Pervasive Factors and Strong Factor Strength

Why Factor Strength Is a Cross-Sectional Question

In a large panel, a factor can affect many variables without having a large effect on any one of them. The relevant question is whether its influence accumulates across the cross-section: does the associated common variation grow in proportion to the number of variables, or is it confined to a small part of the panel? This distinction is central to interpreting factors in high-dimensional data.

Here \(p\) denotes the number of variables and \(T\) the number of time points. The focus is the population scaling as \(p\) grows. It is not a statement about how accurately a factor can be estimated from a finite sample; that also depends on the time dimension and other conditions.

Factor Models and Pervasiveness

For a panel observed at time \(t\), write the \(p\)-vector of measurements as

x_t=\Lambda f_t+e_t,\qquad t=1,\ldots,T.

Here \(f_t\) is an \(r\)-vector of common factors, \(\Lambda\) is the \(p\times r\) loading matrix, and \(e_t\) is the \(p\)-vector of idiosyncratic components. The \(i\)th row of \(\Lambda\), written \(\lambda_i'\), gives the loadings of variable \(i\). For the covariance statements below, assume the factors and errors have mean zero, \(\operatorname{Cov}(f_t)=\Sigma_F\), and \(\operatorname{Cov}(f_t,e_t)=0\). The common component is \(\chi_t=\Lambda f_t\), with covariance

\Sigma_\chi=\operatorname{Cov}(\chi_t)=\Lambda\Sigma_F\Lambda'.

The number \(r\) of factors is held fixed as \(p\) grows. We also use the standard normalization that the eigenvalues of \(\Sigma_F\) are bounded above and bounded away from zero. Without a normalization, factor scales can be shifted between \(f_t\) and \(\Lambda\), so the loading size alone does not have an invariant interpretation.

Definition: A factor system is pervasive (or has strong factors) when its common covariance has \(r\) nonzero eigenvalues of order \(p\). Equivalently, under fixed \(r\) and a well-scaled, nonsingular \(\Sigma_F\), its loading matrix satisfies \(cI_r\leq p^{-1}\Lambda'\Lambda\leq CI_r\) for constants \(0<c\leq C<\infty\) independent of \(p\). Here the inequalities are in the positive-semidefinite matrix order.

“Of order \(p\)” means bounded above and below by positive constant multiples of \(p\) for sufficiently large \(p\). This criterion is collective: a strong factor need not load equally on every variable, and some variables may have zero loadings. What matters is that the loading vectors, taken together, provide nonvanishing information in every factor direction as the panel widens.

Why Loading Strength Determines Common Covariance

The matrix \(\Lambda'\Lambda\) summarizes the cross-sectional distribution of loadings. Its \(j\)th diagonal entry is the sum of squared loadings on factor \(j\), while its off-diagonal entries describe how the loading directions overlap. If \(p^{-1}\Lambda'\Lambda\) remains positive definite with bounded eigenvalues, then no factor direction disappears in the aggregate, and no direction grows faster than the stated scale.

Theorem: Suppose \(r\) is fixed and, for constants \(c,C,a,b>0\), \(cpI_r\leq\Lambda'\Lambda\leq CpI_r\) and \(aI_r\leq\Sigma_F\leq bI_r\). Then the \(r\) nonzero eigenvalues of \(\Sigma_\chi=\Lambda\Sigma_F\Lambda'\) lie between \(acp\) and \(bCp\). In particular, each is of order \(p\).

Proof. Let \(A=\Lambda\Sigma_F^{1/2}\), where the symmetric square root exists because \(\Sigma_F\) is positive definite. Then \(\Sigma_\chi=AA'\). The nonzero eigenvalues of \(AA'\) are the eigenvalues of \(A'A=\Sigma_F^{1/2}\Lambda'\Lambda\Sigma_F^{1/2}\). This equality follows, for example, because if \(A'A v=\mu v\) for \(\mu>0\), then \(Av\neq0\) and \(AA'(Av)=\mu Av\); the converse follows by applying \(A'\) to a positive-eigenvalue eigenvector of \(AA'\).

For every \(u\in\mathbb{R}^r\), the assumed matrix bounds give

cp\,u'\Sigma_Fu \leq u'\Sigma_F^{1/2}\Lambda'\Lambda\Sigma_F^{1/2}u \leq Cp\,u'\Sigma_Fu.

The bounds \(aI_r\leq\Sigma_F\leq bI_r\) then imply \(acp\|u\|^2\leq u'A'A u\leq bCp\|u\|^2\). Applying the Rayleigh-quotient characterization of the smallest and largest eigenvalues, and the corresponding bounds to all eigenvalues, proves the claim. Since \(\Lambda\) has at most \(r\) columns, \(\Sigma_\chi\) has at most \(r\) nonzero eigenvalues; the lower bound on \(\Lambda'\Lambda\) ensures its rank is \(r\). \(\square\)

This result explains the term “strong.” Each of the \(r\) common directions contributes a covariance eigenvalue proportional to \(p\), even though the factor dimension stays fixed. It also implies that the total common variance, \(\operatorname{tr}(\Sigma_\chi)\), is of order \(p\): it is the sum of a fixed number of eigenvalues, each of order \(p\). Thus average common variance per variable is of order one.

There is a converse under the same factor normalization and fixed-\(r\) conditions. If the nonzero eigenvalues of \(\Sigma_\chi\) are all bounded above and below by positive multiples of \(p\), then the eigenvalues of \(\Lambda'\Lambda\) have the same order. Indeed, the eigenvalues of \(\Sigma_F^{1/2}\Lambda'\Lambda\Sigma_F^{1/2}\) are those of \(\Sigma_\chi\) apart from zeros, and multiplication by matrices whose eigenvalues are bounded above and away from zero preserves the scale of these eigenvalues. This is the link between the covariance and loading formulations of pervasiveness.

Worked Panel Examples

Worked Example: One Factor With Equal Loadings

Suppose \(p\) variables follow \(x_{it}=f_t+e_{it}\), where \(\operatorname{Var}(f_t)=2\). Thus every loading equals \(1\), so \(\Lambda=\mathbf{1}_p\), the \(p\)-vector of ones. The common component is \(\chi_t=\mathbf{1}_p f_t\), and its covariance is

\Sigma_\chi=2\mathbf{1}_p\mathbf{1}_p'.

The vector \(\mathbf{1}_p\) is an eigenvector of \(\mathbf{1}_p\mathbf{1}_p'\) with eigenvalue \(p\), because \(\mathbf{1}_p\mathbf{1}_p'\mathbf{1}_p =\mathbf{1}_p(\mathbf{1}_p'\mathbf{1}_p) =p\mathbf{1}_p\). Every vector orthogonal to \(\mathbf{1}_p\) has eigenvalue zero. Therefore \(\Sigma_\chi\) has one nonzero eigenvalue, \(2p\), and the factor is pervasive. Each variable has common variance \(2\), while the total common variance is \(2p\). The strength comes from the factor appearing across all \(p\) variables, not from an individually growing variance.

Worked Example: Two Factors Across Two Large Groups

Let \(p\) be even. For the first \(p/2\) variables set \(\lambda_i=(1,0)'\), and for the remaining \(p/2\) set \(\lambda_i=(0,2)'\). Let \(\Sigma_F=\operatorname{diag}(1,1/4)\), so the two factors are uncorrelated with variances \(1\) and \(1/4\). Summing the outer products of the loading vectors gives

\Lambda'\Lambda = \begin{pmatrix} p/2&0\\ 0&2p \end{pmatrix}, \qquad \Sigma_F^{1/2}\Lambda'\Lambda\Sigma_F^{1/2} = \begin{pmatrix} p/2&0\\ 0&p/2 \end{pmatrix}.

The second matrix has the two nonzero eigenvalues of the common covariance, so both are \(p/2\). Equivalently, the first group has common variance \(1\) from factor one, and the second has common variance \(4(1/4)=1\) from factor two. Each group contains \(p/2\) variables, producing an eigenvalue \(p/2\) in each factor direction. Both factors are pervasive even though neither loads on every variable.

The loading condition also checks directly: the eigenvalues of \(p^{-1}\Lambda'\Lambda\) are \(1/2\) and \(2\), bounded away from zero and infinity for every even \(p\).

Worked Example: A Factor Concentrated on a Shrinking Share

Now let one factor have variance \(1\), with loading \(1\) on the first \(m_p=\lfloor\sqrt{p}\rfloor\) variables and loading \(0\) on the rest. Then \(\Lambda'\Lambda=m_p\), and the common covariance has one nonzero eigenvalue \(m_p\), since it is the outer product of the loading vector with itself.

Although \(m_p\) grows, its ratio to the number of variables satisfies \(m_p/p\leq 1/\sqrt{p}\to0\). Thus the eigenvalue is of order \(\sqrt{p}\), not order \(p\), and this factor is not pervasive. The share of variables directly affected also tends to zero. Merely observing that a covariance eigenvalue diverges is not enough to call a factor strong; its scale must be compared with \(p\).

Interpreting Common Variation Carefully

Pervasiveness describes the factor component, not automatically the entire observed covariance. Under the factor-error uncorrelatedness assumption, \(\operatorname{Cov}(x_t)=\Sigma_\chi+\Sigma_e\), where \(\Sigma_e=\operatorname{Cov}(e_t)\). If \(\Sigma_e\) has bounded largest eigenvalue as \(p\) grows, then the order-\(p\) common eigenvalues remain order \(p\) in the total covariance: adding a matrix with bounded operator norm can shift eigenvalues by at most that norm. This qualification matters because covariance structure in the errors can also affect the observed spectrum.

A further qualification is that “strong” does not mean every variable has a large loading. In the two-group example, each factor is absent from half the variables, yet it is pervasive because it loads on a positive fraction of the panel. Conversely, a factor with a growing number of affected variables can still be nonpervasive if that number is negligible relative to \(p\). The relevant object is the aggregate loading in each factor direction.

Common Mistakes

Using divergence instead of the right scale. An eigenvalue of order \(\sqrt{p}\) diverges, but it is not of order \(p\). Compare the eigenvalue with \(p\), not just with a constant.
Checking only one loading or one diagonal entry. Pervasiveness is a matrix-wide condition. For multiple factors, every direction in the loading space must retain order-\(p\) aggregate strength.
Requiring every variable to load on every factor. That is stronger than needed. A factor may be zero on a subset of variables and still be pervasive if its squared loadings sum to order \(p\).
Ignoring factor scaling. Loading magnitudes depend on how the factors are normalized. State the factor covariance condition before interpreting the loading matrix; the order-\(p\) common covariance criterion is the substantive one.

Key takeaway: With a fixed number of well-scaled factors, pervasiveness means that every factor direction accumulates order-\(p\) loading strength. It produces order-\(p\) eigenvalues and order-\(p\) total common variation, rather than merely a covariance eigenvalue that happens to grow.

Check Your Understanding

Use the order of the loading sums and common covariance eigenvalues to answer each question.

A one-factor panel has factor variance \(3\), and every loading equals \(2\). What is the nonzero common covariance eigenvalue, and is the factor pervasive?
For two factors, the eigenvalues of \(\Lambda'\Lambda\) are \(0.4p\) and \(3p\), while \(\Sigma_F=I_2\). What does this imply about factor strength?
A factor loads with value \(1\) on \(p^{2/3}\) variables and zero elsewhere. Is its common covariance eigenvalue of order \(p\)? Explain.
Why can a factor be pervasive even if it has zero loadings for half of the variables?
What additional condition on the error covariance lets order-\(p\) common eigenvalues remain visible at the same scale in the total covariance?