See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/369053779

Application of Generalized Functions in Optimal Control

Conference Paper · March 2023

CITATION READS

1 82

1 author:

Erik Verriest
Georgia Institute of Technology
307 PUBLICATIONS 4,250 CITATIONS

SEE PROFILE

All content following this page was uploaded by Erik Verriest on 07 March 2023.

The user has requested enhancement of the downloaded file.

 Application of Generalized Functions in
Optimal Control
Erik I. Verriest ∗
∗
Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail:
erik.verriest@ece.gatech.edu).

Abstract: A new spin is given on the classical optimal control problem with piecewise
differentiable dynamics and performance index with respect to the state variables. While in
each domain of differentiability, the necessary conditions for optimality are easily established,
their interpretation at the boundaries between domains is not well-understood. In this paper
we show that in order to make sense of the Euler-Lagrange equation at this interface one needs
to transcend the classical theory of Schwartz distributions and make suitable extensions to
allow for the questionable behavior of impulses multiplied by discontinuities, and the notion
of partial derivatives at a discontinuity. Such a theory has been developed, in the Colombeau,
Oberguggenberger and Rosinger theory of Generalized Functions in 1990, going back to ideas
from Nonstandard Analysis (NSA). We develop an alternative NSA based approach applicable
to impulsive dynamics and optimal control.

Keywords: Generalized Functions, Optimal Control

1. INTRODUCTION The initial condition, x0 , provides n initial conditions

for these equations. Augmenting with the n unknown
At the Boulder AMS Conference in 1990, R. Hermann parameters, λ0 = λ− (0), the solution of these 2n ODE’s
presented a talk on the Colombeau, Oberguggenberger can be specified in parameterized form as x− (t; λ0 ), with
and Rosinger theory of generalized functions and predicted x− (0; λ0 ) = x0 , for the states, and the co-states λ− (t; λ0 ).
that it would revolutionize mathematical physics and Likewise, in the domain Ω+ , the necessary conditions
system theory in the 21-st century (Hermann (1994)). This lead to the 2n coupled ODE’s, parameterized by say
has not happened, perhaps because the details involved λ+ (tf ) = λf , which are as yet unknown. This leads to
on the extended Schwartz distribution theory are not very the parameterized form of the n states x+ (t) = x(t; λf )
transparent. In this paper, the theory is revisited, and it and the n co-states λ+ (t; λf ).
is shown that it has indeed a distinctive merit. This yields two families of trajectories: One starting from
x0 and parameterized by n parameters in λ0 . The other
Assume that the state space, for simplicity embedded in family, x+ (t; λf ), ends at xf and is likewise parameterized
Rn , is partitioned in two domains, Ω+ and Ω− , separated by the n parameters in λf . Continuity of the state at
by the interface Ω0 = { x | g(x) = 0 }. A standard opti- the interface requires that for some time τ , x− (τ ; λ0 ) =
mal control problem is to find the control and trajectory x+ (τ ; λf ), which provides n equations, but introduces an-
for a system governed by the dynamics ẋ = f (x, u), in other variable: τ . However, at the interface xi = x− (τ ) =
going from an initial state, x0 , to a desired final state, x+ (τ ) ∈ Ω0 the equation g(x− (τ )) = 0 determines this τ in
xf , while minimizing some performance index, say J = principle. This still leaves n unknowns in the problem. One
RT
0 L(x, u) dt, with T fixed for simplicity. In the hybrid
solves for these n remaining parameters by optimization.
problem, L and f are smooth functions of the arguments But, a close inspection of the problem reveals that the
in each domain, but are discontinuous at Ω0 . The optimal information present in the Euler-Lagrange equation of the
trajectory from a given state x0 ∈ Ω− to a given final state optimal problem has not been exhausted. By separately
xf ∈ Ω+ must cross the interface an odd number of times. solving the EL equation in both domains, the EL equation
Let’s focus here on the case having one such crossing. across the interface has not been used. Since λ− and λ+
Transversality requires that [∇gf− ]t− [∇gf+ ]t+ > 0, where may differ on g(x) = 0, it means that the derivative of
>
t− = ti − ǫ and t+ = ti + ǫ for ǫ → 0. Crossing time and λ must be impulsive across the interface. Thus the right-
crossing state are a priori unknown. hand-side of the EL is impulsive as well. But there are
Classically, the solution is found stepwise: First, establish several problems with this. The right-hand side of the EL
the necessary conditions for optimality in Ω− : The opti- equation is specified as the gradient with respect to x. If x
mality condition relates u− to x− and λ− implicitly. In were smooth, Schwartz’s distribution theory tells us how
principle one can express u explicitly so that the state to relate the impulse with argument x to the impulse with
argument t, but here ẋ is not continuous across the inter-
equation and the Euler-Lagrange equation (EL), λ̇− = face. Likewise the right hand side also involves a product
 ⊤
− ∂H ∂x
−
, where H− (x, u, λ) = L(x, u) + λ⊤ f is the of distributions with respect to x with functions that are
Hamiltonian, leaves 2n coupled ODE’s for x− and λ− . necessarily discontinuous. Such objects are not allowed in
Schwartz’s distribution theory. This is the essence of the f = g a.e.. In contrast, the evaluation functional, R σt :
Schwartz impossibility theorem, relating to the fact that C0∞ → R : φ 7→ φ(t), is usually denoted as R δt (x)φ(x) dx,
the set of distributions does not have the structure of an however the δt known as Dirac-delta’s are not functions.
algebra. See Oberguggenberger (1992). Consequently, the For D ∈ D′ , the distributional derivative is defined by
needed interpretation of the EL completely fails in the D′ : D′ → D′ : D → (φ 7→ −D(φ′ )). For f˜ ∈ C 1 ∩ L1,loc ,
classical Schwartz sense at the interface. we get (f g ′ ) = (f˜)′ . For smooth functions f ∈ C ∞ and
D ∈ D , we define f D : C0∞ → R : φ 7→ D(f φ), and
′

2. KRYLOV-SPACE note that f D ∈ D′ . Furthermore, (f D)′ = f ′ D + f D′ , and

∀g ∈ L1,loc , f g̃ = (f g g).
The way out of this impasse is to use the framework The support of a distribution is the complement of the
of generalized functions with a well-defined multiplicative largest open set on which the distribution vanishes. Let
′
structure, such as outlined in Grosser et al (2001). We pro- DM = {D ∈ D′ supp D ⊆ M } for any measurable subset
pose a new definition of generalized functions, one that can of R. Let DM denote the restriction of the distribution D.
handle the interpretative problems of δ(x(t)) and φ(t)δ(t) Desirable properties for a restriction are:
respectively when x is not differentiable and φ is not con- i) ∀D ∈ D′ and measurable M , DM ⊆ Dcl ′
M , and DM
tinuous. Non-standard analysis (NSA) provides an answer, linear and idempotent.
but seems more complicated than it needs to be (Goldblatt ii) ∀f ∈ L1,loc and measurable M , if fM = χM f , with χM
(1998). In previous work Hyun and Verriest (2016, 2017) the indicator function of M , then ff ˜
M = (f )M .
we developed an axiomatic approach from the ground up. iii) ∀φ ∈ C0∞ , ∀D ∈ D′ and measurable M it holds that
This was introduced for the purpose of placing causality, supp φ ⊆ M ⇒ DM (φ) = D(φ) and supp φ ∩ M = ∅ ⇒
understood in the sense of a cause leading to an effect, DM (φ) = 0.
back in hybrid system theory. While in the linear theory, iv) Mi P pairwise disjoint with M = ∪M1 , it holds that
the reachability problem is investigated using impulsive DM = DMi and (DMi )Mj = 0 if i 6= j.
inputs (the cause) creating jumps in the state (the effect), If supp D = {t}, then there exists N ∈ N, and αi ∈ R;
in the nonlinear case hybrid systems are modeled at their PN (i) PN (i)
jumps solely by the effects, e.g., x(t+ ) = φ(x(t− ), p), where i = 0, . . . , n such that D = i=0 αi δt and i=0 αi δt =
the p may be some control parameters. In the cited thesis, 0 ⇔ αi = 0, for i = 0, . . . , N .
the objective was to put the impulses back in the continu- As the conditions (i) to (iv) cannot be satisfied together,
ous dynamical equation, but properly model the dynamic Trenn (2009) defined an appropriate subspace of piecewise
behavior in our sense. This theory borrowed also from the regular distributions that allows (i) to (iv) to be satisfied.
X
generalized function theory of Colombeau (1985), which Dpw′ ˜ Dt | f ∈ L1,loc , T ⊂ R, locally finite },
reg = {f +
has found widely acceptance in mathematical physics. See
t∈T
also Oberguggenberger (1992).
′
Within our post-Schwartzian generalized function theory, where Dt ∈ D{t} . It is shown in (Trenn (2009)) that
we were able to obtain a clear understanding of this sin- ′ ′
Dpw reg ⊂ D , the representation is unique, (i) to (iv) are
gular behavior of the EL equation λ̇ = −Hx⊤ , across the P
satisfied, and for measurable M ⊂ R, D = f˜ + T Dt
interface, which we called the generalized Euler-Lagrange P
implies DM = ff M + T χM (t) dt.
equation (GEL). The results then lead to a precise formula ′
For D ∈ Dpw PD[t] = D{t} as Dt if t ∈ T and
reg , define
for the jump ∆λ in the co-states across the interface
g(x) = 0 in terms of the co-states just before and just zero else. Then D[·] = T Dt is the impulsive part of the
after the Ω0 crossing (Zhou and Verriest (2022)). The distribution D.
upshot of this is now that the GEL provides n additional Multiplication with piecewise smooth functions is well-
equations precisely in the remaining parameters. Hence the defined (Trenn (2009)).
GEL can be used instead of solving the classical param- The class of piecewise smooth distributions is then defined
eter optimization problem. Solving nonlinear equations is and with the differentiation and multiplication defines
computationally more straightforward. This information an associative differential algebra. In Trenn (2009) it
has not been exploited in problems involving singularities was argued that χ ^[t,∞) δt = δt for all t corresponds with
in optimal control before. a causality condition, rendering the Cauchy initial value
problem of an algebraic differential equation unique. While
successful in switched behaviors with impulses (Trenn and
3. SCHWARTZ DISTRIBUTION THEORY AND Willems (2012)), the application of this multiplication to
BEYOND the Euler-Lagrange equation in multi-mode optimal con-
trol is inconsistent with the solution method that com-
The Schwartz distribution theory centers on D = C0∞ (R), putes the set of optimal solutions in each domain, param-
the space of C ∞ functions with compact support. The eterized by the interface condition, and then determines
space of distributions is then defined as the dual space D′ . the optimal parameters. This stems from the fact that
A distribution f˜ ∈ D′ is regular if there exists f ∈ L1,loc , the Euler-Lagrange and state equations constitutes a two-
the space of locally integrable functions, such that point boundary value problem and not an initial value
Z problem. Hence causality is not a concern, and the so-
˜
∀φ ∈ D, f (φ) = f (x)φ(x) dx. lutions in the above extension of Dpw ′
reg will not solve the
R problem. The way out is to enlarge the set of distributions
′
Let Dreg denote the space of regular distributions. Then in the nonstandard sense. Indeed, Todorov (1990) proved
Dreg ⊂ D′ and ∀f, g ∈ L1,loc it holds that f˜ = g̃ ⇔
′
the existence of a nonstandard function ∗ δ ∈ ∗ C0∞ , such
that for all φ ∈ C 0 valued functions of the form hh, Sα h, Sα2 h, . . .i denoted
Z
∗
as hSα : hi, where h : R → R and Sα is the scaling
δ(x) ∗φ(x) dx = φ(0). operator, σt (Sα x) = σαt x, and take the closure under
∗R
componentwise addition and multiplication
This means that pointwise evaluation of a generalized
distribution is possible. The next section uses some notions
of non-standard analysis, see Goldblatt (1998). Essentially, hSα ; f i + hSβ ; gi = hf + g, Sα f + Sβ g, Sα2 f + Sβ2 g, . . . i
just as Cantor’s description of the reals is done by consid- hSα ; f i · hSβ , gi = hf g, Sα f · Sβ g, Sα2 f · Sβ2 g, . . . i.
ering equivalence classes of Cauchy sequences of rationals, Let σt be the evaluation functional on the class of regular
the hyperreals or non-standard reals are definable as equiv- functions ∀t ∈ R : σt h = h(t). The evaluation of a
alence classes (under a different equivalence relation) of hyperfunction hhn i at t∗ = hγ; ti ∈ ∗ R is the hyperreal
sequences of reals. A sequence in RN will be denoted by hh1 (t), h2 (γt), h3 (γ 2 t), . . .i The Krylov space is the space of
{rn }n or {rn }. (equivalence classes) of such functions. A regular function,
h is embedded in the space generated by Krylov sequences
4. KRYLOV SPACE as the class of h1; hi. The idea is that a function of the
form f ∗ = h1; f i + hSα ; gi is a hyperfunction converging to
We review the formal construction in Hyun and Verriest f in a precise way determined by the rate α and shape g.
(2017). The above defined hyperreal valued function is said to be
in the halo of f , while f is the shadow of f ∗ . More gen-
4.1 Krylov Hyperreals erally, we have σhβ;ti hf0 , f1 , f2 , . . .i = hf0 (t), f1 (βt), . . .i,
and σh1;t0 i+hβ;s0 i h1; f i = hf (t0 + s0 ), f (t0 + βs0 ), f (t0 +
The set of hyperreals, ∗ R, defined in NSA is too large to
βs20 ), . . .i. The right limit of a regular function at t0 is the
be easily manageable. The redundancy of infinitesimals
shadow, f (t0 +), if s0 > 0 and 0 < β < 1.
prevents the space from having a constructive property
such as having a countable basis generating the hyperreals
as a vector space over a field, R. Therefore, we construct Consider now u ∈ C 1 (−1, 1)), with limt→−1 u(t) = 0,
a countably infinite basis, which generates a reduced limt→1 u(t) = 1, and extend it to a C 0 (R) function, h,
extension of the reals, denoted as K. with h(t) = 0 for t < −1 and h(t) = 1 for t > 1.
Then h is a model for the Heaviside function, which
Definition 1: Let α > 1. A sequence hrn i is called a Ki -
is defined as the generalized function Hh = hSα ; hi. If
sequence if there exists s ∈ R such that rn = ( α1in )s for all
t ∈ R, then Hh (t) is the classical Heaviside function for
n ∈ N.
all h. Evaluation at the infinitesimal time h α1 ; ti gives
This geometric form of the sequence is reminiscent of linear
algebra where it would be called a Krylov sequence. We’ll σh 1 ;ti hSα ; hi = hh(t), . . . , h(t), . . .i = h1; h(t)i. On a micro-
α
denote it by h α1i ; si. A set {x ∈ R | x is a Ki −sequence} is scopic scale, there are infinitely many Heaviside functions.
called Ki space. Thus x ∈ Ki means there exists s ∈ R Consequently, Hh2 and Hh differ in insensible time, as they
such that x = h α1i ; si. respectively give h2 (t) and h(t).
Ki is a one-dimensional vector space over R with property Similarly we can define Hh+ by letting h(t) = 0 for t ∈
K0 = R, Ki ⊂ hal(0), ∀i ∈ N \ {0}, (−1, 0] and h(t) = 1 for t ∈ [0, 1). The derivative of Hh+ is
where hal(0) = { b ∈ ∗ R|b is infinitesimal } is the halo of represented by the sequence DHh+ = hα; 1ihSα ; Dhi. With
zero, as defined in NSA. A convenient basis for Ki is the Dh(t) = h(t) − h(t − 1), we see that in turn this can be
element ei = ( α1 )i ; 1 . With the product defined above, made continuous by a regularization δreg = hSα ; Dhireg =
we have ei · ej = ei+j for all i, j ∈ Z. With addition, the hα, 1ihSα ; h1 i with h1 = Sα (I − T−α )h. In turn, this gives
PN for Dδreg the k-th term in the sequence DSαk (I − Tα )h =
set KN = {x ∈ ∗ R | x = i=0 si ei , {sn }N i=0 ⊂ R } is the
αk Sαk (I − T−α )Dh. Taking again the regularization Dh =
direct sum vector space ⊕N i=0 KP
N
i with basis {ei }i=0 .
∞ Sα (I − t−α )h, we find (Dδreg )reg = hα2 ; 1ihSα ; Sα (I −
The set K = {x ∈ ∗ R | x = s
i=0 i ie , {s } ∞
n i=0 ∈ ℓ1 }
T−α2 )(I − T−α )hi. It is easily shown by induction that
is called the Krylov hyperreal space. In fact K is also
the regularized ℓ-th derivative of the delta corresponds to
a commutative ring, and the representation of x∗ ∈ K Qℓ
with coordinates si is unique up to equivalence in ∗ R. the generalized function hαℓ ; 1ihSα ; Sαℓ i=1 (I − T−αi )hi.
There are elements in the hyperreal space that are not
elements of K. However, the Krylov space is large enough Alternatively, let u ∈ C ∞ ((0, 1)) with again u(1−) = 1
to contain the reals and the infinitesimals. Given t∗ ∈ K and u(0+) = 0, and extend it to a Heaviside function. All
with representation {si }i ∈ ℓ1 , we shall refer to s0 as the derivatives are defined and regularization is no longer nec-
sensible time of t∗ , and sk , k > 0 as the insensible time of essary. In this case the ℓ-th derivative of Hh is represented
t∗ with 1/αk as convergence rate. by hαℓ ; 1ihSαℓ ; Dℓ hi. Its evaluation at ∗ t = h α1k ; ti is the
smooth function h(t) at level k = ℓ, and 0 for the other
4.2 Krylov Functions cases.
Definition 2: A hyperreal function is a mapping F : 5. MAIN RESULTS
∗
R → ∗ R such that there exists a sequence {fn }n ⊂ RN an
def
t∗ ∈ ∗ R and σt∗ F = hfn (tn )i. Let f g ∈ C 0 and F (t) = f (t) + g(t)H(t). The product
The restriction of F to the Krylov space K is called a F (t)δ is not defined as a Schwartzian distribution, but does
Krylov function. We start with sequences of smooth real in the Krylov sense.
Definition 3 The impulsivity of a generalized func- Since L and f are discontinuous when g(x) = 0, the right
tion ∗f at a point t0 ∈ R is given by the integral hand side of the EL is singular (as function of x). Hence
R t0 +h α1 ;1i ∗ also the left hand side must have a singularity. However,
t −h 1 ;1i
f (t) dt = h. . . , fk ( t−t
αk
0
) dt, . . .i.
0 α the latter is with respect to the time t. It is therefore neces-
For δ represented by hα; 1ihSα h′ i, we find sary to express δ(g(x)) in terms of a delta with respect to
Z 1 * Z 1 + time. But precisely because of this dependency, x, and thus
h α ;1i
αk
∗
δ(t) dt = ..., k ′ k
α h (α t)dt, . . . also g(x) have a discontinuous derivative when g(x) = 0.
1
hα ;1i − 1
αk
Thus Theorem 2 applies. Likewise, the singularity in ∂f ∂x is
= h(1) − h(−1) = 1. multiplied by λ⊤ which has a jump when g(x) = 0. Here
It is well known that f is continuous at 0, then f (t)δ(t) = Theorem 1 applies. Combining and integrating over time
f (0)δ(t) for any representation of the Heaviside function. in an infinitesimal interval across the discontinuity yields
the result. ✷
The traditional ”two-domain” method of solving the prob-
0
Theorem 1 The impulsivity of f (t)δh (t) with f (t) ∈ Cpw lem, keeping the cross-over point and time as parameters
def to be optimized, is now replaced by a non-linear equation.
is given by hf i0 = 12 (f (0− ) + f (0+ )).
Proof: It holds that there exists g and k ∈ C 0 and some
ACKNOWLEDGEMENTS
0 < ǫ ∈ R such that f (t) = g(t) + k(t)Hh (t). Thus the
impulsivity of f (t)δh (t) follows from
The author thanks the support from the Julius-Maximilans-
Z 1
hα ;1i Universität Würzburg through the Giovanni Prodi Chair.
(g(t) + k(t)Hh (t))δh (t) dt
1
hα ;1i REFERENCES
* Z 1
+
αk
J.F. Colombeau. Elementary introduction to new gener-
= ..., g(t) + k(t)h(αk t) αk h′ (αk t)dt, . . .
− 1 alized functions. Nort-Holland Elsevier 1985.
αk
R. Goldblatt. Lectures on the Hyperreals: An Introduction
1 to Nonstandard Analysis. Graduate Texts in Mathemat-
= g(0) + k(0)(h2 (1) − h2 (−1)) = hf i0 .
2 ics, Springer, New York, 1998
M. Grosser, M. Kunzinger, M. Oberguggenberger and R.
Corollary On the sensible time scale f (t)δ(t) with f ∈ Steinbauwer.
0
Cpw is equivalent to hf i0 δ(t), and using the shift property R. Herman. C-O-R generalized functions, current algebras,
and control. Math Sci Press, 1994.
f (t)δ(t − t0 ) = hf it0 δ(t − t0 ).
N.-sP.Hyun and E.I. Verriest. Causal impact modeling
of state dependent impulsive affine systems using non-
If x ∈ C 1 with x(t0 ) = 0 and ẋ(t0 ) 6= 0, then it is known standard analy6sis. Proceedings of the 55-th IEEE
that δ(x(t) = δ(t−t 0) Conference on Decision and Control, 2016, 3024–3029.
|ẋ(t0 )| . The following theorem generalizes.
N.-s.P. Hyun, and E.I. Verriest. A causal interpretation
Theorem 2 Let x ∈ Cpw 1
(O(t0 )) where O(t0 ) is a neigh- of nonlinear impulsive systems based on non-standard
borhood of t0 , where x(t0 ) = 0. Let x be transversal, i.e., analysis. Nonlinear Analysis: Hybrid Systems 25, 138-
0) –154.
ẋ(0−)ẋ(0+) > 0, then δ(x(t)) ≈ δ(t−t
h|ẋ|it . 0 M. Nedeljkov and M. Oberguggenberger. Ordinary Differ-
Here, hxit+0 is the average 12 (x(t0 −) + x(t0 +)). ential Equations with Delta Function Terms. Publica-
Proof: Use the fact that x(t) = (t − t0 )(x− (t)H(−t) + tions de ’l Institut Mathématique. Novelle série 91 (105)
x+ (t)H(t)) in a sufficiently small neighborhood of t0 where 125–135.
x± ∈ C 1 have the same sign. ✷ M. Oberguggenberger Multiplication of distributions and
applications to partial differential equations. Pitman
The solution of the ODE ẋ(t) = Ax(t)δ(t) is constant
Research Notes in Mathematics, Vol 259. Longman
for t > 0 and t < 0. Its behavior at t = 0 is retrieved
1992.
from the sequence of equations ẋk (t) = Aαk h′ (αk txk (t)),
T. Todorov. A Nonstandard Delta Function Proceedings
and leads to x(0+) = eA x(0−), which is consistent with
of the American Mathematical Society Volume 110,
Nedeljkov and Oberguggenberger (2012), but differs from
Number 4, December 1990, pp. 1143–1144.
Trenn (2009).
S. Trenn. Regularity of distributional differential alge-
Theorem 3 The standard optimal control problem braic equations. Math. Control Signals Syst. (2009)
sketched in the Introduction, with discontinuous cost rate 21:229–264.
and dynamics w.r.t. x has a jump in the costates when the S. Trenn and J.C. Willems. Switched behaviors with
interface is crossed given by impulses - a unifying framework. Proc. 51st IEEE Con-
 ⊤ ference on Decision and Control Maui, USA, December
(∆L + hλi∆f )m ∂g
∆λ = − ⊤
, m = . 2012, pp. 3203-3208.
h|m f |i ∂x M. Zhou and E.I. Verriest. Generalized Euler-Lagrange
Proof: The Euler-Lagrange equation for the optimal con- Equation: A Challenge to Schwartz’s Distribution The-
trol problem is given by ory. Proc. American Control Conference, Atlanta, GA
⊤ June 2022.
∂L(x, u)
λ̇ = − − λ⊤ f (x, u).
∂x

View publication stats