Consistency hinges on reconstructible boundary signals
The paper asks whether prediction-error identification of a subnetwork inside a dynamic network can yield consistent estimates of that subnetwork's transfer functions. The proposed route is to append the measured boundary signals entering the subnetwork to the exogenous inputs, then apply complete-network prediction-error machinery to the enlarged subnetwork.
Consistency is not automatic. It requires that the component of each measured boundary signal generated by the subnetwork be causally reconstructible from the measured history and the measured exogenous inputs available at time k. If that holds, the boundary signal carries no extra information about the current subnetwork innovation, and the standard prediction-error orthogonality argument can be applied.
Measuring every signal from the subnetwork into the rest of the network is a simple sufficient structural condition, but the paper states it is not necessary. Reconstructibility can permit a smaller subnetwork, and the subnetwork may instead be enlarged until the condition is satisfied.
The boundary-information condition replaces a loop delay
In the standard closed-loop setting, a delay in the loop is a common sufficient condition: the current prediction error cannot travel around the feedback loop and affect the input used to predict the same output. The paper notes this property is not automatic when only part of a dynamic network is measured, because a boundary signal returning to the subnetwork may depend on an unmeasured subnetwork output.
The boundary-information condition is the paper's replacement for that delay argument. It states that the subnetwork-generated component of each measured boundary signal entering the subnetwork is reconstructible from past measurements and the measured exogenous inputs at time k. Under this condition, the predictor's information contains no current subnetwork innovation, giving the orthogonality relations E[epsilon_k x_j^T] = 0 for j < k and E[epsilon_k rbar_j^T] = 0 for j <= k.
The paper presents this as the additional condition needed in the subnetwork setting, beyond what complete-network identification requires.
Four assumptions turn the criterion into a consistent estimator
Theorem 1 states that if the boundary-information condition holds and four further conditions are met, the estimated subnetwork transfer functions converge almost surely to the true ones as the sample size N grows. The four conditions are: the A- and B-polynomials of the ARMAX models are coprime; the C-polynomials have no zeros on the unit circle; the pair (x_k, rbar_k) is informative enough; and the A-part of the network is generically network identifiable.
The proof relies on strict properness of G(q) to prevent the current disturbance in the enlarged A-part from reaching the boundary signal, and on diagonality of H(q) and the noise covariance to separate current disturbances in the enlarged A-part from the remaining B-part. Together with informativity and identifiability, these relations let the complete-network consistency result be applied to the enlarged A-part.
The paper also notes that the subvector parameterizing the subnetwork transfer functions converges almost surely to its true value, and that x_k may need to be replaced by suitable linearly independent linear combinations of its components to satisfy the informativity assumption.
Nonlinear predictors can need fewer internal measurements
The main motivation for the prediction-error formulation is that it permits predictors that are nonlinear in the open-loop transfer functions of the subnetwork. Such predictors can require fewer measured signals inside the subnetwork than linear predictors while still yielding consistent estimates of the subnetwork parameters.
This contrasts with local-module methods whose predictor formulations are linear in the target open-loop module. The paper states that the resulting predictor here may depend nonlinearly on the open-loop transfer functions and can consequently require fewer internal measurements.
One limitation stated alongside the method: in general, the noise model cannot be estimated.
A seven-module example separates fit quality from parameter bias
The numerical example uses a seven-module network and estimates G1, G2 and G3, with the target subnetwork A = {1,2,3} and the remaining modules B = {4,5,6,7}. The measured input signal for estimation is (r1, r2, y6), and the signal from A to B that can return to A is y3, which is included in every choice of measured output.
Three measurement configurations are compared: measuring only y3, measuring (y1, y3), and measuring (y1, y2, y3). The proposed PEM and direct PEM give similar results with the same measured signals. The paper's stated main advantage is that all three open-loop transfer functions can be recovered using fewer internal measurements, with only a small decrease in fit values: measuring only y3 changes each fit by less than one percentage point relative to complete internal measurement.
The fit values themselves are close across configurations. With only y3 measured, the proposed PEM gives F1 = 56.58, F2 = 75.89 and F3 = 60.36. With (y1, y3) it gives 57.10, 75.84 and 60.21. With all three outputs it gives 57.13, 76.86 and 60.43, and direct PEM gives the same three fit values.
Fewer measurements raise parameter bias more than they lower fit
The parameter statistics tell a different story from the fit values. Squared bias rises from 0.1974 with all three outputs measured to 1.3776 with y3 alone, while parameter MSE rises from 1.0475 to 1.7539. The intermediate configuration with (y1, y3) gives squared bias 1.2648 and MSE 1.5090.
The trace of the empirical covariance moves in the opposite direction: it is 0.8501 with all three outputs, 0.2442 with (y1, y3), and 0.3764 with y3 alone. Direct PEM with all three outputs has trace 0.8717, squared bias 0.1400 and MSE 1.0117.
The paper attributes part of the observed parameter bias to the nonconvexity of the objective, since convergence to local minima can contribute to it. These statistics come from 100 disturbance realizations using the same external-input data, so they describe variability conditional on that excitation data.
The method pays off when a small boundary separates the target
The paper states the method is most useful when a small measured or reconstructible boundary separates the target A-part from the B-part. Otherwise, the number of nuisance modules and the optimization dimension can become large. Enlargement only moves modules from B to A and terminates after at most all modules in B have been included, so in the worst case estimation covers the complete network.
Enlargement is needed only to obtain the simple structural condition; direct verification of reconstructibility may permit a smaller A-part. The paper also notes that the initial A-part may contain several target modules, and that if only one transfer function is the target, the initial A-part may contain only that module, making single-module identification a special case.
The paper's future work points to numerical aspects of the optimization problem and further investigation of measurement requirements for subnetwork identification.
Fit values alone understate the cost of sparse measurement
The decision-relevant comparison in this paper is not between the proposed PEM and direct PEM on fit, where they agree when the same signals are measured. It is between measurement configurations within the proposed method. Dropping from three internal outputs to y3 alone costs less than one percentage point of fit on each of F1, F2 and F3, but multiplies squared parameter bias by roughly seven and raises parameter MSE by about two thirds.
For a reader deciding what to measure, the paper's own variables point to the boundary condition first: if a small measured or reconstructible boundary separates the target subnetwork from the rest, the method can recover all three open-loop transfer functions with fewer internal signals. If that boundary is not small, nuisance modules and optimization dimension grow, and the worst case is estimating the complete network. The fit-versus-bias split is the number to watch, because the fit values alone would suggest sparse measurement is nearly free.
Sources used in this answer
Prediction Error Method for Identification of Subnetworks of Dynamical Systems
Under the boundary-information condition plus stability, delay, disturbance-separation, informativity and identifiability assumptions, the prediction-error criterion yields consistent estimates of subnetwork transfer functions, and nonlinear-in-the-open-loop predictors can achieve this with fewer measured internal signals than linear predictors.