Augmented-State Observability Tests for Parameters

Why Augment the State?

A dynamical model produces outputs through changing states, but its behaviour also depends on parameters that are fixed over time. To test whether output trajectories can distinguish those parameters, we can treat each constant parameter as another state variable whose derivative is zero. The resulting augmented system lets us use state-observability calculations to investigate local parameter identifiability.

This is a structural test: it concerns what an ideal, noise-free output trajectory can distinguish under the stated model, inputs, and assumptions. It is not a guarantee that noisy, sparse clinical measurements will estimate the parameters precisely. The earlier material on state observability versus parameter identifiability distinguishes these questions; here the emphasis is the augmented-state rank calculation.

Definition: Consider a smooth autonomous model \(\dot{x}=f(x,\theta)\) with output \(y=h(x,\theta)\), where \(x\) is the state vector and \(\theta\) is a vector of constant unknown parameters. Define the augmented state \(z=(x,\theta)\) and augmented dynamics \(\dot{z}=F(z)=(f(x,\theta),0)\). The output is \(H(z)=h(x,\theta)\). The Lie derivatives are \(L_F^0H=H\) and \(L_F^{k+1}H=\nabla_z(L_F^kH)F\). The observability matrix through order \(r\) has rows \(\nabla_z L_F^kH\) for \(k=0,\ldots,r\).

For several outputs, include the corresponding Lie-derivative rows for each output. The matrix has columns for the augmented coordinates: both the ordinary states and the parameters. A full column rank calculation means that, locally, small changes in any augmented coordinate produce distinguishable changes in the output-derivative information included in the matrix.

The basic assumptions are important. The model equations and output map must be specified, the functions must be sufficiently smooth for the derivatives used, and the input conditions must be fixed or explicitly represented in the model. The parameter values are constant over the observation interval. In the examples below there is no external time-varying input. A calculation at a particular point tests rank there; it does not automatically establish the same rank at every possible state and parameter value.

How the Rank Test Is Built

The output itself gives the first row of the matrix. Differentiating the output along the system gives the next observable quantity, and differentiating again gives further rows. The parameter coordinates enter these derivatives even though their own time derivatives are zero: they affect the rates of change of the ordinary states.

\mathcal{O}_r(z)= \begin{bmatrix} \nabla_z H\\ \nabla_z L_F H\\ \nabla_z L_F^2 H\\ \vdots\\ \nabla_z L_F^r H \end{bmatrix}.

For a scalar output and an augmented state of dimension \(m\), \(m\) linearly independent rows are needed for full column rank. More rows may be required to achieve that rank. For multiple outputs, rows from each output and its Lie derivatives can contribute. Rank is evaluated as a function of \(z\), so a determinant or minor that is nonzero at one point can establish full rank there, while a zero minor alone does not establish that the full matrix is rank deficient.

Rank-test interpretation: If the observability matrix has full column rank at a point, then the augmented state, including its constant parameters, passes the local observability test there. This supports local identifiability of those parameters under the model assumptions. A deficient matrix made from only a finite number of derivatives is not, by itself, proof of nonidentifiability: higher derivatives may add information. Conversely, a rank calculation at one point does not establish global uniqueness or practical estimability.

Worked Calculations

Worked Example: A Generic Full-Rank Point

Consider the hypothetical scalar system \(\dot{x}=ax+b\), \(y=x\), where \(x\) is a concentration in \(\mathrm{mg/L}\), \(a\) is a rate constant in \(\mathrm{day}^{-1}\), and \(b\) is a constant rate in \(\mathrm{mg/(L\,day)}\). Treat \(a\) and \(b\) as unknown constants. The augmented state is \(z=(x,a,b)\), with \(\dot{z}=(ax+b,0,0)\).

The output and first two Lie derivatives are

H=x,\qquad L_FH=ax+b,\qquad L_F^2H=a(ax+b)=a^2x+ab.

Taking gradients with respect to \((x,a,b)\) gives

\mathcal{O}_2= \begin{bmatrix} 1&0&0\\ a&x&1\\ a^2&2ax+b&a \end{bmatrix}.

At the hypothetical point \(x=2.0\ \mathrm{mg/L}\), \(a=0.40\ \mathrm{day}^{-1}\), and \(b=0.30\ \mathrm{mg/(L\,day)}\), the rows are \([1,0,0]\), \([0.40,2.0,1]\), and \([0.16,1.9,0.40]\), using the corresponding coordinate units. Expansion along the first row gives

\det(\mathcal{O}_2) =(2.0)(0.40)-(1)(1.9) =0.80-1.9=-1.1\ne 0.

Thus the matrix has rank 3 at this point, equal to the number of augmented coordinates. Under the stated model, the local test supports joint local recovery of \(x\), \(a\), and \(b\) from ideal output-derivative information near this point. It does not say that measurements collected with noise will estimate the rate constant or rate term precisely.

Worked Example: A Singular Equilibrium and the Correct Derivative

Now consider the hypothetical model \(\dot{x}=-ax+b\), \(y=x\), with augmented state \(z=(x,a,b)\). Its first two Lie derivatives are \(L_FH=-ax+b\) and \(L_F^2H=a^2x-ab\). Therefore

\mathcal{O}_2= \begin{bmatrix} 1&0&0\\ -a&-x&1\\ a^2&2ax-b&-a \end{bmatrix}.

Evaluate this at \(x=2.0\), \(a=0.40\), and \(b=0.80\), in mutually consistent units. The model rate is \(-ax+b=-(0.40)(2.0)+0.80=0\), so this point is an equilibrium. The third-row middle entry must be calculated as

$$ 2ax-b=2(0.40)(2.0)-0.80=1.6-0.80=0.80. $$

Consequently,

\mathcal{O}_2= \begin{bmatrix} 1&0&0\\ -0.40&-2.0&1\\ 0.16&0.80&-0.40 \end{bmatrix}.

The third row is \(-0.40\) times the second row, so it adds no independent direction. The determinant is

$$ (-2.0)(-0.40)-(1)(0.80)=0.80-0.80=0. $$

The matrix has rank 2, not 3. This is a singular point for this derivative test. Here the model also explains why the parameters cannot be recovered from the equilibrium output alone: starting at \(x=2.0\), the output remains at \(2.0\), and the equilibrium condition only constrains \(b=ax\). Different values of \(a\) paired with \(b=2a\) produce that same constant output. This trajectory-specific argument establishes nonidentifiability at this equilibrium; the rank loss alone would not establish it in every model.

Worked Example: A Nonlinear Output With Full Rank

For a third hypothetical system, let \(\dot{x}=k\), \(y=x^2\), and treat \(k\) as an unknown constant. Here \(x\) is a normalized state and \(k\) has units of normalized state per day. The augmented state is \(z=(x,k)\), with \(\dot{z}=(k,0)\). The first two Lie derivatives are \(H=x^2\) and \(L_FH=2xk\). Their gradients are

\mathcal{O}_1= \begin{bmatrix} 2x&0\\ 2k&2x \end{bmatrix}.

At the hypothetical point \(x=1.5\) and \(k=0.60\) normalized state units per day, the matrix is \(\begin{bmatrix}3.0&0\\1.2&3.0\end{bmatrix}\), and its determinant is \((3.0)(3.0)-(0)(1.2)=9.0\ne0\). The rank is 2, so the local test supports joint local observability of \(x\) and \(k\) at this point. Although \(x^2\) alone cannot distinguish \(x\) from \(-x\) globally, this local rank result concerns nearby augmented states, not global uniqueness over the entire state space.

Why Full Rank Is a Local Result

The calculation can be understood through the map from an augmented state to its output derivatives. For a scalar output, define \(\Phi_r(z)=(H,L_FH,\ldots,L_F^rH)\). Its Jacobian is exactly \(\mathcal{O}_r(z)\). If this Jacobian has full column rank at a point, then some square minor is nonsingular. The inverse function theorem applied to the corresponding selected derivative components implies that those components locally distinguish nearby augmented states.

This is the short justification for treating full rank as a positive local test. It does not imply that distant parameter values cannot produce the same output, because a locally one-to-one mapping need not be globally one-to-one. Nor does it guarantee that the derivatives can be estimated accurately from a finite set of noisy measurements.

A deficient finite matrix needs more care. It may reflect a true structural ambiguity, as the constant-output equilibrium example does, or it may simply mean that the derivatives calculated so far have not exposed the missing direction. For analytic models, the standard observability analysis considers the span of the full set of output Lie-derivative gradients; a finite calculation is conclusive for a positive result once full rank is reached, but a finite rank deficiency requires further analysis before a negative conclusion.

Practical Checks, Alternatives, and Limitations

When applying the test, check the following before interpreting a reported rank:

Differentiate the correct expression. Compute each Lie derivative from the preceding one, then take its gradient with respect to every augmented coordinate. A missing parameter contribution can change the rank.
Evaluate at the stated point. Rank can differ between generic points and special points such as equilibria. Report the conditions under which a determinant or minor is nonzero.
Check the full matrix. A zero determinant for one selected square submatrix does not prove that a larger matrix has deficient rank; another minor may be nonzero.
Separate exact rank from numerical conditioning. Very small singular values can make a numerically computed rank sensitive to scaling and tolerances. Rescale coordinates or use a rank-revealing factorization, then report the convention used.
Keep the model assumptions visible. A rank result applies to the specified output, dynamics, parameter constancy, and input conditions. An unmodelled input or a different output can alter the conclusion.

Augmented-state observability is one approach among several. Generic rank tests examine whether a suitable rank condition holds away from exceptional parameter or state sets. Input-output differential equations instead eliminate hidden states to obtain relations among outputs, inputs, and parameters. Transfer-function methods are useful for suitable linear compartmental models. These alternatives can provide complementary calculations, but their assumptions and scope differ; none turns a structural rank result into a promise of precise estimation from real data.

A common mistake is to say, “The first observability matrix I calculated is rank deficient, so the parameter is unidentifiable.” The correction is to specify the derivative order and point, calculate whether further derivatives can add directions, and distinguish a singular-point failure from a structural ambiguity. Another mistake is to call a full-rank result global identifiability. Full rank supports a local conclusion only; global distinctions require additional analysis.

Key takeaway: Add each constant parameter to the state vector with derivative zero, form gradients of the output and its Lie derivatives, and test their rank. Full column rank is a local positive result at the evaluated point; rank deficiency demands further analysis unless the model or trajectory independently demonstrates an ambiguity.

Check Your Understanding

Try each question, then compare your reasoning with the solutions below.

For \(\dot{x}=cx\), \(y=x\), and augmented state \(z=(x,c)\), write the first two rows of the observability matrix.
What does full column rank at one point support, and what does it not establish?
In the equilibrium example, why is the third-row middle entry \(0.80\), and why does the resulting matrix have rank 2?
A finite derivative matrix is rank deficient. Is that alone sufficient to conclude structural nonidentifiability? Explain.

Solutions. (1) Here \(H=x\) and \(L_FH=cx\), so the rows are \([1,0]\) and \([c,x]\); the determinant is \(x\), which is nonzero when \(x\ne0\). (2) It supports local observability of the augmented coordinates, including parameters, under the specified model assumptions. It does not establish global uniqueness or precise estimation from noisy data. (3) Substitution gives \(2(0.40)(2.0)-0.80=0.80\); the third row \([0.16,0.80,-0.40]\) equals \(-0.40[-0.40,-2.0,1]\), so only two rows are independent. (4) No. Higher Lie derivatives may supply additional independent rows; a deficient finite calculation is not by itself a proof of structural nonidentifiability.