You are on page 1of 10

JOURNAL OF OPTICS. 2000. VOL. 29 NO.2 ».

P'dge 95-104

TOTAL HESSIAN IN LEAST SQUARES OPTIMIZATION OF


LENS DESIGN

J. BASlJ ANI) L.N.HAZRA


Department of Applied Physics, University of Calcullil
92. A.P.c. Road. CalcutUt - 700 009.
(Receil'eti JurI/wry !Y. 2nOn)

III le(ISI .f'luares oplimiZa/iOlI oj 1"11.1' de.film II,,' cOIl/rillll/ill/! oj second order
aberralioll cleriwlI;ves ill the Hessian malr;x tII"f! excluded ill cOlI/ptlri.l·ol/ with lilt! sall/e
jmlll 'he .firsl order a/tt'I'I'ClliOIl de,.i\'(lii1·('.~. Hl'eC',,' jor .101111' CII/"S('/~\' /"(!f1wrks 011 Ille
relative IIItigillitlldes (~,. the two contributions. 1I11./('l1Iwl basis just!fving this exclllsioll
has bem rel'0rle(1 ill Ihe fileratl/l~'. III .\"I,,·le (!t" Ihe /l1"II('/icul.l"lfccess ill oplimizalioll lI'ilh
Irwlculed Hrssialllllairix. it i~' gelleraflyjidlllt<lllllllillliwlioll wilillowl H('ss;(/IIII/"'rix
is likdy /0 I,mdm'c! bt'ller r('sulls. Willi lIlt' "h"'JOIIIC'II(1/ ri.\·C' ill COIl1fIUI(ltiolllll sl'eC'l1 ill
tll(' recent past. the primm:" dijficulty in lIIuJcr/akillg .1·.I'.rlemalic im'es1i8"licms on Ihis
I'mhlelll is gradual/y pelering 01/1. n,is IWI}(!/" P"<'.I"c'III.\· .1·Ollle l"f!suflS (!f" 0/11' illl'e.wig"-
liollS OIl least squares oplimiWlioll of lell.l· desigll /Ising tofU! Hessian mat/"i.\~

l.INTRODUCTION
Computer aided lens design involves tackling of a constrained nonlinear
multivariate opimization problem [n. An iterative procedure is usually adopted
to solve this problem by seeking the solution vector that yields an optimum value
of the objective function (more commonly described as merit or defect function
in optics) in the neighbourhood of an initial estimate. Most commonly adopted
procedures in practical nonlinear optimization are essentially variants of New-
ton's method that can attain an asymptotic rate of convergence, ifthe merit func-
tion can be approximated by a quadratic function over the search domain [2].
However. successful implementation of Newton's methods calls for ready avail-
ability of tirst and second deri vati ves of the merit function. In practical lens de-
sign problems. the merit function is usually formulated as a weighted sum of
squares of aberrations and pseudoabcrrations. This way of formulating the merit
function leads to an expression for the Hessian matrix as a sun of two matrices.
one of which involves only the first derivatives of the same. The difficulty in
obtaining fast and accurate values for the second part ofthe Hessian matrix has so
far prompted exclusion of this part in comparison with first part of the Hessian
matrix that could he determined from the values of the first derivatives of aberra-
tions alone [3-8]. An' incidental advantage of this approach is the positiVI;! sem-
idcfiniteness of the truncated Hessian matrix - one of .the key reasons for
pnu.:tical success of the -widely used damped least squares (DLS) algorithm in
lens design optimization. Exclusion of the seconu derivative terms in the Hessian
matrix is primarily motivated from computational considerations. and no other
justification for the same has yet been established. Indeed, for counteracting this
96
J. Ba.lu alld LN. Hazra

total exclusion of second derivative terms, several suggestions are made to incor-
porate the effects ofhomog~neous second derivatives in calculation of the damp-
ing factor in DLS programs, and the latter seem to provide better results [9-12].
The question, therefore, comes to mind : can we obtain better results if the total
Hessian, incorporating both first and second derivatives, is utilized in optimiza-
tion procedure? In view of the ready availability of fast computers, this is no
longer an infeasible proposition. By 'better' results in optimization procedure,
we imply either a faster rate of convergence, or overcoming stagnation, or a final
solution yielding a lower value for the defect function.
With the above backdrop in view, this paper presents some results of our
investigations on least squares optimization of lens design using total Hessian
matrix in the normal equations. The next section presents a brief restatement of
mathematical formulation of the problem. Section 3 presents numerical results of
optimization run with total Hessian, followed by our concluding remarks in the
last section.

2. Mathematical Formulation:

2.1 Merit function


The merit function 'If, is defined as

'1'= :Loo;[~CJ.)~~2 +~ro!(~)2 +oo;(D-D )2]+:LOO!{G -GOgt +L~ (I)


t I IC
O
g
g
h

where the summations over t, g and h indicate respectively the summations


over the field points, the Gaussian controls and the penalty functions Ph for viola-
tions of the boundary conditions on the edge and centre thicknesses, and on the
range of glass types. The quantities W~ are the weights for image quality at the
field points. For each field point, a set of rays are traced to calculate the mono-
chromatic wavefront aberrations Wi and the Conrady chromatic aberrations O\\{,
together with the fractional distortion D; Do is the target value for fractional dis-
tortion at the same field point, and the associated weight is Wd • Finally Wim,Wic are
the weights for the monochromatic wave aberration and the Conrady chromatic
aberration respectively, for the ith ray traced from the same field point. The
Gaussian properties such as position and size of the image are controlled by set-
ting a target value GOg for each of the paraxial quantities, and Gg are the actual
values obtained from the paraxial ray trace. The corresponding weights are wg ,
The scheme for control of boundary conditions with the help of penalty functions
Ph has been reported earlier [8,12] and, for the sake of brevity, are not presented
here.
97
TOTAL HESSIAN IN LEAST SQUARES OP11MIZATION OF LENS DESIGN

For the sake of convenience in representation. the merit function '" of eq.
(I) can be written as
M
\fI=D2 (2)
i=1
Wher!; (XI' x2••••••• , x N), i=I,2,3, ....... , M are the aberrations and pseudoaberra-
lions, and each of them is a function of the N variable Xi' j=1,2,3, ...... , N of the
lens system. Typically, the variables, also called the degrees of freedom of de-
sign, are the surface curvatures, aspheric coefficients, thicknesses, separations
between lens elements, optical materials for the lens elements etc.

2~2 LEAST SQUARES OPTIMIZATION


Using matrix representation, we have the column vectors f and x as

fl XI

f2 Xl

f= X=
(3)

fM xN

The merit function", is given by


'" =fTf,
where fT is the transpose off. (4)
If we define the Jacobian matrix by

J=
(5)

then, the gradient of '" can be written as


g = V", = 2J T f
Further, let T.I be the Hessian matrix of I,,
98
J. Bu.m and L N. Helml

(7)

The complete Hessian matrix of IV can be written as


M
G =211'J +2LtiTi (8)
i=1

Furthermore. defining

(9)
we have,
G = 2JTJ + 2S (10)
Let us consider a change vector p given by

PI
P2
P= .
(11 )
PN
Using Taylor's series expansion of the gradient of 'I' • at the point x+p in the
hyperspace of the variabes,we have
g(x+p) =g+Gp (12)
If the point x + pis to be a mi ni mum of the function'll, then g (x + p) must
be zero. and we have
=
Gp -g. (13)
Substi tuting from Eqs. (6) and ( 10). we obtain
=
(P J + S) p -JTf. (14)
On nonquadratic functions, x.+p wi1l not in general, be tne minimum if pis
given by Eq. (14). and the process has to be performed iteratively. Thus at the kth
ih..~rt.Hion. the nonnal equations are

IJrJ k +Sk)Pk =-J[fk. (15)


In Newton's method, the above is used to dctennine change vector p .. at the
99
TOTAL HESSIAN IN LEAST SQUARES OPTIMIZATION OF LENS DESIGN

kth iteration and the initial estimate for the (k+ 1) th iteration is
Xk+1 = Xk+Pk (16)

3. NUMERICAL RESULTS
Evaluation of total Hessian in each iteration involves detennination of its
two parts, nameJy JTJ and S. Whereas the evaluation of rJ calls for the compu-
tation of (MN) first order aberration derivatives, the evaluation of S neccessitates
the computation of (MN) homogeneous and {(MN(N-I)I2) mixed second order
derivatives. Numerical techniques for reliable computation of these aberration
derivatives have been pre);ented earlier [13-14].
Experiments for total Hessian optimization were conducted on the follow-
ing systems:
i) An f/4.3 cemented doublet of focal length 6.4 inch and semi field an-
glctfl.
ii) An fl-t. air ~paced triplet of focal length 100 mm and semi field angle
15 n.
(inilial system adapted from U,S. Patent No. 1.987,878).
iii) An fl I. 12 Pellval lens of focal length 50mm and semi field angle S().
(initial system adapted from.U.S.Patent No.2. 158.202).
i\') An 171.5 Dmihle Gauss system of focal length I inch and semi field
angll! 2~1I.
(initial system adapted from U.S. Patent No. 2.379, 392).
For the cementeddoubfet the degrees of freedom are the three curvatures
=
and the two thicknesses, i.e .. N 5. Tables lA and IB give the JTJ and S matrices

Table IA.
The 5x5 JTJ matrix. for the cemented doublet.

.2623 E + U{i .2821 E + 06 -.5128 E + (}6 .8153 E+ 04 -.5153 E + 04

.2821 E + 06 .4592 E+06 -.6907 E +06 .1481 E+ 06 -.7461 E + 04

-.5128 E+06 -.6907 E + 06 .1133E+07 -.2130 E + 05 .1203 E + 05

.8153E+04 .1481 E + ()'5 -.2130 E + 05 .5708 E + 03 -.2029 E +03

-.5153 E + 04 -.7461 E+04 .1203 E + 05 -.2029 E + 03 .1645 E +03

respectively for the initial system. It is significant to note that magonitudes of


most c1eml!nts of the S matrix are substantially smaller than the corresponding
clement~ of the (J"J) matrix. but for some of them the magnitude are comparable.
The same' Observation can also be made from Table II that presents twenty four
elements of the first and the last rows of the matrices JTJ and S for the Double
100
J. Basu and L N. Hur.rll

Gauss system where we used twelve curvatures, seven thicknesses and five air
separations as degrees of freedom, i.e., N 24. =
Tablem

The 5x5 S matrix for the cemented doublet .

.1549 E + OS .5063 E +04 -.8896 E +04 -.7537 E +04 -.1096E+05

.5063 E.+ 04 -.1366 E + 05 -.3110 E + 04 -.2850E+ 04 -.2312 E + 04

-.8896 E + 04 -.3110 E + 04 .5258 E+05 .8764 E + 04 .4829 E+04

-.7537 E +04 -.2850 E +04 .8764 E+04 -.4622 E + 02 -.7439E +02

-.1096 E + 05 -.2312 E + 04 .4829 E+04 -.7439 E+02 .0000E+00

Thble II
Twenty four elements of the first and the last rows of the matrices JTJ and S for
the double Gauss system.
r (PJl 1• (S) I. <.PJ )l4. (S) l4r

1 .6824 E+05 .7099E+04 -.3385 E+05 -.9297 E+04

2 -.9455 E+ 05 -.9758 E+04 .1340 E + 05 .8031 E+ 04

3 .4809 E +05 -.3037 E +03 .4762E +05 -.1097 E + 05

4 .9675 E + 04 .1315 E + 03 -.1245 E+05 -.3082 E+03

5 -.5073 E + 05 -.3852 E + 03 .4746 E +03 .7404 E+04

6 .5355 E+ 05 -.2500 E +04 -.5736 E + 05 -.7054 E+04

7 .1646 E +05 -.8680 E + 03 -.1673 E+05 -.1683. E + 04

8 -.5805 E+05 .6318 E + 04 .3963 E+05 .1186E+05

9 .2709 E + 05 -.7525 E+ 04 -.3603 E + 05 -.1176 E+ 05

10 -.2739 E + 05 .7497 E + 04 .2369 E +05 .1153E+05

11 .1548E+05 -.6923 E +04 -.2130E + 05 -.9591 E+04

12 -.2487 E+ 05 .6867 E + 04 .2088 E+ 05 .8809 E + 04


101
TOTAL HESSIAN IN LEAST SQUARES OPTIMIZATION OF LENS DESIGN

13 .2970 E + 05 -.1178 E + 05 .6078 E + 02 .7032 E + 04

14 .5324 E + 04 -.1713 E + 05 -.2705 E + 05 . 1040 E + 05

15 .1903 E + 06 -.7210 E + 04 .2870 E + 06 -.1786E+05

16 .1652E+06 -.8512 E + 04 .3089 E + 06 -.1767 E + 05

17 .1310E+06 -.8911 E + 04 -.1585 E + 06 .1545 E +05


-,
18 .9670 E + 05 -.4090 E + 05 -.1903 E+06 .2574 E + 05

19 -.5394 E + 05 -.6343 E + 05 .3064 E + 06 -.7781 E + 04

20 -.7723 E + 05 -.5343 E + 05 .3124 E + 06 -.4913 E + 04

21 -.1221 E + 06 -.3188 E + 05 .3005 E + 06 .1657 E + 05

22 -.6502 E + 05 -.1662 E + 05 .1553 E + 06 .7733 E + 04

23 -.3385 E + 05 -.9297 E + 04 .1484 E +06 .9248 E + 04

24 -.9455 E + 05 -.9758 E + 04 .90867 E + 05 .2351 E + 04

Figure 1-4 present results of optimization run for the four lens systems men-
tioned above. They demonstrate significantly high rate of convergence
achieved in least squares optimization when the total Hessian is utilised.

900
800
700

..
III
C 600
0
u 500

--
c
:J

'0:::
Q)
::e
400
300
200
100
0
0 5 10 15
No. of iterations

fig. I Results of optimization run using lolal Hessian method with line search on a cemented doublet.
102
.I Da,wl and L. ,IV Ha;m

600

500

c:
0400

I
..
.2 300
' I;
~ 200

100

0
0 2 3 .\ 5 6 7 8
No. of iterations

Fig. 2. Resul!s of optimizalion run lIsing lolal Hessian method with line search on a Tronnier Iriplet.

1900
1880
1860
1840
1820

.. 1800
·c 1780
II
:E 1760
1740
1720
1700
0 2 6 8 10 12 14 16
No. of Iterations

Fig. 3. iteslilts bf 0l'Wnir.ailon nlll ll~ihg tolal Hessian method with lihe scardl oh a Schade Petzval
sY!item.
103
TOTAL IIESSIAN I~ LEAST SQUARES OPTIMIZATION OF LENS DESIGN

3500
3000

roo 2000
1500

1 1000
500

0
0 5 10 15
No.of lleratlon.

Fig. 4. Results of optimi/alion run using total lics.~ian Int'thod with line sear~ h on OJ Double
Gauss ~ ys tem .

4. CONCLUDING REMARKS
It is obvious that the total Hessian method for lens design optimization.
pre~ented above. takes proper account of the relative nonlinearity of the vari-
able~. and this may be one of the reasons for fast convergence at the initial stages
of the optimization run. Another advantage is the nonrequirement of any damp-
ing factor.
Nevertheless. after first few iteration stages. the convergence is slow which
can hardly justify the large amount of additional computation required in the
evaluation of total Hessian at this stage. A hybrid method incorporating Ihe total
Hessian approach along with usual DLS or Gauss-Newton approach seems 10 be
more appropriate for overall optimiz.ation.

REFERENCES
I. R. R. Shannon. The An and Sden~e of Opti~ a l Design. Camoridge l.;niversity Prc s ~. Cam·
bridge ( 1997).
L.E. Scales. IntrodUl:tion to Nonlinerar Optimil.ation. Manniilal!. l.lIndon 11981 ).
J. A . Girard. Rev. Opt 37. 225-241. W7-424! 1958l.
4. C. G. Wynne. Proc. Phys. SO<.' . London 73. 777-7F.7 119591.
5. D.P. Feder. Appl. Opt. 2. 1209-1226 (196.h
6. T H.Jamieson . Optimir.ation Techniques in lens Design, Adam Hilger. London I 1971 )
7. M.1 . Kidgcr, Opt. Eng. jl(8J. 1731-1739 l199J).
104
J. Bam and L. N. Haz",

8. 1. Basu and L.N. Hazra, Optical Engineering, 33(12), 4060-4066 (1994).


9. D.R. Buchele, Appl. Opt. 7,2433-2435 (1968).
10. D.C. Dilworth, Appl. Opt., 17,3372-3375 (1978).
II. G. G. Slyussarev, Aberrations and Optical Design Theory, p. 432, Adam Hilger, Bristol (1984).
12. H.H. Hopkins and A. Kadkly, 1. Mod. Opt. 35, 49-74 (1988).
13. 1.Basu and L.N. Hazra, Journal of Optics, 25(1), 37-60 (1996).
14. 1. Basu and L.N. Hazra, Journal ofOplics, 26(3),157-159 (1997).

You might also like