to Differential Operators

5.1 Sturm–Liouville theory (variational eigenvalue characterization)

Sturm–Liouville operators, common in mathematical physics, often admit a spectral decomposition with real eigenvalues and orthogonal eigenfunctions. Many Sturm–Liouville problems can be expressed through quadratic forms, making the min–max principle directly applicable.

In this setting, eigenvalues correspond to stationary values of the associated energy functional under appropriate boundary conditions. The ordering of eigenvalues and the orthogonality structure of eigenfunctions can be reflected through the nested subspace optimization inherent in min–max.

5.2 Laplacians and boundary value problems (general framework)

For Laplace-type operators on domains, the spectral problem is frequently formulated as \[ -\Delta u = \lambda u \] with boundary conditions such as Dirichlet or Neumann. The min–max principle characterizes the eigenvalues by minimizing the Dirichlet energy (or an analogous form) over suitable Sobolev spaces.

The principle thereby provides a unified mechanism for understanding how geometry, boundary conditions, and constraints influence the spectrum. It is particularly suited to studying how changes in domain or constraints alter eigenvalue locations.

5.3 Eigenvalue bounds via test functions

A central application is obtaining estimates without solving the differential equation explicitly. By selecting a convenient finite-dimensional test subspace and evaluating the Rayleigh quotient over it, one obtains an upper bound on certain eigenvalues (or, using complementary subspaces, a lower bound). This is the variational method in its most computationally oriented form.

The quality of the bounds depends on how well the chosen test functions approximate the true eigenfunctions or capture their dominant oscillation patterns.

5.4 Stability criteria from spectral gaps

In stability analysis, what matters is not only whether eigenvalues are positive or negative, but also how large a “gap” is between spectral regions. Min–max principles help translate energy inequalities into statements about spectral gaps: if one can bound the Rayleigh quotient from above or below uniformly on constrained subspaces, one obtains information about the number of eigenvalues in certain intervals.

This mechanism is used in various contexts where one needs to ensure that perturbations do not push the operator’s spectrum across a critical threshold.

6.1 Courant–Fischer theorem (and equivalences)

The classical finite-dimensional min–max statement is often referred to as the Courant–Fischer theorem. In practice, several equivalent forms exist, including formulations using increasing subspaces, decreasing subspaces, or orthogonal complements.

These equivalences are not merely cosmetic: they allow one to tailor the optimization direction to the available information. For example, when one has a natural “large” subspace capturing approximate eigenvectors, one may use the form that yields upper bounds naturally. When control is better on an orthogonal complement, the complementary version may be more convenient.

6.2 Courant nodal domain connections (overview-level)

Courant’s nodal domain theorem relates eigenvalues of differential operators to the number of nodal domains of eigenfunctions. While it is not the min–max principle itself, nodal information connects to variational characterization: eigenfunctions corresponding to low-lying eigenvalues minimize energy and exhibit constrained oscillation, which in turn limits nodal complexity.

In operator-theoretic approaches, variational tools can help underpin arguments about nodal sets by producing eigenvalue-dependent bounds derived from subspace restrictions.

6.3 Interlacing inequalities from subspace restriction

When one restricts an operator to a subspace or considers a compression, eigenvalues often interlace with those of the original operator. Min–max principles provide a mechanism for proving such interlacing inequalities: optimizing over a smaller set of admissible subspaces generally shifts extremal Rayleigh values in a predictable way, and nested spaces yield monotone comparisons.

Interlacing is widely used in numerical linear algebra and in spectral theory of discrete approximations, where one compares eigenvalues of a full operator with those of reduced models.

The min–max structure is a specific example of a broader family of variational methods in which one seeks critical values of functionals over constrained sets. In many contexts, eigenvalue problems correspond to saddle points of an energy functional. The min–max principle can then be seen as a spectral analogue of saddle-point characterization, with the operator’s self-adjointness ensuring that the optimization produces meaningful, real-valued extrema.

Generalizations include frameworks based on quadratic forms, min–max for spectral projections, and principles adapted to semi-bounded operators.

7 Practical Considerations

7.1 Regularity assumptions and domain issues

For unbounded operators, the Rayleigh quotient and the quadratic form may require careful domain selection. One must ensure that the trial subspace vectors lie in the form domain so that \(\langle Tx,x\rangle\) (or the associated form) is finite. Regularity assumptions on coefficients and boundary conditions in differential settings also affect whether eigenvalues are well-defined and whether variational principles apply directly.

These issues influence both the theoretical correctness and the numerical implementation, since the chosen basis must respect the operator’s domain constraints.

7.2 Handling multiplicities and degeneracies

Eigenvalues with multiplicity complicate interpretation because the min–max principle yields ordered values but does not uniquely select eigenvectors. In computations, numerical eigensolvers may produce any orthonormal basis of the eigenspace. Variational characterizations still hold at the level of eigenvalues, but error assessment may need to be phrased in terms of subspace angles or residual norms rather than pointwise comparison of eigenvectors.

When degeneracies occur, tracking invariant subspaces is often more stable than tracking individual vectors.

7.3 Numerical stability in variational computations

Implementing Rayleigh–Ritz requires forming matrices associated with the operator and the chosen basis. Numerical stability can be affected by ill-conditioning of the basis, near-linear dependence, or rounding errors in orthonormalization. Since the variational bounds rely on accurate computation of Rayleigh quotients and Ritz values, numerical errors can produce misleading monotonicity or apparent violations of bounds.

Using orthonormal or well-conditioned bases and reliable linear algebra routines helps preserve the theoretical guarantees as closely as possible.

7.4 Common pitfalls in implementing min–max bounds

A frequent pitfall is using trial functions that do not satisfy the operator’s boundary conditions or belong to the correct form domain, leading to Rayleigh quotients that correspond to a different operator. Another issue is assuming monotone convergence without verifying that the sequence of subspaces is nested or that the operator convergence assumptions are met.

Additionally, one may confuse indexing conventions (e.g., ordering decreasing vs increasing, or handling zero/accumulation points) and thereby mismatch computed Ritz values with the intended eigenvalue in the min–max formula.

8 Further Reading and Exercises

8.1 Standard references in spectral theory

Classical treatments of the min–max principle and its variants appear in standard texts on functional analysis and spectral theory. For differential operators, dedicated chapters on Sturm–Liouville theory, eigenvalues of elliptic operators, and variational methods provide concrete applications and derivations tailored to common boundary value problems.

When reading, it is helpful to focus on sections that connect quadratic forms, Rayleigh quotients, and the ordering of eigenvalues.

8.2 Guided derivations for special cases

Many special cases yield accessible derivations: finite-dimensional Hermitian matrices, compact self-adjoint operators with explicit eigen-expansions, and Sturm–Liouville problems with well-chosen test functions. Working through these examples helps clarify how orthogonal complements and nested subspaces drive the optimization structure.

Guided derivations often emphasize how to translate the operator problem into a variational one and back.

8.3 Exercise set: compute bounds for model operators

Exercises can be designed around explicit operators, such as diagonal matrices, rank-one perturbations, or one-dimensional differential operators with simple boundary conditions. Students can compute Rayleigh quotient maxima on chosen subspaces and then form the min–max bounds directly.

Typical tasks include improving bounds by enlarging trial spaces and comparing the resulting estimates with known exact eigenvalues.

8.4 Projects: variational convergence experiments

A project-oriented approach is to implement Rayleigh–Ritz numerically for a model operator and run experiments showing convergence as the trial space grows. One can track how the approximate eigenvalues approach the true ones, study the behavior under basis refinement, and test sensitivity to subspace nesting.

To connect back to theory, the project should also record bounds (upper/lower) when monotonicity is expected and examine where numerical errors degrade the theoretical ordering.