Eigenvalues, Normalization, and Spatial Dependence

Why the Spectrum of a Weights Matrix Matters

A spatial weights matrix encodes which units are connected and how strongly each unit is connected to others. Its entries describe local relationships, but its eigenvalues summarize how the matrix acts on broader patterns across all units. They therefore help explain why changing a matrix’s normalization can change the numerical scale of a spatial-dependence coefficient—and the range of values that coefficient can take without creating a singular operator.

Use \(W=(w_{ij})\) for an \(n\times n\) spatial weights matrix, with \(w_{ij}\) representing a link from unit \(j\) to unit \(i\). In particular, rows index the receiving units and columns index the units sending the link. The matrix may be symmetric, as with some undirected weights, or nonsymmetric, as with directed or row-standardized weights. Its spectrum must be interpreted with that distinction in mind.

Definition: An eigenvalue of \(W\) is a number \(\lambda\) for which there is a nonzero vector \(v\) satisfying \(Wv=\lambda v\). The vector \(v\) is an eigenvector. The spectral radius is \(r(W)=\max_j|\lambda_j|\), the largest absolute value among the eigenvalues of \(W\). For a real matrix, eigenvalues can be complex even though all entries of the matrix are real.

An eigenvector describes a pattern across units. If a pattern \(v\) is an eigenvector, applying \(W\) to it produces the same pattern multiplied by \(\lambda\). In a dependence operator such as \(I-\rho W\), where \(I\) is the identity matrix and \(\rho\) is a real spatial coefficient, the corresponding pattern is multiplied by \(1-\rho\lambda\). Thus the spectrum describes which patterns are amplified, reduced, or—at particular coefficient values—annihilated by the operator.

Normalization Changes the Scale

A common starting point is an unstandardized matrix of nonnegative link weights. A common alternative is row-standardization: for each row with a positive sum, divide its entries by that row’s sum. The resulting row sums are one. Each receiving unit’s weights then sum to one across the units sending links to it. Row-standardization changes the matrix entries and generally changes its eigenvalues; it is not merely a change of notation.

For a nonnegative row-stochastic matrix—one whose row sums are all one—the spectral radius equals one. The reason is that the vector of all ones is an eigenvector with eigenvalue one, while no eigenvalue can have absolute value greater than one. A row with sum zero requires a separate convention; if zero rows remain, the matrix is generally substochastic and the conclusion becomes \(r(W)\leq 1\), rather than necessarily equality.

Worked Example: An Undirected Three-Unit Chain

Suppose three units form a chain: unit 1 is linked to unit 2, and unit 2 to unit 3. Give each link weight one in both directions. The unstandardized matrix is

W= \begin{pmatrix} 0&1&0\\ 1&0&1\\ 0&1&0 \end{pmatrix}.

The characteristic polynomial is \(\det(\lambda I-W)=\lambda(\lambda^2-2)\), so the eigenvalues are \(\sqrt{2}\), \(0\), and \(-\sqrt{2}\). Hence \(r(W)=\sqrt{2}\approx1.414\). The row sums are 1, 2, and 1, so this matrix is not row-stochastic.

Row-standardizing gives

W_R= \begin{pmatrix} 0&1&0\\ 1/2&0&1/2\\ 0&1&0 \end{pmatrix}.

Its characteristic polynomial is \(\lambda(\lambda^2-1)\), so its eigenvalues are \(1\), \(0\), and \(-1\), and \(r(W_R)=1\). The central unit divides its weight equally between the two neighbors; each endpoint has only one neighbor. The two matrices have the same link pattern but different scales and different numerical eigenvalues.

Eigenvalues and Coefficient Ranges

The spectral connection to a spatial coefficient can be stated precisely without assuming a particular statistical model. The matrix \(I-\rho W\) is singular exactly when at least one of its eigenvalues is zero. Since an eigenvalue of \(I-\rho W\) corresponding to an eigenvalue \(\lambda\) of \(W\) is \(1-\rho\lambda\), singularity occurs when \(1-\rho\lambda=0\) for some eigenvalue \(\lambda\).

Key result: For real \(\rho\), \(I-\rho W\) is singular if and only if \(\rho=1/\lambda\) for some nonzero real eigenvalue \(\lambda\) of \(W\). Complex eigenvalues do not produce a real singularity at a real \(\rho\), unless they are real themselves. Thus the exact excluded real coefficient values are determined by the reciprocals of the nonzero real eigenvalues.

When \(W\) is real and symmetric, its eigenvalues are all real. If its smallest and largest eigenvalues satisfy \(\lambda_{\min}<0<\lambda_{\max}\), then the interval containing zero with no singularity is

\left(\frac{1}{\lambda_{\min}},\frac{1}{\lambda_{\max}}\right).

The endpoints are excluded because they make one of the factors \(1-\rho\lambda\) equal to zero. This is the interval between the closest negative and positive singularity values. It is not a universal description of the exact admissible set for every nonsymmetric matrix: a nonsymmetric matrix may have complex eigenvalues, and its real eigenvalues may leave gaps between excluded values.

A frequently useful, stronger condition is \(|\rho|r(W)<1\). It implies that no eigenvalue can satisfy \(\rho\lambda=1\), so it guarantees nonsingularity. But it can rule out some values for which the matrix is still nonsingular. In particular, the inequality describes the symmetric interval \((-1/r(W),1/r(W))\); it is a sufficient bound, not always the full nonsingularity range.

Worked Example: A Nonsingular Value Outside the Spectral-Radius Bound

Consider the hypothetical symmetric weights matrix

W= \begin{pmatrix} 0&1.5&1.5\\ 1.5&0&1.5\\ 1.5&1.5&0 \end{pmatrix}.

This is 1.5 times the adjacency matrix of a three-unit complete graph. The all-ones vector has eigenvalue \(3\); any vector whose entries sum to zero has eigenvalue \(-1.5\). Thus the eigenvalues are \(3,-1.5,-1.5\), and \(r(W)=3\).

The singularity values are \(\rho=1/3\) and \(\rho=-2/3\), so the interval containing zero with no singularity is \((-2/3,1/3)\). Take \(\rho=-0.4\), which lies inside this interval. The eigenvalues of \(I-\rho W\) are \(1-(-0.4)(3)=2.2\) and \(1-(-0.4)(-1.5)=0.4\), with the latter occurring twice. None is zero, so the operator is nonsingular.

However, \(|\rho|r(W)=0.4(3)=1.2\), which is not less than one. This calculation shows why the spectral-radius bound cannot be substituted for the exact nonsingularity check: it is sufficient, but it can be more restrictive.

Directed Weights and Complex Eigenvalues

For directed weights, \(W\) need not be symmetric, and eigenvalues can be complex. The spectral radius still summarizes their largest absolute value, but the exact real-coefficient singularity test uses only real eigenvalues. Moreover, a directed matrix can represent cyclical patterns: applying it to some patterns changes both their scale and their orientation across units. The modulus of a complex eigenvalue measures the scale factor, while its angle describes this rotation-like change.

Worked Example: A Directed Four-Unit Cycle

Suppose each unit points to the next unit in a directed cycle, with unit 4 pointing back to unit 1. Under the convention that \(w_{ij}\) represents a link from unit \(j\) to unit \(i\), the row-standardized matrix is

W= \begin{pmatrix} 0&0&0&1\\ 1&0&0&0\\ 0&1&0&0\\ 0&0&1&0 \end{pmatrix}.

For example, \(w_{21}=1\) represents the link from unit 1 to unit 2, and \(w_{14}=1\) represents the link from unit 4 to unit 1. Each row sums to one. This matrix shifts a vector’s entries around the cycle; its eigenvalues are the fourth roots of one: \(1,-1,i,-i\). Therefore \(r(W)=1\).

For real \(\rho\), the real eigenvalues give singularity at \(\rho=1\) and \(\rho=-1\). The complex eigenvalues do not give additional real singularity values. The sufficient condition \(|\rho|r(W)<1\) becomes \(|\rho|<1\), which ensures nonsingularity here. The example illustrates why the spectral radius and the exact real singularity values are related but are not interchangeable concepts.

What Normalization Means for Dependence

A coefficient’s numerical size has meaning only relative to the scale of \(W\). If every weight is multiplied by a positive constant \(c\), the eigenvalues and spectral radius are also multiplied by \(c\). The product \(\rho W\), and therefore the action of \(I-\rho W\), is unchanged if the coefficient is divided by \(c\). This gives a basic reason that coefficient magnitudes cannot be compared across differently scaled weights matrices without accounting for their normalization.

Row-standardization does more than uniformly rescale all weights when row sums differ. It changes the relative contributions of units with different numbers or total weights of links. For a symmetric weights matrix, row-standardization can also make the resulting matrix nonsymmetric, so the symmetric eigenvalue interval formula should not be applied automatically. In some undirected cases with positive degrees, the row-standardized matrix has the same eigenvalues as a related symmetric matrix, but the actual matrix and its assumptions must be checked before using that fact.

The spectrum also makes clear that dependence is not simply “larger \(\rho\) means stronger dependence” in every pattern. Along an eigenvector \(v\), the contribution of \(W\) is scaled by \(\lambda\), so the coefficient acts through \(\rho\lambda\). Positive and negative eigenvalues can therefore produce different directional behavior, while complex eigenvalues describe cyclic patterns. Normalization changes these eigenvalue scales and thus changes the coefficient scale at which those patterns become pronounced.

Common Mistakes

Equating nonsingularity with the spectral-radius bound. The condition \(|\rho|r(W)<1\) is sufficient. It is not necessary: a value can violate it and still avoid every real singularity value.
Using only the largest positive eigenvalue for a symmetric matrix. A large-magnitude negative eigenvalue can determine the lower endpoint of the interval containing zero. Check both \(\lambda_{\min}\) and \(\lambda_{\max}\).
Assuming row-standardization always preserves symmetry. It generally does not when row sums differ. Check the matrix after normalization before using results that require symmetry.
Ignoring the direction convention. With \(w_{ij}\) defined as a link from \(j\) to \(i\), source units are in columns and receiving units are in rows. Reversing that interpretation transposes a directed weights matrix.
Comparing coefficient magnitudes across normalizations as if they were on the same scale. Rescaling \(W\) changes its eigenvalues and the corresponding coefficient scale. State the normalization alongside any interpretation.

Key takeaway: Eigenvalues describe how a weights matrix acts on spatial patterns, and the spectral radius records the largest eigenvalue magnitude. Normalization changes that spectrum. For real coefficients, exact singularity values come from reciprocals of nonzero real eigenvalues; the spectral-radius bound is a useful but potentially stricter guarantee.

Check Your Understanding

Answer using the definitions and distinctions developed above.

What does the spectral radius measure, and how does it differ from the largest eigenvalue when some eigenvalues are negative or complex?
Why does a nonnegative row-stochastic matrix have spectral radius one?
For a symmetric matrix with eigenvalues \(-4\) and \(2\), find the interval containing zero with no singularity for \(I-\rho W\). What is the interval guaranteed by \(|\rho|r(W)<1\)?
Can a real coefficient produce singularity through a nonreal eigenvalue of a real matrix? Explain.
Why should coefficient magnitudes not be compared directly across differently normalized weights matrices?