1 Introduction to Eigenvalue Localization

1.1 Motivation and why bounds matter

For a square matrix, eigenvalues determine many qualitative and quantitative features: long-run behavior of linear dynamical systems, convergence rates of iterative methods, and growth/decay properties in numerical schemes. In many settings, computing eigenvalues exactly is expensive, unstable, or unnecessary. Eigenvalue localization provides alternative information—geometric regions in the complex plane that are guaranteed to contain the spectrum—so one can estimate spectral bounds from readily available matrix data.

1.2 Statement of eigenvalues for complex matrices

Let \(A\in\mathbb{C}^{n\times n}\). The eigenvalues of \(A\) are complex numbers \(\lambda\) for which there exists a nonzero vector \(x\in\mathbb{C}^n\) satisfying \(Ax=\lambda x\). Equivalently, \(\lambda\) is a root of \(\det(A-\lambda I)=0\). While this definition is algebraic, localization aims to translate it into geometric constraints on where such \(\lambda\) can lie.

1.3 From algebraic definitions to geometric regions

Gershgorin’s circle theorem begins with the identity \[ (A-\lambda I)x=0, \] and uses entrywise inequalities to force \(\lambda\) to satisfy certain distance conditions relative to the diagonal entries of \(A\). The result is a collection of disks (often called “circles” in the sense of boundary curves) in the complex plane: if a number fails to lie in these disks, then it cannot be an eigenvalue. Thus, the spectrum is trapped within a union of regions determined by row or column magnitudes of off-diagonal entries.

2 Gershgorin’s Theorem (Core Result)

2.1 Row-based Gershgorin circles

2.1.1 Circle centers and radii from matrix entries

Write \(A=(a_{ij})_{i,j=1}^n\). For each row \(i\), define the center \(a_{ii}\) and the radius \[ R_i=\sum_{j\neq i}\lvert a_{ij}\rvert. \] The associated Gershgorin disk is \[ D_i=\{\lambda\in\mathbb{C}:\lvert \lambda-a_{ii}\rvert\le R_i\}. \] Gershgorin’s theorem asserts that every eigenvalue of \(A\) lies in at least one of the disks \(D_i\).

2.1.2 Interpretation of the union of circles

Let \[ \mathcal{D}=\bigcup_{i=1}^n D_i. \] Then the spectrum \(\sigma(A)\) satisfies \(\sigma(A)\subseteq \mathcal{D}\). Geometrically, the theorem states that the complex plane is partitioned into regions inside these row-based disks and regions outside them; eigenvalues cannot occur outside the union. Although different disks may overlap, their union is the guaranteed localization set.

2.2 Column-based Gershgorin circles

2.2.1 Centers and radii via column sums

A parallel construction uses columns. For each column \(j\), set the center \(a_{jj}\) and the radius \[ C_j=\sum_{i\neq j}\lvert a_{ij}\rvert. \] Define the column disk \[ E_j=\{\lambda\in\mathbb{C}:\lvert \lambda-a_{jj}\rvert\le C_j\}. \] Then \(\sigma(A)\subseteq \bigcup_{j=1}^n E_j\).

2.2.2 Equivalence of row and column formulations

Row-based and column-based versions are consistent because both follow from applying the same underlying argument to \(A\) with different ways of estimating components of an eigenvector relation. In practice, either formulation may yield tighter information depending on how the matrix entries distribute across rows versus columns. Taking intersections of resulting regions can also sharpen localization when combined with further refinements.

3 Proof Ideas and Intuition

3.1 Key inequality using matrix-vector relations

3.1.1 Deriving bounds from the resolvent-like argument

Assume \(\lambda\) is an eigenvalue and \(x\neq 0\) satisfies \((A-\lambda I)x=0\). Let \(k\) be an index where \(\lvert x_k\rvert\) is maximal among \(\{\lvert x_1\rvert,\dots,\lvert x_n\rvert\}\). The \(k\)-th component of \((A-\lambda I)x=0\) gives \[ (a_{kk}-\lambda)x_k+\sum_{j\neq k} a_{kj}x_j=0, \] so \[ (\lambda-a_{kk})x_k=\sum_{j\neq k} a_{kj}x_j. \] Taking absolute values yields \[ \lvert \lambda-a_{kk}\rvert\,\lvert x_k\rvert \le \sum_{j\neq k}\lvert a_{kj}\rvert\,\lvert x_j\rvert \le \left(\sum_{j\neq k}\lvert a_{kj}\rvert\right)\lvert x_k\rvert. \] If \(x_k\neq 0\), divide by \(\lvert x_k\rvert\) to obtain \[ \lvert \lambda-a_{kk}\rvert \le R_k. \] Thus \(\lambda\in D_k\), proving that every eigenvalue is captured by at least one row disk. The column-based result follows from choosing an analogous maximality argument on a transpose-style relation or by applying the same logic to \(A^T\) in an appropriate complex setting.

3.2 Disjoint circles and eigenvalue counting

3.2.1 Counting eigenvalues in isolated regions

Beyond mere inclusion, Gershgorin theory can count eigenvalues in specific separated parts of the plane. If some disks are disjoint, then each disk region contains a fixed number of eigenvalues (counted with algebraic multiplicity) provided separation is sufficient. Intuitively, when disks do not overlap, the maximum-component argument forces eigenvectors corresponding to eigenvalues in a given region to have their “dominant” coordinate aligned with that disk’s row; analytic continuation arguments then prevent eigenvalues from crossing the gap between disjoint regions without violating the eigenvector inequalities. This yields a robust “local count” principle rather than only global containment.

3.3 Special cases and limits

3.3.1 When some radii are zero

If a row \(i\) has no off-diagonal entries, then \(R_i=0\), and the disk \(D_i\) collapses to the single point \(a_{ii}\). In such cases, the theorem indicates that \(a_{ii}\) is either an eigenvalue or lies outside the spectrum depending on how other rows interact. When several radii are zero, the localization becomes sharp: diagonal block structures often emerge, and eigenvalues align with diagonal entries of those blocks.

4 Applications and Consequences

4.1 Diagonal dominance and eigenvalue location

4.1.1 Strict diagonal dominance implies localization

A matrix is (strictly) diagonally dominant if for each row \(i\), \[ \lvert a_{ii}\rvert > \sum_{j\neq i}\lvert a_{ij}\rvert = R_i. \] Geometrically, this means the disk \(D_i\) lies strictly inside the region that excludes the origin after shifting by \(a_{ii}\). More importantly, if \(\lambda=0\) is considered, strict dominance can force \(0\) to lie outside all disks, implying that \(A\) has no zero eigenvalues. Similar reasoning localizes eigenvalues away from certain points or regions by comparing \(\lvert \lambda-a_{ii}\rvert\) to \(R_i\).

As a consequence, strict diagonal dominance is a sufficient condition for nonsingularity: if \(0\) were an eigenvalue, then it would have to belong to some disk \(D_i\), requiring \(\lvert a_{ii}\rvert\le R_i\), contradicting strict dominance. Hence diagonal dominance provides practical, checkable criteria for invertibility and for excluding eigenvalues near specific values.

4.2 Stability and spectral radius bounds

4.2.1 Estimating where eigenvalues can lie

In many linear systems, stability is determined by where eigenvalues reside relative to critical sets in the complex plane. Gershgorin disks give explicit bounds on distances from diagonal entries, which can be converted into bounds on real parts or moduli once the diagonal entries and radii are known. For example, if all disks lie inside a half-plane \(\{\operatorname{Re} z < \alpha\}\), then all eigenvalues satisfy the same constraint, yielding a direct spectral estimate without full diagonalization.

4.2.2 Practical implications for iterations

Iterative methods for linear systems often depend on the spectrum of an iteration matrix. Gershgorin localization can be used to bound eigenvalues of such iteration matrices, leading to conservative but actionable predictions of convergence or divergence. In practice, these estimates guide choices of scaling, preconditioning, or stopping criteria by indicating whether eigenvalues might stray into problematic regions.

4.3 Numerical linear algebra use cases

4.3.1 Preconditioner intuition and scaling considerations

Preconditioning aims to transform a system so that the resulting operator has a more favorable spectrum. Gershgorin’s theorem provides intuition for how scaling affects eigenvalue regions: multiplying by diagonal matrices can redistribute magnitudes of off-diagonal entries and change the radii \(R_i\) or \(C_j\). While Gershgorin bounds are not always tight, they offer a fast heuristic to judge whether a proposed scaling is likely to shrink disk radii and thereby improve spectral behavior.

5 Refinements and Extensions

5.1.1 Comparison with Gershgorin bounds

Brauer’s theorem extends Gershgorin by considering disks determined by selected pairs of rows (or columns) rather than each row alone. The resulting regions can be smaller or more informative in cases where off-diagonal mass is concentrated in particular patterns. Compared with Gershgorin’s original disks, these refined sets can reduce overly conservative overestimation by exploiting structure in how entries interact across indices.

5.2 Improved localization using splitting techniques

5.2.1 Intersections of multiple Gershgorin regions

A common refinement is to intersect localization regions derived from different but related matrices. For instance, applying Gershgorin to \(A\), \(A^*\) (conjugate transpose), or to shifted matrices \(A-\mu I\) yields disks with different centers and radii. Intersections can narrow the possible location of eigenvalues. Another approach uses matrix splittings that separate dominant and residual parts, producing Gershgorin-like bounds for the residual contribution and hence tighter overall localization.

5.3 Structured matrices and sharper estimates

5.3.1 Symmetric/Hermitian improvements

For Hermitian (or real symmetric) matrices, eigenvalues are real, and the theorem’s disks collapse into intervals on the real axis after accounting for Hermitian symmetry. In such situations, Gershgorin localization aligns with classical real-eigenvalue bounds and can be related to norm-based inequalities. Moreover, when the matrix is diagonally dominant in a Hermitian sense, stronger conclusions about eigenvalue separation follow, often improving both interpretability and accuracy.

6 Computation and Practical Workflow

6.1 Constructing circles from a given matrix

Given a matrix \(A\), compute each radius \(R_i=\sum_{j\neq i}\lvert a_{ij}\rvert\) for row disks. Each disk is then centered at \(a_{ii}\) with that radius. Repeat for columns using \(C_j=\sum_{i\neq j}\lvert a_{ij}\rvert\) if a column-based view is preferred. The spectrum must lie in the union of the selected disks.

6.2 Handling complex matrices and magnitudes

For complex entries, localization depends only on magnitudes \(\lvert a_{ij}\rvert\) and the diagonal centers \(a_{ii}\), so the computation is straightforward: take absolute values of off-diagonal entries, sum them per row or column, and use the resulting nonnegative radius. The complex phase of off-diagonal terms does not directly enter the disk construction, which is why the theorem can be conservative even when cancellations occur.

6.3 Visualizing eigenvalue regions

Visualization typically involves plotting disks in the complex plane with center points at \(a_{ii}\) (or \(a_{jj}\)) and radii given by the corresponding sums. Overlaps create thicker “coverage” regions, while isolated disks indicate separated spectral locations. For real matrices, disks become intervals on the real line, which is simpler to interpret in stability studies.

6.4 Assessing tightness of bounds

Tightness varies widely across matrices. Disks may overlap heavily, yielding a large union that contains many points unrelated to actual eigenvalues. Conversely, if radii are small relative to diagonal magnitudes or if the matrix has near block-diagonal structure, disks can be tight and nearly predictive. Comparing row-based and column-based unions can also reveal when a given arrangement of entries produces overestimation in one perspective but not the other.

7 Examples

7.1 A small 2×2 matrix illustration

Consider \[ A=\begin{pmatrix} a_{11} & a_{12}\\ a_{21} & a_{22} \end{pmatrix}. \] Row disks are: \[ D_1=\{\lambda:\lvert \lambda-a_{11}\rvert\le \lvert a_{12}\rvert\},\quad D_2=\{\lambda:\lvert \lambda-a_{22}\rvert\le \lvert a_{21}\rvert\}. \] Therefore, both eigenvalues lie in \(D_1\cup D_2\). For many specific choices of \(a_{ij}\), the disks provide immediate qualitative insight: if \(\lvert a_{11}-a_{22}\rvert\) is large compared to \(\lvert a_{12}\rvert+\lvert a_{21}\rvert\), the disks are separated and each eigenvalue is confined near its corresponding diagonal entry.

7.2 A 3×3 matrix with overlapping circles

Take a matrix with diagonal entries \(a_{11},a_{22},a_{33}\) and off-diagonal magnitudes such that each radius \(R_i\) is comparable to the separation between these diagonal values. Then the three disks overlap substantially. Gershgorin’s theorem still holds—every eigenvalue belongs to the union—but the union can cover a broad region that includes many complex numbers. This example illustrates the theorem’s conservative nature when off-diagonal magnitudes do not allow clean separation.

7.3 Diagonally dominant example

Let \(A\) be diagonally dominant with strict inequalities in each row: \[ \lvert a_{ii}\rvert > \sum_{j\neq i}\lvert a_{ij}\rvert. \] Then \(0\) cannot fall into any disk \(D_i\), so \(A\) has no zero eigenvalues and is nonsingular. Moreover, if the diagonal entries are uniformly bounded away from a target region, disks can be shown not to intersect that region, yielding a direct spectral exclusion statement used in stability arguments and invertibility checks.

7.4 Example showing eigenvalue counting in disjoint circles

Suppose two groups of indices produce disks that are far apart in the complex plane, with no overlap between the corresponding sets of disks. Under sufficient separation, Gershgorin-type counting guarantees that the number of eigenvalues in one group equals the size of that group (counted with algebraic multiplicity). This demonstrates how disjointness upgrades the theorem from inclusion to a “how many” statement, rather than only “where can they be.”

8 Connections to Other Theorems

8.1 Relationship to spectral theorems

Gershgorin’s theorem complements eigenvalue localization results such as those for normal matrices and bounds derived from characteristic polynomials. While some spectral theorems provide exact locations under strong structural assumptions (e.g., normality, Hermitian symmetry), Gershgorin applies universally to complex matrices without requiring special commutativity or symmetry. It is best viewed as a general-purpose first estimate that works with minimal information.

Disk radii come from sums of absolute values of off-diagonal entries, which connect to operator norms induced by vector norms. In particular, the row-sum quantities relate to the behavior of \(A\) under norms such as \(\|\cdot\|_\infty\), and the theorem’s localization can be interpreted as a consequence of bounding \(\|(A-\lambda I)x\|\) for carefully chosen vectors. As a result, Gershgorin localization sits alongside norm-based eigenvalue bounds in the broader framework of matrix inequalities.

8.3 Connections to the pseudospectrum concept

The pseudospectrum generalizes the idea of where eigenvalues “effectively” lie under perturbations. Since Gershgorin disks use only entry magnitudes, they resemble coarse approximations to resolvent behavior: if \(\lambda\) is far from diagonal entries relative to off-diagonal coupling, then \(A-\lambda I\) tends to be “hard to invert,” consistent with small resolvent norms. More refined pseudospectral methods can capture cancellation effects and produce more accurate regions, but Gershgorin’s approach provides a simple outer approximation that often aligns qualitatively with pseudospectral geometry.