Local Identification From Jacobian Rank

Why Examine the Jacobian?

A structural model links its parameters to restrictions on population quantities. Identification asks whether those restrictions distinguish the parameter values: could a different value of the structural parameters satisfy the same population restrictions? The Jacobian provides a local answer by describing how the restrictions change when the parameters move slightly.

The key condition is full column rank of the Jacobian at the parameter value of interest. Under stated regularity conditions, it guarantees that the value is locally identified. The condition is a sufficient test, not a universal equivalence: when the Jacobian is rank deficient, identification may fail, but higher-order changes can still distinguish parameter values.

Definition: Let \(\theta\in\Theta\subset\mathbb{R}^p\) be a vector of \(p\) structural parameters, and let \(g(\theta)\in\mathbb{R}^q\) collect \(q\) population restrictions. A value \(\theta_0\) satisfies the restrictions if \(g(\theta_0)=0\). It is locally identified if there is a neighborhood \(U\) of \(\theta_0\) such that the only \(\theta\in U\) satisfying \(g(\theta)=0\) is \(\theta_0\).

The zero-restriction form is convenient, but not essential. If the population model implies a vector of quantities \(m(\theta)\), and the observed population quantities are \(m_0\), one can set \(g(\theta)=m(\theta)-m_0\). Local identification then means that \(\theta_0\) is the only nearby parameter value satisfying \(m(\theta)=m_0\).

The Jacobian Rank Condition

Assume \(\theta_0\) is an interior point of the parameter space, \(g\) is continuously differentiable in a neighborhood of \(\theta_0\), and \(g(\theta_0)=0\). The Jacobian is the \(q\times p\) matrix of first derivatives:

G(\theta_0)=\frac{\partial g(\theta_0)}{\partial\theta'} = \begin{bmatrix} \frac{\partial g_1}{\partial\theta_1} & \cdots & \frac{\partial g_1}{\partial\theta_p}\\ \vdots & \ddots & \vdots\\ \frac{\partial g_q}{\partial\theta_1} & \cdots & \frac{\partial g_q}{\partial\theta_p} \end{bmatrix}.

Its columns describe the first-order changes in all the restrictions generated by changes in each parameter. Full column rank means \(\operatorname{rank}G(\theta_0)=p\). This requires at least as many restrictions as parameters, \(q\ge p\), and means that no nonzero parameter direction leaves all restrictions unchanged to first order.

Theorem: Suppose \(\theta_0\) is an interior point, \(g\) is continuously differentiable near \(\theta_0\), \(g(\theta_0)=0\), and \(G(\theta_0)\) has full column rank \(p\). Then \(\theta_0\) is locally identified by \(g(\theta)=0\).

The result is local: it rules out other solutions sufficiently close to \(\theta_0\), not necessarily solutions elsewhere in the parameter space. Its proof reduces the possibly rectangular Jacobian to a square nonsingular one.

Select independent restrictions.
Because \(G(\theta_0)\) has column rank \(p\), some \(p\) of its rows form a \(p\times p\) matrix \(B\) with nonzero determinant. Let \(h(\theta)\) be the vector of the corresponding \(p\) components of \(g(\theta)\). Then \(Dh(\theta_0)=B\).

Apply the inverse function theorem.
The map \(h:\mathbb{R}^p\to\mathbb{R}^p\) is continuously differentiable near \(\theta_0\), and its derivative there is nonsingular. The inverse function theorem therefore gives a neighborhood in which \(h\) is one-to-one.

Conclude local uniqueness.
If a nearby \(\theta\) satisfies \(g(\theta)=0\), then it satisfies \(h(\theta)=0=h(\theta_0)\). Since \(h\) is one-to-one in that neighborhood, \(\theta=\theta_0\). Thus \(\theta_0\) is an isolated solution of the full set of restrictions.

Worked Applications

The examples use hypothetical population restrictions. In each case, the question is whether the restrictions distinguish the parameter value, not whether a finite sample estimates it precisely.

Worked Example: A Linear System With Full Rank

Suppose two dimensionless structural parameters, \(\theta_1\) and \(\theta_2\), must satisfy the hypothetical population restrictions \(2\theta_1+\theta_2=8\) and \(\theta_1-\theta_2=1\). Determine the solution and check the rank condition.

Write the restrictions as \(g(\theta)=0\):

g(\theta)= \begin{bmatrix} 2\theta_1+\theta_2-8\\ \theta_1-\theta_2-1 \end{bmatrix}, \qquad G(\theta)= \begin{bmatrix} 2&1\\ 1&-1 \end{bmatrix}.

Solving the second equation gives \(\theta_1=\theta_2+1\). Substituting into the first gives \(2(\theta_2+1)+\theta_2=8\), so \(3\theta_2=6\) and \(\theta_2=2\). Hence \(\theta_1=3\). The determinant of the Jacobian is

\det G=2(-1)-(1)(1)=-3\ne0.

The Jacobian has rank \(2\), equal to the number of parameters. The rank condition therefore guarantees local identification of \((3,2)\). In this particular linear system the solution is also globally unique, but that stronger conclusion follows from the whole system, not from the local rank theorem alone.

Worked Example: Nonlinear Restrictions and Local, Not Global, Uniqueness

Suppose two positive, dimensionless structural parameters satisfy hypothetical population restrictions that their sum is \(3\) and their product is \(2\). Check local identification at \(\theta_0=(1,2)\).

The restriction vector and its Jacobian are

g(\theta)= \begin{bmatrix} \theta_1+\theta_2-3\\ \theta_1\theta_2-2 \end{bmatrix}, \qquad G(\theta)= \begin{bmatrix} 1&1\\ \theta_2&\theta_1 \end{bmatrix}.

At \((1,2)\), the Jacobian is \(\begin{bmatrix}1&1\\2&1\end{bmatrix}\), with determinant \(1-2=-1\ne0\). Its rank is \(2\), so the rank condition establishes local identification.

To see the distinction from global uniqueness, the restrictions imply that \(\theta_1\) and \(\theta_2\) are the roots of \(z^2-3z+2=0\). Factoring gives \((z-1)(z-2)=0\), so the two ordered solutions are \((1,2)\) and \((2,1)\). The second solution is distinct but not arbitrarily close to \((1,2)\); for example, a ball of radius \(0.5\) around \((1,2)\) excludes it. Thus \((1,2)\) is locally identified even though it is not globally unique.

Worked Example: Rank Deficiency With a Nearby Continuum of Solutions

Consider two dimensionless parameters with hypothetical restrictions \(\theta_1+\theta_2=5\) and \(2\theta_1+2\theta_2=10\). The value \(\theta_0=(2,3)\) satisfies both. Does the system locally identify it?

Writing the restrictions as residuals gives

g(\theta)= \begin{bmatrix} \theta_1+\theta_2-5\\ 2\theta_1+2\theta_2-10 \end{bmatrix}, \qquad G(\theta)= \begin{bmatrix} 1&1\\ 2&2 \end{bmatrix}.

The second row is twice the first, so \(\operatorname{rank}G=1<2\). Here the rank deficiency reflects actual nonidentification: for every sufficiently small real \(t\), the distinct parameter pair \((2+t,3-t)\) satisfies both restrictions, since

(2+t)+(3-t)=5, \qquad 2(2+t)+2(3-t)=10.

There are therefore other solutions in every neighborhood of \((2,3)\), so this value is not locally identified. The rank calculation alone did not prove that conclusion; the continuum of solutions did.

Worked Example: Rank Deficiency Does Not Always Mean Nonidentification

For a single dimensionless parameter, suppose the hypothetical population restriction is \(g(\theta)=\theta^3=0\), with candidate value \(\theta_0=0\). Evaluate the rank condition and determine identification directly.

The Jacobian is the scalar derivative \(G(\theta)=3\theta^2\), so \(G(0)=0\) has rank \(0<1\). Nevertheless, the equation \(\theta^3=0\) has exactly one real solution: \(\theta=0\). The candidate is therefore locally identified (indeed, it is the only solution), despite the deficient Jacobian.

At zero, the first-order change vanishes, but the restriction still changes at higher order for every nonzero \(\theta\). This illustrates why full column rank is sufficient rather than necessary.

How to Use the Condition

For a proposed parameter value, the practical sequence is to write all population restrictions as a vector of residuals, differentiate with respect to every structural parameter, evaluate the resulting Jacobian at the value of interest, and calculate its rank. The interpretation depends on the result:

Full column rank: under the theorem’s regularity conditions, the parameter value is locally identified.
Rank below the number of parameters: the theorem gives no conclusion by itself. Check whether exact restrictions have nearby alternative solutions, or whether higher-order terms distinguish the parameters.
Fewer restrictions than parameters: full column rank is impossible. This rules out the full-rank sufficient condition, but does not by itself settle identification in every nonlinear model.

The rank is evaluated at a particular parameter value. A Jacobian may have full rank at some points and lose rank at special values, so a conclusion at one point should not automatically be extended to the whole parameter space. The parameterization also matters: the rank calculation concerns the parameters as defined in the model, after any required normalization.

Common Mistake

A frequent error is to reason, “The Jacobian is rank deficient, so the parameters are not identified.” The linear example above supports that conclusion only because direct substitution showed a nearby continuum of solutions. The cubic example disproves the claim in general: its derivative is zero at the solution, yet the solution is isolated. Rank deficiency says that some nonzero direction produces no first-order change in the restrictions; it does not say that the restrictions remain exactly unchanged along that direction.

The converse also requires care. Full column rank establishes uniqueness only in a neighborhood under the theorem’s assumptions. It does not rule out a distant observationally equivalent parameter value, as the nonlinear sum-and-product example demonstrates. Local identification and global identification are different claims.

Key takeaway: If the Jacobian of the population restrictions has full column rank at an interior solution, the inverse function theorem guarantees local identification. If the rank is deficient, investigate the exact restrictions rather than treating the rank result as a verdict.

Check Your Understanding

Use the definitions and rank result to answer the following questions.

For \(p\) parameters and \(q\) restrictions, what does full column rank of the \(q\times p\) Jacobian mean, and what inequality between \(q\) and \(p\) must hold?
Why does selecting \(p\) independent rows of a full-column-rank Jacobian allow the inverse function theorem to establish local identification?
A model has a full-rank Jacobian at one solution and a different solution far away. Which identification claim has been established, and which has not?
Why does a nonzero vector \(v\) satisfying \(G(\theta_0)v=0\) not, by itself, prove that nearby parameter values satisfy the exact restrictions?
Give the Jacobian rank and identification conclusion for \(g(\theta)=\theta^3=0\) at \(\theta_0=0\), and explain why the two conclusions differ.