Ugo oy
‘Then define
=a. Ade
d= Fk.) 0
tis then clear that 2 ~ u, so along our curve a)
ap
ey = 4 (40), 40.0).
BO =F (AO 4.9.9) as)
o 0 (12), with p(t) defined by (15), says that
Bey = 840), dd 0
Py —FE (40.4.9, i
an _ ab :
Finaty, = p 24.5015)
a -
(aos Ho) kd.) =0
June 1997
‘The system of equations (16), (17), (18), usualy written more
concisely as
oy 2H tp
dp dh
aH aH
ay” ou
a9)
isexactly equivalent to (10), provided that His definedasin (14).
‘We will call the funetion 4 the “contsol Hamiltonian," and re-
ferto(19) asthe control Hamiltonian formot the Euler-Lagrange
‘equations. In our view, Formula (14) isthe definition that Hamil
ton should have given for the Hamiltonian, and Equations (19)
‘are “Hamilton's equations as he should have written them.”
‘What Hamilton actually wrote was (in our notation, not his)
yan de
aap
an
4 20)
where 24q,p.) sa funtion ofp. and tale, defined by the for-
mula 24q,p.1) =(p.4)~ U(q.4.1). which resembles (14), butis
tot all the same. The cifeence is that in Hamilton’ defi
tin, qi supposed to be tated no as an independent variable,
but as a function of g, pt, defined implicitly bythe equation
x,
ae) en
This easy tose tha, the map (qyd1) > (Put) defined by
(21) can be inverted, i. i wecan solve (21) for jaa function
of q,P.1 then (20) i equivalent 0 (19), Indeed, is clear that
ae P.O = Hig, ug, p, 1), 0, where u = u(q, p, #) satisfies
80
Seat)
BoC BH, BH
ay” ay” ay 2)
2H Gi) Oe ep. nest 2 tt
Sine (ap) = Of w= ap wesce at =
along solutions of (19), and then the frst equation of (20) holds
1s well. Similarly the second equation of (20) also holds. The
converse is also easily proved.
Ttshould be clear from the above discussion that the Hamilto-
nian reformulation of the Buler-Lagrange equations in terms of
the “control Hamiltonian” is at least as natural as the classical
‘one, und pethaps even simpler. Moreover, the contol formula
tion has at least one obvious advantage, namely,
(AY) the control version of the Hamilton equations is equiva-
lento the Euler-Lagrange system under completely general con-
ditions, whereas the classical version only makes sense when the
transformation (21) canbe inverted, atleast locally, to salve ford
4s. a function ofa, P
‘We nov show that (1) isnot the only advantage ofthe contol
view over the classical one. To see this, we must take another look
at Legendre's congition (13). Since H(g, usp.) is equal t0-L(g.
2) plus a linear function of u, (13) i completely equivalent to
9 FH (dd. dd.K0.1 50.
aw
ie, 2H (a0, 40.H0.)50
og . ) (23)
Now lets (23)sideby side with he tid equation (19
oy ana 2H
ou <0 and oe <0 4
and lt us stare atthe result for afew seconds
‘These equations unmistakably suggest something! Cleary,
‘what has to be going on heresthatH must have a maximum as a
function of w So we state this a8 a conjecture
‘CONJECTURE M: besides (19) (or the equivalent form (10),
«an adiional necessary condition for optimality is that lash
tO 1), a8 afnetion of u, have a maximum at (1) foreach t
‘Notice that Conjecture M isa natural consequence of rewrit-
ing Hamilton sequations “as Hamilton should have done it" and
itis reasonable to guess that, if Hamilton had actually done it,
tnen he himself, or some other 19th century mathematician,
‘would have writen (24) and be led by ito the conjecture. On the
‘ther hand, its only by using the Hamiltonian of (14), as op-
posed to Hamilton's own form ofthe Hamiltonisn that one can
See tha she Legendre condition has odo with the sign ofthe sec
‘ond u-derivative ofa faction of u whose first derivative has 10
vanish. This function cannot be itself, because the first order
onions dono say that = 0. Nor can it be Hamilton's Ha
niltonian + which isn’t even a function of u, Only the use of the
“control” Hamiltonian leads naturally to Conjecture M.
‘tums out that Conjecture M is trae, an that once its auth
{known then vast generalizations are possible. But before we get,
there, we must move to the next chapter in our tale, and discuss
the work of Weierstrass, who essentially discovered and proved
‘Conjecture M, but did using a language that obscured the sim-
plicty of the sult, and for that reason missed some profound
implications of his discovery.
‘The Second Fork in the Road: Weierstrass
‘Weierstass (1815-1897) considered the problem of minimiz-
ingan integra ofthe form = ['1(q(s),4(s) stor Lagrang
‘ans Lsuch that (9, 4) positively homogeneous with respect to
the velocity (tht. 9,08) =a. 4) for all jana @
2 0)and doesnot depend on time. As will become cleat soon, we
havea good reason for using sather han asthe “time” variable
in the expression for )
Tvasense, one can alway’ make tis assuption on Z.“with-
coutlos of generality"by defining anew funtion Ag, 1.42) =t
L(q, wl, 1), and think of tas a new q variable, say q’, and of tas
ay
where sis a new time variable, or “pseudatime,” not 19 be
de
confused with te true time variable r, However, “without loss of|
generality” isa dangerous phrase, and does nat atall entail *with-
‘ut Toss of insight” We shall argue below that this restriction, in
conjunction withthe dominant view that Hamilton's equations
hhad to be written in the form (20), may have served to conceal
0
feom Weierstrass the true meaning and the far-reeching implica-
‘dons ofthe new condition he discovered,
Weierstrass introduced the “excess function”
eo ya
(6.9.8) = 16g.) EGa. aa
depending on thes es of independent variables gx and, He
then prowedhisside condition: Foracarvesm> qis}a bea sie
tion ofthe minimization problem he fnction € has be »
when evaluated for q= as, t= is) and a completely arb
tray
‘Weierstrass derived his side conan comparing thet
erence curse gs with other ures q¢) that are “val pot
tons of inte sense that) islose toes) oral sbut 9)
need notte clos os) Since Weirstan'contonivoles
comparing L(g-(s) 1) for uclose to .(s), with L(qe(s),ie) near an
anrary vl # posi ver fa om) sts bios
that variaons q wi large” vals of ate needed
‘Novice that for Lagrangians with he homogeneity propery
of Weiertes, 14,1) = 2(g.)-m, So Weiss could
‘equally well have written his excess function as
(a3) = Ha.2)- (qa)
06)
Using » = (4,1) 8 (1, we see dat
8(g. mit) =(U(4.%) (0.2) -(H.4)-(9.4)), apy
Which the reader will immediately recognize as
(4,93) = HCG. p)-H(4% Py 2s)
\where His our “control Hamiltonian," So Weierstrass’ condition,
expressed in ermsof the control Hamiltonian, simply says that
(MAX) along ar optimal curve 1 got, if we define pit) via
(13), then for every the value u = 54 ¢)must maximize the (con:
rol) Hamiltonian H(q-(0), pf), 1) a8. a function oft
In Weierstrass’ formulation, the condition was stated in terms of
the excess function, forthe special Lagrangians satisfying his bo-
‘mogeneity assumption. In that case the resulting His independent
‘of time, a in our equation (28) But, if one rewrites Weierstrass
condition as we have done, in terms of H, then one can take 2 gen-
fal Lagrangian, transform the minimization problem into one in
Weierstrass form, write the Weierstrass condition in the form
(MAX) (soi particular His independent of time) and then undo
the transformation and go back othe original problem. The result
is (MAX), as written, with the contol Hamiltonian ofthe original
problem. So the Weierstrass condition, if reformulated as in
(OMAX), is Valid fora problems, with exactly the same statement.
Moreover, (MAX) can be simplified considerably, Indeed,
the requirement that p() be defined via (1S) is now redundant: if
Hg, p(0, 1), regarded as a funetion of u, has a maximum at
IEEE Control Systems = i) ten (961), (0) oft) asf vans, sp
has to be given by (15). Moreover, the vachiog of
Mao. (1). t)isalsoone ofthe conditions of (19).S0
we can state (19) and (MAX) together:
(NCO) ffa curve r= a) is. solution ofthe minimization prob-
Term (1), then there has to exist a faction 1 p(t) such that the
{following three conditions hold for al
i) = (40)...
«) 9p AO 4A)
1h)
aH,
Fle AI- HD.) i
(dds). AO, 1) = m9911( 40). AD.)
[As a version ofthe nevessary conditions for optimality, (NCO)
encapsulates in one single statement the combined power of the
ler- Lagrange necessary conditions and the Weierstrass side
condition as wel, of course, as the Legendre condition, which
obviously follows from (MAX). Notice the elegance and econ-
‘omy of language achieved by this unified statement there is no
seed to bring in an extra entity called the “excess function.” Nor
docs one need to include a formula specifying how p(t) is de
Fined, since (30) does this automatically. So the addition of the
new Weierstrass condition othe three equations of 19) resultsin
anew set of three, rather than four, conditions, a set mueh “sim
plerthan the sum ofits pants" Notice moreover that (MAX)—or,
more precisely, the Weierstrass side condition part of
(MAX)—is exactly Conjecture M, So we can surmise at this
point that (MAX), as stated, probably could have been discov-
tered soon after the work of Hamilton, since itis strongly sug-
gested by (24), and almost cerainly by Weierswass, if only
Hamilton's equations had been writen in the form (14), (19)
So, we can now add two new items io our list of advantages of|
the “control formulation” of Hamilton's equations over the clas-
sical one:
(A2) Using the control Hamiltonian, it would have been an
“obvious next step 10 write Legenaire's condition in “Hamiltonian
You might also like
