228

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 59, NO. 1, JANUARY 2014

Nu-Gap Model Reduction in the Frequency Domain

Aivar Sootla

AbstractThe nu-gap metric was originally developed to evaluate robustness of a plant-controller feedback loop and it has many attractive properties. For example, performance of the closed loop changes continuously and stability is robust with respect to small perturbations of the plant (or the controller) in the nu-gap metric. In light of these properties, one can state that the nu-gap metric provides a good measure of distance between systems in a closed loop setting. This is very useful in model approximation, which is the focus of this technical note. The presented nu-gap approximation method is based on semidenite programming and frequency response matching, which allows to extend the method to account for frequency-dependent weights in the objective function. The frequency-weighted extension is the major advantage of the presented method in comparison with other nu-gap model reduction methods. This extension is applied to approximation of controllers obtained by loop shaping and illustrated on a numerical example. Index TermsController reduction, model reduction, nu-gap metric, semidenite programming.

I. INTRODUCTION Model reduction of linear systems is well studied in the literature and many efcient methods have been derived. The methods delivering norm are balanced truncathe best approximation quality in the tion [1] and optimal Hankel model reduction [2]. These methods also guarantee stability, however, computationally very expensive. Krylov methods, cf. [3], on the other hand, are considerably cheaper, but generally provide only local approximations. It is important to remark that the Krylov method [4] provides locally optimal approximations in the norm, if the resulting approximation is stable. Most of the standard model reduction methods measure the approxior norms, that is, they measure the distance mation error in the in the open loop setting. In a closed loop setting these norms usually do not reect the distance adequately. An early attempt to create a more reliable metric in the closed loop setting was the introduction of the gap metric in [5], followed by many papers including [6], [7] and [8]. In the last reference the -gap metric is introduced. This metric induces the weakest topology in which stability of a closed loop is maintained for sufciently small perturbations and performance varies continuously. This is a valuable property in controller reduction, since the controllers which are sufciently close in the -gap metric will stabilize the nominal plant and the performance of the loops will be similar as well. Literature in -gap model reduction is not as rich as in the case. An approximation method have been already proposed in the rst -gap paper [8], where low-order models are obtained one step at a time. It means, if the order of the full order model is , then rstly opis obtained. Here, is timal in the -gap approximation of order the multiplicity of the smallest Hankel singular value of a normalized order right graph symbol of the model. After that, optimal order model is calculated and so on down to approximation to
Manuscript received September 26, 2011; revised April 12, 2012; accepted May 10, 2013. Date of publication November 18, 2013; date of current version December 19, 2013. This work was supported by the Swedish Research Council with the Linnaeus Grant for Lund Center for Complex Engineering Systems. Recommended by Associate Editor Z. Wang. The author was with the Department of Automatic Control, Lund University, Lund 221 00, Sweden and is now with the Department of Bioengineering, Imperial College London, London SW7 2AZ, UK (e-mail: a.sootla@imperial.ac.uk). Digital Object Identier 10.1109/TAC.2013.2290817

the required reduction order. In [9] a similar method is proposed, with more freedom to choose optimal approximations. In this note, a single-input-single-output model reduction algorithm is discussed, as well as, its extension to the frequency-weighted problem. The presented method has its roots in [10], [11], which are also related to [12], [13]. As opposed to the previous works in -gap model reduction [8], [9], which use state-space representations to compute low-order approximation, the presented algorithm employs the frequency domain data. The use of frequency domain data has some advantages, e.g., a straightforward frequency-weighted extension. The frequency-weighted version is applied to approximation of loop shaping procedure controllers, which are obtained using the (cf. [14], [15]). The frequency-weighted -gap approximation is able to further reduce the order of the controller in comparison with its non-weighted counterpart without sacricing performance levels. This is done by successfully exploiting the structure of the controller design which is based on adding frequency weights. The rest of the note is organized as follows: in Section II the -gap metric is briey introduced. Section III is devoted to the algorithm derivation. In Section IV extension to the frequency-weighted problem and its application to controller reduction is discussed, which is illustrated on a numerical example in Section V. Notation denote the space of discrete-time by rational Let denote . The norm transfer matrices and , where is the maxis computed as imum singular value of a matrix and is the complex identity. If is stable, then this expression computes its norm . The is called the winding number of and it is equal function encircles the origin. Let to the number of times the curve be the normalized right graph symbol of a plant , where and is the normalized right coprime factorization of , and nally

if stabilizes otherwise

II. PRELIMINARIES Without loss of generality consider only discrete-time models. The continuous-time models are discretized using the bilinear transformation while pre-warping around a specic frequency . Due to bijectivity of the bilinear transformation, one can view the discrete-time variable as a basis function in the continuous-time domain . be rational transfer matrices, and , Denition 1: Let , be their right normalized graph symbols. Let (1) be called the winding number condition. Let be the functions such that , :

if (1) is satised otherwise

0018-9286 2013 IEEE

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 59, NO. 1, JANUARY 2014

229

Note that there is no assumption on the open-loop stability of and , which proves useful in many applications. The function is called -gap and it is a metric. Moreover, the stability margin denes a ball around in the -gap metric, which contains only plants stabilized by the controller ; performance is changing continuously depending on the plant, if the plant is perturbed in the -gap metric. These properties can be formalized in the following statement, which is a combination of results from [15]: Proposition 1: is a metric. 1) The function and 2) Robustness (Remark 3.11 to Theorem 3.10). Given i) is stable for all satisfying if, and only if, ii) is stable for all satisfying if, and only if, 3) Performance (Theorem 3.19). For any , , such that and are stable,

Note by construction. In this formulation is a full order model, and parametrizes its th order approximinimization is rewriting the mation. The usual approach to norm constraint (4) with an innite number of constraints for every frequency (5)

The obtained program is not generally convex even for the scalarvalued functions. The main result of this technical note is rewriting (5) in an equivalent formulation, from which a quasi-convex program , can be obtained. Introduce new variables and yet another formulation as (6)

4) Approximation. This result is based on Theorem 8.6. Let be a be a set of the transfer functions which plant of order and have a minimal realization of order smaller than , then

where is a -th largest Hankel singular value of the right normalized graph symbol of . Proposition 1 can be restated to describe controller properties instead. For instance, it can be shown that the changes in performance are continuous given small perturbations in the controller. III. LOW-ORDER APPROXIMATION IN THE -GAP METRIC Let and the elements of be scalar, not necessarily stable, transfer functions and consider the problem

Let , be the normalized right coprime factorization of . Let , be nite impulse response (FIR) lters with a xed order , i.e., and . Let also and be a right . In the new variables , , and , the coprime factorization of is computed as function

The winding number condition (1) is equivalent to (2) and the equivalence is shown in a straightforward manner using the properties of the winding number function. Finally, -gap model reduction in the new variables is formulated as follows: (3) (4)

Theorem 1: Assume and is the normalized coprime factorization of the plant , then: and of the programs (5) and (6) are equal The minima , are the solutions to (5), if and only if, , , , are and . the solutions to (6) for some Before proceeding with the proof of the theorem, it is important to remark how (6) is going to be used. A quasi-convex formulation is obtained in two steps: removing the coprimeness constraint on , and xing variables , in the program. Removing the coprimeness constraint on and in (6) is a relaxation; however in the authors experience, numerical examples provide coprime and . This can occur due to the following reasoning. For and to be not-coprime, they should have precisely the same zeros, which is a measure zero subset of all admissible and . Therefore numerical solutions typically do not reach not-coprime and . If and are nearly not-coprime, a possible remedy would be lowering the order . The winding number condition in (6) is enforced on the variables and . Therefore after xing those, the norm constraint becomes semidenite and the program becomes quasi-convex. This is a key observation in the presented approach and it is discussed in detail in the next subsection, while now the main result is to be shown using the following lemma: has no poles on Lemma 1: Assume that a rational function is strictly positive for all , then the unit circle. If has no poles on the unit circle and is positive for all , then the function does not encircle the origin and is equal to zero. Using the properties of the winding number function it follows . that: . And, nally, Proof: If

230

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 59, NO. 1, JANUARY 2014

Proof of Theorem 1: First, let us show that and . Let and . Since for any complex number , we have

for all

using the interior-point solver SEDUMI [17] and YALMIP [18]. Since the program (7) is quasi-convex, it is solved using bisection. To save computational effort bisection is performed in variable instead of . This also gives upper and lower bounds for the bisection procedure (1 and 0, respectively). The stopping criterion is based on and convergence in is attained. Since the previous solutions and are used on the next step, is always bigger or equal to . On is bounded from below, which entails convergence the other hand in . Algorithm 1 -gap model reduction

The term appears on both sides of the norm constraint and can be eliminated without affecting the solution. Note that has no zeros on the unit circle according to constraints above, whereas has no poles on the unit circle by construction. Now we can apply Lemma 1 and and on the unit circle with replace the conditions and on the unit circle. This gives us (5) and, nally, . and be an optimal solution to (5). To prove the converse let The functions and also satisfy the constraints of (6) with and . Therefore . The second part of the theorem is shown using the same arguments. , , are the solutions of (6), then by construction , If , , , are the solutions to (5). On the other hand, if , , and are the solutions to (5), then they also satisfy the constraints of (6) with , and . A. Tractable Algorithm and Implementation Let the program (6) without the coprimeness constraints on , and , be called the relaxed program (6), which is the basis for the preand sented below model reduction method. After the functions satisfying the winding number condition are found, the relaxed program (6) with xed and becomes quasi-convex and can be solved using standard tools. Moreover, the proof of Theorem 1 shows that and can be interpreted as a guess on variables and . This is exand on ploited by iterating over and , while setting every step. This approach to the non-convex program (i.e., the relaxed version of (6)) is a standard heuristic, which gives suboptimal solutions to (6). Generally, there is no guarantee that this heuristic provides solutions close to the optimal one. However, in the studied case and are closely related to and , which provides a good reason to perform such iterations. , which An obvious choice for the initial and is computing is as close as possible to in the -gap. However, the relaxed problem (6) is non-convex and there is no guarantee that a better in terms of quality initial guess will provide a better nal result. Therefore, it may not be benecial to compute the initial guess with a method as [9], very close to , but is numerically expensive. which provides Hence a simpler choice of the initial point is proposed and it is motibe any vated by the fourth bullet of Proposition 1. Let optimal Hankel approximation of the right normalized graph symbol of , where , and are FIR lters or order . Now simply set initial equal to and initial to . Frequency dependent constraints in (6) may be imposed for all the frequencies at once using the Kalman-Yakubovitch-Popov lemma (cf. [16]). To provide a computationally cheaper program the constraints , where is big enough may be enforced on a frequency grid to avoid over-t. Since the -gap metric emphasizes the behavior around the gain cross-over frequencies, it is reasonable to make the grid denser around those frequencies. The -gap model reduction method is summarized in Algorithm 1. The algorithm is implemented

Inputs: Full order plant domain of

, approximation order , tolerance ,

Output: Reduced order plant Compute the normalized right coprime factorization of Compute -th order Hankel approximation , where , and are FIR lters. Set repeat Set and solve (7) , and of

where Set until B. Computational Complexity ,

and and ,

Let be the order of the full order model and be the order of and let . There are two main contributors to approximation computational complexity: the computation of the initial guess and the optimization problem. The computation of the initial guess involves normalized coprime factorization and Hankel approximation, which are performed using Riccati and Lyapunov equations. Therefore comoating point operations (ops). plexity is The overall cost of an optimization program when solved with SEDUMI does not exceed ops, where is the number of decision variables and is the number of rows in the inequalities [19]. In the case, when the constraints are enforced on a freand , where is the number or quency grid, frequency points in the grid. Computing the frequency samples costs . Therefore, in this case, the cost of the optimization proat most . If the constraints are enforced using the gram does not exceed KYP lemma for all the frequencies, then and . . This gives the overall cost of In comparison with other methods, the presented one has a competRiccati itive cost. The methods from [9] and [15] require solving or Lyapunov equations, respectively. Therefore the total cost of the al. gorithms is

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 59, NO. 1, JANUARY 2014

231

IV. FREQUENCY-WEIGHTED EXTENSION AND CONTROLLER REDUCTION In this section the -gap model reduction is applied to the controller reduction problem. Specically, the controllers obtained using the loop shaping (cf. [14], [20]) are considered. The rst step in the loop shaping is the introduction of the shaped plant , and are chosen according to desired where the weights , which minimizes specications. After that the central controller , is obtained. The actual controller is computed . In the studied case and are scalar, therefore a as can be considered without loss of generality and one-sided weight . In order to obtain a reduced order controller one can approximate by a controller of order (8) In this case the actual reduced order controller is .A better way to address the low-order controller design is to introduce directly as a variable. Note that only the shaped loop has a conve, which allows the use of the -gap metric nient form for controller approximation. Therefore the only possibility of approxas a variable, is as follows: imating the controller, while having (9) The program (8) can be viewed as (9) with an additional constraint . In (9) the structure of is relaxed and this grants more exibility in the decision variable choice. Hence the frequencyweighted approach can provide controllers of lower order in comparison with the central controller approximation, while keeping similar and performance levels. As a nal remark, note that the problem (9) is essentially the frequency-weighted -gap reduction problem

TABLE I APPROXIMATION ERRORS OF VARIOUS METHODS IN THE -GAP METRIC AND EFFECT ON THE CLOSED LOOP PERFORMANCE

To nalize the algorithm, let , be the normalized right coand change the program (7) accordingly prime factors of (11)

where

and

V. NUMERICAL EXAMPLE. ROBUST STABILIZATION OF CART The main goal of this example (which is taken form [15]) is to illustrate that the proposed method is competitive in comparison with the known techniques in the non-weighted case, but has an additional advantage in the form of the frequency-weighted extension. For the loop-shaping procedure the same weight as in [15] cannot be used, since in (10) and (11) the weight has to be invertible. Hence, two are added, which results in a slightly worse disturbance zeros at are chosen as attenuation. The plant and the shaping weight

A. Frequency-Weighted -Gap Reduction As discussed above, is the problem to solve. In order to account for the frequency-weights in Algorithm 1, it is required to redesign the initial point computation and the program (7), which is solved inside the loop. In preparation of this technical note different approaches to both problems were considered and the listed heuristics provided the most acceptable results. , and Let , be the normalized right coprime factors of , be the normalized right coprime factors of . Similarly to the non-weighted algorithm, the initial guess is found by ap. However, in this case the proximating a right graph symbol of weight in the approximated graph symbol must be preserved, hence is approximated. This transfer mathe transfer matrix , but its norm is smaller or equal to trix is a right graph symbol of one, instead of being equal to one for the non-weighted computation of the initial guess. Finally, the computation of the initial guess is performed as follows: (10) , are stable functions of order , , , where and . and the initial guess is are FIR lters

In comparison with [15] the controller design procedure is also simplied. The controller is obtained using Theorem 16.1 from [20] with , providing . The reader may notice that overall this design is less robust to changes in than the one provided in [15], however, this design will still colorfully illustrate the results of -gap controller reduction. The continuous-time models are discretized while pre-warping around the frequency , which is roughly the cut-off frequency of the closed loop. A. Approximation of the Central Controller The results of -gap approximation of the central controller are presented in Table I. The parameter for [9] is chosen as a zero matrix. The notation [9]+Alg 1 signies that the initial guess in Algorithm 1 is obtained using [9]. Finally, the relative closed loop error is dened as follows:

The tolerance for Algorithm 1 is set to , all the constraints are enforced on an equidistant frequency grid with 0.02 between the samples. The algorithm converges after at most 6 iterations, which occurs for order 4.

232

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 59, NO. 1, JANUARY 2014

The results of [9] and Algorithm 1 with different initial guesses are quite similar for all the orders. Moreover, the results are quite similar in both the -gap approximation quality and the relative closed loop error. Another important remark is that Algorithm 1 with the initial guess [9] does not provide consistently better solutions than Algorithm 1 with a simple initial guess. This happens because the problem is non-convex with possible multiple local minima. Hence while starting with a better initial guess in terms of approximation quality, there is no guarantee that the nal result will still be better in terms of approximation quality. Surprisingly, for order 3 Algorithm 1 provides the same controller (coefcients of the transfer functions match) with both initial guesses. An interesting detail appears when the order of approximation is set to 2. The initial guess provides the low-order controller on the distance 0.45 in the -gap metric from the original one. This second order approximation does not stabilize the shaped plant . However, after applying Algorithm 1 a stabilizing controller is obtained already after the rst iteration. This controller lies on the distance 0.29 in the -gap metric. While the performance of the closed loop is not satisfactory (the relative closed loop error is larger than 4), the loop is stabilized. The method [9] also provides a stabilizing controller with the -gap error 0.29 and the relative closed loop error 2.47. According to Table I, the relative closed loop error of [9]+Alg 1 can be worse than the relative closed loop error of [9]. This means that the performance of the initial guess (model obtained with [9]) is better than the performance of the model obtained with [9]+Alg 1. This seems counterintuitive, however, the minimization objective is the -gap metric not the relative closed loop error. It is certainly posand such that sible that there are controllers , but

than its non-weighted counterpart without sacricing performance or robustness. VI. CONCLUSION This technical note is devoted to the -gap model reduction algorithm. The main asset of the proposed approach is the use of the frequency domain data, which grants more exibility. For example, the frequency-weighted version of the proposed approach is obtained, and it is not clear if it is possible with [9]. The frequency-weighted extension is validated on a numerical example, where reduced order controllers of signicantly smaller orders were obtained without sacricing performance levels. Other extensions are also easier to perform in the frequency domain setting, since then the only obstacle is to formulate the extension as a convex constraint. ACKNOWLEDGMENT The author would like to thank Dr. Sandberg for bringing attention to this problem, Dr. Ranzter for valuable comments, and the anonymous referees, whose comments contributed to the improvement of this technical note.

REFERENCES
[1] B. Moore, Principal component analysis in linear systems: Controllability, observability, and model reduction, IEEE Trans. Autom. Control, vol. 26, no. 1, pp. 1732, Feb. 1981. [2] K. Glover, All optimal Hankel-norm approximations of linear multi-error bounds, Int. J. Control, vol. 39, variable systems and their pp. 11151193, 1984. [3] A. C. Antoulas, Approximation of Large-Scale Dynamical Systems, ser. Advances in Design and Control. Singapore: SIAM, 2005. model reduction for [4] S. Gugercin, A. C. Antoulas, and C. Beattie, large-scale linear dynamical systems, SIAM J. Matrix Anal. Appl., vol. 30, no. 2, pp. 609638, 2008. [5] G. Zames and A. K. El Sakkary, Unstable systems and feedback: The gap metric, in Proc. Allerton Conf., 1980, pp. 380385. [6] T. Georgiou and M. Smith, Optimal robustness in the gap metric, IEEE Trans. Autom. Control, vol. 35, no. 6, pp. 673686, Jun. 1990. [7] M. Vidyasagar, The graph metric for unstable plants and robustness estimates for feedback stability, IEEE Trans. Autom. Control, vol. 29, no. 5, pp. 403418, May 1984. [8] G. Vinnicombe, Frequency domain uncertainty and the graph topology, IEEE Trans. Autom. Control, vol. 38, no. 9, pp. 13711383, Sep. 1993. [9] G. Buskes and M. Cantoni, A step-wise procedure for reduced order approximation in the -gap metric, in Proc. 47th IEEE Conf. Decision Control, Dec. 2008, pp. 48554860. [10] K. Sou, A. Megretski, and L. Daniel, A quasi-convex optimization approach to parameterized model order reduction, IEEE Trans. Comput.-Aided Des. Integr. Syst., vol. 27, no. 3, pp. 456469, Mar. 2008. [11] A. Sootla, Semidenite Hankel-type model reduction based on frequency response matching, IEEE Trans. Autom. Control, vol. 58, no. 4, pp. 10571062, Apr. 2013. [12] D. Henrion, M. ebek, and V. Kuera, Positive polynomials and robust stabilization with xed-order controllers, IEEE Trans. Autom. Control, vol. 48, no. 7, pp. 11781186, Jul. 2003. controller [13] F. Yang, M. Gani, and D. Henrion, Fixed-order robust design with regional pole assignment, IEEE Trans. Autom. Control, vol. 52, no. 10, pp. 19591963, Oct. 2007. [14] K. Glover and D. C. McFarlane, Robust Controller Design using Normalized Coprime Factor Plant Descriptions. Secaucus, NJ, USA: Springer-Verlag, 1989. Loop-Shaping and [15] G. Vinnicombe, Uncertainty and Feedback the -Gap Metric. Singapore: World Scientic Publishing Company, 2000. [16] J. Willems, Least squares stationary optimal control and the algebraic Riccati equation, IEEE Trans. Autom. Control, vol. 16, no. 6, pp. 621634, Dec. 1971.

To conclude this discussion, the performance is changing continuously with respect to small changes of the controller in the -gap metric (according to Proposition 1), however, there is no guarantee that the performance will change monotonically. B. Weighted Approximation of the Controller Since the actual controller for the plant is , the order of the actual controller is plus the order of , which is equal to 6. To obtain an is minimized even lower order design the distance . Now, is the controller for the plant , not for the shaped over . The grid for this example is also equidistant with 0.005 plant . All the obtained between two consecutive samples, and controllers of orders less than 6 destabilize the systems, and for orders larger than 7 the match is almost perfect even after the initial guess computation. For order 6 approximation 33 iterations are required, for order 7 approximation 19. The number of iterations grows signicantly, mainly since the problem is now more complicated due to the weight. For order 6 the presented algorithm provides relative closed loop , and for order 7 . This is a signicant error improvement in comparison with the initial guess values and for orders 6 and 7, respectively. The improvement in against for order the -gap distance is even larger against for order 7. This improvement 6 and is not surprising, since in the weighted case there are no guarantees on the initial guess quality. Note that the relative closed loop error of of order 6 is comparable to of order 6, which results controller . Hence, the frequency-weighted -gap in order 12 controller model reduction is able to provide a much lower order approximation

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 59, NO. 1, JANUARY 2014

233

[17] J. F. Sturm, Using SeDuMi 1.02, A MATLAB toolbox for optimization over symmetric cones, Optim. Methods Softw., vol. 1112, pp. 625653, 1999. [18] J. Lfberg, Yalmip: A toolbox for modeling and optimization in MATLAB, in Proc. IEEE Int. Symp. Comput.-Aided Control Syst. Des., Taipei, Taiwan, 2004, pp. 284289. [19] D. Peaucelle, D. Henrion, Y. Labit, and K. Taitz, Users Guide for SEDUMI Interface 1.04, LAAS CNRS, Tech. Rep., 2002. [20] K. Zhou and J. C. Doyle, Essentials of Robust Control. Upper Saddle River, NJ, USA: Prentice-Hall, 1998.

Modied Bryson-Frazier Smoother Cross-Covariance

William R. Martin, Senior Member, IEEE

AbstractThe Expectation Maximization algorithm can be used to estimate Kalman lter parameter matrices. This requires smoother results and typically the Rauch-Tung-Striebel smoother is used for this purpose. If either of the state transition or plant noise matrices require estimation, a sequence of single time-step smoother cross-covariance matrices, based on the Rauch-Tung-Striebel smoother formulation, are also required. The modied Bryson-Frazier smoother could be used as the smoother for the Expectation Maximization algorithm, but to estimate the state transition or plant noise matrix, this smoother would rst need a method to generate the sequence of single time-step smoother cross-covariance matrices based on the modied Bryson-Frazier formulation. This paper develops an expression for these modied Bryson-Frazier cross-covariance matrices, thus allowing the modied Bryson-Frazier smoother to be substituted for the Rauch-Tung-Striebel smoother in the Expectation Maximization algorithm. The modied Bryson-Frazier smoother formulation requires inversion only of lter innovation matrices, which are generally much smaller than state error covariance matrices, which allows both greater computational speed and better numerical accuracy due to smaller matrix size. This formulation can be used even when covariance and transition matrices are singular. Index TermsEstimation, expectation maximization, Kalman ltering, linear dynamic system, smoothing.

Appendix equations (A8A10) [1, p. 263] as the discrete-time xedinterval Rauch-Tung-Striebel (RTS) [3], [4, pp. 163164] smoother equations. In order to provide an alternate smoothing technique for the EM algorithm, this paper develops the single time-step modied Bryson-Frazier (mBF) cross-covariance matrix equation for the mBF smoother. The paper is organized as follows. Section II investigates parameter estimation using the EM algorithm with the RTS smoother, with Section II-A showing the Kalman lter equations, Section II-B showing the RTS smoother equations, and Section II-C showing the RTS cross-covariance equation and EM algorithm as given in [1]. Section III discusses the mBF smoother equations and Section IV shows the mBF equations with cross-covariance. Section V derives the mBF cross-covariance, with Section V-A developing truth, prediction error, and innovation equations and showing two innovation expectation properties, Section V-B demonstrating the relationship between mBF adjoint vectors and matrices and dening lter cross-covariance, Section V-C reviewing the orthogonality principle, and Section V-D deriving the mBF cross-covariance using results from the previous subsections. Conclusions are given in Section VI. II. PARAMETER ESTIMATION USING THE EM ALGORITHM A. Kalman Filter Equations The time interval over which the ltering and smoothing will take . measurements will occur from times place is denoted by through . Fixed-interval ltering and smoothing will be used exclufor and . sively; dene the Kroneker delta by The system model is given by (1) (2) and are the true system state and measurement where . Vectors , and are independent vectors at time zero-mean Gaussian random variables with plant noise covari, measurement noise covariance ance matrix , a priori state error covariance mamatrix and state with trix , , , is the state to , and is the measurement transition matrix from time matrix relating the state to the measurement at time . The Kalman lter time extrapolation equations are (3) (4) is the extrapolated state estimate at time given meawhere , is the extrapolated surements from time through time and following [1], we take state error covariance of and to simplify notation. The Kalman lter update equations at measurement time are given by (5) (6) (7) (8) (9) where trix, is the innovation vector, is the innovation covariance mais the updated state is the Kalman (optimal) gain matrix,

I. INTRODUCTION Kalman ltering applications occasionally arise where the lter designer has imperfect knowledge of some, many or all of the lter parameters for various lter matrices due to a variety of reasons; the physics is not tractable, there are too many lter states, the cost involved in determining parameters is too high, or perhaps one has only an educated guess for values of Markov time constants, initial state uncertainties, plant or measurement noise, or measurement matrix coefcients. Under these conditions, it is possible to use the Expectation Maximization (EM) algorithm [1] to rene these Kalman lter parameter estimates. The application of the EM algorithm in this context requires the use of smoothing techniques. The Appendix in [1] shows the equations for the lter, smoother, and cross-covariance required for the EM algorithm. We note that although [1] references Jazwinski [2] as the source for the lter and smoother equations, we also recognize
Manuscript received March 31, 2011; revised November 09, 2011 and November 10, 2011; accepted June 27, 2012. Date of publication June 19, 2013; date of current version December 19, 2013. Recommended by Associate Editor Z. Wang. The author is with the Applied Physics Laboratory, Johns Hopkins University, Laurel, MD 20723 USA (e-mail: william.martin@jhuapl.edu). Digital Object Identier 10.1109/TAC.2013.2270074

0018-9286 2013 IEEE