Jacobian-Rank Diagnosis of Local Identification

Why Diagnose Identification With a Jacobian?

A factor model describes the observed covariance matrix through a lower-dimensional set of parameters. Identification asks whether distinct parameter values can produce the same covariance matrix. Parameter counts and indicator patterns can flag possible problems, but a local Jacobian calculation examines how the covariance changes in each parameter direction at a particular point.

The distinction between a parameter point and its neighborhood matters. A model may have isolated alternative solutions far away, or it may have a continuous family of equivalent parameters passing arbitrarily close to the point of interest. The Jacobian is a local diagnostic: full column rank gives a strong sufficient condition for local identification. Rank deficiency signals that some first-order parameter direction is invisible to the covariance map, but additional reasoning is needed to decide whether that direction corresponds to genuine nonidentification.

Definition: Let \(g:\Theta\to\mathbb{R}^{m}\) be a model-implied map on a parameter space \(\Theta\), and let \(\theta_0\in\Theta\). The parameter \(\theta_0\) is locally identified if there is a neighborhood \(U\) of \(\theta_0\) such that \(g(\theta)=g(\theta_0)\) for \(\theta\in U\) implies \(\theta=\theta_0\). In a factor model, \(g(\theta)\) will be a vector containing the distinct entries of the model-implied covariance matrix.

This definition concerns the chosen parameterization, including any constraints used to set factor scale or orientation. If a model is described using \(d\) free coordinates and an observed covariance matrix has \(p\) variables, then the map has \(d\) input coordinates and \(m=p(p+1)/2\) output coordinates. Having \(d\leq m\) is necessary for the Jacobian to have full column rank, but that dimension count alone does not establish identification.

The Covariance Map and Its Jacobian

For a common-factor model with \(p\) observed variables and \(q\) factors, write the model-implied covariance as

\Sigma(\theta)=\Lambda\Lambda^\mathsf{T}+\Psi,

where \(\Lambda\) is the \(p\times q\) loading matrix and \(\Psi\) is diagonal, with uniquenesses on its diagonal. The parameter vector \(\theta\) consists of the free loadings and uniquenesses after any identifying constraints have been imposed. The covariance map \(g\) lists the \(p\) diagonal and \(p(p-1)/2\) distinct off-diagonal entries of \(\Sigma(\theta)\), in a fixed order. Its Jacobian is

J(\theta)=\frac{\partial g(\theta)}{\partial\theta^\mathsf{T}}.

Each column records how all model-implied covariance entries change as one free parameter changes. For a loading perturbation \(d\Lambda\) and a diagonal-uniqueness perturbation \(d\Psi\), the corresponding first-order change in the covariance is

d\Sigma=(d\Lambda)\Lambda^\mathsf{T}+\Lambda(d\Lambda)^\mathsf{T}+d\Psi.

This differential gives a direct way to find Jacobian columns or to check whether a proposed parameter direction lies in the Jacobian’s null space.

Theorem: Suppose \(\Theta\) is open in \(\mathbb{R}^{d}\), the covariance map \(g:\Theta\to\mathbb{R}^{m}\) is continuously differentiable, and \(\theta_0\in\Theta\). If \(J(\theta_0)\) has full column rank \(d\), then \(\theta_0\) is locally identified.

Proof. Full column rank means that some \(d\) rows of \(J(\theta_0)\) form a nonsingular \(d\times d\) matrix. Let \(P\) select the corresponding \(d\) output coordinates, and define \(h(\theta)=Pg(\theta)\). Then the square Jacobian \(Dh(\theta_0)\) is nonsingular. By the inverse function theorem, \(h\) is one-to-one on some neighborhood \(U\) of \(\theta_0\). If \(\theta\in U\) and \(g(\theta)=g(\theta_0)\), then \(h(\theta)=Ph(\text{not applicable})\) is not the relevant expression; rather, by the definition \(h=Pg\), we have \(h(\theta)=Pg(\theta)=Pg(\theta_0)=h(\theta_0)\). Since \(h\) is one-to-one on \(U\), \(\theta=\theta_0\). Thus \(\theta_0\) is locally identified. \(\square\)

The theorem is a sufficient condition, not an equivalence. If the Jacobian is rank deficient, there is a nonzero first-order parameter direction that produces zero first-order change in the covariance. That fact alone does not establish that the covariance remains exactly constant along a curve. The examples below distinguish the two situations.

Worked Jacobian Diagnoses

Worked Example: Full Rank in a Three-Indicator One-Factor Model

Consider a one-factor model for three observed variables with loading vector \((a,b,c)^\mathsf{T}\), diagonal uniquenesses \((\psi_1,\psi_2,\psi_3)\), and the sign constraint \(a>0\). Take the parameter point \((a,b,c)=(1,2,3)\) and \((\psi_1,\psi_2,\psi_3)=(1,2,3)\). All uniquenesses are positive, and the model-implied covariance matrix is

\Sigma= \begin{pmatrix} 1^2+1&1\cdot2&1\cdot3\\ 1\cdot2&2^2+2&2\cdot3\\ 1\cdot3&2\cdot3&3^2+3 \end{pmatrix} = \begin{pmatrix} 2&2&3\\ 2&6&6\\ 3&6&12 \end{pmatrix}.

Order the output entries as \((\Sigma_{11},\Sigma_{22},\Sigma_{33},\Sigma_{12},\Sigma_{13},\Sigma_{23})\) and the parameters as \((a,b,c,\psi_1,\psi_2,\psi_3)\). The covariance map is

g(a,b,c,\psi_1,\psi_2,\psi_3) =(a^2+\psi_1,\ b^2+\psi_2,\ c^2+\psi_3,\ ab,\ ac,\ bc).

Differentiating each component gives

J= \begin{pmatrix} 2&0&0&1&0&0\\ 0&4&0&0&1&0\\ 0&0&6&0&0&1\\ 2&1&0&0&0&0\\ 3&0&1&0&0&0\\ 0&3&2&0&0&0 \end{pmatrix}.

The first three rows contain an identity matrix in the uniqueness columns. After using those rows to eliminate the loading entries in the first three rows, full rank reduces to whether the loading-derivative block for the three off-diagonal covariances has full rank:

D= \begin{pmatrix} b&a&0\\ c&0&a\\ 0&c&b \end{pmatrix}, \qquad \det(D)=-2abc=-12.

Because this determinant is nonzero, \(D\) has rank three. The uniqueness columns supply three more independent directions, so \(J\) has rank six, its full column rank. The theorem therefore guarantees local identification at this parameter point.

Worked Example: A Rank-Deficient Point With a Constant-Covariance Curve

Keep the same three-indicator model and the constraint \(a>0\). At \((a,b,c)=(1,0,2)\) and \((\psi_1,\psi_2,\psi_3)=(1,2,3)\), consider the parameter curve, for \(t\) near 1,

a(t)=t,\qquad b(t)=0,\qquad c(t)=\frac{2}{t},\qquad \psi_1(t)=2-t^2,\qquad \psi_2(t)=2,\qquad \psi_3(t)=7-\frac{4}{t^2}.

At \(t=1\), the curve passes through the stated point. To check that it remains in the parameter space, take \(0.9<t<1.1\). Then \(a(t)>0\), \(\psi_1(t)=2-t^2>2-1.1^2=0.79>0\), and \(\psi_3(t)=7-4/t^2>7-4/0.9^2>0\); also \(\psi_2(t)=2>0\). Every point on this interval is admissible.

Substitution shows that all six covariance entries are constant:

a(t)^2+\psi_1(t)=2,\quad b(t)^2+\psi_2(t)=2,\quad c(t)^2+\psi_3(t)=7,

a(t)b(t)=0,\qquad a(t)c(t)=2,\qquad b(t)c(t)=0.

Thus \(\Sigma(t)\) is the same matrix for every \(t\) in the interval, namely

\Sigma(t)= \begin{pmatrix} 2&0&2\\ 0&2&0\\ 2&0&7 \end{pmatrix}.

Since distinct values of \(t\) arbitrarily close to 1 give distinct parameter vectors with the same covariance, the point is not locally identified. Differentiating the curve at \(t=1\) gives the nonzero parameter direction \((1,0,-2,-2,0,8)\). Because \(g(\theta(t))\) is constant, the chain rule gives \(J(\theta(1))\theta'(1)=0\); hence the Jacobian is rank deficient as well. Here the null direction is accompanied by an exact curve, which establishes nonidentification rather than merely suggesting it.

Worked Example: A Jacobian Null Direction From Unrestricted Factor Orientation

Consider two observed variables and two factors, with no orientation constraint on the loading matrix. Set \(\Lambda(0)=I_2\) and \(\Psi=I_2\), so \(\Sigma(0)=2I_2\). Let

\Lambda(t)= \begin{pmatrix} \cos t&-\sin t\\ \sin t&\cos t \end{pmatrix}.

For every \(t\), \(\Lambda(t)\Lambda(t)^\mathsf{T}=I_2\), so \(\Sigma(t)=2I_2\). The covariance map is therefore constant along this nonconstant curve, and these unconstrained loading parameters are not locally identified. The Jacobian has a nonzero null direction at \(t=0\): the loading derivative is

\Lambda'(0)= \begin{pmatrix} 0&-1\\ 1&0 \end{pmatrix}=K, \qquad K\Lambda(0)^\mathsf{T}+\Lambda(0)K^\mathsf{T}=K+K^\mathsf{T}=0.

The differential formula therefore gives \(d\Sigma=0\) for this nonzero perturbation. In practice, constraints on scale and orientation must be included in the parameterization before interpreting a Jacobian rank as a diagnosis of the constrained model.

Interpreting Rank Deficiency Carefully

A zero or dependent Jacobian column means that at least one parameter change has no first-order effect on the listed covariance entries at the point. It does not always mean that there are other parameter values nearby with exactly the same covariance. For instance, the scalar map \(g(t)=t^3\) is one-to-one around zero, although its derivative at zero is zero. A singular derivative can therefore occur even when local identification holds.

The factor-model curve in the preceding example supplies the extra evidence needed for a definitive conclusion: it verifies equality of the full covariance matrix at every point along an admissible curve. In other cases, a rank-deficient Jacobian calls for further analysis of the model equations or higher-order behavior. Conversely, a full-column-rank Jacobian at an interior point is decisive for local identification under the stated smoothness conditions.

Common Mistakes

Treating parameter counting as a rank calculation. Having at least as many covariance entries as parameters only makes full column rank possible; the actual derivatives may still be dependent.
Calling every rank-deficient point unidentified. Rank deficiency is a first-order statement. Establish exact nonidentification with, for example, a curve of distinct admissible parameters that preserves the covariance.
Ignoring constraints in the parameter space. The Jacobian must be computed with respect to free coordinates after constraints have been imposed. A curve that violates a sign restriction or makes a uniqueness inadmissible cannot demonstrate nonidentification for that model.
Confusing local and global identification. The full-rank theorem guarantees uniqueness only in some neighborhood of the point. It does not rule out a distant parameter value with the same covariance.

Key takeaway: Compute the Jacobian of the model-implied covariance map in the constrained free parameters. Full column rank at an interior point guarantees local identification. Rank deficiency identifies an invisible first-order direction, but an exact covariance-preserving argument is needed to establish nonidentification.

Check Your Understanding

Use the definitions and examples above to test your interpretation of Jacobian rank.

In the full-rank one-factor example, how many free parameters are there, and why does the nonzero determinant of the off-diagonal derivative block matter?
What does full column rank guarantee at an interior point, and what does it not guarantee about distant parameter values?
In the curve example, verify the value of \(\psi_3(t)\) at \(t=1\) and explain why the curve establishes nonidentification.
Why does a rank-deficient Jacobian not, by itself, prove that a parameter point is locally unidentified?
What must be taken into account when computing a Jacobian for a factor model with scale or orientation constraints?