The New Linear Programming Method of Karmarkar

A. Schrijver
Department of Econometrics Ti/burg University
P.O. Box 90153, 5000 LE Ti/burg, The Netherlands
and
Centre for Mathematics and Computer Science
P.O. Box 4079, 1009 AB Amsterdam, The Netherlands

1. INTRODUCTION
In April 1984, at the 16th Annual ACM Symposium on Theory of Computing,
NARENDRA KARMARKAR of AT&T Bell Laboratories presented a new algo-
rithm for linear programming. The algorithm was not only shown to be
theoretically efficient (i.e., its running time is bounded by a polynomial in the
input size}, but was also claimed to be very fast in practice - about 50 times
faster than Dantzig's classical simple~ method, for the largest problems
evaluated.
This news created much excitement among computer scientists and
mathematical programmers, and subsequent reports, inter alia in Science maga-
zine and on the front page of the New York Times, contributed to a further
propagation of the sensation. Linear programming is one of the mathematical
fields most applied in practice. Linear programming problems occur in such
diverse areas as engineering, transportation, agriculture, distribution, schedul-
ing, nutrition, management, and a reduction of the computer time needed
would not only speed up solving linear programming problems, but also would
allow one to solve larger LP-problems than before. In situations like oil pro-
cessing and automatic control, quick, almost forthwith, solution of LP-
problems is essential.
In 1979, L.G. KHACHIYAN published the first polynomial-time method for
linear programming, the ellipsoid method. This method, though theoretically
efficient, turned out to behave rather disappointingly in practice. So
Karmarkar's claim that he now has found a method which is both theoretically
and practically efficient, was much welcomed. Karmarkar's paper was pub-
lished in the December issue of Combinatorica [5]. However, no computational
details were given.
Next, KARMARKAR was invited to give plenary lectures at two international
conferences, the ORSA/TIMS-meeting in November 1984 in Dallas, and the
12th International Symposium on Mathematical Programming in August 1985
at MIT. KARMARKAR described his method and variants, explaining some of
the tricks used in practice, claiming superiority of his method over the simplex
method, and giving a few comparisons, but he refused to give full disclosure of
test problems, computer programs and running times. This has led to much
uncertainty and discussion among mathematical programmers with respect to
the practical value of the new method. It led to a report 'Founding father of
just a footnote?' in the Boston Globe of August 9, 1985:

'This week in Cambridge, the 28-year-old KARMARKAR came under

mounting fire from his colleagues at the 12th International Sympo-
sium on Mathematical Programming. They snorted at his scientific
manners, scoffed at his claims and derided his results as being
everything from 'frisky' to 'majestic'. Mostly, they said that his
accounts of super-fast solutions to difficult problems couldn't be
replicated .....
... 'He may have some wonderful method after all, but I habitually
mistrust all secret mathematics', said E.M.L. BEALE, a pioneer in
the commercial applications of linear programming....
... KARMARKAR himself didn't advance his cause much in a talk
before an unusual plenary session of the MIT meeting of some 800
scientists from around the world. He began by observing that while
mathematicians agree on what constitutes convincing proof of a
mathematical proposition, there is no corresponding consensus as
to what makes a persuasive presentation of experimental results -
a contention that was immediately disputed by many of his
listeners.'

Indeed, there is some generally accepted standard in presenting computional

results. One gives (or makes available) the computer program, the test data,
the type of computer, the output, and the CPU-time. In essence, the results are
replicable, possibly making due allowance for the running time. Generally, one
tries to give as much information as possible within the compass of a lecture
or report.
Although this consensus differs from that holding in mathematics, with its
strict rules for definitions, theorems and proofs, it essentially is comparable
with the praxis in other branches of sciences, such as physics and chemistry,
when reporting on experiments.
Karmarkar's reservedness in presenting computational details may have
respectable reasons, for instance that Bell Labs has propriety of the actual
computer program, which might not yet be ready as a marketable package, but
the scientific community turns out to sputter if despite that a similar reserved-
ness in making claims is not observed.

3
 In this account of the new method I will restrict myself to the theoretical
aspects.
In Section 2 and 3 we briefly discuss the simplex method and Khachiyan's
ellipsoid method. In Section 4 and 5 we describe Karmarkars method, while
in Sections 6 and 7 we show that the method has polynomially bounded run-
ning time.

2. LINEAR PROGRAMMING AND THE SIMPLEX METHOD

The linear programming problem (or LP-problem) is as follows:
given A El.m Xn, b Elm, c El.n, (1)
find a vector x eon attaining max {CTx IAx .;;;;b }.
So it is asked to maximize the linear function c T x where x ranges over the
polyhedron {x IAx .;;;;b}. The practical relevance of this problem was revealed
in the 1940s by the work of L.V. KANTOROVICH, TJ.C. KOOPMANS and G.B.
DANTZIG.
In 1947 DANTZIG designed his famous simplex method for solving ( 1). The
idea is to make a trip over the polyhedron P: ={x IAx.;;;;; b} from vertex to ver-
tex along edges, on which c T x increases, until an optimum vertex is attained.
The correctness of this algorithm is based on the property that if a vertex x 0
of a polytope P does not maximize c Tx over P, then there exists a vertex x 1
adjacent to-x 0 for wich cTx 1 >cTx 0 • (Here x 1 adjacent to x 0 means that the
segment x 0 x 1 forms an edge of P.)
Roots of this idea occur already in FOURIER [3], describing a method for
minimizing llAx - b 11 00 (where II* 11 00 denotes the maximum absolute value of
the entries in *):

'Pour atteindre promptement le point inferieur du vase, on eleve en

un point quelconque du plan horizontal, par exemple a l'origine
des x et y, une ordonnee verticale jusqu'a la rencontre du plan le
plus eleve, c'est-a-dire que parmi tous les points d'intersection que
l'on trouve sur cette verticale, on choisit le plus distant du plan des
x et y. Soit m 1, ce point d'intersection place sur le plan extreme.
On descend sur ce meme plan depuis le point m 1 jusqu'a un point
m2 d'une arete du polyedre, et en suivant cette arete, on descend
depuis le point m 2 jusqu'au sommet m 3 commun a trois plans
extremes. Apartir du point m 3 on continue de descendre suivant
une seconde arete jusqu'a un nouveau sommet m 4 , et !'on continue
l'application du meme precede, en suivant toujours celle des deux
aretes qui conduit a un sommet moins eleve. On arrive ainsi tres-
prochainement au point le plus bas du polyedre.'

According to FOURIER, this description suffices to understand the method in

4
more dimensions. DE LA VALLEE PoussIN [9) gave a similar method.
DANTZIG [2] algebraized the method, obtaining an attractive compact
scheme (simplex tableau) and iterative procedure (pivoting), which facilitates
computer implementation. This simplex method turns out to be very efficient
in practice and enables one to solve LP-problems in several thousands of vari-
ables.
However, it could not be proved theoretically that the simplex method is
efficient. That is, no proof has been found that the running time of the simplex
method is bounded by a polynomial in the size of the problem, i.e. in
2: log( Iaij I+ 1) + 2: log (I b; I + l) + 2: log (I c1 I + 1). (2)
0 j

In fact, KLEE and MINTY [7] showed, by giving a bad class of LP-problems,
that with Dantzig's pivoting rule, the simplex method can require exponential
running time. Their examples have as feasible regions a deformation of the n-
dimensional cube (described by 2n inequalities), for which Dantzig's rule leads
to a trip along all 2n vertices. Several alternative pivot selection rules have
been proposed, but none of them could be proved to yield a polynomial-time
method.
On the other hand, BORGWARDT [l] recently gave a pivoting rule which he
showed to yield a polynomial-time algorithm on the average, in a certain
natural probabilistic model. His result very much agrees with practical experi-
ence, where data seem to be more 'random' than structural.

3. THE ELLIPSOID METHOD

It has been an open question for a long time whether linear programming is
solvable in polynomial time. Although the simplex method works well for
present-day practical problems, one never knows whether the barycenter of
practical problems will change, and it would then be good to have a method
which can be proved to perform well always.
It was a big surprise when in 1979 the Soviet mathematician L.G. KHACHI-
YAN answered this question affirmatively, showing that the ellipsoid method for
nonlinear programming has polynomially bounded running time when applied
to LP-problems. Also this result was reported on the front page of the New
York Times.
Khachiyan's method can be described by application to the following prob-
lem:
given
A ezmxn, bezm, find xeOn such that Ax.;;;;,b. (3)
This problem is polynomially equivalent to problem (1), i.e., any polynomial-
tirne algorithm for problem (1) yields a polynomial-time algorithm for problem
(3), and conversely. Indeed, (3) easily reduces to (I) by taking c-o. Con-
versely, by the Duality theorem of linear programming, solving ( 1) is

5
 I
I
I
I
I
I
I
I
I
I
I
I

\
\
\
\
\
\
\
\
\
\
\
i
A deformation of the n-dimensional cube, with a simplex path along 2n vertices
(n =3)

equivalent to solving the following system of linear inequalities:

Ax:;;;;;b, yT;;;.o, yTA = c, yTb,;;;;cT x (4)
(clearly, equations can be split into two opposite inequalities). This is a special
case of (3).
To sketch Khachiyan's method, we assume that the polyhedron {x IAx,;;;;b}
is bounded and full-dimensional (KHACHIYAN showed that we without loss of
generality may restrict ourselves to this case). Let T be the maximum absolute
value of the entries in A and b (w.l.o.g. T;;;.n ;;;.2). With Cramer's rule, one
may show that the components of the vertices {x IAx,;;;; b} are at most n n Tn
in absolute value. Hence {xlAx,;;;;b} is contained in the ball £ 0 : = B(O,R)

6
around the origin of radius R : = n n+I rn.
E o is the first ellipsoid. Next ellipsoids E 1> E 2 , ••• are determined with the
following rule. If Ek has been found, with center say zk> check if Azk';;;;,b holds.
If so, we have found a solution of Ax ..;;,b as required. If not, we can choose an
inequality, say aTx..;;,b; in Ax:s;;;;.b violated by zk. Let Ek+I be the ellipsoid
such that
Ek+1~Ekn(xlafx:s;;;;.aTzk} (5)
and such that Ek+ 1 has smallest volume (there exist simple updating formulas
for obtaining the parameters describing Ek+ 1 from those describing Ek and
from a;). Since (xlAx..;;,b}~(xlaTx..;;,aTzk}, it follows by induction on k
from (5) that
Ek~ (x IAx..;;,b }. (6)
Moreover, it can be proved that
volume Ek+J <e-l/ 4n ·volume Ek. (7)
Since one easily sees that volume E 0 ..;;,(2Rf<n 2n'rn', inductively from (7) we
have:
(8)
On the other hand, with Cramer's rule, using the boundedness and fulldi-
mensionality of {x !Ax ..;;,b }, we know:
volume {x IAx..;;,b};;;;i.n- 2n'r-n'. (9)
(6), (8) and (9) imply:
(10)
i.e.,
(11)
So after a polynomially bounded number of iterations we will have found a
solution of Ax :s;;;;.b . Updating the ellipsoid parameters can be done in 0(n 2 )
arithmetic operations, while all calculations have to be done with a precision
of 0(n 3 log T) digits. Altogether this amounts to 0(n 8 log2 T) bit operations
(excluding data-handling, which takes 0(loglog T · logloglog T) for each bit
operation).
Although KHACHIYAN showed the polynomial solvability of the linear pro-
gramming problem, his method turned out to perform badly in practice. This
is caused, among others, by the facts that the upper bound (11) of iterative
steps, though polynomial in the input size, can be rather big also for moderate
problems, and that the precision required to describe the successive ellipsoids
is huge. (The ellipsoid method has implications in combinatorial optimization
- see [4].)

7
 Thus the question remained if there is a method for linear programming
which is both practically and theoretically efficient. K.ARMARKAR claims that
the following method is so.

4. l<ARMARKAR'S FORM OF THE LINEAR PROGRAMMING PROBLEM

Karmarkar's method applies to problems of the following form:
given A Ezmxn, cEZn such that A 1 = 0, (12)
findxEQnsuchthatx~O, Ax= 0, 1Tx = 1, cTx~O.

(Here 1 denotes an all-one column vector of appropriate dimension.) This is a

problem equivalent to (1) and (3). Indeed, (12) clearly is a special case of (1),
as (12) amounts to finding x attaining max{-crxlx~O,Ax =O, 1Tx =l}.
Conversely, (3) can be reduced to solving a system of linear equations in non-
negative variables:
given A Ezmxn, b E"ll.m, ( 13)

find a vector x EO!n such that x ~O, Ax = b.

This follows by replacing Ax ~b by the system:
Ax'-Ax" + x"' = b, x',x",x'"~O. Now (13) can be reduced to (12) as fol-
lows. Let A,b as in (13) be given. Let T be the maximum absolute value of the
entries in A and b. With Cramer's rule we can prove that if x ~O, Ax= b has a
solution, it has one satisfying 1Tx~nn+ 1 yn. So we wish to solve:
(14)
By adding one extra variable we may assume we must solve:
x~O, Ax=b, 1Tx=nn+Iyn. (15)

By subtracting multiples of the last equation in (15) from the equations in

Ax =band by scaling equations, this is equivalent to:
x~O, Ax= 0, 1Tx = 1. (16)
If A 1 = 0 then n - 1 1 is a solution. Otherwise, by elementary changes of the
system we may assume A 1=1. So we wish to find a solution x, A. for:
x~O, A.;;?;0, Ax-IA= 0, 1Tx+A. = 1, such that A.~O. (17)
Since Al-1-1=0, this is a special case of (12) (taking
A:=[A,-1], c:=(O, ... ,O,l)TEzn+1).
So Karmarkar's method applied to (12) solves linear programming in gen-
eral.

8
5.KARMARKAR's METHOD
Karmarkar's method consists of constructing a sequence of vectors x 0 ,x 1 ,x 2 , ...
converging to a solution of (12) (provided (12) has a solution). The essence of
the method is to replace the condition x ;;;;.o by a stronger condition x EE, for
some ellipsoid E contained in ~"+-. As we shall see, minimizing c T x over
{xlxEE,Ax=O,lTx=l} is easy, while minimizing c7 x over
{x Ix EIR "+-,Ax =O, 1T x = l} is the original problem ( 12).
Let A and c as in (12) be given. We may assume without loss of generality
that the rows of A are linearly independent, and that n ;;;;.2. Throughout we use

r:=~ (18)

Let
1
XO : =-·1. (19)
n
So Ax 0 =0, 1Tx 0 =I, x 0 >0. Next a sequence of vectors x 0 ,x 1 ,x 2 , ... such
that Axk =O, 1Txk=I, xk>O is determined, with the following recursion:
denote xk = :(x\k>, ... , x~k))T, and let D be the diagonal matrix:
D : = diag(x\kl, ... , x~k >). (20)

Define zk+l and xk+l as:

z k + 1 is the vector attaining (21)

min {(cTD)zl(AD)z = O;lTz=n; zEB(l,+r)},

xk+l : = (lTDzk+1)-1·Dzk+1.
I
Note that if we replace in the minimization problem the condition z EB(l, 2r)
by the weaker condition z ;;;;.o, then we would obtain a minimization problem
with optimum value at most 0 if and only if min
{ c Tx IAx =O, 1Tx=1, x ;;;;.Q} ,,;;;;O, which is our original problem (12).
As z k + 1 minimizes (c TD )z over the intersection of a ball with an affine
space, we can write down a formula for z k + 1 :

PROPOSITION 1.
(J-DA T(AD 2 A T)- 1AD-n - 1·1 · lT)Dc
zk+1 - 1 - -I r .~-=----=-:...::_~=-.::...:.......L-_.:;_=------'--
- 2 11(1-DA T(AD 2 AT)- 1AD-n - 1·l · l7)Dcll.

PROOF. The minimum in (21) can be determined by projecting De onto the

space {z I(AD)z =O, t7 z =0 }, thus obtaining the vector:
p: = (J-DAT(AD 2 AT)- 11·1T)Dc. (22)
(Indeed, ADp = 0, tTp = 0, as one easily checks (using ADl=Axk=O), and

9
c TD -p T is a linear combination of rows of AD and of the row vector 1T.)
Then zk+I is the vector reached from 1 by going over a distance ; r in the
direction - p, i.e.,
zk+I =1- l.r_E_ 0 (23)
2 llpll"

This method describes Karmarkar's method.

6. A LEMMA IN CALCULUS
In order to show correctness and convergence of the algorithm, we use the fol-
lowing lemma in elementary calculus. For x=(xJ. ... ,xnf, we denote:
(24)

LEMMA. Let nEN, H: ={xEIRn llTx=n}. Then:

(i) HnB(l,r) cHnlR'+ cHnB(l,(n-l)r);
(ii) if XEHnB(l, ~r), then IIx~+(l+ n~l r- 1.
(i) Let x =(xi. ... ,xnl eH nB(l,r). To show x EIR'+, suppose without loss
of generality x 1 <0. Since x EB(l,r) we know:
(x2 -1) 2 + ... +(xn-1)2 .;;;;r 2 -(x 1 - I)2<r 2 - l = 1 / (n-1). Hence, with
Cauchy-Schwarz:
(x2- l)+ ... +(xn-1).;;;; y;;=-}. V(x2 -1)2 + ... +(xn -1)2 <I.
Therefore, x 1+ ... + Xn <x 2 + ... + Xn <n, contradicting the fact that x belongs
to H.
Next let xEHnlRr+. Then
(xi-1)2 + ... +(xn-1)2 =(xt + x~)-2(x1 + ... +xn)+n
.;;;;(x1 + ... +xn)2-2(x1 + ... +xn)+n
= n 2 -2n+n = (n-1) 2 r 2 .
(The last inequality follows from the fact that x ~O.) Sox EB(l,(n - l)r).

(ii) We first show an auxiliary result:

let A.,µelR; if x*,y*,z* achieve (25)
min {xyz lx+y+z=A., x 2 +y 2 +z 2 =µ},and
x*.;;;;y*.;;;;z*, theny*=z*.
By replacing x,y,z by x-fA., y-fA.,z-~A., we may assume
A.=O.Then it is clear that the minimum is nonpositive, and hence

10
 x' ~O~v' ~='. Therefore .
"' " "' .
.
.. . r --+-·::. ~ _1!:_ \---.
x r :: ~x 1-----.--r (26)
18
Thi'.' !irsl inequality here follows fmm
quality (note that x • and the that if
x--rr+;; O.x 2 ..... r 2 +::==µ..
On the other hand. (x. satisfies
x ~.1· +:: =: O..x2 + +::: =.u. . = - Ll8 \
:c:: Hence \Ve haH~ equal it\.
throughout in (2.6t Therefore. r' '"':: •. proving (25 ).

We now prove of the Lemma. The case n = 2 heing easy. assume 1;

Let x attain

mm l Ilx Ix EH B(l. 0

Without k)SS of generality. \ i ~x: ~ ... ~xn. Then for all l ~1 < </... ~n.
vector ) attains
nun{xr::ix+_r-r;; (281
(otherwise we could replace the componenb .t, ..\.X1. of x by better 1values).
Hence bv (25). x,-='x,. Therefore x 2 =x1 = ... -=x,.. A:-. xEHi"IB(l.-::;r). this
. 1· . l , .I . . l Th' h ,-.-
imp !CS X1 =2· anu X: = ... =xn =(I +2 (n ·-- l l). lS s OWS LJ

7. OPERATIVENESS OF THE Al GORlTHM

The operativeness of the algorithm now follows from the following proposi-
tion:

PROPOSITION 2. ~f (1:;) hus lJ solutwn Ihen al! k ~O:

(c r_...x· ; )n < l. ~c Tx'- t (29)

nx" TI e Ilx'

PROOF. First note:

(crx•~1)" rJxk (c 1 D::i.•l)" llxA
(30)
nx' + l (cr_\.i. )" - m.D::' • i l (cr,.i. l"

I er D::' •I n
1
t'rDI 11.:p\1-"l.

using (21) and ll(D.::' ~ 1 l = (Hx' )(II::'· 1 ) and x' = Dl.

We next show that, if ( 11) has a s•.)lution. then

II
 1/2
-- -
n-1"
(31)

Indeed, if(12) has a solution then ADz=O, z;;;.O, cTDzo;;;,O for some z=f:O. We
may assume lrz=n. Hence,
Q;;a.min{(crD)zlzER'!i.,ADz=O, lrz=n} (32)
;;a.min{(crD)z lz EB(l,(n - l)r),ADz =O, lr z =n}
(the last inequality follows from (i) of the Lemma).
The last minimum in (32) is attained by the vector 1-(n - l)r~, as

1- ~ r~ attains the minimum in (21), (cf. (22)).

Therefore, cTD(l-(n - l)r~,,;;;,o.

This implies

CT Dzk+ I =cTD(l-l.r_p__) (33)

2 llpll
=(I - ___12_ )c TD 1 + _Zi_c TD(l -(n - 1)r_p__)
n-1 n-1 llpll
o;;;,(l - ___'2_l) (cTD)l,
n-
proving (31 ).
Therefore, as Ilzk+ 1 ;;a.f(l ++ / (n - l)t- 1• by (ii) of the Lemma,

[ crDzk+1 Jn 1 .;;;;;(1 - _V2_r. < 1. (34)

cTDl IIzk+I n-1 'h(l+_Zi_)n-1 e
n-1
(as (l-x)/(l+x),,;;;,e-2x for x;;a.O, since the function e- 2x-(l-x)/(l+x)
is 0 for x =O, and has nonnegative derivative if x ;a.Q). (30) and (34) combined
give (29). 0

By induction on k, Proposition 2 implies, if c r x 0 ;a.O, c T x 1 ;;;.o, ... , c T xk ;a.O,

(using
l n I
(IIxk)l/n.;;;;;- ~ xf = -o;;;,1):
n i=I n

cTxk,,;;;, c
Txk
< [2]k/n
- · CTXO
:s;::::
[2]k/n ·nT
- (35)
(IIxk)l/n e (IIxO)l/n"""' e '

where T denotes the maximum absolute value of the entries in A and c (w.
I.o.g.T;;a.n).
This gives, if we take

12
 N : = I [ l -2/n2] n2ln(nT)l, (36)

the following theorem.

THEOREM. If (12) has a solution, then CT xk <n --n T-n for some k =O, ... , N.

PROOF. Suppose cTx 0 , . . . ,cTxN~n-nT-n. Then (35) holds for k=N,

implying c TXN <(21 et In ·nT~n -n T-n. Contradiction. D

So supfose (12) has a solution. Then with Karmarkar's method we find a vec-
tor x satisfying xk~O,Ax=O, lTx=I, cTxk<n-nT-n. By elementary
linear algebra, it is easy to find a vertex x • of the polytope
{x~OIAx=O, lTx=l} with cTx·~cTxk. Hence, cTx*<n-nT-n. By
Cramer's rule, the entries in x • have a common denominator at most nn T".
As c is integral, this implies c T x * ~O. So x • is a solution of ( l 2).

Karmarkar's method consists of O(n 2 logT) iterations, each consisting of l9(n 3 )

arithmetic operations (due to the updating formula given by Proposition 1).
All calculations have to be made with a precision of O(n 2 log T) digits. Alto-
gether this amounts to 0(n 7 log2 T) bit operations (excluding data-handling,
which takes l9(loglog T · logloglogT) for each bit operation).

Parts of the description above are taken from the forthcoming book [8].

REFERENCES
1. K.-H. BoRGW ARDT ( 1982). The average number of pivot steps required
by the simplex method is polynomial. Zeitschrift fur Operations Research
26, 157-177.
2. G.B. DANTZIG (1951 ). Maximization of a linear function of variables sub-
ject to linear inequalities. TJ.C. KOOPMANS (ed.). Activity Analysis of Pro-
duction and Allocation, John Wiley & Sons, New York, 339-347.
3. J.B.J. FOURIER (1826). Analyse des travaux de l'Academie Royale des
Sciences, pendant J'annee 1823, Partie mathematique. Histoire de
f'Academie Royale des Sciences de l'lnstitut de France 6, xxix-xli.
4. M. GR6TSCHEL, L. LovAsz, A. SCHRIJVER (1986). The Ellipsoid Method
and Combinatorial Optimization, Springer-Verlag, Berlin.
5. N. KARMARKAR (1984). A new polynomial-time algorithm for linear pro-
gramming. Combinatorica 4, 373-395.
6. LG. KHACHIYAN (1979). A polynomial algorithm in linear programming
(in Russian). Doklady Akademii Nauk SSSR 244, 1093-1096.
7. V. KLEE, G.J. MINTY (1972). How good is the simplex algorithm?
0. SHISHA (ed.). Inequalities, Ill. Academic Press, New York, 159-175.
8. A. ScHRIJVER (1986). Theory of Linear and Integer Programming, John

13
 & Sons. Chichester.
9. Cu. Dr LA VAUEr Pot:SSIN ( 1910). Sur la methode de l'approximatwn
Scient~fique de Bruxelles 35 (.!}, 1-16.
rmrmnum. Annales de 111 ."io;:i<'tE;

14