You are on page 1of 21

A subspace approach to balanced truncation for model reduction of nonlinear control systems

Sanjay Lall , Jerrold E. Marsdeny , Sonja Glavaskiz

Submitted to
International Journal on Robust and Nonlinear Control

29 August 2000

In this paper we introduce a new method of model reduction for nonlinear control systems. Our approach is to construct an approximately balanced realization. The method requires only standard matrix computations, and we show that that when it is applied to linear systems it results in the usual balanced truncation. For nonlinear systems, the method makes use of data from either simulation or experiment to identify the dynamics relevant to the input-output map of the system. An important feature of this approach is that the resulting reduced-order model is nonlinear, and has inputs and outputs suitable for control. We perform an example reduction for a nonlinear mechanical system.


Balanced truncation, model reduction, Karhunen-Loeve expansion, total least-squares, Hankel norm.

1 Introduction
One of the di culties in designing controllers for complex physical systems is in the problem of modeling. Very often, we have mathematical models which are highly nonlinear partial di erential equations, for which control design techniques are little understood. An example is given by formation ight; accurate models are known for the uid dynamics, and for the aircraft motion, which together can be accurately simulated. However, on paper, the equations of motion consist of partial and ordinary di erential equations coupled via their boundary conditions, a formulation which o ers little to the control designer. It is therefore a problem

Correspondence to: Sanjay Lall, Department of Aeronautics and Astronautics, Stanford University, Stanford CA 94305-4035, USA. y Control and Dynamical Systems 107-81, California Institute of Technology, Pasadena, CA 91125, USA. z Honeywell Technology Center, 3660 Technology Drive, Minneapolis, MN 55418, USA.


of considerable interest to construct explicit low-order models; once a control design has been performed using such a low-order model, it can be tested by simulating it in feedback with the full high-order simulation. In this paper we present a new method of model reduction for nonlinear control systems. The goal is to develop an intuitively-motivated and systematic procedure for construction of low-order models for complex high-dimensional nonlinear systems. The focus is on preserving those features of the dynamics which are most relevant to the control design, in a similar way to that in which standard methods for model reduction of linear systems, such as balanced truncation, preserve the most controllable and observable states. This is one of the main di erences between modeling in the sense used by control systems engineers, and modeling as used for simulation or mechanical design. In the latter case, one is typically interested in the behavior of an autonomous dynamical system, that is one with no external inputs driving the system. However, in the case of controlled systems, we are concerned with modeling the relationship between the system behavior and the system inputs and outputs, or actuators and sensors. Linear input-output systems enjoy several such input-output model reduction techniques, including balanced truncation and optimal Hankel, H1 and H2 norm methods. In general, these methods cannot be applied to complex high-dimensional nonlinear systems. Our approach is to make use of data from either simulation or experiment to identify those states of the system which are most a ected by actuators, and which most a ect the sensors. We construct an empirical balanced realization for nonlinear systems, which coincides with the usual balanced realization for linear systems. A Galerkin projection is then used on the balanced realization to construct a low-dimensional nonlinear model. This reduced-order model may then be used for e cient controller synthesis, design optimization, and realtime control algorithms, where having a reduced-order model brings signi cant computational complexity bene ts. An example is given by Linear-Parameter-Varying (LPV) control design 43], where the required optimization for control synthesis involves solving a linear matrix inequality feasibility problem where the number of variables grows as the square of the state-dimension of the system. A signi cant reduction in overall computational e ort required is achieved by constructing a reduced-order model via simulation of the full-order model, and then using this reduced-order model for the control synthesis.

1.1 Prior work

Two of the most well-known methods of model reduction used in control are the Karhunen-Loeve decomposition and the method of balanced truncation. The approach in this paper combines the ideas and methodology from both of these. The Karhunen-Loeve expansion is a method of least-squares approximation which goes back to Karl Pearson 32]. The method is known in the literature 2

1.1.1 The Karhunen-Loeve expansion.

By making use of data. where it was used to model complex uid ows. 9]. and has been very successfully used in control design. and have been compared with the KarhunenLoeve method in 30]. 19. Another approach to balanced model reduction of nonlinear systems has been to make use of linear parameter-varying (LPV) systems theory. 39]. this approach is databased. 35]. based on energy functions. Moore also proposed using principal components analysis for nonlinear systems 27]. factor analysis. direct optimization techniques have been used 33]. 6. 17]. 11]. Other approaches have also been used for nonlinear model reduction. proper orthogonal decomposition (POD). the KLE has also been used to design controllers for PDEs 3. The method provides an immediate computable procedure for 3 1. including principal component analysis. For linear systems. 2. 4]. The KLE has also been used for model reduction of solids and structures 21. we avoid some of the computational di culties involved in computing energy functions for exact balanced truncation of nonlinear system 36]. and for chemical plants. The method of balanced truncation has been extensively developed for nonlinear systems by Scherpen 36. for modular interconnected nonlinear systems. 37]. A closed-loop approach is also presented in 31]. For control systems. the approach requires only matrix computations.3 Our approach . thereby developing a new method of model reduction for nonlinear systems. Computational approaches for these methods have been investigated in 29]. methods have been proposed in 34]. and is now a welldeveloped method of model reduction that appears in standard textbooks 10]. and balancing methods for bilinear systems also been developed 1. 5. and total-least-squares estimation. In fact. time-varying methods have been developed. see 22] for details. 1. making use of data from simulation or experiment in the construction of the reduced-order model.1. 38. This has found wide applicability and led to a method of understanding the important dynamical features seen in the uid 7. Along trajectories. where the linear fractional transformation is used to construct uncertain multidimensional models capturing the nonlinear dynamics 42.2 Balanced truncation The method of balanced truncation for model reduction of linear systems was proposed by Moore 28] in the context of realization theory. A-priori error bounds in the induced 2-norm are known for the error between the original and the reduced system 14. 23]. although this was not directly connected to the balanced truncation approach he had proposed in 28]. Our approach in this paper has been to combine the features of the KarhunenLoeve expansion and balanced truncation. for analysis of chaotic systems 40] and for analysis of systems with dynamical symmetries 13.1. The use of the Karhunen-Loeve expansion for model reduction was pioneered by Lumley 26].by several names. In particular.

The optimization need not be performed analytically. and a way of incorporating nonlinear a-priori mathematical models for the physics into the identi cation process. and the decomposition provides a way of nding the best approximating low-dimensional subspace. 4 . Once such a subspace is found. In this sense. where data is used for construction of models. This has the advantage of requiring only linear matrix computations. When applied to linear input-output systems. Once an appropriate measure is assigned to the space of trajectories of the system. it is a well-used method in control. This approach and a variation of it have also been used for modeling of a continuously stirred tank reactor (CSTR) in 18]. The philosophy in this paper is also similar to that of identi cation. of the form x _ (t) = f (x(t)): (1) This equation describes the system behavior as the evolution of the state x(t) in a high dimensional state space Rn . The central idea used when applying the Karhunen-Loeve expansion for model reduction is to search for a low dimensional a ne subspace of the state space. a Galerkin projection can be applied to project the dynamics onto it. We conclude the paper with an example to illustrate the application of the method to a nonlinear mechanical system. instead data from experiment or simulation can be used directly. despite its usefulness in application to nonlinear systems. in which the dynamics of interest of the original system are contained. so that the high-dimensional system is approximated by a small number of nonlinear ordinary di erential equations. The dynamics of such a physical system are often modeled by a high-dimensional nonlinear di erential equation. which is both intuitively motivated by realization theory and known to perform well in engineering practice. with the accompanying a-priori error bounds in the H1 and Hankel norms. Note that the method of balanced truncation is not optimal. in that in general it does not achieve the minimal possible error in any known norm even for linear systems. 2 Model reduction for autonomous systems In this section we give an outline of existing techniques using the KarhunenLoeve decomposition for model reduction of nonlinear autonomous systems. In general a subspace containing the interesting state trajectories may not have low (or nite) dimension. The motivation and intended use of the methods in this paper is to construct deliberately approximate reduced-order models motivated by the ideas of balanced truncation. the method of subspace identi cation 41] is relevant here. our approach for model reduction can alternatively be viewed as a new approach to identi cation. This paper is a longer version of the conference paper 25]. However. In particular. it results in the usual balanced truncation of the system. the problem of nding the subspace can be formulated as a 2-norm optimization problem. requiring only standard matrix computations.model reduction of nonlinear input-output systems.

taken either from experiments or from numerical simulation.The method makes essential use of empirical data. x(N ) g of x(t). corresponding to the nonzero i . if the given data lies within a strict subspace of Rn . de ned by R := and let 1 2 N X i=1 x(i) x(i) (3) n be the ordered eigenvalues of R. We would like to nd Q to minimize H (Q) = kx i ? Qx i k ( ) ( ) 2 2 (2) the total squared perpendicular distance of the points from the k-plane. The following result is standard. We can characterize a subspace S Rn by the projection operator Q mapping Rn onto S . : : : . 2 . Let s = rank R. so that PP = I . (Total Least Squares) Let R be the correlation matrix of the data. to nd how well it may be approximated by projection onto a k-dimensional subspace of the original n-dimensional state space. and h i . Then n X min H ( Q ) = i Q i=n?k+1 (4) where the minimum is over all rank k projections Q. s be orthonormal eigenvectors of R. and y = Px is a representation in terms of the new coordinates i on S . Theorem 1. The projected approximant to x is given by P Px 2 S . The Karhunen-Loeve decomposiN X i=1 tion provides a method for nding this best approximating subspace. j i = ij . k . j i. In general. R may not have rank n. consisting of sampled measurements fx(1) . The optimal k-dimensional subspace approximant is given by x ^(i) = k X j =1 aij j (6) Denote by P the k n matrix whose rows are 1 . Each x(i) can be written as x(i) = s X j =1 aij j (5) where aij = hx(i) . The Karhunen-Loeve decomposition. and let 1 . : : : . This subspace approximant is 5 . The next step is to perform a principal component analysis of this data. : : : .

we would expect that the more eigenvectors we keep. to allow a ne variation we make use of the fact that the optimal a ne subspace passes through the mean of the data. 6 . by projecting the vector eld onto the tangent space of S . The eigenvalues of R now 1 where x = N i=1 provide us with information as to how close an approximation of the data is provided by a k-dimensional subspace.then optimal. If we keep all of the eigenvectors corresponding to non-zero eigenvalues. Using the coordinates de ned previously. de ning e = x ? x to arrive at e _ (t) = f (e(t)+x). We would like to approximate e(t) in this equation by P Pe(t) 2 S . The idea is to replace the given dynamics by an associated dynamics on a kdimensional subspace S Rn of the original state space. we construct the correlation matrix as Rij = P N X p=1 (p) p) (x( i ? xi )(xj ? xj ) (8) N x(i) is the mean of the data. Clearly the subspace obtained by the Karhunen-Loeve method depends upon the inner product on the Euclidean space in which the state space is embedded. Hence. Clearly this will not always be possible. the goal is to choose k such that the fraction of the total `energy' in the subspace k X i=1 i n X i=1 i (9) is close to one. then this subspace will contain all of the dynamics seen in the experiment. Galerkin projection. the resulting reduced order approximation is given by y _ (t) = Pf (P y(t) + x) (10) The Karhunen-Loeve method therefore projects the dynamics onto the subspace containing most of the `statistical energy' of the system. We now subtract the mean x from the dynamic equation. The above procedure nds the optimal subspace. yet k is su ciently small. In general. with models which are better approximated by low-dimensional systems having relatively few large eigenvalues. the better approximation we will obtain. in the sense that the total `energy' (2-norm) in the subspace is given by N X i=1 kPx i k = ( ) 2 2 k X j =1 j (7) and this is the maximum achievable by any k-plane. It can also be used to construct lower order mathematical models of a given dynamical system. The Galerkin projection has been used extensively to construct numerical solutions to partial di erential equations. Model reduction.

and the function z as an output signal. For control. That is. and z (t) 2 Rq . taking explicit account of the input-output connection of the system. we would like the outputs z of the two systems to be close. The goal of model reduction is to construct another nonlinear system of di erential equations q_(t) = f^(q(t). This lends physical intuition to the procedure. despite its application to nonlinear systems. and we now proceed with this. 3 Model reduction of controlled systems We now turn to the main problem addressed in this paper. for states in some speci c region of the the state space. that of model reducing nonlinear input-output systems of the form x _ (t) = f (x(t). In this paper. but taking explicit account of the input-output connection of the system. and k < n. based on the approach used for autonomous systems. w(t)) z (t) = h(x(t)) (11) Here x(t) 2 Rn is the state of the system. We can expect such a procedure to work well for model reduction of the system within some given region of state space. and it is within such a speci c region that data should be collected. a severe limitation is that the input-output behavior of the system is so far not taken into account.physical meaning is an important indicator of the appropriate choice of inner product. before proceeding further. w(t) 2 Rp . Since the 2-norm of the states is maximized. for a class of inputs w. w(t)) (12) ^ (q(t)) z (t) = h where q(t) 2 Rk . Computational application of this method requires only standard matrix computations. 7 . Also the method has the feature that it separates the model reduction into two parts. The construction is designed to generalize standard methods from linear model reduction theory. that of nding a suitable subspace and that of performing the projection. In this section we develop an approach for model reduction of systems of the form (11). an inner product for which this norm corresponds to an approximation of the physical energy in the system may be an advantage. such that the input-output behavior of the two systems is similar. we need some preliminary de nitions from the theory of linear systems. We construct a new method. we combine the databased approach of the Karhunen-Loeve reduction procedure with the inputoutput focus of balanced truncation. The function w is regarded as an input signal to the system. and generalizing standard methods from linear model reduction theory. and allows alternative projection schemes to be used.

in which case Y > 0 and equation (15) has a unique solution. and hence Ox0 = CeAt x0 : Lemma 3. C : L2 0.1 Linear systems For linear systems. in which case X > 0 and equation (16) has a unique solution. 1) by y+ (t) = Ox0 . de ne the observability operator. Then. 1) ! Rn by x0 = C u?. and A. the equations (11) become x _ (t) = Ax(t) + Bw(t) z (t) = Cx(t) (13) where x(t) 2 Rn . Then Y is the smallest semipositive solution to the Lyapunov equation AY + Y A + BB = 0: (15) The system is called controllable if Im C = Rn . The linear system is called stable if the eigenvalues of A all have real part strictly negative. Write X := O O.3. and C are matrices of appropriate dimension. Lemma 2. the state at time zero x(0) = x0 is given by x0 = Z 0 1 eAs Bu?(s) ds: (14) This de nes the controllability operator. Both X and Y are n n matrices. and they are given by the following integral formulae. O : Rn ! L2 0. Suppose the system in equation (13) is stable. Write Y := CC . 1). Denote the future output by y+ (t) 2 L2 0. Then X is the smallest semipositive solution to the Lyapunov equation A X + XA + C C = 0: (16) The system is said to be observable if ker O = f0g. As is standard. The following standard results can be found in 20]. Y= X= Z 1 0 eAt BB eA t dt eA t C CeAt dt 8 Z 0 1 (17) . Then. The matrix X is called the observability gramian. there is a similar notion for the output. The matrix Y is called the controllability gramian. for u 2 L2 (?1. 0] and denoting u? (t) = u(?t). B .

For given initial conditions. 9 . fT1. fc1. and let M be a set of s positive constants. : : : .2 Gramians and principal component analysis The method of principal component analysis relies on the use of data to construct the correlation matrix. we are not restricted to samples. Let T n be a set of r orthogonal n n matrices. we can instead make the assumption that the initial state of the system is zero. Since we are now considering a controlled system with inputs. Let T p.3. Implicit in this analysis is that the trajectories from which data is sampled are parametrized by x0 . For theoretical purposes. cs g. en. and can simply construct the correlation matrix using the integral R= Z 0 1 (x(t) ? x)(x(t) ? x) dt (18) where x(t) is the state of the system at time t. and x is the mean state. standard unit vectors in Rn g Given a function u 2 `1. : : : . : : : . E p and M be given sets as described above. ^ of Lemma 5. For the ^ by system (11). these assumptions are satis ed for exponentially stable systems. Tr g. de ne the mean u by 1 (19) (20) Z T 1 u(t) dt u := Tlim !1 T 0 We make the standing assumptions that x. and parametrize the trajectories for principal component analysis with respect to the system input u. Further de ne E n = fe . z 2 `1 and z 2 `2 . the empirical controllability gramian Y the stable linear system x _ (t) = Ax(t)+ Bw(t) is equal to the usual controllability gramian Y . with the underlying assumption that this data is collected from `typical' system trajectories. De nition 4. For any sets T p and M. the initial state of the system. de ne the empirical controllability gramian Y ^= Y p r X s X X 2 l=1 m=1 i=1 rscm 1 Z 0 1 ilm (t) dt (21) where ilm (t) 2 R is given by ilm (t) := (xilm (t) ? xilm )(xilm (t) ? xilm ) (22) and xilm (t) is the state of the system (11) corresponding to the impulsive input w(t) = cm Tl ei (t).

For the linear system. The empirical controllability gramian is a computable generalization of this ^ correspondto nonlinear systems. and M be given sets as described above. simply studying the input-state behavior is not enough. z (t) = Cx(t) is equal to the usual observability gramian X . it has the property that the eigenvectors of Y n ing to non-zero eigenvalues span a subspace O R which contains the set of states reachable using the chosen initial impulsive inputs. one might perform a small eigenvalues of Y Galerkin projection onto the subspace spanned by the eigenvectors corresponding to the largest eigenvalues. E n. de ne the empirical observability gramian X ^= X r X s X 2 l=1 m=1 rscm 1 Z 0 1 Tl lm (t)Tl dt (25) where lm (t) 2 Rn n is given by lm (t) := (z ilm (t) ? z ilm ) (z jlm (t) ? z jlm ) ij (26) and z ilm (t) is the output of the system (11) corresponding to the initial condition x 0 = c m Tl ei . A variant of this construction was shown by Moore 28]. truncate those states corresponding to ^ . Let T n. For the ^ by system (11). The next de nition is the analogue of the previous one for the output behavior. for controlled systems. In the context of nonlinear systems. We are therefore led to a methodology for model reduction motivated by a standard idea from realization theory. the empirical observability gramian ^ of the stable linear system x X _ (t) = Ax(t) + Bw(t). De nition 6. for impulsive inputs distributed in the above sense on the unit ball in Rp . ilm (t) = c2 (eAt BTl ei )(eAt BTl ei ) m 2 At = cm e BTl ei ei Tl B eA t (23) hence ^= Y = Z 0 p r X s X 1X 1 eAt BT e e T B eA t dt l i i l s 1 1X eAt BB eA t dt = Y m=1 s l=1 m=1 i=1 Z 0 rs (24) which is the desired result. However. 10 . For any nonempty sets T n and M.Proof. Lemma 7. the Karhunen-Loeve decomposition of the states leads to a construction of the controllability gramian for linear systems.

in which case de ning y = Tx leads to 11 . Rather than searching for exact controllability and observability submanifolds within the state space. 3. equal to the state space dimension of the system. all that is required is the solution of standard linear matrix eigenvalue problems. We may change state coordinates via any non-singular linear transformation T without a ecting the input-output behavior. We now have the tools we need for empirical analysis of the input-output behavior of the nonlinear system. The balanced realization gives a way to nd which particular states correspond to which Hankel singular value. the Hankel operator. Denoting as before u? (t) = u(?t). The Hankel operator of a linear system is the operator which maps (the re ection of) past inputs to future outputs. These are the eigenvalues of OCC O . we now have two important subspaces of the state space. The advantage of this approach lies in the computation. and make use of the ideas of balanced realization theory to decide on which subspace to project. and their corresponding eigenvalues. which are the same as those of the n n matrix O OCC = XY . is given by the map ? := OC so that y+ = ?u? . called the Hankel singular values.Proof. ?. A singular value decomposition of this linear operator leads to n singular values. Unlike in the autonomous case. assuming the future input is zero. hence and ^= X lm (t) = c2 (CeAt Tl ei ) (CeAt Tl ej ) ij m At At = c2 m ei Tl e C Ce Tl ej lm (t) = c2 T eA t C m l s X m=1 (27) (28) (29) CeAt Tl 1 Z 1 eA t C CeAt dt = X s 0 which is the desired result.3 Model reduction via balanced truncation In this section we give a brief overview of the method of balanced truncation for linear systems. and are a property of the system as an input-output map. it is immediate that the Hankel operator has at most rank n. our approach is to search for subspaces which approximate these manifolds. which are independent of the particular state space realization of the system. For the linear system. Since x0 = C u? and y+ = Ox0 . We can proceed in the same manner as for linear systems. The Hankel singular values are the eigenvalues of ?? . We therefore return to a description of linear systems theory.

with respect to the inputs and outputs of the system. and perform a 12 3. TB. there will always exist a nonsingular state transformation T such that the new realization (TAT ?1. controllable. The procedure is to truncate those states from the balanced realization corresponding to small Hankel singular values i . We propose to use these for model reduction of nonlinear systems in the same way as for linear systems. C ) is stable.y _ = TAT ?1y + TBw z = CT ?1 y: Y ?! TY T X ?! T ?1 XT ?1 (30) The above Lyapunov equations (15) and (16) imply that the gramians transform correspondingly according to (31) Even though the choice of T does not a ect the Hankel singular values. De nition 8. The balanced realization gives a way to nd which particular states correspond to which Hankel singular value. This leads to a method of model reduction known as balanced truncation. the truncated system is given by y _ (t) = PAP y(t) + PBu(t) z (t) = CP y(t) (32) The reduced-order model constructed via balanced truncation is independent of the original state space realization. introduced by Moore 28] in the context of realization theory. B. where P = I 0 : If (A. C ) is balanced. this is equivalent to applying a Galerkin projection to the balanced realization. If the states are ordered according to decreasing singular value. it does a ect the eigenvalues of X and Y separately. C ) is called balanced if the controllability gramian Y and the observability gramian X are equal and diagonal. In this case. CT ?1) is balanced. nd a linear change of coordinates such that the empirical gramians are balanced. B. The Hankel singular values correspond to states through which the input is transmitted to the output.4 Empirical balanced truncation . which are the eigenvalues of the product XY . The realization (A. We assume throughout the following that the system (A. Suppose A is stable. The empirical gramians give a quantitative method for deciding upon the importance of particular subspaces of the state space. B. and observable.

and ^ 2 value decomposition of Z XZ . w(t)) (33) z (t) = h(T ?1P y(t)) This is the nal reduced order model. and results in a new nonlinear model. X Since for linear systems. and increase the allowable sample time. Then T Y ? 1 ^ ?1 T XT = . First. Let U 2U be a singular 1 ? 1 ^ T = . although in the case of uid dynamics. These conditions may be satis ed for certain mechanical systems. and is directly computable from data. a crucial feature of this procedure is that in general we should expect it to work well only in limited regions of state space. The choice is motivated by the magnitudes of the inputs and states actually seen in experiments. as desired. since this corresponds to using both positive and negative inputs (or initial states) on each input separately. the empirical gramians are exactly the usual gramians. with Z lower triangular with non-negative diagonal entries. All that is needed is numerical approximation of the above integrals. that is TY T = T ?1 XT ?1 = . Let T be the change of coordinates such that the system is balanced. ?I g. We do not expect to be able to approximate global behavior of the system well using a linear projection. A reasonable simple choice is T = fI. It is also possible to take advantage of knowledge of f to calculate x _ (t) in this computation to achieve higher order accuracy. using sampled-data from both x(t) and z (t). it is perhaps more reasonable to expect to be able to construct low order models for systems within some given operating region. To perform the required experiments. The sets M specify the size of inputs and states we are interested in. As for Karhunen-Loeve reduction of autonomous systems. we are forced to rely on data from simulation. apply Cholesky factorization 15] to Y so that Y = ZZ . We also need to choose the sets T and M. and let P = I 0 be the k n projection matrix. When applied to a nonlinear system. however. 4 Computation We can directly apply De nitions 4 and 6 to nonlinear systems for construction of the empirical gramians from data. larger sets may be justi ed.Galerkin projection onto the states corresponding to the largest eigenvalues of ^Y ^. 13 . The data may be taken either from simulation or from experiment. for example. this method is exactly balanced truncation when it is applied to a linear system. For distributed actuators. it requires only matrix computations. This empirical balanced truncation gives a reduced order model which takes into account the input-output behavior. Applying the previous analysis leads to a reduced order model given by y _ (t) = PTf (T ?1P y(t). and to apply approximately impulsive inputs. it is necessary to be able to set the initial state of the system. and let T = U Z . to measure the state at all times. ^ and Y ^ A simple numerical technique for balancing the empirical gramians X ^ ^ is as follows.

a torque about the lowest pin joint.We can now change state coordinates of the nonlinear system. the horizontal displacement of the end of the last rod from the vertical symmetry axis. The system has a single input. The potential energy V of this system is 1 k( 1 )2 + 1 k X( i ? i?1 )2 V=2 2 i=2 and the kinetic energy T is 1 T=2 n X n ml2 ( _i )2 + 1 X ((x _ i )2 + (y _ i )2 ) n (34) i=1 2 2 i=1 (35) 14 . It consists of ve rigid rods in the plane. mildly nonlinear mechanical system. and a single output. and truncate using the Galerkin projection. connected via torsional springs and dampers. The lowest rod is pinned to ground with a torsional spring. and in this section we give examples for a simple. so that the system has a stable equilibrium in the upright vertical position. 5 Mechanical links example Not only are high-order systems hard to control. These mode shapes often are physically meaningful. but it is di cult to develop intuition as to their behavior. One of the advantages of the procedure developed here is that it can be viewed as a selection of appropriate mode shapes on which to project the nonlinear dynamics. θ 5 θ 4 θ 2l 3 θ 2 θ 1 Figure 1: Example mechanical system The system is shown in Figure 1. The rows of T may be thought of as giving the `modes' of the system associated with the Hankel singular values.

n. _) := ?2l n X i=1 sin i (40) The constants are given by b = 0:5. and the equations of motion are then where F is the forcing term containing dissipative forces and the external force term w. and n = 5. We rst analyze the linearization of the system about its stable equilibrium. k = 3.493 0.010 0.where xi . l = 1. The singular values have been normalized so that they sum to one. with the corresponding singular value of each mode. if i = 2.1 Linearized model 0. Figure 2 shows the con guration part of the rst four of the ten mode shapes and the corresponding singular values of the balanced realization. given by i xi = ?l sin i i?1 sin p ?l sin ? 2l Pp =1 ( i yi = l cos i Pi?1 p l cos + 2l p =1 cos ( if i = 1. The measurement equation is d @L @L dt @ _i ? @ i = Fi (38) z = h( . : : : . yi are the Cartesian coordinates of the center of mass of the i'th rod. : : : .009 Figure 2: The balanced modes of the linearization. if i = 2. (39) ?b( ? ) if i = 2. if i = 1. ( _i Fi = ?b _i+ w_i?1 if i = 1. : : : . n.485 0. n. 5. (36) (37) The Lagrangian L = T ? V . m = 1. 15 .

In this case. The dynamics of a mechanical system also satisfy a variational principle. this is a consequence of the underlying correspondence 16 . Lagrangian systems conserve energy as well as quantities associated with the symmetries of the system.535 0.069 0.5. We choose M = f0:4g and T = fI g for both controllability and observability gramians. where each mode shape appears twice.2 Empirical balancing Application of the above empirical model reduction procedure to the nonlinear Lagrangian model of this system leads to a set of corresponding modes.523 0.146 0. Methods for model reduction which take account of this underlying geometry and preserve it are developed in 21. for ci = 0:4.264 0. Since this system is mechanical.387 0. so we are concentrating on the dynamics resulting from larger force inputs. All of these properties can be viewed as fundamental to a model of a mechanical system. the dynamics have a Lagrangian structure. 24].023 Figure 4: The empirical balanced modes of the nonlinear model. Figure 4 shows the mode shapes using M = f1g and T = fI g for both controllability and observability gramians. In the absence of forcing and dissipation. 0. and the evolution maps are symplectic transformations.011 Figure 3: The empirical balanced modes of the nonlinear model. for ci = 1. The importance of this is evidenced by Figure 2. we can see that the rst two mode shapes have split into three. Here the mode shapes become signi cantly di erent from those of the linearization. Figure 3 shows these modes. 0.

Many other methods might be suggested. 24] and constructing the empirical gramians on the con guration space rather than the phase space. and may be applied directly to nonlinear systems. There are several other possibilities that might be suggested. The choice of impulsive inputs. the underlying geometric structure is preserved by the reduction. Similarly. and the resulting reduced-order system is itself a Lagrangian system. leading to reduced computational requirements and producing more accurate reduced-order models. and initial states distributed uniformly on the unit sphere. for example vortex methods are often used to simulate incompressible ow. 35] and other special structure present in the original system can be of great use. 17 . For mechanical systems. the reduced order model constructed is exactly the usual balanced truncation of the system. Acknowledgments We thank Petr Krysl. with the control engineering methodology of balanced truncation. The resulting method of model reduction is based on data. such as subspaceidenti cation procedures to compute both empirical gramians simultaneously. these heuristics can be incorporated into the control design process. 6 Conclusions In this paper we have presented new methods for model reduction of nonlinear systems. as they have been applied to analysis of complex ows. our focus here has been the construction of a computable extension of balanced truncation for nonlinear systems. and for control a very reasonable strategy would be to apply a Karhunen-Loeve decomposition of the states seen during simulation of closed-loop behavior. It requires only simple matrix computations. and Clancy Rowley for helpful comments and inspiration. Peter Schroder. In this way. One advantage of the empirical procedure described in this paper is that often the simulation itself has been implemented using known heuristics. are perhaps somewhat arbitrary methods of parametrizing trajectories for nonlinear systems. taking account of symmetry 13.between con guration variables and their generalized momenta. Using data from such a simulation for model reduction. EPRI/DOD CIN/SI project WO8333-06 and NSF/DARPA OPAAL grant DMS-9874082. These methods combine the physical intuition of Karhunen-Loeve techniques. This research was partially supported by AFOSR MURI grant F49620-96-1-0471. When applied to linear systems. often the inputs are forces which appear in a structured way in the dynamics. and for these systems an improved reduced-order model can be achieved by combining the techniques in this paper with those of 21.

and K. Al-Saggaf. and Hamiltonian systems. and R. A. Estimation. M. E. Glover. L. PhD thesis. Systems and Control Letters. December 1999. and the Karhunen-Loeve expansion. King. 11] D. 2000. 6] C. 41:1466{1477. October 1999. In Proceedings of the IEEE Conference on Decision and Control. Using Karhunen-Loeve decomposition to analyze the vibro-impact response of a rotor. Dullerud and F. A Course in Robust Control Theory: a Convex Approach. L. 6:103{108. In Proceedings of the IEEE Conference on Decision and Control. International Journal of Systems Science. 25:1631{1642. Computational aspects of reduced order feedback controllers for spatially distributed systems. IEEE Transactions on Automatic Control. New model reduction scheme for bilinear systems. 18 . Annual Review of Fluid Mechanics. Atwell and B. 10] G. In Seventh Conference on Nonlinear Vibrations. 1994. California Institute of Technology. P. Glavaski. Lumley. Vakakis. 44:1802{ 1813. M. M. centering. In Proceedings of the IEEE Conference on Decision and Control. 25:539{575. F. E. The proper orthogonal decomposition in the analysis of turbulent ows. 9] A. and U. L. Model reduction. 1998. A. 1993. Model reduction of multidimensional and uncertain systems. 1:179{190. 5] C. 8] A.A. C. A new model reduction scheme for kpower bilinear systems. volume 37. Bettayeb. J. Marsden. A. Model reduction for control system design. VA. Paganini.References 1] S. PhD thesis. and J. pages 4301{4306. October 1996. pages 2071{2076. Stability. Theoretical and Computational Fluid Mechanics. Beck and J. C. Murray. E. Beck. IEEE Transactions on Automatic Control. 1989. Bettayeb. F. 7] G. Azeez and A. Bloch. Al-Baiyat and M. July 1998. Glavaski. 1984. M. Doyle. Blacksburg. 1993. Doyle. Stanford university. F. and Dynamics of Structures. Berkooz.B. Marsden. 4] M. 2] S. J. Controlling homoclinic orbits. 3] J. Al-Baiyat. May 1998. 1985. Enns. A necessary and su cient minimality condition for uncertain systems. Springer-Verlag. Robust system analysis and nonlinear system model reduction. M. Holmes. 13] S. Bloch and J. principal components. 12] S.

J. volume F. Van Loan. 1967. Lall. 19] P.14] K. S. Lumley. Model reduction of complex systems in the linearfractional framework. Lumley. Lall. E. 1996. Turbulence. 25] S. 22] S. 15] G. and symmetry.F. 21] P. 1979. Yaglom and V. Empirical model reduction of controlled nonlinear systems. M. Atmospheric turbulence and wave propagation. Prentice-Hall. 16] W. Moscow. Edgar. Gray and J. Marsden. Johns Hopkins University Press. 27] B. submitted to IEEE Transactions on Automatic Control. Principal component analysis in nonlinear systems: preliminary results. Moore. Mesko. In Proceedings of the American Control Conference. F. 18] J. 39:1115{1193. Energy functions and algebraic gramians for bilinear systems. All optimal Hankel-norm approximations of linear multivariable systems and their L-in nity error bounds. Krysl. coherent structures. International Journal for Numerical Methods in Engineering. 23] S. S. pages 473{478. Marsden. Berkooz. pages 1057{1060. P. Structure-preserving model reduction of mechanical systems. 1997. S. E. 24] S. General input balancing and model reduction for linear and nonlinear systems. Error bounds for model reduction via balanced truncation of linear time-varying systems. 1996. In A. 1999. Tatarski. Beck. pages 166{178. 1980. Dimensional model reduction in nonlinear nite-element dynamics of solids and structures. Krysl. In Proceedings of the 4th IFAC Nonlinear Control Systems Design Symposium. Matrix Computations. Glover. Marsden. In Proceedings of the IEEE Conference on Decision and Control. Lall and C. Nauka. In preparation. pages 2864{2868. Glavaski. December 2000. In IEEE Conference on Computer-Aided Control System Design. In Proceedings of the IFAC World Congress. Lall. L. 1984. Cambridge University Press. J. Linear Systems. 19 . H. E. Kailath. and J. International Journal of Control. Golub and C. Beck. L. 1998. Lall and C. Mesko. To appear. Holmes. and J. Reduction of nonlinear models using balancing of empirical gramians and galerkin projections. and S. and G. The structure of inhomogeneous turbulence. dynamical systems. Gray and J. Hahn and T . 20] T. I. 17] W. C. editors. 26] J. In Proceedings of the European Control Conference. July 1999.

J. 142:1{ 19. J. 32] K. Petzold. 2:609{629. Subspacebased signal analysis using singular value decomposition. Model reduction for chemical kinetics: An optimization approach. Pearson. parts I-III. Scherpen. International Journal of Robust and Nonlinear Control. J. Computation for nonlinear balancing. D. A. February 1981. 29] A. Balancing for nonlinear systems. 31] L. A. 2000. December 1994. Marsden. September 1993. and A. M. 37] J. International Journal of Robust and Nonlinear Control. August 1993. 1996. and model reduction. E. 7:475{505. Swindlehurst. AiChe Journal. 20 . Newman and P. Philosophical magazine. On lines and planes of closest t to points in space. October 1987. 1998. 2000. van der Schaft. M. NJ. Coherent structures and chaos: a model problem. Rodriguez. 33] L. Princeton. Petzold and W. 45(4):869{886. Scherpen. Newman and P. 120(5):211{214. 38] J. S. Principal component analysis in linear systems: controllability. E. 35] C. Sirovich. Reconstruction equations and the Karhunen-Loeve expansion for systems with symmetry. Zhu. Controller reduction for nonlinear plants { an L2 approach. 6:645{668. In Proceedings of the 32nd Conference on Information Sciences and Systems. February 1987. Quarterly of Applied Mathematics. 21(2):143{153. 45(3):561{590. Physics Letters A. A.28] B. 41] A-J. S. In Proceedings of the IEEE Conference on Decision and Control. W. Normalized coprime factorizations and balancing for unstable nonlinear-systems. Physica D. Pavel and F. W. In Proceedings of the IEEE Conference on Decision and Control. 1997. M. International Journal of Control. Sirovich and J. 60(6):1193{1222. Moore. F. Krishnaprasad. 81(9):1277{1308. L. Systems and Control Letters. H-in nity balancing for nonlinear systems. observability. 36] J. Rathinam and L. C. van der Veen. Nonlinear model reduction for RTCVD. Deprettere. 30] A. IEEE Transactions on Automatic Control. March 1998. 40] L. Proceedings of the IEEE. Scherpen and A. Turbulence and dynamics of coherent structures. Fairman. Rowley and J. 39] L. 1901. 34] M. An iterative method for simulation of largescale modular systems using reduced order models. 26(1):17{32. April 1999. Krishnaprasad.

In Proceedings of the IEEE Conference on Decision and Control. H. Packard. Becker. J.42] G. Induced L2 -norm control for lpv systems with bounded parameter variation rates. Glover. and K. Yang. pages 406{411. Wood. and G. Approximation of linear parameter-varying systems. X. 43] F. P. Goddard. A. D. 6(9{10):983{998. International Journal of Robust and Nonlinear Control. 21 . November 1996. 1996. Wu.