Vincent

Nonlinear Analysis: Modelling and Control, 2005, Vol. 10, No.
4, 297304
A Comparative Study of Two Real Root Isolation

Methods
2
A. G. Akritas1 , A. W. Strzebonski
1
University of Thessaly, Department of Computer and Communication Engineering

GR-38221 Volos, Greece
akritas@uth.gr
2
Wolfram Research, Inc., 100 Trade Center Drive, Champaign, IL 61820, USA
adams@wolfram.com
Received: 07.10.2005
Accepted: 10.11.2005
Abstract. Recent progress in polynomial elimination has rendered the
computation of the real roots of ill-conditioned polynomials of high degree (over
1000) with huge coefficients (several thousand digits) a critical operation in
computer algebra.
To rise to the occasion, the only method-candidate that has been considered by
various authors for modification and improvement has been the Collins-Akritas
bisection method [1], which is a based on a variation of Vincents theorem [2].
The most recent example is the paper by Rouillier and Zimmermann [3], where
the authors present
... a new algorithm, which is optimal in terms of memory usage and as
fast as both Collins and Akritas algorithm and Krandick variant ... [3]
In this paper we compare our own continued fractions method CF [4] (which
is directly based on Vincents theorem) with the best bisection method REL
described in [3]. Experimentation with the data presented in [3] showed that,
with respect to time, our continued fractions method CF is by far superior to
REL, whereas the two are about equal with respect to space.
Keywords: root isolation, Vincents theorem, coninued fractions method,
bisection (or Collins-Akritas) method.
1 Description of the two algorithms

For completeness we briefly describe our implementation of both the continued
fractions method CF and algorithm REL. The correctness of the first algorithm
297
A. G. Akritas, A. W. Strzebonski
along with an analysis of its computational complexity can be found in the literature [58]. A discussion of the second algorithm can be found elsewhere [3].
1.1
Description of the Continued Fractions Algorithm CF
Let us first introduce the notation used in the algorithm. Let f Z[x] \ {0}.
By sgc(f ) we denote the number of sign changes in the sequence of nonzero
coefficients of f . For nonnegative integers a, b, c, and d, such that ad bc 6= 0,
we put

intrv(a, b, c, d) := a,b,c,d (0, ) ,
where
a,b,c,d : (0, ) 3 x
ax + b
(0, )
cx + d
and by interval data we denote a list

{a, b, c, d, p, s},
where p is a polynomial such that the roots of f in intrv(a, b, c, d) are images of
positive roots of p through a,b,c,d , and s = sgc(p).
The value of parameter 0 used in step 4 below needs to be chosen empirically. In our implementation 0 = 16.
Algortihm Continued Fractions (CF)
Input: a squarefree polynomial f Z[x] \ {0}.
Output: the list rootlist of positive roots of f .
1. Set rootlist to an empty list. Compute s sgc(f ). If s = 0, return an
empty list. If s = 1, return {(0, )}. Put interval data {1, 0, 0, 1, f, s} on
intervalstack.
2. If intervalstack is empty, return rootlist, else take interval data {a, b, c,
d, p, s} off intervalstack.
3. Compute a lower bound on positive roots of p.
4. If > 0 , set p(x) p(x), a a, c c, and 1.
298
A Comparative Study of Two Real Root Isolation Methods
5. If 1, set p(x) p(x + ), b a + b, and c c. If p(0) = 0, add

[b/d, b/d] to rootlist, and set p(x) p(x)/x. Compute s sgc(p). If s = 0,
go to step 2. If s = 1, add intrv(a, b, c, d) to rootlist and go to step 2.
6. Compute p1 (x) p(x + 1), and set a1 a, b1 a + b, c1 c,
d1 c + d, and r 0. If p1 (0) = 0, add [b1 /d1 , b1 /d1 ] to rootlist, and set
p1 (x) p1 (x)/x, and r 1. Compute s1 sgc(p1 ), and set
s2 s s1 r, a2 b, b2 a + b, c2 d, and d2 c + d.
1
7. If s2 > 1, compute p2 (x) (x + 1)m p( x+1
), where m is the degree of p. If
p2 (0) = 0, set p2 (x) p2 (x)/x. Compute s2 sgc(p2 ).
8. If s1 < s2 , swap {a1 , b1 , c1 , d1 , p1 , s1 } with {a2 , b2 , c2 , d2 , p2 , s2 }.

9. If s1 = 0, goto step 2. If s1 = 1, add intrv(a1 , b1 , c1 , d1 ) to rootlist, else
put interval data {a1 , b1 , c1 , d1 , p1 , s1 } on intervalstack.
10. If s2 = 0, goto step 2. If s2 = 1 add intrv(a2 , b2 , c2 , d2 ) to rootlist, else
put interval data {a2 , b2 , c2 , d2 , p2 , s2 }on intervalstack. Go to step 2.
In the present paper we also address the issue of memory usage by the continued fraction method. We show that the algorithm can be so structured that the
maximal number of transformed versions of the polynomial that need to be stored
at any given time is bounded by 1 + log2 n, where n is the degree of the input
polynomial. This bound is based on the following conjecture, which we have not
proven, but which we have extensively tested.
Conjecture. Let f be a polynomial of degree n, and let sgc(f ) denote the
number of sign changes in the sequence of nonzero coefficients of f . Then

1
) sgc f (x)
sgc f (x + 1) + sgc (x + 1)n f (
x+1
Hence the number of sign changes in any interval on intervalstack is at
least equal to the total number of sign changes in all intervals above it. Since the
number of sign changes in the top interval is at least 2, and the total number of
sign changes in all intervals on stack is at most the degree n of f , the number
of possible levels in intervalstack is at most log2 n. Therefore the maximal
number of transformed polynomials we need to keep at any given time is at most
1 + log2 n.
299
1.2
Description of the Algorithm REL
As in [3], let
Ha (p)(x) = p(ax),
Ta (p)(x) = p(x + a).
The algorithm calls subprocedure DesBound which for a polynomial p of
1
))). This is done by computing
degree m returns min(2, sgc((x + 1)m p( x+1
m
subsequent coefficients of the Taylor shift of x p( x1 ), returning 2 as soon as we
get two sign changes.
Algortihm REL
Input: a squarefree polynomial f Z[x] \ {0}.
Output: the list rootlist of positive roots of f .
1. Set rootlist to an empty list. Set p f . Compute an upper bound B on
positive roots of p (a nonnegative power of 2). If B > 1, set p(x) p(Bx).
2. Compute s DesBound(p). If s = 0, return an empty list. If s = 1,
return {(0, B)}. Put pairs (1, 1) and then (1, 0) on intervalstack. (Pair
).) Set k 0 and c 0.
(k, c) corresponds to interval [B 2ck , B c+1
2k
3. If intervalstack is empty, return rootlist, else take pair (k 0 , c0 ) off intervalstack.
0
4. Compute p 2n(k k) H2kk0 (T2kk0 c0 c (p)). ( [3] proves that the translation is either the identity or the Taylor shift.)
0
5. If p(0) = 0, add [B 2ck0 , B 2ck0 ] to rootlist, and set p(x) p(x)/x.

0
6. If k 0 k, compute s sgc(p). If s = 1, add (B 2ck0 , B) to rootlist. If s = 0

or s = 1, return rootlist.
7. Set c c0 , k k 0 , and compute s DesBound(p).
8. If s > 1, put pairs (k + 1, 2c + 1) and then (k + 1, 2c) on intervalstack. If
s = 1, add (B 2ck , B c+1
) to rootlist. Go to step 3.
2k
300
2 Empirical results
We compare performance of our continued fraction algorithm CF, and the algorithm REL described in [3]. We have implemented both algorithms as a part of
Mathematica kernel. They both use the same implementation of Shaw and Traubs
algorithm for Taylor shifts (see [9]). As benchmark examples we use Chebyshev,
Laguerre, Wilkinson, and Mignotte polynomials used in [3], as well as three types
of randomly generated polynomials used in [4].
All computations were done on a 850 MHz Athlon PC with 256 MB RAM.
The memory used data was obtained using Mathematica MaxMemoryUsed command, so is includes the total memory used by Mathematica kernel. The startup
size of Mathematica kernel is 1.6 MB.
In case of special polynomials, Table 1, CF is faster by factors ranging from
around 3 for Chebyshev polynomials to 50000 for Mignotte polynomials. The
case of Mignotte polynomials is especially advantageous for our continued fractions method, because there is a point with a very simple continued fraction
expansion (namely 15 ), which lies between the two close roots. For Chebyshev
polynomials we used the fact that the polynomials are even and so with both
methods we isolated only the positive roots.
Table 1. Special polynomials
Polynomial
Degree
No. of roots
Chebyshev
Chebyshev
Laguerre
Laguerre
Wilkinson
Wilkinson
Wilkinson
Mignotte
Mignotte
Mignotte
1000
1200
900
1000
800
900
1000
300
400
600
1000
1200
900
1000
800
900
1000
4
4
4
CF
T (s)/M (MB)
2172/9.2
4851/12.8
3790/8.7
6210/10.4
73.4/3.24
143/3.66
256/4.1
0.12/1.75
0.22/1.77
0.54/1.89
REL
T (s)/M (M B)
7368/8.5
15660/11.8
22169/14.1
34024/17.1
3244/10
5402/12.5
8284/15.1
803/7.7
3422/15.8
26245/49.1
The results given for random polynomials, Table 2, were averaged over sets
of 5 random polynomials each, both methods were tested on the same sets of
301
randomly generated polynomials. When all coefficients were randomly generated

integers CF was faster by factors between 1.5 and 5.
Table 2. Polynomials with randomly generated coefficients
Coefficients
(bit length)
10
10
10
1000
1000
1000
Degree
500
1000
2000
500
1000
2000
No. of roots
(average)
3.6
4.4
5.6
3.2
3.6
6.0
CF
T (s)/M (MB)
0.78/2.2
6.67/3.75
215/11.4
0.56/2.28
12.7/5.1
329/14.2
REL
T (s)/M (MB)
1.66/2.81
34.2/7.5
562/22.8
2.19/2.97
31.4/6.5
510/24.3
The case of monic polynomials, Table 3, with randomly generated large integer coefficients, at lower terms proved to be especially hard for REL. In this
case CF was several thousand times faster. This is because such polynomials tend
to have both very large and small roots, so an isolation method based on interval
bisection starts with a very large interval, and needs to bisect it many times before
it isolates the small roots. CF does not have this problem, because the size of its
each step is based on an estimate of how far the next root is.
Table 3. Monic polynomials with randomly generated coefficients
Coefficients
(bit length)
10
10
10
1000
1000
1000
1000
Degree
500
1000
2000
100
200
500
1000
No. of roots
(average)
5.2
4.8
6.8
4.4
6.0
5.6
6.0
CF
T (s)/M (MB)
1.43/2.48
7.12/3.74
263/11.4
0.01/1.75
0.086/1.93
0.57/2.28
25.5/5.2
REL
T (s)/M (MB)
8.84/3.84
80.7/10.1
1001/37.1
56.8/5.5
252/17
1917/96.8
>5000/?
For polynomials with all roots being randomly generated integers, Table 4,
CF was up to 25 times faster for small roots, but REL was up to 4 times faster for
large roots. The latter being the only case when we found CF to be slower than
REL.
302
Table 4. Products of factors (x-randomly generated integer root)

Coefficients
(bit length)
10
10
10
1000
1000
1000
Degree
No. of roots
100
200
500
20
50
100
100
200
500
20
50
100
CF
T (s)/M (MB)
0.8/1.82
2.45/2.07
33.9/3.34
0.12/1.88
16.7/3.18
550/8.9
REL
T (s)/M (MB)
0.61/1.92
10.1/2.64
878/8.4
0.044/1.83
4.27/2.86
133/6.49
3 Conclusions
We have shown that our continued fraction root isolation algorithm CF is almost
always faster than the algorithms based on interval bisection. Its bound on memory usage, given in terms of the number of transformed polynomials it needs to
keep, is not much worse then for the algorithm REL presented in [3], and in
practice its memory usage is often smaller than that of REL.
The link http://members.wolfram.com/webMathematica/Users/
adams/RootIsolation.jsp gives access to both isolation methods with one
caveat: the memory comparison does not work too well. Probably due to the fact
that webMathematica is using a kernel shared by several users, if somebody had
run a memory intensive computation before, MaxMemoryUsed will return the
memory used by that computation, and will not change after our test examples.
References
1. G. E. Collins, A. G. Akritas. Polynomial real root isolation using Descartes rule of
signs, in: Proceedings of the 1976 ACM Symposium on Symbolic and Algebraic
Computations, Yorktown Heights, N.Y., pp. 272275, 1976.
2. A. J. H. Vincent. Sur la resolution des quations numriques, Journal de Mathmatiques Pures et Appliques, 1, pp. 341372, 1836.
3. F. Rouillier, P. Zimmermann. Efficient isolation of polynomials real roots, Journal
of Computational and Applied Mathematics, 162, pp. 3350, 2004.
4. A. G. Akritas, A. V. Bocharov, A. W. Strzebonski. Implementation of real root
isolation algorithms, in Mathematica, in: Abstracts of the International Conference
303
on Interval and Computer-Algebraic Methods in Science and Engineering

(Interval94), St. Petersburg, Russia, March 710, pp. 2327, 1994.
5. A. G. Akritas. An implementation of Vincents Theorem, Numerische Mathematik,
36, pp. 5362, 1980.
6. A. G. Akritas. The fastest exact algorithms for the isolation of the real roots of a
polynomial equation, Computing, 24, pp. 299313, 1980.
7. A. G. Akritas. Reflections on a pair of theorems by Budan and Fourier, Mathematics
Magazine, 55, pp. 292298, 1982.
8. A. G. Akritas. Elements of Computer Algebra with Applications, Wiley, New York,
NY, 1989. Available also in Russian, MIR Publishers, Moscow, 1994 (with new
material).
9. J. von zur Gathen, J. Gerhard. Fast Algorithms for Taylor Shifts and Certain
Difference Equations, in: Proceedings of ISSAC97, Maui, Hawaii, U.S.A.,
pp. 4047, 1997.
304
Author: Alkiviadis G. Akritas

Title: A Correction on a Theorem By Uspensky
Creator: HDML
Anistoriton Journal, vol. 13 (2012-2013) Essays
Alexandre Joseph Hidulphe Vincent

on George Gemistos Plethon
Abstract
George Gemistos Plethons
work in all its dimensions has attracted
many scholars across the ages. One of
those scholars was Alexandre Joseph
Hidulphe
Vincent,
a
French
mathematician and erudite, who in the
first and the only critical edition of
Plethons Book of Laws by C.
Alexandre in the nineteenth century,
added three notes on his calendar,
metrics and music, as he could
reconstruct them from the ancient text.
Vincents calculations were dictated by
the main scientific thought of his time,
which was Positivism, and through this
he thus contributed to the elucidation of
some practical aspects of Plethons
metaphysics. The results of meticulous
calculations of the Plethonian calendar,
metrics and musical modes show that the scientific spirit, which started appearing in
the last days of Byzantium and during the Renaissance, was not only a revival of
Antiquity, but an innovative attempt at explaining and being in accordance with the
social demands and the physical reality of the time, both being understood as an
extension of the metaphysical order. Vincents positivistic approach allows
considering the impact of Plethons system in a postmodern perspective.
Keywords: Metaphysics, Mathematics, Positivism, Calendar, Metric, Music,
Byzantium
Photo: Alexandre Joseph Hidulphe Vincent
Contemporary rediscovery of A. J. H. Vincent
While the work of Alexandre Joseph Hidulphe Vincent remained in obscurity
for a century and a half, there has been, in the past few decades, a renewed interest in
the special contribution made by him to mathematics, which helped modern scientists
to find an efficient solution to an old mathematical problem. Especially, his study
Sur la resolution des equations numeriques in the Journal de Mathematiques Pures
et Appliquees in 1836 was used again in 1976, when the continued fractions real root
isolation algorithm was introduced, for a new solution to the polynomial equation
(Akritas, 1976-2010, Vigklas, 2006-2010). The use of Vincents solution to this
algorithm brings the old French mathematician and polymath into the spotlight again.
http://www.anistor.gr/index.html
But who was Alexandre Joseph Hidulphe Vincent? He was born in Hesdin
(Pas de Calais) in France on 20 November 1797, and studied in the college of
Donai and Amien. He passed the examinations for the Ecole Normale in 1816, from
which he graduated in 1824, with the title of agrg. Then he taught physics,
chemistry and natural history in the Royal College of Reims. His first publication was
in 1824 in the journal Annales de mathematiques, a remarkable study under the title:
Considerations nouvelles sur la nature des courbes exponentilles et logarithmiques.
His book, Cours de gometrie lmentaires, was famous and was approved by the
University, and went on to produce five further editions. In 1825 he published the
Dialogue sur la loterie, which was honored by the Socit des Morale Chrtienne.
Soon he was called to Paris to teach in different colleges, in Rollin in 1826, Bourbon
in 1830, and Saint-Luis in 1831 (in the latter he was named as a professor in special
mathematics). In general, he published more than ninety studies on Geometry,
Mathematics, Physics, Music, Archeology, Philology, Prosody, History, Geography,
Philosophy, Criticism of Literature, and other scientific areas. He became a member
of the Legion d Honneur, of the Acadmie des Inscriptions de Belles Lettres in 1850,
and of many other orders, societies, and many academies of foreign countries, before
he died, on 26 November 1868 (Vapereau, 1858, 1737-1738, Resbecq, 1869).
Though the main subjects of his investigations were included in the field of
the history of science, he elucidated many aspects of the Ancient, Neoplatonic and
Byzantine philosophy, as he studied the works of Heron the Alexandrian, Pappus the
Alexandrian, Neron the Alexandrian, Bacchius, Proclus, George Pachymeres, Michel
Psellos, Aristide Quintialianus, Plato, Augustine, Pythagoras, Boethius, Julius the
Africanus, Eratosthenes, Euclides etc. In what follows, Vincents mathematical
calculations of the Byzantine scholars calendar, metrics and music, as are included in
the above book, are outlined, in order to show their historical significance for
Byzantine thought and the history of science.
Vincents three notes on the Plethonian calendar, metric and music
In the important critical edition of Plethons Book of Lows by Charles
Alexandre, published in 1858, Vincent contributed with three notes on Plethons
calendar, rhythmics and metrics, and musical modes (Alexandre Pelissier, 1858,
444-470, Woodhouse, 1986, 316-379). Also, he cited a small annotated and translated
text of Plethon, regarding the metrics (kefalaia atta logn mousikn), which he had
discovered in the Bibliothque Impriale of France (Vincent, 1849, 234-241, Hunger,
1994, 428, Benakis, 2002, 179-186). Vincents analysis does not remain within the
metaphysical elements of Plethons Book of Laws, but describes the positivistic
structure of time, metric and music, inherent in his metaphysics. This structure is not
useless or secondary, but constitutes the practical foundation for metaphysics
application to everyday life. Without the calendar, the rules of metrics, and the
harmony of music for the recitation of hymns, the Plethonian liturgy could not
function properly, and the beneficial communication with the divine would fail. Those
practical rules for the calculation of time and liturgy were as vital links in the
hierarchy of the genera of beings, as they connected the sensible with the intelligible.
Without this practical structure, Plethons metaphysics would be incomplete, and the
salvation of his country could not be accomplished according to his final inapplicable
plans.
In the first note, Vincent presented the Plethonian calendar (Anastos, 1948,
183-205, Kandz, 1950), 199-210), which he supposed was based on the Athenian
calendar, except that in the former, the solstice starts in the winter. The years have
twelve months, but under the condition that one thirteenth month is inserted seven
times every nineteenth year (the metonic cycle). He examines in detail the innovations
of this calendar, which is solar-lunar, the days, the weeks, the months and the years.
The weeks of Plethon start constantly every month, with the addition of one or two
complementary days after four weeks have elapsed, to reach the beginning of the
following month, and related in that manner approximately to the phases of the moon.
The first day of the month, begins at midnight after the new moon, and for that it is
named the noumeny (the new moon), which was hieromeny (a holiday) with six
days 2, 3, 4, 5, 6, 7 which are followed by a second hieromeny, starting by the eight
day of the month, and after the first week, which was called the beginning week.
From the second hieromeny the second week started, called the average week, of
which every day could be counted by downgrading from the eighth day, and
numbering 7, 6, 5, 4, 3, 2 and 1. The dichomeny or the middle moon includes the
second and the third week, which is called the waning week, before the month ends
with the last week. That month is a hollow month, of twenty nine days, and the last
day was called , that is moon both old and new. If there was a full moon
of thirty days, that is the new moon after the twenty nine days, that name is
transported to the thirty-day name.
As for the calculation of months, it always starts from the midnight of the sunmoon conjunction. For the succession of the months, if each lunation was exactly 29
days and a half, the months would be perpetually of 30 and 29 days alternately. But
the lunation average is in reality of 29 days 12 h 44 2, 9, or in round numbers, 29
days 12 h 45, giving three quarters of an hour longer, but one small error of 57, 1,
less than a minute, which if it was considered in isolation should be repeated 1,500
times to make an entire day, and consequently could produce one day's error on the
calendar at the end of 1,500 lunar months or after about 120 years. If the lunation is
supposed to be three-quarters of an hour longer than 29 days and a half, the full
months will be more frequent than the hollow months. In fact three quarters of an
hour, form exactly the thirty-second part of a day, so, it follows that 32 lunar months
have an exact number of days, namely 945, which are decomposed into 17 times 30
plus 15 times 29.
As a complement to this, Vincent produced a table with twenty months, to
which he added these calculations, pushing them for up to 20 months. He defined the
first two consecutive months of 30 days, because he assumed that the fist new moon,
from which the origin must be dated, falls at midnight, which always happens for the
rest of the new moon in a certain place on the Earth or rather for a certain determined
meridian. After these first two months of thirty days, the others are alternately full and
hollow up to the sixteenth and seventeenth months (that is to say to the 3o and the 4o
of the second year), which are also both full, and from those two later, the regular
alternation would continue until the thirty-second month (or the 7o of the third year),
which is a hollow month, and at the end of which the new moon occurs at midnight,
as at the beginning of the first month. The thirty-third month begins as in a series as in
the first and so on indefinitely or for 120 years at least, without alteration or error
other than the determination of the initial point.
The calculation by Vincent of the Plethonian year is made in comparison to

the Julian year. The Plethonian cycle of the nineteenth years starts from the year
which would start the day after the solstice, and the leap years are the 4, 8, 12, 16, 20
etc. The first year starts at midnight, some moments after the solstice, and the twelve
next lunations form an ensemble of 354 days and 9 hours; the twelfth conjunction
falls eleven days before the solstice of the next year, and it cannot consequently start
in a new year. The thirteen year consists of a) the eleven days that remain between the
end of the twelfth month and the solstice, and b) the days that remain after the solstice
in order to reach the new month or the first month of the next year, which may not
begin until January 10, the day after the new moon on the thirteenth and twentieth day
after the solstice. These, later, are the intercalary days. But off course, as Vincent
observes, the movement of the moon is far from having reached the regularity that the
writer describes in his tables, especially due to the variation of the distance of the
Earth from the Sun. Thus, the Plethonian lunar-solar calendar is established on the
value of the lunar month average, as the most suitable method. So, according to the
metonic circle, since there is a thirteen month, among the nineteenth years of the
circle, seven embolimic years should be added, and the new moon which defines the
end of the nineteenth year falls 20h and in the same day that begins the New Year.
The Plethonian calendar was never practiced in reality, thus the calculations
concerning it are characterized by a degree of approximation.
As one can see, Vincent accepted Plethons intercalation of the metonic circle
of the enneakaidekateris, which was introduced by the Athenian astronomer Meton,
in 432/2 B. - C., and which required the intercalation of a thirteen month for seven
years in every nineteen. At this point Milton Anastos raised an objection, saying that
Plethon is not implausible as he would have rejected the inaccurate octaeteris, of the
8-year circle, the Callippean 76-year circle, or the Christian form of
enneakaidekateris (the 532-year circle = 19 x 28). But if he had studied Ptolemy's
Mathematike Syntaxis, it is difficult to see why he would have repudiated the
Hipparchan 304-year cycle, with its extraordinarily accurate computations for the
length of the solar year and of the lunar month (Anastos, 1948, 202). The lost sections
of Plethon's chapter on the calendar would undoubtedly have been explicit on this
specific problem. As it is, however, since the intercalation of a thirteenth month
approximately every three years would have been required whether Plethon had
depended upon his own astronomical observations or upon any one of the principal
lunar-solar cycles, he cannot be sure which of these two expedients he preferred.
Consequently, Anastos reaches the conclusion that Plethons New Year date, based
upon observation, would not differ appreciably in any one or two
enneakaidekaeterides which Vincent calculated (Anastos, 1948, 204).
Regarding the metrics and rhythmics, Vincent remarked on the acceptance that
the long syllable can have more than two musical times, in accordance with the
doctrines of antiquity. In the heroic verse the foot starts with the thesis (foot down),
and finishes with the arsis (foot up), in contrast to Aristoxenus, Diomedes, Sergius
etc. for whom the opposite was more reasonable. Vincent approaches the doctrines of
Plethon through the theory of Saint Augustine, who added at the end of the hexameter
verse a silence of two times, equal in durance with a long syllable, and which
reunifies this silence with the last syllable of the verse in order to compose one
rhythmic spondee; then, it forms one anapest with two short syllables of the last

dactyl, and the first long not of the last spondee, continuing in the same way, and
following a retrograde course. Vincent, in order to answer the question as to whether
the arsis corresponds to the fort sound, and the thesis to the light sound, or vice versa,
goes back to Marius Victorinus, who says to the contrary that arsis is soundless, and
that thesis is full of sound. The same correspondence was applied to Virgil, and the
poets of the School of Alexandria.
The musical modes of the Plethonic liturgical system, according to Vincent,
are phrygian, dorian, hypophrygian and hypodorian, which correspond respectively to
the octave of re, mi, sol and la. The French mathematician tried to find the real
correspondences of Plethons musical modes to the ancient music, but also to the
ecclesiastical music, which was known in his era. Thus, according to him, the real
ancient notes that Plethon used, and which are not in correspondence with the
preceding four modes, are the note re for the hypodorian, the note mi for the ancient
mixolydian (which Vincent identifies with the dorian of Plethon, irrelevantly if there
are now two notes under the same name), the note sol for the Phrygian, and finally,
the note la for the dorian. The four modes have a special function, since they are
attributed to different levels of being: 1) the hypodorian is attributed to Zeus, which is
our minor mode; 2) the hypophrygian is attributed to the Olympian gods, and is
equivalent to the major mode without a sensible note; 3) the Phrygian is attributed to
the inferior gods, and the minor mode with a major sixth; 4) the dorian, which
coincides with the ancient mixodylian, refers to the mystic mode as it is known in the
Latin Church. According to Vincent, Plethon rejected the chromatic and enharmonic
genera of the Ancient Greeks, as the nuances of the diatonic genus, which are
different to the diatonicity, the only one admitted to by Plato. The general historical
conclusion is that Plethons system was composed by the noblest modes of Ancient
music.
Anastos, ninety years after Vincents notes were formalized, parallelized
Plethos hymns included in the Book of Laws, with the pedantic hymns of Proclus,
pseudo Orpheus, and Julian. He agreed that each of the hymns was to have been
accompanied by music, the range of which was limited to four modes, bearing the
ancient names in adverbial form (, , , ).
Anastos repeated Vincent observations that the nomenclature of Plethons music was
derived from his studies of Aristoxenus and Aristides Quintilanus, and had some
affinity with the section of the Republic (3, 398E 399C) of Plato. But he concluded
that Plethon imitated the Greek rites, as was the Byzantine liturgical practice, of the
eight sounds of the Christian liturgy, which Manuel Bryennius and John Cucuzeles in
the fourteen century had already equated with the eight modes of their pagan forbears
(Anastos, 1948, 267-8 and note 428).
The aspects of the Plethonean work elucidated by Vincent were necessary for
the function of the pagan liturgy which Plethon had imagined and composed via the
use of many Greek sources. This is also proved from the fragment of the Book of
Laws containing the calendar which occurs in the chapter entitled
. This fact in itself shows that Plethon regarded, as did the ancient Greeks,
the calendar as intimately bound up with the cycle of religious festivals (Anastos,
1948, 192). The defining of five daily prayers and twenty-eight hymns, guided by the
example of pagan literature, in which prayers and hymns abound, were composed for
the use of the subjects of the ideal state. These propositions were never applied, as it

is well known, on a general level, because of the dangerous and fluid social and
political situation of the last years of Byzantium. However, it is testified by historical
sources, that a secret society () (Wilson, 1994, 106-108) had probably been
established, a characteristic member of which was Iouvenalios, whose life had a tragic
end, and who can be considered as a martyr of the movement (Constatelos, 1999).
Thus, this reformation of the calendar, metrics and music in the last years of
Byzantium could have been applied on a small scale inside this secret society of the
philosophers followers.
An interpretive perspective on Vincents notes
Vincents notes, observations and calculations aimed at clarifying the above
fields of knowledge, as were developed by the philosopher of Mystra. His scope was
of course scientific, and this is why he meticulously examined Plethons systems of
the calendar, metrics and music. Being a mathematician and a scholar at the same
time, Vincents interest falls on the aspects of Plethons work which were close to the
scientific and positivistic spirit, as spread widely during his century. The way in
which he analyses the structures of Plethons efforts to constitute a corpus of
knowledge is very important, as he tries not only to imitate the erudition of the ancient
philosophers, but to revive them in his time. It is well known from Plethons
biography of the tireless and repeated efforts he made to find a solution to the
problems the Byzantine State had to deal with. It was from those exact efforts that his
scientific vision was encapsulated, that brought about the real enlightenment of the
Byzantines, and the path out of their historical, social and economic crisis was
through their baptism in the reviving waters of this ancient wisdom.
Vincent does certainly remain holding this plausible conclusion. He agreed
that Plethon aimed at the revival of ancient Greek culture, and he examined Plethons
work from the inside, working as much as a historian of science, following the
scientific ideals of his time, and showing in part the differences and the similarities
with Plethons predecessors as Milton Anastos would do later. Vincents special
interest was in the way Plethon solved the abstract and mathematical relations
between the elements of reality, for the creation of a sustainable calendar, a new
metric for the expression of poetic meaning, and more successful musical modes.
Though the Neoplatonic philosopher used the ancient terms and ideas for the reform
he tried to bring into his society, he applied to them the new content dictated by the
conditions of his era. In this sense, the work of Plethon was an attempt to give an
answer to the real social and spiritual problems of the Byzantines, who were
confronted with a serious and catastrophic crisis after one thousand years of their
Empires history. Auguste Comte, who was a positivist philosopher in the nineteenth
century, influenced Vincents thought, and as Plethon, intended to bring a social
reform, based on scientific criteria. Also, Comte had a project for the Religion of
Humanity, a Positive Church that would provide the scientific-humanist equivalent to
what systematic theology had been in the high Middle Ages (Wernick, 2003, 2);
Plethon analogously planned to implement a religion of ratio through the common
notions (Bargeliotis, 1980, 43). Vincents positivist approach may reinforce the view
that Plethon's goal was not just antiquarian, but practical as well. That Plethon aimed
at finding a correspondence of the Neo-platonic hierarchies with the sensible realities,
via the forming of new scientific tools, cannot be ignored; he gave a boost not only in
the way of perceiving the real, but also of handling it. This connection is useful for

establishing a pragmatist perspective (Kolakowski, 1976, 174) on the way Plethons
work should be interpreted.
Vincent in his calculations, even if he focused on the relations and analogies in
Plethons structures, succeeded in illuminating some aspects of the history of ideas
and science, which refer to the changes that Plethon intended to apply to the society of
Byzantium. Plethon tried to combine the Neo-platonic pagan believes and knowledge
with the scientific spirit and method. Consequently, Plethons plans for reform were
not consisted only of idealistic constructions. Plethons intellectual changes implicate
a dramatic historical situation, when everyone and everything changed quickly. If the
Byzantine people had returned to the pagan past of Ecumeny, they would have found
a path out from their problems, only by reinterpreting the ancient knowledge through
a new perspective. The Plethonian work lucks intellectual vacuum and emptiness, it is
not even a romantic and regressive ideology, filled with nostalgia of a mythical pagan
era. Vincents analysis shows a scientific innovative spirit which is in dialogue with
reality, and for which the return to the past is the only solution, but only under certain
presuppositions; according to Plethon, the real solution for the Byzantines was to
confront their difficult problems, having coolness of mind, without emotional
entanglements in the doctrine, especially if these were coming from the Christian
Church (Dedes, 1975-76, 424-441).
The Plethonian combination of intellectualism and pragmatism was an intense
and superhuman effort to awake the Byzantines, to surpass the religious defeatist
fatalism, according to which the end of the Byzantines was coming due to the sins
before God. Plethon avoided being trapped in the belief of an inescapable fate since
human being is in the methorion, in the border between the sensible and the
intelligible realm, maintaining the hope to ascend in the ontological and axiological
scale. In the end of the Middle Ages, the ideal of anagoge () of the whole
sensible world to the intelligible, did not remain tied to the religious otherworldliness. It changed towards an art of life, when most of the metaphysical and
scientific theories dealt with the cultivation and the development of the individual
(Burckhardt, 1961 (1860), Part II). If for the Byzantine thought the distance between
the two levels, the sensible and the intelligible, grew larger, for the Westerns
diminished. The effort for the unification of these two levels is evident in Plethons
work, if inside his normative and educational program the reformation of the calendar,
music and rhythm is included. Thus, our point of view is that Plethon aimed at
reviving the ancient Greek doctrines and the ancient way of life, but not as they were
exactly, and in distinction to Christianity, in connection with a scientific, educational
and normative transformation of life. Consequently, utopianism, as an indispensable
element of his thought, is not so much enlarged as it has been written, and it has
meaning only in a pragmatist framework. As the Plethonian thought was on the
boundary of transition from the closed worldview of Antiquity and the Middle Ages
to the open universe of Modern Times (Koyre, 1957, ch. V), primarily referred to a
metaphysical ontology. But Plethons platonic Realism was not a groundless
metaphysics; it could be tested in relation to the sensible reality, in the interpretation
and change of which it intended.
Plethons philosophical goal is the Platonic Good as the true knowledge, but
without neglecting the change of the totality of the way of living. The concept of
infinity which had no position in the Platonic world of ideas, more intensely and

openly in the work of Plethon, found its culmination, not only in the Neo-platonic
One or in the Christian God, but in the possibility of attainment of truth and
perfection. The possibility of practical implementation can be regarded as an
expression of the will for power in the sense of the transfiguration of human
existence. Vincents interpretation revealed not only the way Plethon conceived the
world, but also his planning for the organization and normalization of life. Behind his
hierarchical paganism there is an ontological level that generates and forms the
sensible world, which is destined to return through the climax of Being into the place
of truth. But in order this truth not to be conceived as a private reverie, has to find its
immediate correspondence and similarity in the world of action. The world of spirit in
order to be materialized should include every aspect of human life, to spiritualize it,
conquering the reality. The perishable human nature cannot stay as it is; it has no right
to be perished. The big challenge is to become a collective consciousness, finding a
similarity with the metaphysical and astral order. Every aspect of this world should
reflect the supersensible world, neither as a substitute, nor as duplication, but through
an ontological verification: its value is dependent on the way it imitates the ideal state
of being. Nothing can exist for itself, the human actions have to be rational, and this is
the only way to sociability. The scientific goal of Plethon is the transformation of life
according to the eternal ideas, but the homogeneity of action is not an end in itself, it
always has to refer to the infinity of possibilities opened to the world of senses.
Man is included in this organization of the way of life in order his
consciousness to be controlled. The process of individualization passes through the
complexity of the Plethonian structural hierarchy, and the action is its necessary
outcome, as it is the formation of interiority. Through his positivism and erudition,
assisted by his mathematical methodology, Vincent interested in the empirical
realities of Plethons work and perhaps unintentionally unfolded the ideal
development of the subject in the pagan intellectualism of the latter. The human
existence was determined externally in order to be determined internally. It is not just
essentialism, definition of the general types in the most perfect degree (
) (Plethon and Tambrun-Krasker, 1987, 1); the aim is the
elimination of the differentiation. Although as regards the human subject is given the
promise of openness to the diversity and multiplicity that implies the world of matter,
at the same time in absentia of the subject its objectives is established.
The mathematization of the Plethonian work could so easily lead to
mechanization, as happened at the time of Vincent. Thus, the individuals, although
they are ensouled entities, are confused with the objects, and they can see themselves
mediated by the world of a hierarchy and a logical structure. The natural and social
beings in their becoming are deprived of their autonomy, obtain meaning as
othernesses. The ultimate otherness for Plethon is God or Zeus, which allows the soul
to move within an organized framework, and gives the raison d'tre, the cause and the
purpose, without which the world would be meaningless. The genesis and the end of
the world are opened to eternity. The infinity is the horizon of sensible beings, and at
the same time it is beyond the intelligible reality. Now, as regards the axis of
Vincents methodology were the classification and an endless chain of calculations,
not the contemplation and the reflection on the metaphysics and the ultimate. He was
not interested in connecting the results of his detailed calculations with a meaning of
existence (Kolakowski, 1976, 11-19). Only the human beings can give meaning to the
results of positivistic research, and to the events of the historical becoming. Although

according to Positivism any positive existence is always in reference to something
else, this should be also positive, empirically checked. In this respect, the Positivism
of Vincent meets the postmodern project of a history of problematization (Foucault,
1984). The classification and the organization of forms are sine qua nons in his
epistemology. He is not interested in an ontology, but in a morphology. The existence
of beings does not emanate from the Forms or Ideas, platonically understood, but it is
considered as a system of relationships between factors, numbers, distances, modes,
natural realities, phenomena.
Consequently, the difference between the subject and the object has no
meaning. If the real Being in Plethons work is the intelligible world, in Vincents
work it is the world of numbers. For Plethon the truth of individual consciousness is
related to the pantheon of paganism, which is viewed as an objective order of
meaningful types, in the extension of, and inside the sensible world; for Vincent, the
reality is truthful, if it is reduced to its abstract components. The point of their
convergence is not only the historical, the scientific, the human affairs, and the
metaphysics; it is the effort to define a symbolic and timeless order, where there is no
difference between the way of thinking and the way of life, in a utopia where there is
no difference between the thinking subject and the object of thought.
Thus, Plethons Platonic Realism and Vincent's Positivism converge at
another point, in the postmodern era, in the light of the symbolic order (Deleuze,
1972). In the space-time, these three points that are the fifteenth century of Plethon,
the nineteenth century of Vincent, and todays later Postmodernism, can be connected
in a form that would resemble a trihedron, each one epoch corresponding to each of
its sides. In the symbolic realm, what counts is not only the distinction between the
sensible and the intelligible or transcendental, but the structural relations, the typical
characteristics, the production of forms; it is a condition under which the essence has
no meaning in itself, but only in connection with the form, or the form remains empty.
Epilogue
Vincents work suggests his intense attention to detail and the responsible way
in which he interpreted the above-mentioned scientific subjects. Some unknown
dimensions of Vincents works have only come to the surface today, such as his
polynomial real root isolation algorithm along with other information of the distant
and near past. Though today there are an incredible number of studies on the
disciplines with which Vincent occupied his mind, due to an increasing specialization,
he gave to the scientific community the valuable results of his erudition, and he can be
considered as a pioneer of the history of science in Byzantium. The French erudite
and mathematician undertook the effort to study some scientific aspects of Plethonian
work, which were as significant for Plethon as they were for his own era, and they are
still significant today. He was a scholar, whose reception of the Plethonian work, was
influenced by the historical conditions of his time and his place. Under the influence
of the positivistic spirit of the nineteenth century, he brought to light the calendar, the
metrics and rhythmics, and the musical modes described by Plethon, in which we can
see, except for a revival of the ancient Greek culture, a development of science and of
the individual, the outcome of which was the intellectual atmosphere of Modernity,
and of the later Postmodernity.
Katelis S. Viglas, PhD (History of Philosophy)

Aristotle University of Thessaloniki, Greece
kviglas@hol.gr
BIBLIOGRAPHY
Akritas, 1976, A. Akritas, Polynomial real root isolation using Descartes' rule of
signs. Proceedings of the 1976 ACM.
Akritas, 1978a, Vincent's theorem and Akritas' polynomial real root isolation
algorithm. Proceedings of the North Carolina Academy of Sciences. In: Journal of
the Elisha Mitchell Scientific Society, Vol. 94, No. 2, 95, 1978.
Akritas, 1978b, A new method for polynomial real root isolation. Proceedings of the
16-th annual southeast regional ACM conference, Atlanta, Georgia, 39-43, April
1978.
Akritas, 1978c, A correction on a theorem by Uspensky. Bulletin of the Greek
Mathematical Society. Vol. 19, 1978, 278-285.
Akritas, 1978d, On the forgotten theorem of Mr. Vincent. Historia Mathematica, Vol.
5, 1978, 427-435.
Akritas, 1978e, A short note on a new method for polynomial real root isolation.
ACM-SIGSAM Bulletin, Vol. 12, No. 4, 1978, 12-13.
Akritas, 1978f, Vincent's theorem in algebraic manipulation. Ph.D. thesis, Operations
Research Program, North Carolina State University, Raleigh, N.C., 1978.
Akritas and Vigklas, 2010, A. Akritas P. S. Vigklas, Counting the Number of Real
Roots in an Interval with Vincent's Theorem. Bulletin Mathematique de la Societe des
Sciences Mathematiques de Roumanie, Vol. 53 (101), No. 3, 2010, 201-211. See:
http://inf-server.inf.uth.gr/~akritas/publications.htm
Anastos, 1948, Milton Anastos, Plethos Calendar and Liturgy, Dumbardon Oaks
Papers, Vol. 4, (1948), 183-205.
Bargeliotis, 1980, . . ,
.
, - . 1.
. , 1980.
Benakis, 2002, Linos Benakis, in Texts and Studies on Byzantine Philosophy,
, 2002, 179-186.
10

Burckhardt, 1961 (1860), The Civilization of the Renaissance in Italy, S.G.C.
Middlemore, New American Library, New York, Part II. The development of the
Individual
Constatelos, 1999, D. Constatelos, Christian Hellenism. Essays and Studies in
Continuity and Change Publishing by Aristide D. Caratzas, New Rochelle, New York
& Athens 1999, Chapt:. A Conflict between Ancient Greek Philosophy and Christian
Orthodoxy in the Late Greek Middle Ages.
Dedes, 1975-6, . , , ,
5-6 (1975-76), 424-441.
Deleuze, 1972, A quoi reconnait-on la structuralisme? , F. Chtelet (ed.), Histoire
de la Philosophie, tom. 8: Le XXe sicle, Hachette, Paris 1972, 299
Hunger, 1994, H. Hunger, .
, . , ..., 1994.
Foucault, 1997 (1984), M. Foucault, Interview: Polemics, Politics and
Problematizations, in Essential Works of Foucault, 19541984, Vol. 1, Edited by
Paul Rabinow, Transl. Lydia Davis, The New Press, New York 1997
Kandz, 1950, Solomon Kandz, The Calendar Reform of Pletho: Its Significance
and its place in the History of the Calendar, Osiris, vol. 9, (1950), 199-210.
Koyre, 1957, Alexandre Koyre, From the Closed World to the Infinite Universe, The
John Hopkins Press, Baltimore 1957
Kolakowski, 1976, L. Kolakowski, La Philosophie Positiviste. Trand. Claire Brendel,
Denol/Gonthier, Paris 1976
Plethon, 1858, Plthon Trait des Lois, par C. Alexandre et A. Pelissier, Librairie de
Firmin Didot Frres, Fils et Cie, Paris 1858, 124. New edition by R. Brague, J. Vrin,
Paris 1982.
Plethon and Tambrun-Krasker, 1987, , ,
Georges Gmiste Plthon, Trait de Vertues, Edition Critique avec Introduction,
Traduction et Commentaire par Brigitte Tambrun-Krasker, Corpus Philosophorum
Medii Aevi: Philosophi Byzantini - 3,
, Athens- The Academy of Greece. E.J. Brill, Leiden- New York
Kbenhavn Kln 1987
Resbecg, 1869, Comte Eugene Hippolyte Marie Theodore Fontaine de Resbecq,
Notice sur la vie et les travaux de feu M. Vincent, L. Danel, Lille 1869.
Vaperau, 1958, Gustave Vapereau, Dictionnaire Universel des Contemporains,
contenant toutes les personnes notables de la France et des pays trangers. Librairie
de L. Hachette et Cie, Paris 1858, 1737-1738.
11

Vincent, 1847, A. J. H. Vincent, Notice sur divers manuscrits grecs relatifs la
musique, la deuxime partie du tome XVI des Notices et Extraits de Manuscrites de la
Bibliothque du Roi et autres Bibliothques, Institut Royal de France, Imprimerie
Royale, Paris M DCCC XLVII, 234-241.
Wernick, 2003, Andrew Wernick, Auguste Comte and the Religion of Humanity. The
Post-Theistic program of French Social Theory. Cambridge University Press,
Cambridge 2003
12
Bull. Math. Soc. Sci. Math. Roumanie

Tome 53(101) No. 3, 2010, 201211
Counting the number of real roots in an interval with

Vincents theorem
by
Alkiviadis G. Akritas and Panagiotis S. Vigklas
Dedicated to the memory of Laurentiu Panaitopol (1940-2008)
on the occasion of his 70th anniversary
Abstract
It is well known that, in 1829, the French mathematician Jacques Charles
Francois Sturm (1803-1855) solved the problem of finding the number of
real roots of a polynomial equation f (x) = 0, with rational coefficients and
without multiple roots, over a given interval, say ]a, b[. As a byproduct,
he also solved the related problem of isolating the real roots of f (x). In
1835 Sturm published another theorem for counting the number of complex
roots of f (x); this theorem applies only to complete Sturm sequences and
was recently extended to Sturm sequences with at least one missing term.
Less known, however, is the fact that Sturms fellow countryman and
contemporary Alexandre Joseph Hidulphe Vincent (1797-1868) also presented, in 1836, another theorem for the isolation (only) of the positive
roots of f (x) using continued fractions. In its latest implementation, the
Vincent-Akritas-Strzebo
nski (VAS) continued fractions method for the isolation of real roots of polynomials turns out to be the fastest method derived
from Vincents theorem, by far outperformes the one by Sturm, and has
been implemented in major computer algebra systems.
In this paper we use the VAS real root isolation method to count the
number of real and complex roots of f (x) as well as the number of real
roots f (x) has in an open interval ]a, b[.
Key Words: Root counting, real roots, polynomial, real roots isolation, Vincents theorem, Sturms theorem, Sturm sequences, Sylvesters
matrix.
2000 Mathematics Subject Classification: Primary 12D10, 12E05,
12E12; Secondary 26C10.
202
Introduction
The famous theorem by Sturm appeared in 1829, [12], and it solved the problem
of counting the number of real roots of a polynomial equation f (x) = 0, over
a given interval, say ]a, b[. Using this theorem Sturm was also able to isolate
the real roots of f (x); that is, he was able to find open intervals such that each
contains one real root and each real root is contained in some interval. Before we
state this theorem, we need the following definition:
Sturms Sequence or Chain. Let f (x) = 0 be a polynomial equation
of degree n, with rational coefficients and without multiple roots. The Sturm
sequence or chain of f (x) is
Sseq (x) = {f (x), f (x), r1 (x), r2 (x), . . . , rk (x)},
(1)
where f (x) is the first derivative of f (x) and the polynomials ri (x), 1 i
k n1, are the negatives of the remainders obtained by applying the Euclidean
gcd algorithm on f (x) and f (x).
That is, we have:
f (x) = f (x)q1 (x) r1 (x)
f (x) = r1 (x)q2 (x) r2 (x)
..
.
rk2 = rk1 qk (x) rk (x)
When k = n1, that is, when there are no missing terms, the Sturm sequence
is called complete, and when k < n 1, it is called incomplete. The condition
that f (x) has only simple roots is required for the real root isolation process and
it does not restrict the generality of our discussion.
The computation of the Sturm sequence poses several problems when performed in Z[x], since the latter is not a Euclidean domain. In this case we have
to do pseudo-divisions, that is we have to pre-multiply each dividend times the
leading coefficient of the divisor raised to some power for the results to be in
Z[x]. This is the classical Euclidean polynomial remainder sequence (prs) algorithm.
However, pseudo-division leads to explosive coefficient growth, which has to
be controlled. This can be done in various ways; namely, we can use one of the
following:
the primitive prs algorithm, which means dividing out the content of each
pseudo-remainder before using it, [4]. Since this process involves computing
the gcd of the coefficients of each remainder we would like to avoid it.
Counting the Number of Real Roots
203
the Sylvester-Habicht subresultant prs algorithm. Sylvester initiated this

process back in 1853, [14], for complete Sturm sequnces (reduced subresultant prs algorithm) and Habicht wrapped it up in 1948, [9], for incomplete
Sturm sequences (subresultant prs algorithm). In this case without computing any gcds we divide out of each remainder a certain quantity,
knowing a priori that the division will be exact!
the matrix-triangularization subresultant prs algorithm which is equivalent
to the one by Sylvester-Habicht. This was initiated by Van Vleck in 1900,
[15], for complete Sturm sequences, and was wrapped up (within signs)
by Akritas in the 1990s, [1], [2], [3], [4] for incomplete Sturm sequences;
finally, the exact computation of the signs of the polynomials in the Sturm
sequence was achieved by Akritas, Akritas and Malaschonok in 1995, [5],
with the introduction of a new type of resultant matrix.
We can now present Sturms theorem, which makes use of the sequence defined
above:
Theorem 1. (Sturms Theorem of 1829 for real roots) Let f (x) = 0 be a polynomial equation of degree n, with rational coefficients and without multiple roots.
Then the number of its real roots in the open interval ]a, b[ satisfies the equality
= a b ,
(2)
where a , b is the number of sign variations in the Sturm sequence Sseq (a),
Sseq (b), respectively.
This is the theorem for which Sturm is mostly remembered for and its proof
can be found in almost all texts of Numerical Analysis. To isolate the real roots of
f (x) Sturm suggested to first isolate the positive roots and then the negative ones
(by replacing x by x in f (x)), not forgetting to check if 0 is a root. To isolate the
positive roots, all we have to do is to: (a) compute the Sturm sequence Sseq (x),
(b) compute an upper bound ub, on the values of the positive roots of f (x), and
(c) bisect the interval ]a, b[=]0, ub[ until root isolation has been accomplished.
However, there is yet another theorem by Sturm, published in 1835, [13],
which deals with the number of pairs of complex roots. This theorem whose
proof is also in the literature can be stated as follows:
Theorem 2. (Sturms Theorem of 1835 for complex roots) Let f (x) = 0 be a
polynomial equation of degree n, with rational coefficients and without multiple
roots. Then the number of pairs of complex roots of f (x) is equal to the number
of sign variations in the sequence of the leading coefficients of the polynomials in
the tail of the Sturm sequence, where
tail(Sseq (x)) = {f (x), r1 (x), r2 (x), . . . , rn1 (x)}.
(3)
204
The above theorem is true only when tail(Sseq (x)) is complete. As explained
in section 2, a technique was introduced by a group of Chinese mathematicians,
[10], [17], [18], so that Theorem 2 can be used when the tail sequence (3) is
incomplete; they have also extended Theorem 2 in such a way that it not only
counts the number of complex roots of f (x) but can be also used as a criterion
for the number of positive or negative real roots.
It is obvious that with the two theorems by Sturm mentioned above we can:
a. compute the exact number of the real roots f (x) has in an open interval ]a, b[,
b. isolate the real roots of f (x),
c. compute the number of the real and complex roots of f (x).
The first two items, (a) and (b), depend only on Theorem 1, whereas the last
item, (c), depends on both Theorem 1 and Theorem 2.
To count the number of complex roots using Theorem 1, all we have to do
is compute ub, an upper bound on the absolute values of the roots, use (2) to
evaluate the exact number of real roots in the interval ] ub, ub[ and subtract it
from n, the degree of f (x).
On the other hand, using Theorem 2 we can count the complex roots only
in the case when the Sturm sequence is complete; if the sequence is incomplete,
we have to resort either to the process mentioned above using Theorem 1, or to
subresultants, mentioned in section 2.
The rest of the paper is structured as follows:
In section 2 we first introduce Sylvesters matrix and subresultants and then
describe the technique introduced by the Chinese group of mathematicians, [17],
[18] and [10], with the help of which the signs of the missing leading coefficients
in (3) are filled in in an easy to remember manner. In this way, incomplete
tail sequences can be now handled by Theorem 2.
As was pointed out in the literature, [17], [18], [10], subresultants along with
the Chinese technique and extended versions of Theorem 2 are faster and better
suited for polynomials with symbolic coefficients.
In section 3 we present Vincents theorem of 1836 in its original continued
fractions form; with this we can isolate the real roots of f (x), and, hence,
we can compute the number of its real and complex roots. In the sequel we
present the bisection version of Vincents theorem, which was presented in 2000
by Alesina and Galuzzi, [8]; using the transformation mentioned in this theorem
we then present an algorithm to count the number of real roots of f (x) in any
real open interval ]a, b[.
2
The subresultant version of Sturms theorem and the extension of

Theorem 2
As it was stated in the Introduction, it is well known that all the coefficients of
the polynomials in the tail of the Sturm sequence, tailSseq (x), can be computed
205
as subresultants of Sylvesters matrix, [14], [15], [9], [2], [4] and [5]. Since our
purpose is to explain the Chinese technique for making Theorem 2 work with
incomplete sequences, we will concentrate only on the leading coefficients of the
polynomials in tailSseq (x).
Given the polynomial
f (x) = 0 xn + 1 xn1 + . . . + n ,
without multiple roots, and its derivative
f (x) = 0 xn + n 0 xn1 + (n 1)1 xn2 + . . . + n1 ,
the Sylvesters matrix of f (x) and f (x) (which is also referred to as the discrimination matrix, Discr(f ), of f (x), [10], [17], [18] ) is the following 2n 2n
matrix:
0 1
2
0 n0 (n 1)1 n1
0
1
n1
n
0
n
0
n2
n1
0
1
n
0
n0 n1
If by dk (f ) we denote the determinant of the submatrix formed by the first
2k rows and the first 2k columns of Sylvesters matrix, for k = 1, 2, . . . , n, then
the leading coefficients of the polynomials in the tail sequence (3) are
{d1 (f ), d2 (f ), , dn (f )},
(4)
which in the sequel will be called the leading-coefficients sequence our terminology differs slightly from that of our Chinese colleagues, who call it the
discriminant sequence. From (4) we form the corresponding signs sequence,
{s1 , s2 , , sn }, as
{sign(d1 (f )), sign(d2 (f )), , sign(dn (f ))},
(5)
where sign is the known signus function:
if x > 0,
1
0
if x = 0,
sign(x) =
1 if x < 0.
If the signs sequence (5) contains no zero, then Theorem 2 can be used to
count the number of complex roots. Note that because of our assumption that
206
f (x) contains no multiple roots, zeros cannot appear at the end of the signs
sequence.
However, if zeros do appear in (5) then in order to apply Theorem 2 we
have to construct the revised signs sequence, {1 , 2 , , n }, of the corresponding signs sequence {s1 , s2 , , sn }, [10], [17], [18].
The construction of the revised signs sequence proceeds as follows:
If {si , si+1 , , si+j } is a section of the given sequence, with {si 6= 0; si+1 =
si+2 = = si+j1 = 0; si+j 6= 0}, then replace the subsection of the zero
terms {si+1 , si+2 , , si+j1 } by {si , si , si , si , si , si , , }
otherwise make no changes for the other terms.
Example: For the polynomial f (x) = x9 7x + 7 we obtain the leadingcoefficients sequence
{9, 0, 0, 0, 0, 0, 0, 15543853645824, 1556380841389577}
or the signs sequence
{1, 0, 0, 0, 0, 0, 0, 1, 1}.
Obviously, due to the presence of zeros, Theorem 2 cannot be used. However,
the revised signs sequence is
{1, 1, 1, 1, 1, 1, 1, 1, 1}
with 4 sign variations; hence, from Theorem 2 we deduce that f (x) has 4 pairs
of complex roots and 9 2 4 = 1 real root.
Note that the Sturm sequence, Sseq (x), of f (x) = x9 7x + 7 is
{x9 7x + 7, 9x8 7, 504x 567, 1556380841389577},
where the signs of the leading coefficients differ from those of the leadingcoefficients sequence. As a result, we cannot use (1) to obtain the revised signs
sequence!
As stated in the Introduction, when the coefficients are symbolic, it is best to
use the subresultants of the Sylvester matrix as described above.
3
The two versions of Vincents theorem
Vincents theorem, in its original form, can be stated as follows:

Theorem 3. (Vincents Theorem of 1836 the continued fractions version) If
in a polynomial, f (x), of degree n, with rational coefficients and without multiple
roots we perform sequentially replacements of the form
207
1
1
1
, x 2 + , x 3 + , . . .
x
x
x
where 1 0 is an arbitrary non negative integer and 2 , 3 , . . . are arbitrary
positive integers, i > 0, i > 1, then the resulting polynomial either has no sign
variations or it has one sign variation. In the first case there are no positive roots
whereas in the last case the equation has exactly one positive root, represented by
the continued fraction
x 1 +
1 +
1
2 +
1
3 +
(6)
..
In his paper of 1836, [16], Vincent proved this theorem and presented several
examples to demonstrate the concepts involved. However, his theorem appeared
several years after Sturms theorem on counting the number of real roots in
an interval and isolating them. Due to Sturms fame and priority, Vincents
theorem was almost totally forgotten. An interesting account of the history of
this theorem, along with an overview of its various proofs etc, can be found
elsewhere, [6].
The termination of the process described in Vincents theorem is guaranteed
by the following theorem, [11]:
Theorem 4. (Obreschkoffs Theorem of 1920) If a real polynomial has one positive simple root x0 and all the other possibly multiple roots lie in the cone
or sector
S3 = {x = + | > 0 and 2 32 }
then the sequence of its coefficients has exactly one sign variation.
As can be seen from its statement, Theorem 3 can be used to isolate the
positive roots of a polynomial equations f (x) = 0; call VAS positive roots the
corresponding algorithm, [7]. For the negative roots we simply replace x x
and repeat the process.
To isolate the positive roots Vincent computed each partial quotient ai by a
series of unit increments ai ai + 1 which are equivalent to substitutions of the
form x x + 1. This approach resulted in an exponential method and for a
discussion on how this problem was solved we refer the reader to the literature,
[6].
From the above it becomes clear that the continued fractions version of Vincents theorem can be used to isolate the real roots of f (x) and, as a byproduct,
to count the number of its real and complex roots; but it cannot be used to count
the number of the real roots of f (x) in an open interval ]a, b[. For the latter we
need the bisection version of Vincents theorem, due to Alesina and Galuzzi, [8]:
Theorem 5. (Vincents Theorem the bisections version of 2000) Let f (x), be
a polynomial of degree n, with rational coefficients and without multiple roots. It
208
is possible to determine a positive quantity so that for every pair of positive

rational numbers a, b with |b a| < , every transformed polynomial of the form
(x) = (1 + x)n f (
a + bx
)
1+x
(7)
has exactly 0 or 1 variations in the sequence of its coefficients. The second

case is possible if and only if f (x) has a simple root within ]a, b[.
In the sequel we will refer to the transformation mentioned above, as the
(x)-transformation. With the help of the above theorem whose proof can
be found in the literature we can easily count the number of roots in a given
interval. All we have to do is to perform the transformation and then to isolate
the positive roots of the transformed polynomial (x). We are not interested in
the roots themselves of (x); we simply want to know their cardinality.
Note, however, that both endpoints of the open interval ]a, b[ have to be positive; a different approach is stated elsewhere ([18], Theorem 5, p. 145).
As Alesina and Galuzzi pointed out, [8], the substitution of x, in f (x) = 0,
by the continued fraction (6) or, equivalently, the (x)-transformation (7) result
in all roots outside the three circles being placed in Obreschkoffs cone or sector,
shown in Figure 1.
Figure 1: Obreschkoffs cone or sector.

Below is the algorithm for counting the number of real roots in any interval:
209
Algorithm to count the number of real roots in any open interval:

1. If both endpoints are positive, i.e. a > 0 and b > 0, then perform the (x)transformation and use VAS positive roots to isolate the positive roots of
(x); return their cardinality.
2. If both endpoints are negative, i.e. a < 0 and b < 0, then replace x x,
in f (x), along with a a and b b, for the endpoints; after that,
perform the (x)-transformation and use VAS positive roots to isolate
the positive roots of (x); return their cardinality.
3. If a < 0 and b > 0, perform the (x)-transformation with ]a, b[ = ]0, b[
and use VAS positive roots to isolate, and count the cardinality of the
positive roots of (x); next replace x x, in f (x), perform once again
the (x)-transformation with ]a, b[ = ]0, a[, and use VAS positive roots
to isolate, and count the cardinality of the positive roots of (x); return
their total cardinality, adding one if 0 is a root of f (x).
Obviously, the third case is the most time consuming, since it requires two
(x)-transformations.
4
Conclusions
From the above we see that in order to count the number of real roots of f (x) in a
given interval ]a, b[ there are alternatives to Sturms method. The first author was
astonished to see how students exposed to the theorems by Sturm and Vincent,
never think of using the latter for counting the roots. Hopefully our presentation
will bring some change.
References
[1] Alkiviadis G. Akritas, A simple proof of the validity of the reduced prs
algorithm, Computing, 38 (1987), pp. 369372.
[2] Alkiviadis G. Akritas, A new method for computing polynomial greatest
common divisors and polynomial remainder sequences, Numerische Mathematik, 52 (1988), 119127.
[3] Alkiviadis G. Akritas, Exact algorithms for the matrix-triangularization
subresultant prs method, In Erich Kaltofen, Steven Watt (Eds), Proceedings
of the Conference on Computers and Mathematics, Boston, Massachusetts,
June, (1989), pp. 145155.
210
[4] Alkiviadis G. Akritas, Elements of Computer Algebrawith Applications, Wiley, New York, NY, 1989. Available also in Russian, MIR Publishers, Moscow, 1994 (with additional material).
[5] Alkiviadis G. Akritas, Evgenia K. Akritas and Genadii I.
Malaschonok, Matrix computations of subresultant polynomial remainder
sequences in integral domains, Reliable Computing, 1 (1995), pp. 375381.
[6] Alkiviadis G. Akritas, Vincents Theorem of 1836: Overview and Future
Research, In N.N. Vassiliev, A.M Vershik (Eds.) Zap. Nauchnyh Sem. S.Peterburg. Otdel. Mat. Inst. Steklov. (POMI), Teor. Predst. Din. Sist. Komb.
Metody. XVII, 373 (2009), pp. 533. (To be published by Springer Verlag,
in the Journal of Mathematical Sciences.)
ski and Panagiotis S.
[7] Alkiviadis G. Akritas, Adam W. Strzebon
Vigklas, Improving the Performance of the Continued Fractions Method
Using new Bounds of Positive Roots, Nonlinear Analysis: Modelling and
Control, 13:3, (2008), pp. 265279.
[8] Alberto Alesina and Massimo Galuzzi, Vincents Theorem from a
Modern Point of View, In R. Betti and W.F. Lawvere (Eds.), Categorical
Studies in Italy 2000, Rendiconti del Circolo Matematico di Palermo, Serie
II, 64, (2000), pp. 179191.
[9] Walter Habicht, Eine Verallgemeinerung des Sturmschen Wurzelzaelverfahrens, Commentarii Mathematici Helvetici, 21 (1948), pp. 99116.
[10] Songxin Liang and Jingzhong Zhang, A complete discrimination system for polynomials with complex coefficients and its automatic generation,
Science in China (Series E), 42:2 (1999), pp. 113128.
[11] Nikola Obreschkoff, Verteilung und Berechnung der Nullstellen reeller
Polynome. VEB Deutscher Verlag der Wissenschaften, Berlin, (1963).
[12] Jacques Charles Franc
ois Sturm, Resolution des equations algebriques,
Bulletin des Sciences de Ferussac, 11 (1829), pp. 419425.
[13] Jacques Charles Franc
ois Sturm, Memoire sur la resolution des
equations numeriques, Memoires des Savants Etrangers, 6 (1835), pp. 271
318.
[14] James J. Sylvester, On a theory of the syzygetic relations of two rational
integral functions, comprising an application to the theory of Sturms functions, and that of the greatest common measure, Philosophical Transactions,
143 (1853), pp. 407548.
[15] Edward Burr Van Vleck, On the determination of a series of Sturms
functions by the calculation of a single determinant, Annals of Mathematics,
Second Series 1 (1899-1900), pp. 113.
211
[16] Alexandre Joseph Hidulphe Vincent, Sur la resolution des equations

numeriques, Journal de Mathematiques Pures et Appliquees, 1 (1836), pp.
341372 .
[17] Lu Yang, Xiaorong Hou and Zhenbing Zeng, A complete discrimination system for polynomials, Science in China (Series E), 39:6 (1996), pp.
628646.
[18] Lu Yang and Bican Xia, Explicit Criterion to Determine the Number
of Positive Roots of a Polynomial, MM Research Preprints, 15 (1997), pp.
134145.
Received: 06.06.2010
University of Thessaly,
Department of Computer and
Communication Engineering,
E-mail: akritas@uth.gr
E-mail: pviglas@uth.gr
Note on Vincent's Theorem

A. M. Ostrowski
The Annals of Mathematics, 2nd Ser., Vol. 52, No. 3. (Nov., 1950), pp. 702-707.
Stable URL:
http://links.jstor.org/sici?sici=0003-486X%28195011%292%3A52%3A3%3C702%3ANOVT%3E2.0.CO%3B2-Z
The Annals of Mathematics is currently published by Annals of Mathematics.
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained
prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in
the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/journals/annals.html.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is an independent not-for-profit organization dedicated to and preserving a digital archive of scholarly journals. For
more information regarding JSTOR, please contact support@jstor.org.
http://www.jstor.org
Thu Mar 22 04:35:04 2007
HISTORIA
MATHEMATICA
(1978),
427-435
ON THE FORGOTTEN
THEOREM
OF MR, VINCENT
NORTH
BY ALKIVIADIS
G, AKRITAS
AND STYLIANOS
D, DANIELOPOULOS
CAROLINA
STATE UNIVERSITY,
RALEIGH,
NC 27607
SUMMARIES
A little
known theorem concerning
the isolation
of
roots
of polynomial
equations,
published
in 1836 by a
mathematician
known only as Mr. Vincent,
is discussed.
Mr. Vincent's
method is of historical
and practical
interest
because it requires
fewer computations
than
Sturm's
method.
The advantages
afforded
by this
theorem are particularly
relevant
to software
systems
for computerized
algebra.
Certain
computational
results
which offer
an empirical
comparison
of the
two methods are also presented.
Ci'examins
est un thgoreme peu connu relatif
a
l'isolement
des racines
des Equations
polynomes,
publig
en 1836 par un mathdmaticien
connu simplement
sous le nom de Monsieur
Vincent.
La m&hode de Vincent
est d'un int&&
historique
aussi
bien que pratique
puisqu'elle
exige moins de calculs
que celle
de
Sturm.
Les avantages
admis par ce th&or&me sont
particulikrment
applicables
a des programmes
de
manipulations
algebriques.
Certains
rCsultats
qui
offrent
une comparaison
empirique
des deux m&thodes
sont Ggalement pr&sent&s.
Diskutiert
wird ein wenig bekannter
Lehrsatz,
der die Isolation
von Wurzeln polynomischer
Gleichungen
betrifft
und 1836 von einem nur als Herr Vincent
bekannten
Mathematiker
verijffentlicht
wurde.
Herrn
Vincents
Methode hat historisches
und praktisches
Interesse,
weil
sie weniger
Berechnungen
als Sturms
Methode verlangt.
Die Vorteile,
die dieser
Lehrsatz
hat, sind fir Software
Systeme fttr Computeralgebra
besonders
wichtig.
Eingeschlossen
sind such gewisse
Berechnungsergebnisse,
die einen empirischen
Vergleich
der beiden Methoden darbieten.
0315-0860/78/0054-0427$02.00/O
Copyright 0 1978 by AcademicPress,Inc.
All rightsofreproductionin
any form
reserved.
428
Akritas
1.
and Danielopoulos
HMS
INTRODUCTION
In 1834 a certain
Mr. Vincent published
a "note"
(of thirty
pages) in the Mboires
de la SociGt& royale
de Lille
concerning
the solution
of polynomial
equations
with numerical
coefficients.
The same memorandum appeared two years later,
with a few
additions,
under the title
"Note sur la resolution
des
Bquations
numi5rique"
in the October issue,
1836, of the Journal
de Mathgmatiques
Pures et Appliqu6es
[Vincent
18361.
According
to a footnote,
the article
was reprinted
"for the benefit
of the
professors".
Nevertheless,
the article
and the remarkable
method described
therein
were consigned
to oblivion
for more
than a century
[l],
although
it seems that several people had
dealt with variations
of this method; Mr. Vincent even mentions
that a similar
"note" was included
in the sixth edition
of
Bourdon's Algebra
[Bourdon 18311.
We may attempt
to explain
the fact that Vincent's
theorem
was forgotten
by noting
the careful
manner in which he pays
tribute
to Sturm and notes the "beauty"
and usefulness
of Sturm's
celebrated
theorem on the location
of the roots of equations.
In 1834, the same year in which Vincent first
published
his
paper, Sturm published
his work on second order differential
equations,
known today as the Sturm-Liouville
theory,
for which
he received
the "Grand Prix des Sciences Mathematiques"
from
the AcadGmie des Sciences.
Two years later,
when Vincent's
paper was reprinted,
Sturm was elected
to the Academic des
Sciences.
It is not surprising
that Sturm's method outshone
all the others.
Vincent uses in his approach a theorem on the
isolation
of roots,
given by the French physician
F. D. Budan
in 1807.
This theorem was enunciated
in a somewhat different
:orm by J. B. J. Fourier
and included
in his dnalyse des
Equations
published
posthumously
by C. L. M. H. Navier.
Vincent
indicates
his surprise
that Fourier did not try to go further
and prove the proposition
that was the main subject
of Vincent's
paper.
He states,
however, the belief
that such a proof may
exist in other manuscripts
which were not published
because of
the untimely
death of Navier.
The Budan-Fourier
theorem on the isolation
of roots uses a
sequence of derivatives,
f(x),
f(')(x),
. . . . f(n)(x).
The
procedure
is similar
to that used later
(1829), by Sturm, to
calculate
the number of variations
of sign (as defined by
Descartes'
rule)
in the two sequences of real numbers obtained
when x is replaced
by the real numbers p and q (p < q).
This
method, however, involves
less computation
than Sturm's method.
Another possible
reason that Vincent's
original
method may
not have proved popular
is his somewhat inefficient
procedure
for obtaining
the transformed
equation
for the substitution
x = y + cl. He obtains
the coefficients
of this equation
as
coefficients
of the Taylor expansion of the polynomial.
As it
HM5
On the Forgotten
Theorem
429
will be seen in the next section,

Uspensky simplified
the method
considerably
by using synthetic
division
(Homer's
rule in
essence) to obtain
the coefficients
of the transformed
equation
[Uspensky 1948, 127-1371.
Moreover,
for a = 1 the synthetic
division
does not include
any multiplication.
With this
modification
Vincent's
theorem becomes a powerful tool for the
isolation
of real roots of equations,
a tool that represents
an
essential
improvement
over methods based on Sturm's theorem.
So far as we have been able to determine,
Vincent's
theorem
and the implied
method for isolating
roots are not mentioned
by
any author with the exception
of Uspensky [1948, 1281 and
Obreschkoff
[1963, 248-2491.
Uspensky notes that even such a
capital
work as the Enzyclopadie
der mathematischen
Wissenschaften
ignores it.
We will show that Vincent's
contribution
has
practical
significance
as well as historical.
A statement
of the theorem and of the propositions
on which
it is based will be given in the following
sections,
together
with a description
of its use and an assessment of the related
method for root isolation.
2.
Most methods for the
with numerical
coefficients
According
to this rule,
polynomial
equation
with
ao9 al, . . . . an, and if
VINCENT'S
THEOREM
isolation
of the roots of polynomials
rely on Descartes'
rule of signs.
the number of positive
real roots of a
real coefficients
is never greater
than
in the sequence of its coefficients,
less, the difference
is an even number.
A variation
is defined
as a change of sign in two consecutive
Zero coefficients
are
terms of the sequence of coefficients.
disregarded
in counting
the number of variations.
It must be
noted that Descartes'
rule gives the exact number of roots only
if there is either
one or no variation.
In the first
case
there is one positive
real root; in the second there is no root.
As mentioned
previously,
Vincent states in his paper that
he based his proposition
on a result
which was formulated
somewhat differently
by both Budan and Fourier.
Vincent states
Budan's theorem as follows:
If in an equation
in x, f(x) = 0, we make two
transformations
x = p + x1 and x = q + x", where p
and q are real numbers such that p < q, then
(i)
the transformed
equation
in xl = x - p cannot
have fewer variations
than the transformed
equation
in x" = x - q;
(ii)
the number of real roots of the equation
f(x) = 0 located
between p and q can never be
more than the number of variations
lost in
passing from the transformed
equation
in
Akritas
430
(iii)
Fourier's
version
If
= x - p to the transformed
= x - q;
when the first

the difference
of this
in the sequence
HM5
and Danielopoulos
equation
in
number is less than the second,

is always an even number,
result
is stated
of (m+l)
as follows:
functions
f(x),
f(l)(x),
. . . . f(m)(x),
we replace x by any two real numbers p, q (p < q), and
if we represent
by P and Q the two resulting
sequences
of numbers, then
(i)
the sequence P cannot present fewer variations
than the sequence Q;
(ii)
the number of real roots of the equation
f(x) = 0, located between p and q, can never
be more than the number of variations
lost
in passing from the substitution
x = p to
the substitution
x = q;
(iii)
when the first
number is less than the second,
the difference
is an even number.
Using this result,
Vincent carries
out several consecutive
transformations
in order that the transformed
equation
will have
only one or zero variations;
in this case the number of roots
can be determined
urnambiguously.
He states his proposition
as
follows:
If in a polynomial
equation
with rational
coefficients
and without multiple
roots,
one makes successive
transformations
of the form
1
1
1
x=a+xl-b+x" 3 x" = c + -x""
*.*)
x' '
where a, b and c are any positive
numbers greater
than
one, then the resulting
transformed
equation
either
has
zero variations
or it has a single variation.
In the
second case the equation
has a single positive
real
root represented
by a continued
fraction
1
a+---bd-----(2)
c + ... '
in the first
case there is no root.
It is obvious that Vincent's
method relies
heavily
on
transformations
which consist
of a translation
and an inversion.
While the inversion
can be easily obtained
by reversing
the order
the translation
operation
requires
of the polynomial
coefficients,
a computation
which Vincent does not perform in the easiest
On the Forgotten
HMS
possible
way. He uses Taylor's
expansion
coefficients
of the transformed
polynomial.
tution
x = y + a is made in the polynomial
f(x)
= CO + clx + . . . + cnxn = 0, then

c
of the transformed
f(k)(a)
k!
431
Theorem
theorem to obtain
If the substiequation
the
the coefficients
,k=O,l,...,n,
polynomial
in y may be expressed
c; = j;k (jykCj,
by
k = 0, 1, . . . . n.
The computation
is somewhat simplified
when a = 1. Since Homer's
method had been known from 1819, it is rather
surprising
that
Vincent did not employ it for polynomial
evaluation
[Horner
1819, 308-3351 in order to perform translations.
The applicability
and superiority
of this method can be seen from the
following:
consider a polynomial
f(x)
= coxn + clxn-l
. . . + cn
produces
and the transformation
x = y + a. This substitution
a polynomial
in y having coefficients
bi which satisfy
the
relations
; cixn-i
= ; ~~(y+a)~-~
= ; biyn-i
= j,
bi(x-a)n-i.
i=O
i=O
i=O
From the first
and last expressions
in (4), we deduce that the
coefficients
bi of the transformed
polynomial,
can be obtained
(4)
with a sequence of applications

of the synthetic
division
algorithm
(i.e.,
Homer's
rule for the evaluation
of polynomials).
Indeed,
(4) suggests that we can write
f(x) = (x-a)g(x)
+ R ,
(5)
n
where g(x) is a polynomial
of degree n-l, and Rn is the
coefficient
bn in
(4).
If we express
as a polynomial
of degree
and then
we obtain
the coefficients
(6)
equate
a0 = co
aj
n-l
with
= cj + oraj-1
f(x)
as in
(3),
and g(x)
a o, al, . . . . anqlj
powers of x in (S),
coefficients
of equal
j = 1, 2, . . . . 11,
We notice
that the
which is the synthetic
division
algorithm.
last coefficient
an is precisely
the remainder
Rn in (S), or
equivalently,
the coefficient
bn in (4).
Further,
application
Akritas
432
of the same process
where
h(x)
f(x)
the process
f(x)
in which
n-2.
= (x-a)
is
gives
= (x-a)
is of degree
(8)
If
to g(x),
g(x)
(7)
HM5
and Danielopoulos
repeated
= RO(X-~)~
the coefficients
h(x)
+ R
n-l
Combining
h(x)
(5)
+ R,&-)
n times
(7),
we obtain
+ Rn.
we obtain
+ Rl(~-a)~-
R. are
and
+ ... + R
n
equal
to the
coefficients
b.,
and appear as remainders

inreach
application
of algorithm
(6;.
In particular,
for a = 1 this algorithm
does not require
any
multiplication,
so that the transformation
x = y + 1 can be
performed
in a very efficient
manner.
To find the exact number
of positive
roots
of a polynomial,
we consider
separately
the
positive
roots
of a polynomial
which are > 1 or < 1, excluding
the case when 1 is a root.
The positive
roots
that are greater
than 1 may be written
in the form x = 1 + y, while
those less
than 1 may be written
in the form x = l/(1
+ y) where y > 0.
We can therefore
transform
a polynomial
equation
by the
substitutions
x = 1 + y and x = l/(1
+ y) and examine the
number of variations
of the transformed
equations
[Uspensky
1948,
1281.
If the equation
obtained
by the transformation
x = 1 + y
has no variations,
it means that the original
equation
has no
the presence
of one variation
indicates
roots
> 1, whereas
precisely
one root > 1 of the given equation.
Similar
conclusions
hold for the equation
resulting
from the transformation
If one or both of the transformed
equations
have
x = l/(1
+ y).
more than one variation,
we transform
them again by the substicontinue
to
tutions
y = 1 + z, y = l/(1
+ z), and if necessary
make similar
substitutions
until
the transformed
equations
have
no more than one variation.
This necessarily
must happen after
a finite
number of steps.
The negative
roots
are investigated
by replacing
x by -x in the original
equation
and by investigating
the positive
roots
of the transformed
equation.
Uspensky
has given a proof of Vincents
theorem which
For the sake of
includes
also the concept
of root separation.
completeness
we state
this
theorem in the words of Uspensky:
Let
Nk be the
k-th
term
of
the
series
1, 1, 2, 3, 5, 8, 13, 21, . . .
in which
each term is the sum of the preceding
two and
where
A > 0 is the smallest
distance
between
any two
roots
of the equation
f(x) = 0 of degree
n and
without
multiple
roots.
Let the number m be so
chosen
that
On the Forgotten
HM5
433
Theorem
where
Then the substitution
.+-g-
1
m
presented
arbitrary
+- 1
5
in the form of a continued

positive
integral
elements
fraction
a 1, a2,
with
. . . . am
transforms
the equa,tion f(x)
= 0 into the equation
which
has
not
more
than one variation.
F(5) = 0,
[Uspensky
1948,
298-2991
3. SOME APPLICATIONS AND ASSESSMENT OF VINCENT'S METHOD

It is obvious that this method for isolating
the roots of
polynomials
is of great significance
for numerical
mathematics.
But its most significant
application
is in the area of
computerized
algebra otherwise
known as symbolic
and algebraic
manipulation,
in which the exact computation
of the roots of
Among the existing
major algebra
polynomials
is required.
systems, SAC-l is well known for its polynomial
algebra
Most algorithms
for root isolation
capabilities
[Collins
19711.
have been based on Sturm's theorem [2].
Further,
the decision
methods for elementary
algebra developed by Tarski
[1951] and
by Seidenberg
[1954] involve
root isolation
techniques
which use
Sturm's theorem [Akritas
19731.
Such techniques
are likewise
important
in quantifier
elimination
algorithms
[Collins
19751.
Heindel
[1971] has developed
an algorithm
for root
isolation,
based on Sturm's theorem,
for which he shows that the
computing
time is bounded by n 13L(d)3,
where n is the degree of
the polynomial
and L(d) the length
(number of bits in binary
representation)
of the sum of the absolute
values of the
polynomial
coefficients.
It is obvious that algorithms
which
use more efficient
methods, such as Vincent's,
are very
desirable.
The superiority
of Vincent's
method is confirmed
by
empirical
comparisions
of the computation
time for root isolation
performed with algorithms
based on the two theorems [Collins
and Akritas
19761.
The table below shows the computation
time
in seconds for randomly generated
polynomials
of degrees 5-25.
Akritas
434
and Danielopoulos
DEGREE
STURM
5
10
being
0.58
6.83
HM5
VINCENT
0.07
0.21
15
28.8
0.32
20
89.6
0.68
25
208.2
0.74
A theoretical
analysis
of Vincent's
prepared by the authors.
method
is currently
NOTES
1. One of the authors of this article
(AGA) came across
Vincent's
theorem while reviewing
methods for the isolation
of
real roots of equations
as presented
by Uspensky [Uspensky 19481.
2. Recently,
another algorithm
based on Rolle's
theorem
has been developed
[Collins
and Loos 19761.
REFERENCES
Akritas,
A G 1973 Entscheidungsverfahren,
eine Literatursuche
Bericht
22-3-73,
Institut
fur angewandte Mathematik,
Universitgt
Heidelberg,
B.R.D. (Unpublished
Report)
Akritas,
A G Fr G E Collins
1976 Polynomial
real root isolation
Proceedings
of the 1976 ACM
using Descartes'
rule of signs
Yorktown
Symposium on Symbolic
and Algebraic
Computation
Heights,
New York,,272-275
Bourdon, L P M 1831 Elemens d'lllgebre
Paris (Bachelier,
Pere
et Fils)
6 r$d.
An Introduction
and Survey
Collins,
G E 1971 The SAC-l System:
Proceedings
Manipulation
of the Second
Symposium
on Symbolic
and Algebraic
Los Angeles,
144-152
1975 Quantifier
elimination
for real closed fields
by
Lecture
notes in
cylindrical
algebraic
decomposition
Computer Science
Berlin
(Springer
Verlag)
Vol. 33, 134-183
Collins,
G E 8 R Loos 1976 Polynomial
real root isolation
by
differentiation
Proceedings
of the 1976 ACM Symposium on
Symbolic
and Algebraic
Computation
Yorktown Heights,
New
York, 15-25
Heindel,
L E 1971 Integer
arithmetic
algorithms
for polynomial
real zero determination
Journal
of the ACM 18 (4), 533-548
Horner, W G 1819 A new method of solving
numerical
equations
of all orders,
by continuous
approximation
Royal Society
of London Philosophical
Transactions
109, 308-335
HM5
On the Forgotten
Obreschkoff,
1963
Verteilung
435
und Berechnung
Polynome
(VEB Deutscher
der Nullstellen
der
Wissenschaften)
Seidenberg,
A 1954 A new decision
method for elementary
algebra
Annals of Mathematics
60, 365-374
Tarski,
A 1951 A decision
method for elementary
Algebra
and
Geometry
University
of California
Press
Uspensky, J V 1948 Theory of Equations
New York (McGraw-Hi 11
Co.)
Vincent
1836 Sur la resolution
des equations
numeriques
Journal
de Mathgmatiques
Pures et Appliquges
1, 341-372
reeller
Berlin
Tneorem
Verlag
Serdica J. Computing 2 (2008), 89104
ON THE VARIOUS BISECTION METHODS DERIVED

FROM VINCENTS THEOREM
Alkiviadis G. Akritas, Adam W. Strzebo
nski, Panagiotis S. Vigklas
Dedicated to Professors Alberto Alesina and Massimo Galuzzi. 1

Abstract. In 2000 A. Alesina and M. Galuzzi presented Vincents theorem
from a modern point of view along with two new bisection methods derived from it, B and C. Their profound understanding of Vincents theorem is
responsible for simplicity the characteristic property of these two methods. In this paper we compare the performance of these two new bisection
methods i.e. the time they take, as well as the number of intervals they
examine in order to isolate the real roots of polynomials against that of
the well-known Vincent-Collins-Akritas method, which is the first bisection
method derived from Vincents theorem back in 1976. Experimental results
indicate that REL, the fastest implementation of the Vincent-Collins-Akritas
method, is still the fastest of the three bisection methods, but the number
of intervals it examines is almost the same as that of B. Therefore, further
research on speeding up B while preserving its simplicity looks promising.
ACM Computing Classification System (1998): G.1.5, F.2.1, I.1.2.
Key words: Vincents theorem, real root isolation method, bisection method, continued
fraction method, Descartes method, modified Uspenskys method.
1
For presenting Vincents theorem in their own, unique way!
90

1. Introduction. We begin with a review of Descartes rule of signs [15].

Consider the polynomial p(x) R[x], p(x) = a n xn + + a1 x + a0 and let
var(p) represent the number of sign variations or changes (positive to negative
and vice-versa) in the sequence of coefficients a n , an1 , . . . , a0 .
Descartes rule of signs: The number % + (p) of real roots multiplicities counted of the polynomial p(x) R[x] in the open interval ]0, +[ is
bounded above by var(p); that is, we have var(p) % + (p).
According to Descartes rule of signs if var(p) = 0 it follows that % + (p) = 0.
Additionally, according to Descartes rule of signs, the Mean Value Theorem and
the fact that the polynomial functions are continuous, if var(p) = 1 it follows
that %+ (p) = 1.
Therefore, Descartes rule of signs yields the exact number of positive
roots only in the two special cases mentioned above. 2
These two special cases of Descartes rule are used in Vincents theorem
of 1836, [28], which we present in a way, due to Alessina and Galuzzi [13].
Theorem 1 (Vincents theorem bisection version).
Let f (z),
be a real polynomial of degree n, which has only simple roots. It is possible to
determine a positive quantity so that for every pair of positive real numbers a, b
with |b a| < , every transformed polynomial of the form
n
(z) = (1 + z) f
a + bz
1+z
has exactly 0 or 1 variations. The second case is possible if and only if f (z) has
a simple root within ]a, b[.
We call this the bisection version of Vincents theorem, since all three
bisection methods studied below are derived from it; see also [11], [12] and [21].
Also note that throughout this paper we are interested in polynomials with rational coefficients; see also [20], [23].
Moreover, the Vincent-Akritas-Strzebo
nski (VAS) continued fractions method is also derived from Theorem 1, since the Mobius transformation M (z) =
a + bz
, where a, b Q, is equivalent to a continued fraction of the form
1+z
2
These two special cases were known to Cardano; in other words, what Descartes did was
to generalize Cardanos special rule of signs. This detail is mentioned in [5].
On the Various Bisection Methods Derived from Vincents Theorem
1 +
1
2 +
1
3 +
91
..
.+
1
1
` + z
where 1 0 is an arbitrary non-negative integer and 2 , 3 , . . . , ` are arbitrary

positive integers, i > 0, i > 1, [1], [2], [3], [4], [14].
It is well known, [7], [8], [9], [10], [24], and [26], that the VAS continued
fractions method is, so far, the fastest method for the isolation of the real roots
of polynomials; however, no data exists on the number of intervals it examines
during the isolation process. This is achieved in this paper, where the VAS continued fractions method is used as a benchmark in the empirical results presented
in the sequel.
The rest of the paper is structured as follows:
In Section 2 we present the three bisection methods derived from Vincents
theorem; to wit,
the first bisection method: Vincent-Collins-Akritas (VCA), [15], developed
in 1976 by Collins and Akritas [16],
the second bisection method: B, suggested by Alesina and Galuzzi [13], and
the third bisection method: C, suggested by Saeli [13].
In Section 3 we present experimental data comparing the performance
of the three bisection methods mentioned above juxtaposing it with that of the
(VAS) continued fractions method.
2. The Three Bisection Methods Derived from Vincents

Theorem. The fact that bisection methods can be derived from Vincents
theorem is missing from major works such as ([29], pp. 470478) and almost
every paper on the subject; to our knowledge, it appears only in [6] and the
papers by Alesina and Galuzzi [11], [13].
2.1. The First Bisection Method: Vincent-Collins-Akritas (VCA).
This method seems to be the most complicated of the three. It was originally
presented in a way that obscured its relation to Vincents theorem. Due to that
obfuscation it was originally called modified Uspenskys method and recently
Descartes method ([19], [18], [17], [25]), which is totally misleading. Akritas
set the record straight, [6], by presenting the algorithm in a way that reveals its
92

relation to Vincents theorem, justifying thus the name Vincent-Collins-Akritas

given to it in France, [15].
A brief description of the VCA bisection method is presented below according to [6]; see also [27]. However, we first need the following definition:
Uspenskys test: The number %01 (p) of real roots in the open interval
]0, 1[ multiplicities counted of the polynomial p(x) R[x] is bounded above
by var01 (p), where
(1)

var01 (p) = var (x + 1)deg(p) p
1
x+1

and we have var01 (p) %01 (p).

As in the case of Descartes rule of signs if var 01 (p) = 0 it follows that
%01 (p) = 0 and if var01 (p) = 1 it follows that %01 (p) = 1.
Therefore, Uspenskys test yields the exact number of positive roots only
in the two special cases mentioned above; to wit, whenever var 01 (p) = 0 or
var01 (p) = 1.
1
Please note in equation (1) that, after the substitution x
, the
x+1
positive roots of p(x) that were in the interval ]0, 1[ are now in ]0, +[, in which
case Descartes rule of signs can be applied.
Finally, we note that Uspenskys test is a special instance of the powerful
Vincents test, which is based on Theorem 1, applies to any interval ]a, b[ and
is described in Section 2.2.
We are now ready for the VCA bisection method.
Let p(x) be the polynomial whose roots we want to isolate and let ub be
an upper bound on the values of its positive roots. Then all the positive roots of
p(ub x) lie in the interval ]0, 1[ and the VCA method isolates them by repeatedly
bisecting the interval ]0, 1[, while using in the process Uspenskys test to make
inferences about the number of positive roots certain transformed polynomials
have in the interval ]0, 1[. Finally, the isolating intervals of the roots of p(x) are
easily computed from the bijection:
(2)
]0,ub[ = a + ]0,1[ (b a),
that exists between the roots ]0,1[ ]0, 1[ of the transformed polynomial p(ub x)
and the roots ]0,ub[ ]a, b[=]0, ub[ of the original polynomial p(x).
93
Below is a recursive description of the VCA bisection method as was originally presented by Collins and Akritas:
The first bisection method: VCA
1
2
3
4
5
6
7
8
9
10
11
Input: A univariate, square-free polynomial p(ub x) Z[x], p(0) 6= 0, and

the open interval ]a, b[=]0, ub[, where ub is an upper bound on the
values of the positive roots of p(x). (The positive roots of p(ub x) are
all in the open interval ]0, 1[.)
Output: A list of isolating intervals of the positive roots of p(x)

1
deg(p)
;
var the number of sign changes of (x + 1)
p
x+1
if var = 0 then RETURN ;
if var = 1 then RETURN {]a, b[};

x
1
p0 21 2deg(p) p
// Look for real roots in 0,
;
2
2
a+b
1
m
// Is
a root? ;
2
2

x+1
1
deg(p)
p
p 21 1 2
// Look for real roots in
,1 ;
2
2

1
if p
6= 0 then
2

S

VCA p 21 1 , ]m, b[
RETURN VCA p0 12 , ]a, m[
else

S

S
RETURN VCA p0 12 , ]a, m[
{[m, m]} VCA p 12 1 , ]m, b[
end
Algorithm 1. The original version of the VCA(p, ]a, b[) bisection algorithm, where
the second argument is the open interval ]a, b[ associated with p(x). The isolating
intervals of the roots of p(x) are computed directly, without using bijection (2).
To obtain the isolating intervals of the positive roots of p(x) we could

have also used the interval ]a, b[=]0, 1[ along with bijection (2). An excellent
discussion of this algorithm can be found in [15]. Please note the following:
Uspenskys test is a crucial component of the VCA(p, ]a, b[) bisection algorithm lines 13.
The substitutions in lines 4 and 6 are performed only on the polynomial
94

p(x), whereas at the same time in line 5 the interval ]a, b[ is divided
into two equal parts ]a, m[ and ]m, b[, to be used in line 8 (or 10).
To isolate the real roots of p(x) in the open interval ]0, 1[ we proceed as
follows:

1
we first isolate the real roots in the interval 0,
lines 4 and 8
2
(or 10),
1
we then deal with the case where is a root of p(x) lines 5, 7 and
2
10,

1
and, finally, we isolate the real roots in the interval
, 1 lines 6
2
and 8 (or 10).
The isolating intervals are directly obtained from line 3 except for those
roots that happen to coincide with the midpoint of an interval that gets
bisected, in which case they are computed in lines 5 and 10.
Algorithm 1 does not make clear the relation of the VCA bisection method
with Vincents theorem. That relation is revealed if instead of the interval ]a, b[
we associate with the polynomial p(x) the Mobius transformation M (x) = ub x
or M (x) = x if we use bijection (2). For details see [6].
Please note that the fastest implementation of the VCA bisection method,
REL, is due to Rouillier and Zimmermann, [22].
2.2.The Second Bisection Method: B. This method is a direct implementation of Theorem 1. However, we first need the following definition, which
is also used in Section 2.3:
Vincents test: If a 0 and b > a then the number % ab (p) of real
roots in the open interval ]a, b[, multiplicities counted of the polynomial
p(x) R[x] is bounded above by varab (p), where
(3)

a + bx
deg(p)
p
,
varab (p) = var (1 + x)
1+x
and we have varab (p) = varba (p) %ab (p).
95
Vincents test is a direct consequence of Theorem 1, presented by Alesina

and Galuzzi [13]. It is a very powerful test and can be applied in all cases
including the case ]a, b[=]1, 0[, from which we obtain Uspenskys test. 3
As in the case of Descartes rule of signs if var ab (p) = 0 it follows that
%ab (p) = 0 and if varab (p) = 1 it follows that %ab (p) = 1.
Therefore, Vincents test yields the exact number of positive roots only
in the two special cases mentioned above; to wit, whenever var ab (p) = 0 or
varab (p) = 1.
Below is a recursive description of the second bisection method derived
from Vincents theorem; its simplicity is unsurpassed, but we pay for it by using
a much more complicated test. Obviously, there is a trade off between simplicity
of the method and complexity of the termination test.
The second bisection method: B
1
2
3
4
5
6
7
8
9
Input: A univariate, square-free polynomial p(x) Z[x], p(0) 6= 0, and the

open interval ]a, b[=]0, ub[, where ub is an upper bound on the values
of the positive roots of p(x).

a + bx
var the number of sign changes of (1 + x)deg(p) p
;
1+x
a+b
// Subdivide the interval ]a, b[ in two equal parts ;
2
if p(m) 6= 0 then
S
RETURN B(p, ]a, m[) B(p, ]m, b[)
else
S
S
RETURN B(p, ]a, m[) {[m, m]} B(p, ]m, b[)
end
m
Algorithm 2. The B(p, ]a, b[) bisection algorithm, proposed by Alesina and Galuzzi
[13]; the second argument is the open interval ]a, b[, whose endpoints a, b are used in
Vincents test in line 1. The isolating intervals of the roots of p(x) are computed
directly, without using bijection (2).
Please note the following:

3
By comparison, Uspenskys test in Section 2.1, is rather weak as it applies only in the case
]a, b[=]0, 1[.
96

Vincents test is a crucial component of the B(p, ]a, b[) bisection algorithm
lines 13.
In line 4 the interval ]a, b[ is divided into two equal parts ]a, m[ and ]m, b[, to
be used in lines 6 (or 8). Note that there are no polynomial transformations
at all; only polynomial evaluations in line 1.
To isolate the real roots of p(x) in the open interval ]a, b[ we proceed as
follows:

a+b
lines 6 (or 8),
we first isolate the real roots in the interval a,
2
a+b
we then deal with the case where
is a root of p(x) lines 5
2
and 8,

a+b
and, finally, we isolate the real roots in the interval
, b lines
2
6 (or 8).
2.3. The Third Bisection Method: C. As Alesina and Galuzzi mention

in their paper ([13], p. 189), it was Donato Saeli who suggested that they look
at the connection between Farey sequences and Vincents theorem. We quote
from their paper: It is possible that Vincents theorem might be connected with
Farey series in the same fruitful way it was connected with continued fractions.
Saelis proposal is evaluated in Section 3.
c
a
Recall that if a 0 and b, c, d > 0 a mediant of the two fractions <
b
d
a+c
a
a+c
c
is given by
and we have <
< .
b+d
b
b+d
d
A Farey series of order N , denoted by F N , is the set of all reduced fractions
between 0 and 1 whose denominators are N or less, arranged in increasing order,
[13]. We can obtain FN from FN 1 by inserting mediants whenever it is possible
to
so without getting a denominator greater than N . We begin with F 1 =
do
0 1
,
.
1 1
Instead of using Farey sequences in ]0, 1[, along with bijection (2), what
we have done with the third bisection method derived from Vincents theorem
97

0 ub
,
and keep inserting mediants until we isolate the
1 1
positive roots of a polynomial.
Below is a recursive description of the third bisection method derived
from Vincents theorem:
is to start with ]a, b[=
The third bisection method: C
1
2
3
4
5
6
7
8
9
Input: A univariate, square-free polynomial p(x) Z[x], p(0) 6= 0, and the

open interval ]a, b[=]0, ub[, where ub is an upper bound on the values
of the positive roots of p(x).

a + bx
;
var the number of sign changes of (1 + x)deg(p) p
1+x
num
mf
// The mediant of the interval ]a, b[, as in Farey
den
sequences ;
if p(mf ) 6= 0 then
S
RETURN C(p, ]a, mf [) C(p, ]mf , b[)
else
S
S
RETURN C(p, ]a, mf [) {[mf , mf ]} C(p, ]mf , b[)
end
Algorithm 3. The C(p, ]a, b[) bisection algorithm, proposed by Donato Saeli, [13];
the second argument is the open interval ]a, b[, whose endpoints a, b are used in
Vincents test in line 1. The isolating intervals of the roots of p(x) are computed
directly, without using bijection (2).
To obtain the isolating intervals of the positive roots of p(x) we could

have also used the polynomial p(ub x) and the interval ]a, b[=]0, 1[ along with
bijection (2). Please note the following:
Vincents test is a crucial component of the C(p, ]a, b[) bisection algorithm
lines 13.
In line 4 the interval ]a, b[ is divided according to Fareys sequences
into two parts ]a, mf [ and ]mf , b[, to be used in lines 6 (or 8). Note
that, again, there are no polynomial transformations at all; only polynomial
evaluations in line 1.
98

To isolate the real roots of p(x) in the open interval ]a, b[ we proceed as
follows:
we first isolate the real roots in the interval ]a, m f [ lines 6 (or 8),
we then deal with the case where mf is a root of p(x) lines 5 and 8,
and, finally, we isolate the real roots in the interval ]m f , b[ lines 6
(or 8).
3. Empirical Results. In this section, we present some statistics using

various classes of polynomials, as in [7]. We compare the performance of REL, the
fastest implementation of Vincent-Collins-Akritas (VCA) bisection method, [22],
with the B and C bisection methods described in Sections 2.2 and 2.3, respectively;
as a benchmark we use the Vincent-Akritas-Strzebo
nski (VAS) continued fractions
method.
In Table 1, we present statistics on some special classes of polynomials
regarding: (a) the number of intervals (#I) the methods use in order to isolate
the real roots of the given polynomial, (b) the time in seconds (T (s)) they need
for doing so, and (c) the number of the roots (#R).
In Table 2, we present statistics regarding some types of random polynomials; rp indicates a random polynomial with randomly generated coefficients, rmp indicates a monic polynomial with randomly generated coefficients;
whereas rr denotes products of factors (x-randomly generated integer root).
From Table 1 we see that bisection method C examines the greatest number of intervals save for the case of Mignotte polynomials, whose roots are
extremely close. The number of intervals examined by VCA and B is of the same
order. By contrast, of all methods VAS examines the smallest number of intervals.
With respect to time, VCA is the fastest of the three bisection methods as
it examines the smallest number of intervals.
Table 2 confirms the observations made in Table 1.
4. Conclusion. In this paper we have presented three bisection methods

derived from Vincents theorem for the isolation of the real roots of polynomials
99
Table 1. Comparison of various root isolation statistics on special polynomials
with rational coefficients and we have compared their performance against the
Vincent-Akritas-Strzebo
nski (VAS) continued fractions method the benchmark.
The three bisection methods VCA, B and C use either Uspenskys or
Vincents test4 :
the more complex VCA method uses the simpler of these two tests, viz.,
Uspenskys,
whereas
the very simple methods B and C use the more complex and very powerful
Vincents test.
The simplest of all termination tests is Descartes rule of signs, which is
employed by the VAS method, the most complex method derived from Vincents
theorem.
4
Remember that the former is a special instance of the latter.
100

Table 2. Comparison of various root isolation statistics on random polynomials
From the empirical results presented we have the following rule of thumb:
the speed and structural complexity of the real root isolation algorithms derived
from Vincents theorem are inversely proportional to the structural complexity
of the termination test they employ.
Therefore, it is no longer a surprise that VAS is the fastest real root isolation method, whereas VCA comes second and is followed by B and C.
The fact that methods VCA and B examine approximately the same number
of intervals indicates that further research on speeding up B is warranted.
101
REFERENCES
[1] Akritas A. G. Vincents theorem in algebraic manipulation. Ph.D. Thesis,
Operations Research Program, North Carolina State University, Raleigh,
NC, 1978.
[2] Akritas A. G. An implementation of Vincents Theorem. Numerische
Mathematik, 36 (1980), 5362.
[3] Akritas A. G. The fastest exact algorithms for the isolation of the real
roots of a polynomial equation. Computing, 24 (1980), 299313.
[4] Akritas A. G. Reflections on a pair of theorems by Budan and Fourier.
Mathematics Magazine, 55, No 5 (1982), 292298.
[5] Akritas A. G. Elements of Computer Algebra with Applications. John
Wiley Interscience, New York, 1989.
[6] Akritas A. G. A new Look at one of the Bisection Methods Derived
from Vincents Theorem or There is no Descartes Method. In: Proceedings
of ACA 2007 session: Applications of Computer Algebra in Education.
Barnes, Nobles and Borders, ISBN: 978-0-9754541-9-0, to appear.
ski. A comparative study of two real root
[7] Akritas A. G., A. Strzebon
isolation methods. Nonlinear Analysis: Modelling and Control, 10, No 4
(2005), 297304.
ski, P. Vigklas. Implementations of a New
[8] Akritas A.G., A. Strzebon
Theorem for Computing Bounds for Positive Roots of Polynomials; Computing, 78, (2006), 355367.
ski, P. Vigklas. Advances on the Contin[9] Akritas A. G., A. Strzebon
ued Fractions Method Using Better Estimations of Positive Root Bounds.
In: Proceedings of the 10th International Workshop on Computer Algebra
in Scientific Computing, CASC 2007 (Eds V. G. Ganzha, E. W. Mayr, E. V.
Vorozhtsov) Bonn, Germany, September 1620, 2007, LNCS 4770, Springer
Verlag, Berlin, 2430.
ski, P. Vigklas. Improving the Perfor[10] Akritas A. G., A. Strzebon
mance of the Continued Fractions Method Using New Bounds of Positive
Roots, submitted.
102

[11] Alesina A., M. Galuzzi. A new proof of Vincents theorem.

LEnseignement Mathemathique, 44 (1998), 219256.
[12] Alesina A., M. Galuzzi. Addentum to the paper A new proof of Vincents theorem. LEnseignement Mathemathique, 45 (1999), 379380.
[13] Alesina A., M. Galuzzi. Vincents Theorem from a Modern Point of
View. (Betti, R. and Lawvere W.F. (eds.)), Categorical Studies in Italy 2000,
Rendiconti del Circolo Matematico di Palermo, Serie II, n. 64, (2000), 179
191.
[14] Bombieri E., A.J. van der Poorten. Continued fractions of algebraic
numbers. In: Computational Algebra and Number Theory, (Sydney, 1992),
Math. Appl. 325, Kluwer Academic Publishers, Dordrecht, 1995, 137152.
[15] Boulier F. Systèmes polynomiaux : que signifie resoudre?. Lecture
Notes, Universite Lille 1, 8 janvier 2007.
http://www2.lifl.fr/ boulier/RESOUDRE/SHARED/support.pdf
or
http://www.fil.univ-lille1.fr/portail/ls4/resoudre
[16] Collins G. E., A. G. Akritas. Polynomial real root isolation using
Descartes rule of signs. In: Proceedings of the 1976 ACM Symposium on
Symbolic and Algebraic Computations, Yorktown Heights, N.Y., 1976, 272
275.
[17] Eigenwillig A., V. Sharma, C.K. Yap. Almost Tight Recursion Tree
Bounds for the Descartes Method. In: Proc. Intl Symp. Symbolic and Algebraic Computation (ISSAC06), July 912, 2006, Genova, Italy, 2006, 7178.
[18] Johnson J. R., W. Krandick, K. M. Lynch, D. G. Richardson, A. D.
Ruslanov. High-Performance Implementations of the Descartes Method.
Technical Report DU-CS-06-04, Department of Computer Science, Drexel
University, Pliladelphia, PA 19104, May 2006.
[19] Krandick W., K. Mehlhorn. New Bounds for the Descartes Method.
Journal of Symbolic Computation, 41 (2006), 4966.
[20] Kyurkchiev N. A note on a method for the localization of the roots of
algebraic equations. C. R. Acad. Bulgare Sci., 44, (1991), 57.
[21] Obreschkoff N. Verteilung und Berechnung der Nullstellen reeller Polynome. VEB Deutscher Verlag der Wissenschaften, Berlin, 1963. (An English
103
translation of a related book by Obreschkoff has also appeared: Zeros of

Polynomials, Bulgarian Academic Monographs (7), Sofia, 2003).
[22] Rouillier F., P. Zimmermann. Efficient isolation of polynomials real
roots. Journal of Computational and Applied Mathematics, 162, (2004), 33
50.
[23] Sendov Bl. A method for simultaneous approximate calculation of all positive roots of a polynomial. Izv. Vyssh. Uchebn. Mat., 5 (1974), 185187.
[24] Sharma V. Complexity of Real Root Isolation Using Continued Fractions.
ISAAC07 preprint, 2007.
[25] Sharma V. Complexity Analysis of Algorithms in Algebraic Computation.
Ph.D. Thesis, Department of Computer Sciences, Courant Institute of Mathematical Sciences, New York University, 2007.
[26] Tsigaridas E. P., I. Z. Emiris. Univariate polynomial real root isolation:
Continued fractions revisited. (Eds Y. Azar and T. Erlebach), ESA 2006,
LNCS 4168, (2006), 817828.
[27] Uspensky J.V. Theory of Equations. McGraw-Hill, New York, (1948).
[28] Vincent A. J. H. Sur la resolution des equations numeriques. Journal de
Mathematiques Pures et Appliquees, 1 (1836), 341372.
[29] Yap C. K. Fundamental Problems of Algorithmic Algebra. Oxford University Press, 2000.
104

Alkiviadis G. Akritas
University of Thessaly
Department of Computer
and Communication Engineering
e-mail: akritas@uth.gr
Panagiotis S. Vigklas
University of Thessaly
Department of Computer
and Communication Engineering
e-mail: pviglas@uth.gr
Adam W. Strzebo
nski
Wolfram Research, Inc.
100 Trade Center Drive
Champaign, IL 61820, USA
e-mail: adams@wolfram.com
Received January 18, 2008

Final Accepted March 10, 2008
Reflectionson a Pair of Theorems by Budan and Fourier

ALKIVIADISG. AKRITAS
University
ofKansas
Lawrence,
KS 66045
real,disjoint
equationis theprocessof finding
Isolationof thereal rootsof a polynomial
intervals
suchthateach containsexactlyone realrootand everyrealrootis containedin some
the
because,as J.B. J.Fourierpointedout,it constitutes
interval.
Thisprocessis quiteimportant
thanfour,thesecondstep
firststeptowardthesolutionof generalequationsof degreegreater
ofrootsto anydesireddegreeof accuracy.
beingtheapproximation
twodifferent
F. D. Budanand J.B. J.Fourierpresented
In thebeginning
of the19thcentury
themaximum
possiblenumberof real
whichenableus to determine
(but equivalent)theorems
rootsthatan equationhas withina giveninterval.
des
Budan'stheorem
appearedin 1807in thememoir"Nouvellemethodepourla resolution
in 1820in "Le
was firstpublished
[10,p. 219],whereasFourier'stheorem
equationsnumeriques"
de Paris,"pp. 156,181[10,p. 223].Due to the
Bulletindes sciencespar la SocietePhilomatique
In his
priority
rights.
regarding
of thesetwotheorems,
therewas a greatcontroversy
importance
us that
men,"p. 383, F. Aragoinforms
book (1859) "Biographiesof distinguished
scientific
of earlystudentsof the
Fourier"deemedit necessaryto have recourseto the certificates
in orderto provethathe had taughthis
of theUniversity"
Polytechnic
Schoolor Professors
in 1829
C. Sturmpresented
in 1796,1797and 1803[10].Basedon Fourier'sproposition,
theorem
ofrealrootswhicha polynomial
an improved
theorem
whoseapplication
yieldstheexactnumber
zeroshas withina realinterval;thushe solvedtherealrootisolation
multiple
equationwithout
problem.Since 1830 Sturm'smethodhas been the only one widelyknownand used, and
can
Budan'stheorem
waspushedintooblivion.To ourknowledge,
Budan'stheorem
consequently
appearsin almostall textson the
be foundonlyin [16] and [61whereasFourier'sproposition
becauseit constitutes
meritsspecialattention
of equations.We feelthatBudan'stheorem
theory
ofourmethod
of 1836which,in turn,is thefoundation
theorem
thebasisofVincent'sforgotten
fortheisolationof thereal rootsof an equation[1],a methodwhichfarsurpassesSturm'sin
efficiency
[2],[3].
and withoutproofs,the classical
In the discussionwhichfollowswe presentseparately,
real root
theorems
by Fourierand Budanand we indicatehow theylead to thecorresponding
forcomparison.
Someempirical
resultsarealso presented
isolationmethods.
Fourier'stheorem
Fourier'stheorem,
firstpublishedin 1820,was also includedin his Analysedes Equations,
of
byC. L. M. N. Navierin 1831.Foundin almostall textson thetheory
published
posthumously
or evenBudan[9],[17].Hurwitz
it is sometimes
givenunderthenameBudan-Fourier
equations,
[13, pp. 76-87]
[12] presentsit as a specialcase of a moregeneraltheoremand Obreschkoff
it forcomplexroots.The statement
byVincent
givenbelowis thewayit is rendered
generalizes
notion
of
variation.
must
first
define
the
[16,p. 342].We
sign
DEFINITION. We say thata signvariationexistsbetweentwo nonzeronumbers
cp and cq
holds:
or infinite
sequenceofrealnumbers
(p < q) of a finite
cl, C2, C3,..., ifthefollowing
forq =p + 1,cpandcq haveoppositesigns;

forq 2p + 2, thenumbers
cp+I,. . , Cql areall zeroand cpand Cqhaveoppositesigns.
THEOREM 1 (Fourier1820).If in thesequence
P(x), P(')(x), ... ,P(M)(x)
ofthem+ 1functions
we replacex by anytworeal numbers
p, q (p < q) and if we
(whereP(i) = theithderivative),
thetworesulting
ofnumbers
byP and Q, then
sequences
represent
292
MATHEMATICSMAGAZINE
(i) thesequenceP cannot

present
fewersignvariations
thanthesequenceQ;
(ii) thenumber
ofrealrootsoftheequation
P(x) = 0, locatedbetween
p andq, canneverbemore
thanthe numberof sign variationslost in passingfromthesubstitution
x =p to the
x = q;
substitution
is lessthanthesecond,thedifference
(iii) whenthefirstnumber
is an evennumber.
The sequenceof them+ 1 derivatives
is calledFourier'ssequence.In (iii) the"firstnumber"
meansthenumberof therealrootsofP(x) = 0 locatedbetween
p and q; the"secondnumber,"
on theotherhand,refersto thenumberof signvariations
lostin passingfromthesubstitution
x = q. Obviously,
x =p to thesubstitution
Fourier'stheorem
givesan upperboundon thenumber
ofrealrootswhichtheequationP(x) = 0 (ofdegreem) has insidetheinterval
(p, q).
We remindthereaderthatthetwomainsubjectsofFourier'slifeworkwerethetheory
ofheat
ofthesolution
ofnumerical
and thetheory
equations.Bothofthesesubjectswerecarriedforward
whohad personaland scientific
relations
withFourier[8].The manuscript
ofFourier's
by Sturm,
treatiseon thesolutionof numerical
to severalpersons
equationswas by 1829communicated
it had on hisownwork.
whata greatinfluence
including
whomentions
Sturm,
explicitly
WhatSturmdidwas to replaceFourier'ssequenceby
P(x), P(l)(x), RI(x),. . . 9RJ(X)

whichis called Sturm'ssequenceor chain.This new sequenceis obtainedby applyingthe
to thepolynomials
Eucidean algorithm
P(x) and P(1)(x), and takingRi(x), i = 1,... ,k as the
thatis, thesequenceis defined
of theremainder
negative
relations:
polynomial;
by thefollowing
P(x) = P(')(x)Q1(x) -R(x),
P(')(x) =
Rk-2(X)
R1(x)Q2(x)-R2(X)g
= Rk-I(X)Qk(X)-
Rk(X).
The advantageof Sturm'ssequenceis thatwe can now obtaintheexactnumberof real roots

Thisis formally
statedas follows:
whichtheequationP(x) = 0 has withina giveninterval.
THEOREM 2 (Sturm1829).If theequation
ofits
P(x) = 0 has onlysimpleroots,thenthenumber
realrootsin theinterval
(p, q) is equaltothedifference
v(p) -v(q)
thenumber
wherev(4) denotes
ofsignvariations
in Sturm'ssequence
forx =.
was merelya by-product
of his extensive
Sturmhimselftellsus [8] thattheabove theorem
on thesubjectoflineardifference
investigations
equationsof thesecondorder.The requirement
ofthegenerality
thatP(x) = 0 has onlysimplerootsis no restriction
becausewe can firstapply
factorization
square-free
[4],[15]and thenuse Sturm'stheorem.
can be used in theisolationof thereal rootsof an equation.The
ClearlySturm'stheorem
processitselfis quitesimplebecauseall we haveto do,onceSturm'ssequencehasbeenobtained,
is to computean absoluteupperrootboundb so thatall therootslie withintheinterval
(- b, b).
thisinterval
untilin eachsubinterval
We thensubdivide
thereis at mostoneroot;thatis, Sturm's
withina
methodis actuallya bisectionmethod.Quiterecently,
thismethodwas implemented
and its computing
time
computer
algebrasystem[11] usingexactintegerarithmetic
algorithms
was thoroughly
analyzed.(Computeralgebrasystems
usuallydeal onlywithinteger(rational)
so thattheuserdoes nothaveto worryaboutroundoffand truncation
errors.For a
numbers,
ofsuchsystems
univariate
see [14].)It was shownthatifP(x) = 0 is an integral-coefficient
survey
timeof Sturm's
polynomial
equationof degreen> 0 without
multiple
roots,thenthecomputing
methodis
N
O
5L5(,
VO(Ln
in bits,of themaximum
of theabsolutevaluesof thecoefficients
whereL(I PI,,) is thelength,
of
P. Thislengthy
computing
timeshowsSturm'smethodleavesa lot to be desired;it has been
thatitsslowness
is due to thecomputation
of theSturmsequence.
determined
Budan'stheorem
AlthoughBudan's theoremappearedmuchearlierthan Fourier's,it seemsto have been
it doesnotappearin anyof thestandardtexts
ignored;as faras we havebeenable to determine
ofequations.
statement
ofthetheorem
on thetheory
Thefollowing
is fromVincent's
paper[16,p.
342].
THEOREM 3 (Budan 1807).If in an equationin x, P(x) = 0, we maketwotransformations,
x = p + x' andx = q + x", where
suchthatp < q, then
p andq arerealnumbers
(i) thetransformed
equationin x' = x -p cannothavefewersignvariations
thanthetransinx" = x - q;
formed
equation
(ii) thenumber
ofrealrootsoftheequation
P(x) = 0, locatedbetween
p andq, canneverbe more
variations
thanthenumber
lostinpassingfromthetransformed
inx' = x -p
ofsign
equation
inx" = x - q;
tothetransformed
equation
is lessthanthesecond,thedifference
is alwaysan evennumber.
(iii) whenthefirstnumber
LikeTheorem1,Budan'stheorem
also givesus an upperboundon thenumber
ofrealrootsof
it onlymakesuse of thetransformatheequationP(x) = 0 insidetheinterval
(p, q). However,
tionsx =p + x' and x = q + x" and doesnotdependon anysequenceofpolynomials.
1 and3 areequivalent;
thisfactcanbe easilyseenifin Fourier'ssequencewereplace
Theorems
x by anyreal numbera. The m+ 1 resulting
numbersare proportional
to thecorresponding
ofthetransformed
coefficients
polynomial
equationP(x + a) = 0, obtainedbyTaylor'sexpansion
theorem.
Budan'stheorem
thebasisofthefollowing
constitutes
statement
[16],[3].
THEOREM 4. LetP(x) = 0 bea polynomial
equation
ofdegreen> 1,withrational
and
coefficients
without
multiple
roots,and letA> 0 be thesmallest
distance
between
anytwoofitsroots.Let m be
thesmallest
indexsuchthat
I
1Fm_l A>
I and Fm-jFm/\> I +-, 1n
whereFk is thekthmember
oftheFibonacci
sequence
1,1,2,3,5,8,13,21,...,
and
n
n)
Let a,, a2,... ,am be arbitrary
Thenthetransformation
positiveintegers.
x=al +
a+
a3+
(1)
+1
m
294
MATHEMATICSMAGAZINE
oftheformx = ai + 1/4,i = 1,2,..., m)

transformations
totheseriesofsuccessive
(whichisequivalent
theequationP(x) = 0 intotheequationP(y) = 0, whichhas notmorethanonesign
transforms
variation.
by Vincent[16],[4].The
presented
versionof theone originally
is an extended
Thistheorem
thatFourierdid
hissurprise
hintedbyFourierand,in hispaper,Vincentindicates
latterwasfirst
article.He
thatwas themainsubjectofVincent's
andprovetheproposition
nottryto go further
whichwerenot
thebeliefthatsuch a proofmayexistin othermanuscripts
states,however,
deathofNavier.
becauseof theuntimely
published
Theorem4 can also be usedin theisolationof therealrootsof an equation.To see roughly
whyit is trueand also howit is applied,observethefollowing:
as
(1) can be also written
transformation
fraction
(i) The continued
X Q ?Y+ Q
',
(2)
fraction
to thecontinued
wherePk/Qk is thekthconvergent
aI +~~~~
a2+
a3+
and,as we recall,
Pk+lIak+lPk+
Qk+1
ak+lQk
Pk-19
+ Qk-1I
is
convergents
(ii) The distancebetweentwoconsecutive
Pm_-I
Pm
Qm-I
Qm
1
Qm-IQmi
valuesoftheQi occurwhenall oftheai = 1.ThenQm = Fm,the

It canbe proventhatthesmallest
and
betweentheFibonaccinumbers
Thisexplainswhythereis a relation
mthFibonaccinumber.
thedistanceA in Theorem4.
of theform
(iii) Let P(y) = 0 be theequationobtainedfromP(x) = 0 aftera transformation
whoseunordered
endpoints
0 <y < oo ontothex-interval
(2). Observethat(2) mapstheinterval
haslengthlessthanA,
Pm-I/Qm- I andPm/Qm.If thisx-interval
convergents
aretheconsecutive
equationP(y) = 0 has at
thenit containsat mostone rootof P(x) = 0, and thecorresponding
mostonerootin (0, oo).
rootx of P(x) = 0 couldbe easily
(iv) If 9 was thispositiveroot,thenthecorresponding
substitutingy
that9liesin theinterval
(0,oo); therefore,
obtainedfrom(2). We onlyknowthough,
whose
in (2) onceby 0 and onceby o0, we obtainforthepositiverootx its isolatinginterval
a
unordered
endpointsare Pm-i/Qm-l and Pm/Qm.To each positiveroottherecorresponds
of
for
the
isolation
to
be
have
m
computed
at most partialquotients
continued
fraction;
different
anypositiveroot.(Negativerootscan be isolatedifwe replacex by -x in theoriginalequation.)
m fromtheinequalities
REMARK. It is clearthatifwe knewthevalueofA, we couldcompute
anytests,we couldobtainP(y) = 0. However,in our algorithmic
of Theorem4. Then,without
procedure(to be describedbelow),we do not initiallyknowA. Thus we need the stronger
testforroot
in orderto havean effective
thatP(y) = 0 has at mostone signvariation
conclusion
For detailssee [3].
in ourtheorem.
theadditionalcomplexities
isolation.Thisis whatrequires
VOL. 55, NO. 5, NOVEMBER 1982
295
Fromtheabovediscussion
itis obviousthatthecalculation
ofthepartialquotients
a1,a2,... ,am
(foreachpositiveroot)constitutes
therealrootisolationprocedure.
(FromBudan'stheorem
we
knowthatthevalueof a particular
partialquotientai has beencomputedif P(x + ai) = 0 has
in thesequenceof its coefficients
thanP(x + ai + 1) = 0.) Thereare two
moresignvariations
Vincent's
to thetwodifferent
methods,
andours,corresponding
of
waysin whichthecomputation
As we willsee, thedifference
betweenthesetwomethodscan be
theai's maybe performed.
betweentheintegrals
ofas beinganalogousto thedifference
ofRiemannand Lebesgue.
thought
Thatis, it is wellknownthatthesum 1 + 1 + 1 + 1 + 1 can be computedin thefollowing
two
? = 2, 2?+ = 39,3 + = 49
, 4 + = 5 (Riemann)and (b) 5 1 = 5 (Lebesgue).
ways:(a) 1 +
a particular
Vincent's
methodbasicallyconsistsofcomputing
ai bya seriesofunitincrementationsai <- ai + 1 (replaceai byai + 1),witheachoneofwhichwe haveto perform
thetranslation
equationP(x) = 0) and checkfora changein thenumber
P(x) <- P(x + 1) (forsomepolynomial
in a methodwithexponential
ofsignvariations.
This"bruteforce"approachresults
behaviorand
As an example,let us isolatetherootsof thepolynomial
henceis of littlepracticalimportance.
equation
0
P(x) = (x-a)(x-f)
fora, whichis 5 109.
wherea = 5 - 109+ E and,6= a + 1. Considera(a), thefirst
partialquotient
UsingVincent'smethodwe seta(a) <- 1, P(x) <- P(x) and computeP(x) P- P(x + 1). Sincethe
in thesequenceofcoefficients
ofthetransformed
number
ofsignvariations
polynomial
P(x) has
not changed,we set a(a) - a(') 1 and computea new P(x) <- P(x + 1), checkingagain the
This processis repeated5 109 timesand, on thefastestcomputer
numberof signvariations.
thatVincent'smethodcan be quite
available,it wouldtakeabout six years!(Note,however,
whenthevaluesof thepartialquotients
efficient
aresmall;forexamplessee [15].)
ofcomputing
on thecontrary,
a particular
Ourmethod,
basicallyconsists
ai as thelowerbound
b on thevaluesof thepositiverootsof a polynomial
equation.(It is assumedthatb= [as (the
orgreatest
whereas is thesmallest
floorfunction
integer
function),
positiveroot.)Thisis achieved
withthehelpof
CAUCHY'S
RULE.Let P(x) = x' + c, 1x'- 1 + * ?c x + co= 0 be a polynomial
equationof
at leastoneofwhichis negative.
degreen withintegral
coefficients,
If X is thenumber
ofnegative
coefficients
ofP(x), then
b
max IXcn lk|/

Cn-k<0
is an upperboundon thevaluesofthepositiverootsofP(x) = 0.
we concludethat
Proof.Fromthewayb is defined
b >XAlCn-kl
foreachk suchthatc__k< 0; forthesek's thelastinequality

can also be written
as
bnaXACn-klbnk
overall theappropriate
k's we obtain
Summing
n
Xb > X
Cfnklbn-k
k=1
Cn-k<O
or
n
bn
2
k=1
|Cn-k I b
nk
Cn-k<O
Fromthelastinequality
we concludethatifwe substitute
b forx in P(x)
296
0, thefirstterm,
i.e.,
MATHEMATICSMAGAZINE
thanor equal to thesumof theabsolutevaluesof all thenegative

bn, willbe greater
coefficients.
Therefore,
P(x) > 0 forall x > b.
Observethatcomputing
thelowerboundb of P(x) = 0 is equivalent
to computing
theupper
boundon thevaluesof thepositiverootsofP(1/x) = 0. It mightbe thought
thatCauchy'srule
a greatamountofcomputation,
requires
sinceit seemsthatthecalculation
ofkthrootsis needed.
This, however,is not truebecause insteadof computingeach kth root we compute,very
thesmallest
efficiently,
integer
m(k) suchthat
(k)
|IXcf_|lk
and thenwe setb= 2K+ 1,whereK is themaximum

ofthem(k)'s. For detailssee [5].
Once we have computeda1i- b, b l1, we need to perform
onlyone translation,
namely,
withour
P(x) <- P(x + b) whichtakesthesame amountof timeas P(x) <- P(x + 1) [7]; therefore,
methodwe haveenormous
savingsof computing
time,and thepreviousexampleis solvedin a
ofa fewseconds.In whatfollows
matter
wepresent
a recursive
definition
ofourmethodas found
in [3]:
Let
P(x) = 0
(3)
in thesequenceofits
rootsand withv signvariations
be a polynomial
equationwithout
multiple
coefficients.
integer
If v = 0 orv = 1: FromtheCardano-Descartes
ruleofsignsweknowthatv = 0 impliesthat(3)
has no positiveroots,whereasv = 1 indicatesthat(3) has exactlyonepositiveroot,in whichcase
in eithercase,no transformation
of(3) is necessary,
and themethod
(0,oo) is itsisolating
interval;
terminates.
We first
If v > 1: In thiscase (3) has to be further
investigated.
computethelowerboundb on
thevaluesof thepositiverootsand thenwe obtainthetranslated
equationPb(x) = P(x + b) = 0,
whichalso has v signvariations
rootof
providedP(b) #0 (ifP(b) = 0, we havefoundan integer
theoriginalequationand v is decreased).The equationPb(x) = 0 is now transformed
by the
substitutions
x <- x + 1 and x l- I/(x + 1), and theprocedure
is appliedagaintwice,oncewith
Pb(l/(x + 1)) = 0 in place of (3) and once withPb(x + 1) = 0.
We haveimplemented
of
ourmethodin a computer
algebrasystem
(fora detaileddescription
thealgorithms
see [2]) and havebeenable to showthatitscomputing
timeboundis
0(n'L(I p
1. ) )
whichis thefastest
obtainedso farwhenexactinteger
arithmetic
areused.
algorithms
TABLES 1 and 2 showtheobserved
timesforthemethodsof Sturm,
and
computing
Vincent,
All timesare in secondsand wereobtainedusingthe
oursforcertainclassesof polynomials.
on theIBM S/370computer,
locatedat theTriangleUniversities
SAC-1computer
algebrasystem
Center(NorthCarolina),wherea subroutine
CCLOCK is availablewhichreadsthe
Computation
clock[3].TABLE 1 clearlyindicates
methodis
Sturm's
computer
that,forthisclassofpolynomials,
out of therace,whereasTABLE 2 makesclear theexponential
natureof Vincent's
completely
method.
withRandomlyGeneratedCoefficients
Polynomials
____________
Computation
Time
Degree
Sturm
Our Method
5
10
15
20
2.05
33.28
156.40
524.42
.26
.48
.94
2.36
TABLE 1
VOL. 55, NO. 5, NOVEMBER 1982
297
ofDegree5 withRandomlyGeneratedRoots
Polynomials
Time
Computation
Roots are
in theInterval
(0,102)
(0,103)
(0,104)
(0,105)
Vincent
Our Method
.45
1.61
16.43
175.62
.16
.71
2.01
4.81
TABLE 2
References
[1]
[2]
[3]
[4]
[5 ]
[6]
[7]
[8 ]
A. G. Akritas,
A newmethod
forpolynomial
realrootisolation,
Proceedings
ofthe16thAnnualSoutheast
Regional
ACM Conference,
Atlanta,
GA (1978)39-43.
, Thefastest
fortheisolation
oftherealrootsofa polynomial
exactalgorithms
equation,
Computing,
24 (1980)299-313.
, Animplementation
ofVincent's
theorem,
Numer.
Math.,36 (1980)53-62.
, Vincent's
forgotten
theorem,
its extension
and application,
Internat.
J. Comput.Math.with
Applications,
7 (1981)309-317.
, Exactalgorithms
fortheimplementation
of Cauchy'srule,Internat.
J.Comput.
Math.,9 (1981)
323-333.
A. G. Akritas
andS. D. Danielopoulos,
On theforgotten
ofMr.Vincent,
theorem
Historia
Math.,5 (1978)
427-435.
, On thecomplexity
ofalgorithms
forthetranslation
24 (1980)51-60.
ofpolynomials,
Computing,
M. B6cher,
Thepublished
andunpublished
works
ofCharlesSturm
on algebraic
anddifferential
equations,
Bull. Amer.Math. Soc., 18 (1911) 1-18.
L. E. Dickson,
FirstCourseintheTheory
ofEquations,
NewYork,1922.
Wiley,
ofthearithmetical
[10] F. Cajori,A history
methods
ofapproximation
ofone
totherootsofnumerical
equations
unknown
ColoradoCollegePublications,
quantity,
GeneralSeriesNo. 51, ScienceSeriesvol.XII, no. 7
(1910)171-215,ColoradoSprings,
CO.
arithmetic
forpolynomial
realzerodetermination,
J.Assoc.Comput.
[11] L. E. Heindel,Integer
algorithms
Mach.,18(1971)533-548.
UberdenSatzvonBudan-Fourier,
[12] A. Hurwitz,
Math.Ann.,71 (1912)584-591.
undBerechnung
derNullstellen
VEB Deutscher
reeller
Verlagder
[13] N. Obreschkoff,
Verteilung
Polynome,
Berlin,
1963.
Wissenschaften,
[14] S. R. Petricle,
ed.,Proceedings
ofthe2ndsymposium
on symbolic
andalgebraic
manipulation,
ACM,1971.
ofEquations,
NewYork,1948.
[15] J.V. Uspensky,
Theory
McGraw-Hill,
[16] A. J.H. Vincent,
Surla resolution
desequations
numeriques,
J.Math.PuresAppl.,1 (1836)341-372.
Introduction
to theTheory
ofEquations,
NewYork,1938.
[17] L. Weisner,
MacMillan,
[ 9]
The Last Duel

Hector'spaces
werevectorspaces.
-RICHARDMooRE
298
MATHEMATICSMAGAZINE
Vincents Theorem from a modern point of view

Alberto Alesina - Massimo Galuzzi
June 9th 1998.
Introduction
In this paper, after having summarized the main results we obtained in [2], we
suggest some ideas which may lead to future developments.
The reader may wonder at the very particular nature of our subject, and
whether it is inconsistent with the generality of category theory; but as emphasized in [6] the peculiar nature of mathematics resides exactly in the force it
gains by contrasting general ideas to facts in a never ending dialectics.
Andre Weil loved to quote Eulers maxim: nihil est in numerico quod non
est in algebraico.
In fact, even the most trivial numerical identity may be the starting point for
a deep understanding of the mathematical structure upon which it may depend
in a subtle and unforeseeable way. On the other hand, no abstract mathematical
structure is meaningful if it isnt able to generate concrete and particular results.
Vincents theorem originally appeared as a sort of complement to Lagranges
method to approximate the roots of algebraic equations via continued fractions.
We described in great detail this aspect of the theorem in [2]. In this paper we
underline its geometrical features which, in principle, make it applicable also in
other situations to obtain different kinds of algorithms.
We conclude our work by giving an example in terms of Farey sequences
(which are very similar to continued fractions...) to emphasize the independence
of the theorem from the particular kind of approximation we devise for the roots.
The polynomials considered throughout the paper have real coefficients and,
for the sake of simplicity, they are assumed to have simple roots, even if (as we
have shown in [2]) this is not a real limitation.
Dipartimento
di Matematica, Universit`
a Statale di Milano.
Positive real roots and variations
The number of sign variations of a polynomial1 gives precise information about

its roots only in the cases it has the value 0 or 1.
The value 0, the absence of variations, points out that the polynomial has
no positive roots, while the value 1 indicates the presence of a single positive
root.
But how can it happen that a polynomial has a number of variations greater
than the number of its positive roots ? In this case, how are its complex roots
located in the complex plane ?
The examination of the number of variations of a third degree polynomial
gives a precise suggestion about a general situation.
Consider the polynomial p (x) which has the real positive root a and the
roots i.
Then
p (x) = x3 (2 + a) x2 + 2 + 2 + 2a x a 2 + 2 .
(1)
The possibility that p (x) has 3 variations corresponds to
2 + a > 0
2 + 2 + 2a > 0.
(2)
Let us look at (2) from a geometrical point of view, with reference to the
following figure:
1 Given
the sequence of the real coefficients of a polynomial

0 , 1 , 2 , . . . ,
we say that a sign variation exists between two coefficients p and q if one of the following
conditions holds:
1) q = p + 1 and p and q have opposite signs;
2) q > p + 1 and the numbers p+1 , p+2 , . . . , q1 are all zero while p and q have opposite
signs. We will say a variation of the polynomial to mean a variation in the sequence of its
coefficients.
y = -sqrt(3) x
-a
-a/2
y = sqrt(3) x
Fig. 1
The possibility that the polynomial (1) has 3 variations corresponds to the
fact that the points i are on the right of the line parallel to the imaginary
axis through the point a2 and at the same time to the fact that they are exterior
to the circle having equation
|z + a| = a.
The points
P andQ, in which the line and the circle intersect, have co
a
ordinates 2 , 23 a . Hence independently of the value a they are on the
lines
Im z = 3 Re z.
(3)
by
Suppose that the points i are in the interior of the sector S3 defined
n
S3 = z | Re z < 0
|Im z| <
o
3 |Re z| .
Then the polynomial (1) has exactly one variation.

This sector is the particular case for r = 1 of Obreschkoffs definition:
S = z |z = (cos + i sin ) , > 0, || <

r+2
The result for the third degree polynomials suggests a general behavior.
(4)
(5)
A Lemma contained in [7, p.81], implies that a real polynomial which has
only one positive root while all the other roots are in the sector S3 defined by
(4) has exactly one variation. We gave a simple and constructive proof of this
result in [2].
More generally we proved the
Lemma 1 A polynomial which has r positive real roots and all the other roots
within the sector (5) has exactly r variations.
Proof. See [2, section 8.2].
The properties of a simple geometrical transformation
We expose some geometrical properties of the transformation T : C C, defined by

za
T (z) =
.
(6)
bz
where a and b are positive real numbers.
1) The circle
z a + b = |b a|
2
2
whose diameter lies on the real axis, with endpoints a and b, is mapped by the
transformation (6) onto the imaginary axis. The exterior points are mapped
into the half-plane Re(z) < 0.
2) The lines
Im (z) = s Re (z) (s R)
are the images of the circles having center
c =
|b a|
a+b
i
,
2
2s
and radius
|b a|
r=
2
r
1+
1
.
s2
It easily follows (see Fig. 2) that the sector Ss of the complex plane defined
by
Ss = {z | Re (z) < 0 and |Im (z)| s |Re (z)|}
(7)
is the image of the exterior of the eight-shaped figure R given by the union of
the two disks
z c r.
Im(y) = -s Re(y)
T(x)
Im(y) = s Re(y)
Fig. 2
How the geometrical transformation influences

Given the real polynomial f (z) , without multiple roots, let be the minimum
distance of its roots z1 , z2 , . . . , zn i.e.:
= min |zi zk | .
i<k
Let the numbers a and b of the previous section be such that:
3
|b a| <
.
2
(8)
Then, in particular, |b a| < , and the circle whose diameter is (a, b)

contains no complex roots and at most one positive root of the equation
f (z) = 0.
Let now R be the union of the two disks centered at
c =
|b a|
a+b
i
2
2 3
(9)
|b a|
,
3
(10)
with radius
r=
which correspond to the half-planes bounded by the lines
Im (z) = 3 Re (z) .
It follows from (8) that R contains at most one real root: indeed, the maximum distance between points of R and points of the interval (a, b) is
|b a|
2r = 2
< .
3
The following alternative is then possible:
all the roots of f (z) = 0 lie outside the circle whose diameter is (a, b);
if a (necessarily unique and real) root lies inside this circle, then all the
other roots lie in the complement of R.
In the first case T maps all the roots of f (z) = 0 into the left complex halfplane, while in the second case the image of the positive root is still positive
and all the other roots are mapped into the sector:
o
n
S3 = z | Re z < 0 |Im z| < 3 |Re z| .

The inverse transformation of T is:
S (z) =
a + bz
,
1+z
hence the polynomial
(z) = (1 + z) f
a + bz
1+z
has no variations in the former case, while it has exactly one variation in the
latter.
Example 2 The polynomial
f (z) = z 3 z 2 2z + 2
has the roots 2, 1. The least distance is
= 2 1 0.414 21,
and
3
1
0.3587 > .
2
3
The length of the interval 1 18 , 1 + 18 = 78 , 89 is less than 13 and it contains

only one root. We see that
7 9
+ z
47 3
83 2
51
79
3
=
z
z +
z+
(1 + z) f 8 8
1+z
512
512
512
512
has only one variation. The interval 28 , 38 does not contain any root and we
see that
2 3
+ z
595 3 967 2 521
93
3
(1 + z) f 8 8
=
z +
z +
z+
1+z
512
256
128
64
has no variations.
Vincents Theorem
We observed at the beginning that Vincents theorem was originally formulated

in terms of continued fractions. But, as we ourselves learned by experience, its
original formulation looks rather enigmatic.
We prefer to formulate Vincents result in the following form:
Theorem 3 (Vincent) Let f (z) be a real polynomial of degree n which has only
simple roots. It is possible to determine a positive quantity so that for every
pair of positive real numbers a, b with |b a| < , every transformed polynomial
of the form
a + bz
n
(z) = (1 + z) f
(11)
1+z
has exactly 0 or 1 variations. The second case is possible if and only if f (z) has
a single root within (a, b) .
Proof.
Let denote the least distance of the roots of f (z), and set
3
= 2 .
Remark 4 Usually, every algorithm to isolate the real roots of a polynomial
equation depends on a scan of an interval, which contains all the roots, by subintervals of decreasing amplitude.
Lagranges famous equation au carre des differences may be used (in principle,
but it is a highly impractical tool) to find an upper bound for the least distance of
the roots, and hence to divide the original interval into subintervals of amplitude
less than the least distance. The presence of a real root is marked by the fact that
the polynomial must have opposite signs at the endpoints of every sub interval
containing a root.
Vincents theorem, without any need of knowing a priori the least distance of the
roots, gives a test to determine when an arbitrary method based on a subdivision
into subintervals reaches its goal.
Remark 5 Let us look at the form of the polynomial (x) in (11). Since
ab
n
,
(z) = (1 + z) f b +
1+z
by the help of the Taylor formula we get
)
(
2
f 0 (b) a b f 00 (b) (a b)
n
+
+ ...
(z) = (1 + z)
f (b) +
1! 1 + z
2! (1 + z)2
7
(12)
From (12) it is quite easy to obtain a vector representation of the coefficients of

. Let us write
(z) = 0 + 1 z + . . . + n z n .
Then
...
n
n
0
n

1
n
=
2

...
n
n
n1
n2
n1
n2
...
n1
n2
...
...
...
0
...
0
...
...
0
0
...
0
...
ab
...
...
0
...
0
...
0
...
...
n
. . . (a b)
f (b)
f 0 (b)
f 00 (b)
2!
...
f (n) (b)
n!
A look at the previous formula shows that the vector of the coefficients is obtained
by the product of the matrix

n n1 n2
. . . 00
0
0
0
n n1 n2
... 0
1
1
1
n n1 n2
(13)
,
.
.
.
0
2
2
2
...
...
...
... ...
n
0
0
... 0
n
which depends only on the degree n, by the matrix
1
0
...
0
0 a b ...
.
.
.
0
...
...
...
...
n
0
0
. . . (a b)
(14)
which depends only on a and b, and at last by the vector
f (n) (b)
f (b) , f 0 (b) , . . .
n!
which collects all the information about the polynomial f (z) at b.
This representation of the transformed polynomial, which was the one originally
used by Vincent to prove his theorem (see [2, section 4]), might be exploited to
obtain the same kind of results which are usually obtained by the help of Sturms
theorem. Just to give the most obvious example: Newton rule to find an upper
bound of the positive roots appears as an obvious corollary.2
Suppose
b<a
2 In fact, we use a more general result. If the polynomial (z) has no variations, the
polynomial f (z) cannot have roots in (a, b).
and that
f (b) > 0, . . . , f (n) (b) > 0.
Then the matrices (13) and (14) have only positive coefficients as well as their
product. The polynomial (z) has no variations and hence f (z) has no roots
between b and a. Since a can be chosen arbitrarily, f (z) has no roots greater
than b.
An idea of the future developments we devise may be given by the following
Proposition 6 Suppose f (z) has only real roots and let a, b be positive real
numbers. Then the number of variations of the polynomial (12) is exactly the
number of real roots of f (z) contained in the interval (a, b).
Proof. Since all the roots are real, the eight-shaped figure R of Fig. 2
does not contain any root. Hence all the roots exterior to the interval (a, b)
are mapped by T (z) onto the negative real axis. The application of Lemma 1
concludes the proof.
Corollary 7 The number of variations of the polynomial (12) is greater than
the number of real roots in (a, b), and the difference is an even number.
Proof. It is enough to decompose f (z) as f1 (z)f2 (z) where f1 (z) contains
all the real roots of f (z). The result follows from Lemma 1(observe that the
zero degree coefficient of f2 (z) is positive. The parity of the number of roots of
f (z) is the same as the one of f1 (z)).
We give an example to show how the proposition and the corollary may be
used.
Example 8 Consider the symmetric matrix
1 0 2
0 2 1
2 1 0
3 5 7
3
5
.
7
1
The roots of the characteristic polynomial

f (z) = z 4 4z 3 83z 2 + 216z 36,
are all real. We want to find the number of the roots in (0, 1) .
Considering what we explained in Remark 5, we set b = 0, a = 1 and consequently we simply consider the matrix product
1 1 1 1 1
36
94
4 3 2 1 0 216 334
6 3 1 0 0 83 = 349 .
4 1 0 0 0 4 72
1 0 0 0 0
1
36
Since the resulting vector has one variation, we have one root in (0, 1).
9
Example 9 Consider the polynomial

f (z) = z 6 5z 5 + 7z 4 5z 3 + 9z 2 15z + 17.
(15)
We want to determine the maximum number of its roots in (1, 4).
1 0 0
0
0
0
1 1 1 1 1 1 1
6 5 4 3 2 1 0 0 3 0
0
0
0
15 10 6 3 1 0 0 0 0 9
0
0
0
20 10 4 1 0 0 0 0 0 0 27 0
0
15 5 1 0 0 0 0 0 0 0
0
81
0
6 1 0 0 0 0 0 0 0 0
0
0 243
0 0 0
0
0
0
1 0 0 0 0 0 0
1 3 9 27 81 243 729
6 15 36 81 162 243 0
15 30 54 81 81
0
0
20 30 36 27 0
0
0
.
15 15 9
0
0
0
0
6 3 0
0
0
0
0
1
0
0
0
0
0
0
We have
0
0
0
=
0
0
729
Since the sequence of values

f (1) ,
f 0 (1)
f (6) (1)
,...,
1!
6!
is
9, 3, 1, 7, 3, 1, 1,
and the product
1 3 9
6 15 36
15 30 54
20 30 36
15 15 9
6 3 0
1
0
0
27 81
81 162
81 81
27 0
0
0
0
0
0
0
243
243
0
0
0
0
0
729
0
0
0
0
0
0
(16)
9
3
1
7
3
1
1
459
27
603
495
189
63
9
gives a vector which has two variations, we may have at most 2 roots in (1, 4).
On the other hand, the sequence of values
f (4) ,
f (6) (4)
f 0 (4)
,...,
1!
6!
is
549, 1353, 1261, 587, 147, 19, 1.
(17)
Since the difference of the number of variations of the two sequences (16) and
(17) is 4, the theorem of Budan and Fourier allows to conclude that there are
10
at most 4 roots in (1, 4). The polynomial (15) actually has two roots in (1, 4),
hence the estimate given by Vincents algorithm is, at least in this case, more
precise. But the comparison of the two algorithms will be a matter for future
developments.
Algorithms
A Farey series of order N , which we denote by FN (we take the definition

from [5, p. 118]), is the set of all reduced fractions between 0 and 1 whose
denominators are N or less, arranged in increasing order.
For example, for N = 3 we have
0 1 1 2 1
F3 =
, , , ,
.
1 3 2 3 1
p
m+p
Recalling that a mediant of two fractions m
n < q is given by n+q and that
m+p
p
m
n < n+q < q , it is evident that we can obtain FN from FN 1 by inserting
mediants whenever it is possible to do so without getting a denominator greater
than N (see [5, ibidem]).
Hence
0 0+1 1 1 2 2+1 1
1 1 1 2 3
F4 =
,
, , , ,
,
= 0, , , , , , 1 .
1 1+3 3 2 3 3+1 1
4 3 2 3 4
The Farey series, or even better the Stern-Brocot tree, gives an interesting
way to obtain all the reduced fractions within (0, 1).
It is possible that Vincents Theorem might be connected with the Farey
series in the same fruitful way it was connected with continued fractions. This
may be a direction of future research, but for the moment we content ourselves
to present an example of its use to separate the roots.3
Example 10 Consider the polynomial
f (z) = 12z 3 + 54z 2 34z + 5.
(18)
We want to find its roots

inthe interval (0, 1)
We begin with F2 = 0, 12 , 1 Then
0 + 12 z
11
= 3z 3 z 2 2z + 5
z+1
2
1
+z
47
3
(z + 1) f 2
= 37z 3 + 56z 2 + z + 3.
1+z
2
3
(z + 1) f
3 We are indebted to Donato Saeli, who, at the end of a conference where we described the
contents and applications of Vincents theorem, suggested to look at its connections with the
Farey series.
11
The second polynomial has no variations, hence there are no roots in 21 , 1 while
the presence
of
two variations in the first polynomial shows the possibility of two
roots in 0, 21 .
Now F3 = 10 , 13 , 12 , 23 , 11 and we have to consider the intervals 0, 13 and 13 , 12 .

We have
0 + 13 z
5
11
1
3
(z + 1) f
= z 3 z 2 + z + 5,
1+z
9
3
3
1 1
+ z
4
1
25
3
(z + 1) f 3 2
= 3z 3 + z 2 + z + .
1+z
6
3
9
0 1
We are reduced to the consideration of the interval 1 , 3 . The further
intervals
we need of F4 are simply obtained by the insertion of the mediant: 0, 41 , 14 , 13 .
(z + 1) f
0 + 14 z
1+z
1 3 11 2 13
z + z + z + 5,
16
8
2
+ 13 z
1
1
5
1
= z3 z2 z + .
1+z
9
6
24
16
1 2 2 1
Now we have to try
4 , 7 , 7 , 3 , since there are no fractions of F5 and
1with
F6 in the interval 4 , 13 .
3
1
4
(z + 1) f
(z + 1) f
3
(z + 1) f
+ 27 z
1+z
+ 31 z
1+z
9 3
1
1
1
z z2 + z + ,
343
14
56
16
1 3
1
13
9
z + z2
z
,
9
21
147
343
Each polynomial has a single variation, hence every interval

tains exactly one root.
2
4, 7
2 1
, 7 , 3 con-
References
[1] A.G. Akritas, Elements of Computer Algebra with Applications, John Wiley
& sons, New York, etc., 1989.
[2] A. Alesina - M. Galuzzi, A new proof of Vincents theorem, L Enseignement
Mathematique, second series, vol. 44 (1998), pp. 219-256.
[3] E. Bombieri - A.J. van der Poorten, Continued fractions of algebraic numbers, Computational algebra and number theory, (Sydney, 1992), Math.
Appl., 325, Kluwer Acad.. Publ., Dordrecht, 1995, pp. 137-152.
ements dAlgèbre, Bachelier père et fils, Paris, 1831,
[4] L.P.M. Bourdon, El
sixième edition.
12
[5] R.L. Graham - D.E. Knuth - O. Patashnik, Concrete Mathematics, Reading,

Mass. &c.,1989.
[6] F.W. Lawvere - S.H. Schanuel, Conceptual matematics. A first introduction
to categories, Cambridge University Press, 1997.
[7] N. Obreschkoff, Verteilung und Berechnung der Nullstellen reeller Polynome,
VEB Deutscher Verlag der Wissenschaften, Berlin, 1963.
[8] A.J.H. Vincent, Sur la resolution des equations numeriques, Memoires de
la Societe royale de Lille, 1834, pp. 1-34. Also in Journal de mathematiques
pures et appliquees, 1 (1836), pp. 341-372.
[9] A.J.H. Vincent, Addition à une precedente note relative à la resolution des
equations numeriques, Memoires de la Societe royale de Lille, 1838, pp. 5-24;
also in Journal de mathematiques pures et appliquees, 3 (1838), pp. 235-243.
13
Conrp. & MO/~. with A&.

Printed in Great Britain.
0097-1943/81/lM030%09SO2.0#0
Pergamon Press Ltd.
Vat. 7. pp. 309-317. 1981
VINCENTS FORGOTTEN THEOREM, ITS EXTENSION

AND APPLICATION
ALKIVIADISG. AKRITAS
University of Kansas, Department of Computer Science, Lawrence, KS 66045,U.S.A.
Communicated by F. L. Bauer
(Received September 1980; in revised form November
1980)
Abstract-Vincents theorem of 1836,which was only recently discovered by the author of this article, is of
extreme importance because it consitutes the basis of the fastest method existing for the isolation of the
real roots of a polynomial equation (using exact integer arithmetic). In this paper this forgotten theorem is
presented both in its original form and in an extended version, and is followed by a general discussion of
its application.
1. INTRODUCTION
In the theory of equations it is well known that in the beginning of the 19th century the
attention of the mathematicians had been focused on numerical methods for the solution of
algebraic equations. During this period Fourier conceived the idea to proceed in two steps; that
is, first to isolate the real roots and then to approximate them to any desired degree of accuracy.
Isolation of the real roots of a polynomial equation is the process of finding real, disjoint
intervals such that each contains exactly one real root and every real root is contained in some
interval. Since 1830 the only method widely known and used for this purpose is that of Sturm; it
has been implemented in a computer algebra system-using exact integer arithmetic-and
proven to be[l]
where n is the degree of the square-free polynomial equation P(x) = 0 and L((PI,) the length, in
bits, of the maximum coefficient in absolute value.?
Quite recently, in Uspenskys Theory of Equations ([2] pp. 127-137) the author of this
article discovered Vincents forgotten theorem of 1836[3,4], according to which, if a univariate
polynomial equation with rational coefficients and without any multiple roots is successively
transformed by transformations of the form x = ai +(1/t), for arbitrary, positive, integer
elements a, one eventually obtains an equation with at most one sign variation in the sequence
of its coefficients. As we will see, this theorem can also be used for the isolation of the real
roots. However, as the reader observes, the statement of Vincents theorem is incomplete
because it does not provide a bound on the number of transformations of the form x =
ai + (l/t), which have to be performed in order to obtain the equation with at most one sign
variation. Such a bound is given, though, by the extended Vincent theorem, which was
presented in a somewhat erroneous manner by Uspensky ([2], pp. 298-304),[5].
Two root isolation methods result from the above theorem, Vincents and ours, corresponding to the two different ways of completing the oiS[6,7]. It has been shown[7] that Vincents
method behaves exponentially, whereas ours has the polynomial computing time bound
which in fact is the best one achieved thus far using exact integer arithmetic181.
In what follows, Vincents forgotten theorem is presented both in its original form and in an
extended version and is followed by a general discussion of its application.
tFor a survey of computer algebra systems see Proceedings of fhe Second Symposium on Symbolic and AIgebruic
(Edited by S. R. Petrick), March, 1971,available from ACM.
Manipulation,
309
A.G.AKRITAS
310
2.VINCENT'STHEOREMANDITSEXTENSION
Before we state Vincents theorem we begin with the following:
Definition 1
We say that a sign variation exists between two numbers c, and cP(p < q) of a finite or
infinite sequence of real numbers
if c, and c, are not zero and have opposite signs, and in case q 1 p f 2 (that is, cg does not
immediately follow cP) the numbers c~+~,. . . , q-1 are all zero.
THEOREM
1 (Cardano-Descartes rule of signs)
The number p of the positive roots of a polynomial equation with real coefficients
c&x +
C,X- + * * * + C,_IX
c, = 0
is never greater than the number u of sign variations in the sequence of its coefficients
co, Cl, C2r * * *, c,; if u -p > 0 then it is an even number.
The proof of the above theorem can be found in any text on the theory of equations.
Subsequently we may say that a polynomial has or presents v sign variations, instead of
using the lengthier terminology of Theorem 1.
A closer examination of Theorem 1 reveals that it is a rather weak proposition; it gives us
the exact number of positive roots only in the following two special cases: (i) if there is no sign
variation there is no positive root, and (ii) if there is one sign variation there is one positive
root. As we will subsequently see, these two special cases are of great importance. Moreover,
the converse of (i) is also true because we have:
LEMMA
1. (Stodolu [9] p. 105)
If the polynomial equation
P(x) = C& + c,P
+ * . - + c, = 0
(co 0)
with real coefficients ck, k = 0,1,2,. . . , n, has only roots with negative real parts, then all its
coefficients are positive, and hence, they present no sign variation.
Regarding the second special case of Theorem 1, we observe that the converse is not in
general true as can be seen from the polynomial x3-x2+2x - 2 = (x - 1)(x - l&)(x + ai).
However, under more restrictive conditions the desired proposition is true; formally this is
stated as follows:
LEMMA
2. ([6] pp. 63-66)
Let P(x) = 0 be a polynomial equation of degree n > 1, without multiple roots, which has
one positive real root ,$# 0 and n - 1 roots &, 6,. . . , &_, with negative real parts-the complex
roots appearing in conjugate pairs- and which can be expressed in the form
b=-(l+(~~),
j=l,2
,...,
n-l
with Iail< E,, where

(Ih-I))
e, =
( >
1+;
- 1.
Then P(x), in its expanded form, presents exactly one sign variation.
Having thoroughly analyzed the two special cases of Theorem 1, we can now state Vincents
theorem which depends heavily upon them.
Vincents forgotten theorem, its extension and application
311
THEOREM
2 (Vincent 1836[4])
If in a polynomial equation with rational coefficients and without multiple roots one makes
successive transformations of the form
x=0,+1
x
x=
a,+Jxn=
x
,,(
u3 +-
1
x
,,I 1 * * * ,
where each a,, a2, u3,. . . is any positive integer, then the resulting, transformed equation has
either zero or one sign variation. In the latter, the equation has a single positive real root
represented by the continued fraction
1
Cl,+a,+
03 -+
whereas in the former case there is no root.

The proof of this theorem can be found in Vincents original paper[4]. Vincent himself
states that Theorem 2 was hinted in 1827 by Fourier who never did give any proof of it, or if he
did it was never found. As Uspensky notes ([2], p. 298) Vincents theorem-which is based on
an earlier theorem by Budan[lOl_was so totally forgotten that even such a capital work as the
Enzyclopaedie der mathematischen Wissenschuften ignores it. As far as we have been able to
determine, Vincents theorem is not mentioned by any authors with the exception of
Uspensky[2] and Obreschkoff [91. The author discovered it while reviewing methods for the
isolation of the real roots of equations as presented by Uspensky.
The question naturally arises as to the maximum number of transformations of the form
x = ai + (l/t), necessary to obtain the polynomial with at most one sign variation. Uspensky ([2],
pp. 298-304) extended Vincents theorem in order to obtain an answer to this question. His
treatment though contains certain errors, in the statement and the proof, which were corrected
in[2]. In what follows we give a new, corrected version of the extension of Vincents theorem;
for completeness we also add its proof, which is much shorter than the one by Uspensky ([2]
pp. 298-304) due to the fact that we use Lemma 2.
THEOREM
3 ( Vincent-Uspensky-Aktitas
181)
Let P(x) = 0 be a polynomial equation of degree n > 1, with rational coefficients and without
multiple roots, and let A > 0 be the smallest distance between any two of its roots. Let m be the
smallest index such that
F,,,-ii>
F,_,F,,,A>
1 and
1 +i
(1)
where Fk is the kth member of the Fibonacci sequence

1,1,2,3,5,8,13,21,...
and
1
E,
(IhI-1))
( >
1+;
- 1.
(2)
Then the transformation

1
x=a,+-g+
+l
1
a,+2
CAMWA
Vol. 1. No. Cc
(3)
312
A. G. AKRITAS
(which is equivalent to the series of successive transformations of the form x = ai + (l/t),

i = 1,2,. . . , m) presented in the form of a continued fraction with arbitrary, positive, integral
elements al, uz, . . . a,, transforms the equation P(x) = 0 into the equation P(t) = 0, which has
not more than one sign variation.
Proof. In order to prove the theorem, it suffices to show that, after the m successive
transformations of the form x = Ui+ (l/l), the real parts of all complex roots, as well as all real
roots except for at most one, become negative.
Indeed, let (PJQJ be the kth convergent to the continued fraction
a,+-
u2+1
03 +
..
From the law of convergents we have:

pk+l
uk+,pk
9-1,
Qk+I= uk+,Qk
+ Q/t-l.
Since Q, = 1 and Q2 = u2 L 1, it follows that Qk Z Fk. Further, the relation (3) can be expressed
in the form
from which it follows that

f=-
Pm-,- Qm-IX
pm-Q&
(4)
Clearly, if x0 is any root of the equation P(x) = 0, the quantity to, determined by (4), is the
corresponding root of the transformed equation P(t) = 0.
(a) Assume that x0 is a complex root of P(x) = 0; that is x0 = u ? ib, bit 0. In this case the
real part of the corresponding root to is
r.p.(50)
(pm-~- Qm-,a)Um- Qma)+ Qm-,Qmb'.

(pm- Q,d2+ Q,Zb2
This is certainly negative if

(P,,,-; - Qm-,u)(P, - Q,u) 2 0.
If, on the contrary
(P,n-,- Qm-,aNPm - Qma)< 0,

then clearly the value of a is contained between the two consecutive convergents
pin-l pm
Qm-1'
Qm'
whose difference in absolute value is
(5)
313
Hence,
andI2-U
I<Qm,Q$
I& -aI<Qm-fiQm
from which it follows that
l(Pd - Qm-taWm
- Q,a)l
c&C m
1.
(6)
From (5) and (6) we conclude that the r.p. (&,) will be negative if
To prove that this is true in our case, first observe that, since A is the minimum distance
between any two roots of P(X) = 0, we have
/(a + ib) - (a -
ib)l= )2ibI= 2/b/B A,
from which we obtain /b/L (A/2); morover, we known that Q,,, 2 Q,,_, Z F,,,_,, and, from (l),
F,,,_, (A/2) > 1. Then clearly F,_,Jbl> 1, which implies Q,,Jb( > 1 and &lb1 > 1. From the last
two inequalities we obtain Q,_&,,~* > 1, proving thus, that the r.p. (&) < 0; this is obviously
true for all complex roots of the transformed equation P(g) = 0.
(b) Assume now that x0 is a real root of P(x) = 0. Suppose first that for all real roots Xi,
From (4) it follows that all real roots of the transformed equation &) = 0 will be negative;
moreover, we know from (a), that all the complex roots of P(t) = 0 have negative real parts.
Consequently, due to Lemma 1, P(Z) presents no sign variation. Suppose, now, that for some
real root x0
(pm-~-
Qm-,xoW,,,
Q,,,xo)s 0.
(7)
Then, clearly, xo is contained between the two consecutive convergents
and hence,
Let x~, k# 0, be any other root, real or complex, of P(x) = 0, and & the corresponding root of
the transformed equation. Then, keeping in mind that
PmQm-1
- Pm-IQ,,, = (- l),
it follows from (4) that
+L
(- 1)
QmUm- Qmxd
A. G. AKRITAS
314
or
where
Qm_,;
&lXk).
Uk
=
Now,
I&&l =I&
(+-x,1-
x+x-&
I&xl
Ph-&>O.
and consequently
from the last expression and the second inequality of (1) we deduce that
Thus, the roots 4, k = 1,2,. . . , n - 1, of the transformed equation, corresponding to the roots
xk, k = 1,2,. . . , n - 1, of the equation P(x) = 0, which are all different from x0, are of the form
4=-Q-lQ
(l+akh
(8)
bkl<%;
that is, the roots of the transformed equation have negative real parts and are clustered together
around -1. If we make the substiutions
k=O,l,...,
n-l,
where,
&>O
and
&=-(ltak),
k=1,2
,...,
n-l,
the transformed polynominal P(g) can be written in the form
m =(~)Pw=c &)(u - &)(u- f,). . .(u - ,g_,).

m
Since P(u) satisfies all the assumptions of Lemma 2, it presents exactly one sign variation, and,
obviously, the same is true for the transformed polynomial p(t). The last thing to consider now
is the case when (7) holds as an equality; that is
U,,,-, - Qm-,xoVm - QJO) = 0.
Vincentsforgottentheorem, its extension and application
315
If P,-r - Q,,,_,xO= 0 then we see, from (4), that &,= 0, and clearly the transformed equation
P(g) = 0 has no sign variation (Lemma 1). In the case P, - Q,,xo = 0 we have &,= m and the
transformed equation reduces to degree n - 1. Since again all the roots have negative real parts,
we conclude, from Lemma 1, that P(t) = 0 presents no sign variation. Thus we have proved the
theorem completely.//
From the above theorem we clearly see that m is the desired bound on the number of
transformations of the form x = Ui+ (l/t) which have to be performed in order to obtain the
equation with at most one sign variation in the sequence of its coefficients.
3. GENERAL DISCUSSION
We first show that the generality of Theorem 3 is in no way restricted by the assumption
that the polynomial equation P(x) = 0 should not have multiple roots. (For convenience we
consider P(x) to be a primitive polynomial; that is the greatest common divisor (g.c.d.) of its
coefficients is 1.) The following theorem will be used.
THEOREM
4
Let G be a Gaussian ring (or unique factorization domain) of characteristic zero and P a
primitive, nonconstant polynomial in G[x]. Let P = P;I . . . P$ be the unique factorization of P
into irreducible factors and P its derivative. Then g.c.d. (P, P) = P;I- . . . P2-.
The proof of this theorem is quite obvious and is left as an exercise for the reader. Note that
the integral domain I of the integers is a Gaussian ring.
Let now P be an integral, primitive, univariate polynomial of positive degree and let
P = P;' . . . Pff be the unique factorization of P into irreducible factors Pi, where for all i,
ei > 0. Let e = max (er, . . . , e.) and for 1 I i 5 e define Ji = G: ej = i} and
There follows that P = ;i, $, where some of the Sis may be 1. This is called the square-free
factorization of P. (A polynomial S is called square-free in the case where there is no
polynomial Q of positive degree such that Q* divides S.) Each of the square-free factors Si,
1 I i 5 e has simple roots, which are of multiplicity i for the polynomial P. Therefore, if we
wish to isolate the real roots of P(x) = 0 in the case where there are multiple roots present, all
we have to do is to obtain the square-free factors of P and then apply Theorem 3 to each one of
them.
The square-free factors-of the polynomial P mentioned above-are obtained with the help
of Theorem 4. Indeed, notice that
R = g.c.d. (P, P) = fi PY- = fi Si-
The greatest square-free divisor of P is
T = g = fi Pi = Q Si.
and hence
V = g.C.d. (R, T) = ir Si.
i=2
As a result of the above we have
316
A.G.AKRITAS
Repeating the process with R in place of P we can compute SZ and eventually obtain ail the
square-free factors of P. The algorithm for the above process is quite obvious and a detailed
description of it can be found elsewhere ([6], pp. 30-31); it has been shown that its computing
time bound is
OW+n4uIpI,)2),
where lP/, is the sum-norm.
We can now focus our attention on how Theorem 3 is used in order to isoIate the real roots
of an integral polynomial equation P(x) = 0, which does not have multiple roots. From the
statement of the theorem we know that a transformation of the form (3), with arbitrary,
positive, integer elements a,, a2,. . . , a, transforms P(x) = 0 into an equation P(t) = 0, which
has at most one sign variation; this transformation can be also written as
(9)
where the Pis and QiS (i = m - 1, m) are defined in the beginning of the proof of Theorem 3.
Since the elements al, u2,. . . a, are arbitrary, there is obviously an infinite number of
transformations of the form (3). However, with the help of Budans[lO] theorem we can easily
determine those that are of interest to us; namely, there is a finite number of them (equal to the
number of positive roots of P(x) = 0) which lead to an equation with exactly one sign variation.
Suppose that p(g) = 0 is one of these equations; then from the Cardano-Descartes rule of signs
we know that it has one root in the interval (O,m). If i was this positive root, then the
corresponding root f of P(x) = 0 could be easily obtained from (9). However, we only know
that llies in the interval (0,~). Therefore, substituting tin (9) once by0 and once by ~0we obtain for
the positive root .? its isolating interval, whose unordered endpoints are (Pm_,/Q,,_,), and (Pm/Q,,).
In this fashion we can isolate all the positive roots of P(x) = 0. If we subsequently replace x by - x
in the original equation, the negative roots will become positive and hence they, too, can be isolated
in the way mentioned above. Thus we see that we have a procedure for isolating all the real roots of
P(x) = 0.
As we mentioned in the Introduction the calculation of the quantities al, us,. . . , u,-for the
transformations of the form (3) which lead to an equation with exactly one sign variationconstitutes the polynomial real root isolation procedure. Two methods actually result, Vincents
and ours, corresponding to the two different ways in which the computation of the Ui'Smay be
performed; the difference between these two methods can be thought of as being analogous to
the difference between the integrals of Riemann and Lebesgue (think of the addition).
Vincents method basically consists of computing a particular ui by a series of unit
incrementations, i.e. Ul+Ui + 1, which corresponds to the substitution X+X + 1. This brute
force approach results in a method with an exponential behavior (in the length of the biggest
coefficient in absolute value) and hence of little practical importance. Examples of this
approach can be found in[2] and [4].
On the contrary, we think that our method is an aesthetically pleasing interpretation of
Theorem 3; basically it consists of immediately computing a particular Ui as the lower bound b
on the values of the positive roots of a polynomial, i.e. ai+b, which corresponds to the
substitution xtx + b (performed on the particular polynomial under consideration). (An
unsuccessful treatment of the big values of the Ui'Scan be found in ([2], p. 136)) Since the
substitutions xtx + 1 and x+x + b can be performed in about the same time[ll], we easily see
that our method results in enormous savings in computing time. We have implemented our
method in a computer algebra system and have been able to show that its computing time
bound is
which is the best one achieved thus far: empirical results also verify the superiority of our
method over all others existing[8]. In order to obtain this computing time bound we needed-
317
among other things-a lower bound for A, the smallest distance between any two roots. This is
given by the following:
(Mahler [ 121)
If P(x) is an integral, univariate polynomial of degree n 2 2 then
THEOREM5
where IPI, is the sum-norm and A = min ((Yi- ~~1,

if aI, a2,. . . , ak are the k distinct roots of
Isi<jsk
in case k = 1, A = 00.
More details regarding the computing time analysis of our method can be found in[6] and [8].
It should be pointed out that the algorithms described in these references use exact integer
arithmetic, which is equivalent to exact rational arithmetic. Our method has not yet been
implemented using machine numbers, and, therefore, we cannot say anything about its behavior
in such an environment or about rounding errors.
P(x);
REFERENCES
1. L. E. Heindel, Integer arithmetic algorithms for polynomial real zero determination. 1. ACM l&533-548 (1971).
2. J. V. Uspensky, Theory of Equations. McGraw-Hill; New York (1948).
3. A. G. Akritas and S. D. Danielopoulos. On the forgotten theorem of Mr. Vincent. Hisroriu Mathemmotico5.427435
(1978).
4. A. J. H. Vincent, Sur la rCsolution des tquations numbriques. J. Math. Pures el AppliquCes 1, 341-372(1836).
5. A. G. Akritas, A correction on a theorem by Uspenskv. Bull. Greek Math. Sot. 19, 278-285 (1978).
6. A. G. Akritas, Vincents theorem in algebroic.monipulation. Ph.D. Thesis, Operations deseaich Program, North
Carolina State University, Raleigh, North Carolina (1978).
7. A. G. Akritas, A new method for polynomial real root isolation. Proc. 16th Annual Southeast Regional ACM
Conference, pp. 39-43. Atlanta, Georgia, 13-15 April (1978).
8. A. G. Akritas, The fastest exact algorithms for the isolation of the real roots of a polynomial equation. Computing 24,
299-313(1980).
9. N. Obreschkoff, Verteilunnund Berechnuna der Nullsrellen reeller Polvnome. VEB Deutscher Verlaa der Wissenschaften, Berlin G.D.R. (1963).
IO. A. G. Akritas. The two different wavs of exnressine the Budan-Fourier Theorem and their conseauences. In: 5th
Volume of Lbctures given al the Geierol Maihematt~calSeminar of fhe Universityof Patras, pp. 32&i (1979) (In
Greek.)
I I. A. G. Akritas and S. D. Danielopoulos, On the complexity of algorithms for the translation of polynomials. Computing
24.51-60 1980).
12. K. Mahler, An inequality for the discriminant of a polynomial. Michigan Mulh. L 11,257-262 (1964).

!"# $%$&' (& $'&#& !& () *"+&'& ,&'#")# )- ."/
0& 1# !&)'&2 )- 345 %* *"#06## !& ,%'")6# '&%7 ')) "#)7% ") 2& !)*#
*&'",&* -')2 !&28 )& 6#"9 0) "6&* -'%0 ")# %* () 6#"9 :"#&0 ")#
; !& -)'2&' :&"9 !& -%# &# '&%7 ')) "#)7% ") 2& !)* <&9%'*"9 !&
=) "6&* >'%0 ")# 2& !)* (& ?'# #!)( !)( @ 6#"9 % '&0& 7A *&,&7/
)$&* B6%*'% "0 0)2$7&C" A :)6* ) !& ,%76&# )- !& $)#" ",& ')) # )$)7A)2"%7# ; " # $&'-)'2%0& !%# :&& "2$'),&* :A % %,&'%9& )- DE
),&' " # "" "%7 "2$7&2& % ") %* !& (& "*"0% & *"'&0 ")# -)' -6 6'&
'&#&%'0!

!

"

#

$%

&'( )

# $* #

# !

! +,

+,
$-' &..(
"
& % %*!%%( &-(
#
# /'

0
1
2
&(
#
#

')) "#)7% ") 0) "6)6# -'%0 ")# 0)2$7&C" A ."0& 1#
!&)'&2
F!"# $%$&' "# :%#&* ) !&
!& %6 !)' 9%,& % G=G 4 !& &'%/
")%7 =)-&'&0& ) G$$7"0% ")# )- =)2$6 &' G79&:'% !&7* % <H=/I"J K%9&:&'9
G6# '"% LM67A / 4N
O
G P GQ<FGH

#
#

3

+ 4
5 *%' -
6 7 8 9

#

!
&%:(

+,
1, #

;
&%.(

= =

R&(
45 <
4 ;5

!" #$ %&

4
4 5 ! ;

R&(
(* =&
#
4 5 4 5
2

!

45
5>

3 1,
4
45
4 5

4
5 < * #

5 < *
3 1,

4
5 < #

4
5 < &%.(
2
1,

#
2
# 1,

+,
$-' #

#
2

#
% #
# +,

?

- # "
# +,

#

. #

"

F!&#& () #$&0"%7 0%#&# (&'& R)( ) =%'*%)S " ) !&' ()'*# (!% T&#0%' &#
*"* (%# ) 9&&'%7"J& U=%'*%)1# '67& )- #"9#V F!"# *& %"7 "# 2& ")&* " W5X
.Y=ZYF1H FKZ[<Z\ [> 345

#
#

@ #
+, #

#

4$-'5

A

'#( 4+,

! $-'5

45

=

*

=

=

2

# !
! #
#

45 2

45

#
; B

;
2

# ? A + C ?
? .$
&.%( #
#

0
1 2
&(

";

D &%*!%%(>
E#?
&--( #

E
?D 4-% $5 @ # ?

#

%*** 3 7BB

+,
&%*!%%( 2
#
G P GQ<FGH
'#( ) 4+,

! %***5
45

| |

=
45 < 4 = 5
=

*

45
( &
' #$ *##$
2
3 7BB

F

!
"
+,

2

3 7BB
#

45 2
42
%5
#
#

2
#

#
E
?D
%*!%-

#

#
"

&-%(
'#( 4E
?D, G
%*;%-5

!

<
{ <
= |
*

-
}

" #
$

! % &
F!"#
WX
"# %0 6%77A [:'&#0!R)+1# !&)'&2 -)' !& #$&0"%7 0%#& (!&'& !& 62:&' )#"9 ,%'"% ")# &B6%7# )&
.Y=ZYF1H FKZ[<Z\ [> 345

H B
# 1, !

* !
+,

45 2

45 <

#
IJ

45 < 4 = 5
=
45
=
#

(* =& !

#

4*5

45
!

45 #

2

45

2

#

!

!

IJ

45 < 4 = 5

45 <

;

5 #

4
! " #+ ,#&

3 # #

IJ

45
<

4 +,

;
5 #

K
F

3? B9
? F #

+!3?!B9
?
F

%

K
#
G 3?

+!G!
3?

#

# ;
3 7BB

+!3!
7BB

G# 0% :& #&& &7#&(!&'& WX !& &*$)" # 2%A %7#) :& 0)2$6 &* -')2
LN %*
L3N "- (& ()'R " !& " &',%7 X 3WS " !% 0%#& T&#0%' &#1 '67& )- #"9# *)&# )
%$$7A %* (& 6#& ]#$&#RA1# &# "# &%*
F) *"# "96"#! " -')2 ) !&' 0) "6&* -'%0 ") 2& !)*# #60! %# W3 5 DX 0" &*
" LWDO $$ D;D4XN WX '()* "# '&-&''&* ) %# !& UGR'" %#1 0) "6&* -'%0 ")#
2& !)*V H&& %7#) WX %* WX
Z'')&)6#7A '&-&''&* ) " !& 7" &'% 6'& &" !&' %# U2)*"?&* ]#$&#RA1# 2& !)*V
L35;345N WOX )' %# UT&#0%' &#1 2& !)*V L345;5N WX
F) *"# "96"#! !&2 -')2 H 6'21# :"#&0 ") 2& !)* WDOX
3
G P GQ<FGH

+G3;
#
! /' G 3? &%/( F

;
# %**. L M &-.( 2

+3;
# ! /$ 3?;
&% -( . 3? H
B9
? &*(

# %**$ 3? B9

?
+? &(
3

? #

2
#
# K
F

>

"
F
# #
# "

-(*" G

*
4
5
#
# # 4
#

< 5
F
@
" D

K
: 'K

** ? :' F
:*

*%*
? $***

2

# ? .

#
F

/K

3 %

I / ? %--

# #
-*

# **:

"
-*
E

-

N

)

#

.Y=ZYF1H FKZ[<Z\ [> 345
33
" + ,
-. /! 0& ,!
1
!,

" + ,
.2 3 ()
,4

! 4 4 ,
" 4.
! !& /! 0
' #&&!. $ #& ('#. .,. $#(

& & '/
#(
2

+;
,
# + $-' &..(
" > +

= #

=

;
/$

# 3?;
#

0
1 2
&(

#

!

# #

= #

?
= &%!.(

&'(>
# #
#

&' G
/(
3
G P GQ<FGH

# "

#
;

'
B9
? &.(

' #

# "

&.( #
#

&-.(
#

# 2 8 &.(>
&/( %**/

#

&-' -/(N
3
# #

@ # &%.( #

K
5 ()
,4
! 4
3

D
O
5

4

3
33
3
3
G 6",%'"% & #B6%'&/-'&& $)7A)2"%7LN ZWXLN ^

%* !& \_):"6# '%#-)'2% ") LN ^ ^ Z
G 7"# )- "#)7% "9 " &',%7# )- !&')) # )-LNS

!& 62:&' )- #"9 0!%9&# )-LNS
"-
^

S
"-
^ 3

{XW} ^ 2"L LN LNN
^ 2%CL LN LNNS
% 7)(&' :)6* ) !& $)#" ",& ')) # )- LNS
"- 3{ L ` N L ` N}S
L ` 3N
L N L N
I))R -)' '&%7 ')) # " X 3WS
L3N# 3 % ')) aS
L ` 3N L ` 3NI))R -)' '&%7 ')) # " X3 `WS
"- L3N ^ !&

|

()L N ()L
N
&7#&

|

()L N {WX} ()L
N
&*
N 66

$$ !&

4 ,! /7 , !
L N
4

L N " !
1& )89 $
,
,44
5 () L

;

# #

I

4 .5
0

1,
2

.Y=ZYF1H FKZ[<Z\ [> 345
3
" " / 1 ! ()

,4

! 4 3
! (:

! 4 ' " 0*
! !- F

#

! .

B

#
!
"

4 .5 I
#K
F
# . :
4 5 #
#
" ;

F 3

IJ

F 2

45
45
F : ' $
45
(* =& #
#K
#

(* & ! '
* 4 %5
#
#

#

45 ! /
%
#

( =& !
$ * 4 %5
F 2

IJ
;
- F "
#

/ %
C

#

#

B 7
7
&%(

3D
G P GQ<FGH

4 5

B & .(

)#

# #

#

#

K

" N 2

#
F

#

4
;
&-.( 5

4 #5 &.
.(
F
I

# :****
4
;
&-.( 5

&. .(
5
F
4 %*5 4 *
#

. #
4

&-.( 5

&. .(>
2
3
G
, 4

# # # .5
#

N
2

%**/ #

#

4 5
#

B &-' -/( )#

#

#
#

!

) ' 0# + #& ('#. .,. $#(

& & '#(
2
#
+,
D
K 45

45

2

+,
#
/' G 3? &%/(
.Y=ZYF1H FKZ[<Z\ [> 345
3O

"
+, ;

C ?, 4" #5

(* &
6
45
#
# #

2

4 5
(* &

(* & #

?

(* &

K

<
45

=

4 5
4%5
(* &

( &<(* &

45

" #

4 5

2

(* &

;

(* =& 2
;
#
D # #
0
1,
(* &
,

>
,

( & &%*(
)

(* &> " #
C ? #
#

*& 1

2

4 5

(* & ! !

#

45
R&(

4 5 #

=
4 5<

4 5
4 5
3
1,
4 5 < *

4 5 < #

=

4 5 < *
4 5 <
4-5
#

2
C ?,

# > # #
4 5 <
*

4 5 <
]#$&#RA1#
*&#0'":&* :&7)(
&# "# % #$&0"%7 "# %0& )- !& $)(&'-67 U."0& 1# &# V (!"0! "#
35
G P GQ<FGH
0 8 4-5

;
45
#
(* & # (* =&
#

1,

C ?, #

H,
&.(
#

45
Z&(

# #
K
F
4 455
<
4 4 = 55
#

45
(* &

F
4 455 4 4 = 55 <

#
45
(* &

+ # # H,

! '
3
> #
4 455 4 4 = 55 %
#

" (* &

# +
:
" ; ;,4 $ !& (

44
4
1
2

# #
#
C ? @

;

4 455
<
4 4 = 55
! # H,

!

(* &>
#

4&.% %$(5

=

#
45

4 5 < *

;

# C ?
'
[
!& ) !&' !%* ]#$&#RA 6#&* 0)''&0 7A %* ) !"# %*,% %9& !& ) !&' #$&0"%7
0%#&
LN ^ 3 %# (&77 %# !& 0%#&
LN
.Y=ZYF1H FKZ[<Z\ [> 345
3
" - ;,4 $ !& + 1

44 !
,
1 !
2
C ? #

4 5 < *

(* &>

2

# C ?
#

&: %*(
H#

2

K

;

#

? ! /

0

C ?,
! !- !

3 "

&%.(>

# #
&/(
I
#K
F 2
. '
;

45 #

! : F
( &
G00)'*"9 ) b')-&##)' G7&C&" ] &#!&, WDX )- H /b& &'#:6'91# H % & ]",&'#" A

!& '&%#) -)' ]#$&#RA1# 6#600&##-67 % &2$ (%# !& -%0 !% !&
#%( ."/
0& 1# %0 6%7 $%$&' )- 345 (!&'& c6*%1# !&)'&2 "# # % &* '"9! % !& :&9""9
# &%* ]#$&#RA '&7"&* ) !& <6##"% '%#7% ") )- M/G H&''& 1#
WOX *&&* " H&0 35 $ 3O )- de f
gh ij

kilm n O (& '&%*8
Uk o p qirst ql q M)6'%7 *& \% !u&2% "B6&#
$6'&# % %$$7"B6&# kq pivi hq

ojh

i p oir
ir tq
h r j
qs/
th o p j wlsq p j xvV
b7&%#& ) & !% H&''& $'&#& # 1# !&)'&2 6*&' !& %2& Uc6*%V
34
G P GQ<FGH
5 (:
! 4 3
3

D
O
5

4

3
33
G 6",%'"% & #B6%'&/-'&& $)7A)2"%7L N ZWX LN ^

%* !& )$& " &',%7 XW^XW (!&'& "# % 6$$&' :)6*
) !& ,%76&# )- !& $)#" ",& ')) # )- LNLF!& $)#" ",& ')) # )-L N
%'& %77 " !& )$& " &',%7 X 3WN
G 7"# )- "#)7% "9 " &',%7# )- !& ')) # )-LN

!& 62:&' )- #"9 0!%9&# )- L ` 3N
L NS
"-
^

S
"-
^ 3

{XW}S

L NI))R -)' '&%7 ')) # " X WS
# % ')) aS

L NI))R -)' '&%7 ')) # " X 3WS
"- L N ^

|

(:L XWN
(:L XWN
&7#&

|

(:L XWN {WX}
(:L XWN
&*

3 (:L X WN 66
$$
!&
4 ,! 3 X W

4
L N 5 3 L N
!,4 4
1& , , <
'*
5

" 0 " / 1 ! (:

! 4
3 , 4 ! ()
,4
! 4
' " *
# ( & ( & $
*
F 2

#K
45
(* & #
.Y=ZYF1H FKZ[<Z\ [> 345
3
#

(* & ! .
$ *
45 ! : /
#
#

#

*
#

( & !
' $ *
F 2
- ! "

#

#

: *
2

4 5
#

B &-/(
2

# 3 7BB
%*** &%* %%( +, #

( &<(* & #

2

;
2
%
& &
*

4
5

( & ! !

45
R&(
4

4
#

4
5<
5<

4
=
5 #

=

=
4
4.5
5
O

( &<( *&
#

# C ?,
3
1,
4
5 < *
4
5 < #

2
+,

4 5 < *
4 5 <
#

# > # #

4
cA
4
5 < *
5 <
0)2$%'"#) ]#$&#RA1# &# "# '% !&' (&%R %# " %$$7"&#

,%7 XW^X 3W

"
!& " &'/

G P GQ<FGH
5 (
=,88
! 4 ;
3

D
O
5

4

G 6",%'"% & #B6%'&/-'&& $)7A)2"%7LN ZWX LN ^

%* !& )$& " &',%7 XW^X W (!&'& "# % 6$$&' :)6*
) !& ,%76&# )- !& $)#" ",& ')) # )- LN
G 7"# )- "#)7% "9 " &',%7# )- !& ')) # )-LN

!& 62:&' )- #"9 0!%9&# )- L3 ` N
L NS
"-
^

S
"-
^ 3

{XW}S
H6:*","*& !& " &',%7 XW " () &B6%7 $%' #S
"- LN ^ !&

|

';*L X WN L;*L XWNN
&7#&

|

';*L X WN {WX} L;*L XWNN
&*

X WN 66
$$ !& 4 1
4 =,88 WX>
4 ,! 3 X W&
4
,4 (
$ 5
3 L N
!,4 4
1&
, , <
'*
5 ;L

H#

+,
> #

E
D
#

"

0
#K
F +,

4 ( &5
;

! !-
F
.
( & # ( &
( &
' $ O

>
F 2

45
( & #
#K
#

(
$
#
#

#

$

&

! '
45 ! :
&
#

(
' $
.Y=ZYF1H FKZ[<Z\ [> 345
3
F 2
- ! "

#

#

: $
3 #
# ;

#
&$(

+!
3!7BB

#

3 #

;

#

#

3

45

4 5

<

#

45

/$
# #

G
&-%( 2 # ;

45 <
#

=

*5
4:5

;

'#( 2 4G
,
5

45
" 4:5
*

45

#
<*
<
"
{ }

P, &-*( #

QO30
2 8 %**' &.(
P,

G

#

G P GQ<FGH
'#( 3
4P,
5
" 4:5
*

45

*

<%

45

"
}
)#

!

G
, P, !
#

3
,
#

# . #

!

L
M &-.( 2 #

42 .5 &.(
+"

! "#$ %&' ( ! ( )#* %
%+ %
( , %
L
1
O
*
**
**
*
%**
%**
*
:**
:**
***
%*
%*
4
5
***
:*
:*
***
**
**

45 4 5 45 4 5
*$$%
*'%
%.:%*/
*%'.
----.
$/$$.
*%$$
**..$'/-$
.%/%$'
::*$
--'.
2

2
#?

2

# K #

#

R
#

#

#
#

S T
,

%**: &-$(
.Y=ZYF1H FKZ[<Z\ [> 345
'#( 4S T,

%**:5

45

45 <

=

&(
,&
=
* *

< "

! 45
&(

S T
,

= ! 45

'#

#

#
>
#
S T
,
#?
#

# I

#

#

@ B
S T
,

2
/ ;
# #
&:( 2

#

;
#
#

&'(
'#( 4 43?!B9?!+? %**'5

45 <
= =

4
45

4
*5
5
4
5

$

45 <
45
" 45 " 45 = " 45 " 45 = = " 45 " 45 = ! 45

" 45
< %
% ! 45
< %

" 45 <
= =

G- &' )6' ()'R W35X Hy &-z%&#06 %7#) &C &*&* !"# F!&)'&2 5 WX
4'5

D
G P GQ<FGH

" 45 <

= =
# < 4" 5 # < 4" 5
" 45

" 45
< %

4"
4" 5

45

<
"

{ }

$ < %

4"
5 %
4"
4/5

4" 5
+ !*
" 45 4" 5 4" 5 =

4" 5 < 4" 5

4/5

4 &'(5 0

" 2
' &:(
4" 5
4" 5
! " + ,#& )
0 #

?

!

!

?

4/5

#
K 45 #

" 45 " 45 % 45 # ?

#

#
C 2
/ #
# G
,
P,
K
.Y=ZYF1H FKZ[<Z\ [> 345
O
G
, ; 2
/

45 8 4:5 #
G

?

#

P, ; 2

/

45 8 4:5 P

45 #

%
P, ; 2

/ D

G
,

# ?

#

D # %

2
/

#
"
2
/

2 (*#,&% 0' "& #(*"-1 +#!&. .,. $#(

'#( 4
2

G
, P,

"
' %&" . #$ ' "& #(*"-1 +#!&.
2

#K
F

#

>
F
"
?

C 2
/ #
; # "

;

6 7 ; 2

/
45
# ?
#

4" 5 4" 5 ?

" 45 4" 5 4" 5 =
@

45 #
4'5 #
" #

5
G P GQ<FGH
2

# G

P

# "

-(*" ) G

= *

*

#

<
{ *

..% *
F G
,
#

F P,
#

{ } %
F

#

{ *
*

{ *
*
}
2
#
} { }
{*

}
}

#
"

6 "# "/(- 2

/
45
45 !

8 4:5 ! #

#

#

4:5

-(*" )8 #&&!.

#

*

%
*

*

%

%

@
" !
F "

# &'(K
F P, 4
5

G
,>

?

#

?

F E
;

4
.Y=ZYF1H FKZ[<Z\ [> 345

5

P,

# #

?

2

" 4& ' ' 5
'
&' =

"

&'(

2 % # #
2

(

-

4& ' ' 5 & ' = ' 4

5 45

45 K 45

G
, 4

5 45

42 %
42 %5
&/(5
1
D

-
45
#
D

3

2 % #
2

(
#

! # -
#
)#

2 % #

- # #

! #
:

@
# ! ;
!

( #

-
#
; :U &(
2) (*#,&% 0' 9!. #(*"-1 +#!&. .,.

$#( '#( 4
2

#

"

"

' %&" . #$ ' 9!. #(*"-1 +#!&.
2

#K

F

#

?>
F
"

?

4
G P GQ<FGH
?4,
!
4! 5
,
!, = 8 ? ,! /

& , , 3 , !
0 =; ./
<)) # T&9
3
3
3

3
3
3

O

O
3
<)) # T&9
3
3
3

3
3
33

O

O
3
()':,
1* LN
()'@!* LN
G,&'%9& L\"{\%CN G,&'%9& L\"{\%CN
3DLD4N OL44N
3DL3DN
3O3L3D355N
35L3534N
3DL35343N
55LD4DN 3LDDN
35L3DODDN
5LO34DN
OL543N
3L333D5N
(:'*LN
G,&'%9& L\"{\%CN
D5L44DN
LDON
3L33DN
4L4DDN
3L353N
3L35334N
\&2)'A L\cN
)7*{&({'&7
DD5DD4DO5
DDOO
545OD334
DOD5DO3
O45OOOO
3D333

"

;
"

D

@ #
" "

4
5 " &$(> P,
" # " ) $ &%$( @ #

& '(
;

&( )# #

#

"

"
&(
#!)67* :& ) &* !% "2& "# ) #) "2$)' %0& " )6' 0%#& #"0& ; %# 0% :&
#&& " !& *&#0'"$ ") )- !& () %79)'" !2 7"& D ; !&#& :)6*# %'& &# "2% &*
% '%#7% ") )- 0)2$7&C" A % L N "# &C&06 &* WX
.Y=ZYF1H FKZ[<Z\ [> 345
6: ;6# "/-< :!.

/

* #

%
?
" 2

;
45 4:5

;
F

#
"
" >

!
?
$>
2
#
K

?

!

"

< "

{ } { }

2 6IV

; .*U

@ 2 - ! 2 $ &( ! #
#

"
2

#

"
& ' = '

6IV

4!
5 3 4, , A
?4,
!
c" /7&9 ! T&9'&&

)- ')) #
3
3
3
3
3
3
3

O

O
3
()L=%60!AN
LN
G,9L\"{\%CN
D5L4DN
3D5L3D34ON
343L35O34N
LD3DN
5L455N
D4L5O5N
()L|`72N
LN
G,9L\"{\%CN
DL344N
3DL3435N
343L355344N
LN
43L534N
34L33N
()L72BN
LN
G,9L\"{\%CN
DLD3N
3DL335N
3L34DN
LDN
43LO333N
3O4L333N
! "

@
+,
$-' #

#

G P GQ<FGH
#

#

#
! #

" !

.*U
@ #

K
F
,

;
)

# #

F 2

#?

) # # # #

;

#

#
)#

#

F 2

#
*****
) # # # #
;
"

F 6 #

#

3 G P GR'" %#

b!T F!&#"# [$&'/
% ")# <&#&%'0! b')9'%2 Y)' ! =%')7"% H % & ]",&'#" A <%7&"9! Y= L34N
G P GR'" %#

! @ Y62&'"#0!& \% !&/
2% "R L34N O;5
G P GR'" %# ! "

#
@ =)2$6 "9 L34N ;3
D G P GR'" %# $%

&

@ \% !/
&2% "0# \%9%J"& O L34N ;4
O G P GR'" %# !
'(
) @ 8 b')0&&*"9# )- !& 345
HA2$)#"62 ) HA2:)7"0 %* G79&:'%"0 =)2$6 % ") }% &'7)) [ %'") =%%*%
L345N $$ 44;
5 G P GR'" %# *

M)! }"7&A /
&'#0"&0& Y&( ~)'R 34
G P GR'" %# !
'+ ) @ \ M }&# &' %* \
c&%6*" LZ*#N =)2$6 &' G79&:'% " Z*60% ") G677)%b'&## ]HG L4N
$$ 3;O
4 G P GR'" %# ,

# "

@ LH6:2" &*N
G P GR'" %# G G'9A'"# G } H 'J&:)u#R" ,-. -
" &

/ $ /
@ H&'*"0%
M)6'%7 )- =)2$6 "9 L4N 3DO;35
.Y=ZYF1H FKZ[<Z\ [> 345
3
3 G P GR'" %# G c)0!%'), G } H 'J&:)u#R" 0

1

2 @ G:# '%0 # )- !& &'% ")%7 =)-&'&0& )
&',%7 %* =)2$6 &'/G79&:'%"0 \& !)*# " H0"&0& %* Z9"&&'"9 L &',%7
1DN H b& &'#:6'9 <6##"% L\%'0! /3N L3DN ;
33 G P GR'" %# H T T%"&7)$)67)# 3

2

@
K"# )'"% \% !&2% "0% L34N D;DO
3 G P GR'" %# H T T%"&7)$)67)#

@ P%" %/c!%'% " Lc677& " )- !& *"% H)0"& A -)' K"# )'A
)- \% !&2% "0#N L34N D3;D
3 G P GR'" %# Q K Y9 *"
"

@ =)2$6 "9 L34N 5;5
3D G P GR'" %# G H 'J&:)u#R"
@ Y)7"&%' G%7A#"#8 \)*&77"9 %* =) ')7 D LON ;D
3O G P GR'" %# b ."9R7%#
2

&
/ $ /
@ M)6'%7 )- ]",&'#%7 =)2$6 &'
H0"&0& D LN DOO;D5
35 G P GR'" %# G H 'J&:)u#R" b ."9R7%# 0

4 !

&
/ $ /
@ =)2$6 "9 ! L5N
OO;5
3 G P GR'" %# G H 'J&:)u#R" b ."9R7%#

2 (
& *
/ $ &
8 b')0&&*"9# )!& 3 ! &'% ")%7 })'R#!)$ ) =)2$6 &' G79&:'% " H0"& "?0 =)2$6 /
"9 =GH= LN $$ D; c) P&'2%A H&$ &2:&' 35; IY=H
D H$'"9&' .&'7%9 c&'7" Z*" &* :A . P P%J!% Z } \%A' %* Z .
.)')J! #),
34 G P GR'" %# G H 'J&:)u#R" b ."9R7%# 3
&
2
+

! @ H&'*"0% M)6'%7 )- =)2$6 "9 L4N 4;
3D
3 G P GR'" %# G H 'J&:)u#R" b ."9R7%# 0
/

2 (

&
/ $ @ Y)7"&%'
G%7A#"#8 \)*&77"9 %* =) ')7 LN L4N 5O;
G G7&#"% \ P%76JJ"

@ I1Z#&"9&2&
\% !&2% !"B6& L34N 3;O5
3 G G7&#"% \ P%76JJ"
'

1
) @ I1Z#&"9&2& \% !&2% !"B6& L3N ;4
G G7&#"% \ P%76JJ"

! 2
/
@
Lc& " < %* I%(,&'& } > L&*#NN =% &9)'"0%7 H 6*"&# " %7A <&*"0) "
*&7 ="'0)7) \% &2% "0) *" b%7&'2) H&'"& LN 3;33
Z c)2:"&'" G M ,% *&' b))' &

@ 8 =)2$6 % ")%7 G79&:'% %* Y62:&' F!&)'A LHA*&A 3N \% ! G$$7
O Q76(&' G0%*&2"0 b6:7"#!&'# T)'*'&0! 3O $$ 3;3O
D > c)67"&'
"5 #
6 ') 7 @ I&0 Y) &# ]"/
,&'#" u& I"77& 3 4 %,"&'
, .B)#+C.B)D.BC,4 )'
, 3 ,4
O > c)67"&' /

[0 ):&'

G P GQ<FGH
5 T P =% )' b K P%7A&% K P "22&'

$ 4 @ \% !&2% "0# )- =)2$6 % ") L33N L3N
4O;3
P Z =)77"# G P GR'" %# /

+

@ 8 b')0&&*"9# )- !& 35 G=\ HA2$)#"62 ) HA2:)7"0 %* G79&:'%"0
=)2$6 % ")# ~)'R )( K&"9! # Y~ L35N $$ ;O
4 K K)9 &

@ M
HA2: =)2$6 LON L34N O3;O4O
M ,) J6' P% !& M P&'!%'* !

+1

*#
@ 8 b')0&&*"9# )- HHG=1 \%6" K%(%"" ]HG L3N
$$ D;D
c Q")6# &7"*"# &

@ M =)2$6 G$$7
\% ! :-35 L345N D3;DD
3 Z Q I7)A* 3

2.

@ K"# )'"% \% !&2% "0% L3N
DD4;DO
Y [:'&#0!R)+

&

4
/
@
.Zc T&6 #0!&' .&'7%9 *&' }"##&#0!%- & c&'7" L35N
G \ [# ')(#R" 4

! @ F!& G%7# )- \% !&2% "0#
* H&'"&# L3ON ;
D > <)6"77"&' b "22&'2% *

@
M)6'%7 )- =)2$6 % ")%7 %* G$$7"&* \% !&2% "0# LDN ;O
O M/G H&''& .)7 3 b%'"#8 P%6 !"&';."77%'# L3455N
=)$"&# )- !&#& ,)762&# 0% :& *)(7)%*&* -')2

34
,4,%

5 . H!%'2% " $ $ 0

(

@
HGG= $'&$'" LN
. H!%'2% "

b!T
F!&#"# T&$%' 2& )- =)2$6 &' H0"&0&# =)6'% # " 6 & )- \% !&2% "0%7
H0"&0&# Y&( ~)'R ]",&'#" A
4 T Hy &-z%&#06 4

@ M)6'%7 )- ]"/
,&'#%7 =)2$6 &' H0"&0& L3N LON 3;3D3
T Hy &-z%&#06 &
$ $

3
/
1
@ 8 . P P%J!% Z } \%A' %* Z . .)')J! #), LZ*" )'#N8 /

89 0

:

6 1

=GH= $$ ; 3 c) P&'2%A H&$ &2:&' 35/ IY=H
D H$'"9&' .&'7%9 c&'7" K&"*&7:&'9

D Q F!677 "

/
$ $ @
b')0&&*"9# )- !& 34D G=\ HA2$)#"62 ) HA2:)7"0 %* G79&:'%"0 =)2$6 %/
")# IY=H L34DN 5;
D3 Z b F#"9%'"*%# Z2"'"# (

5

@ L~ GJ%' %* F Z'7&:%0! LZ*#NN ZHG 5 IY=H !
L5N 43;44
D M . ]#$&#RA ! *#
\0P'%(/K"77 Y&( ~)'R 3D4
D G ~6 ] &#!&, /

H&$ &2:&'
>)'
% Z97"#! '%#7% ") )- % :))R (" ! #"2"7%' 0) & #&&8 [:'&#0!R)+ Y8
U&')# )- b)7A)2"%7#V c679%'"% G0%*&2"0 \))9'%$!# LN H)?%
.Y=ZYF1H FKZ[<Z\ [> 345

DD G M K ."0&

#

# @ M)6'%7 *&
\% !u&2% "B6&# b6'&# & G$$7"B6u&&# L345N D3;
DO = Q ~%$

/ [C-)'* ]",&'#" A
b'&##
T&$%' 2& )- =)2$6 &' %* =)226"0% ")
Z9"&&'"9 ]",&'#" A )- F!&##%7A P'&&0&
*1 8 %R'" %#6 !9'
qi 3D qrjr

Vincent

Uploaded by

Copyright:

Available Formats

Vincent

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Vincent

Uploaded by

Copyright:

Available Formats

Nonlinear Analysis: Modelling and Control, 2005, Vol. 10, No.

A Comparative Study of Two Real Root Isolation

University of Thessaly, Department of Computer and Communication Engineering

1 Description of the two algorithms

Description of the Continued Fractions Algorithm CF

and by interval data we denote a list

A Comparative Study of Two Real Root Isolation Methods

5. If 1, set p(x) p(x + ), b a + b, and c c. If p(0) = 0, add

8. If s1 < s2 , swap {a1 , b1 , c1 , d1 , p1 , s1 } with {a2 , b2 , c2 , d2 , p2 , s2 }.

Description of the Algorithm REL

5. If p(0) = 0, add [B 2ck0 , B 2ck0 ] to rootlist, and set p(x) p(x)/x.

6. If k 0 k, compute s sgc(p). If s = 1, add (B 2ck0 , B) to rootlist. If s = 0

A Comparative Study of Two Real Root Isolation Methods

randomly generated polynomials. When all coefficients were randomly generated

A Comparative Study of Two Real Root Isolation Methods

Table 4. Products of factors (x-randomly generated integer root)

on Interval and Computer-Algebraic Methods in Science and Engineering

Author: Alkiviadis G. Akritas

Anistoriton Journal, vol. 13 (2012-2013) Essays

Alexandre Joseph Hidulphe Vincent

Anistoriton Journal, vol. 13 (2012-2013) Essays

Anistoriton Journal, vol. 13 (2012-2013) Essays

Anistoriton Journal, vol. 13 (2012-2013) Essays

The calculation by Vincent of the Plethonian year is made in comparison to

Anistoriton Journal, vol. 13 (2012-2013) Essays

Anistoriton Journal, vol. 13 (2012-2013) Essays

Anistoriton Journal, vol. 13 (2012-2013) Essays

Anistoriton Journal, vol. 13 (2012-2013) Essays

Anistoriton Journal, vol. 13 (2012-2013) Essays

Anistoriton Journal, vol. 13 (2012-2013) Essays

Katelis S. Viglas, PhD (History of Philosophy)

Anistoriton Journal, vol. 13 (2012-2013) Essays

Anistoriton Journal, vol. 13 (2012-2013) Essays

Bull. Math. Soc. Sci. Math. Roumanie

Counting the number of real roots in an interval with

Alkiviadis G. Akritas and Panagiotis S. Vigklas

Counting the Number of Real Roots

the Sylvester-Habicht subresultant prs algorithm. Sylvester initiated this

Alkiviadis G. Akritas and Panagiotis S. Vigklas

The subresultant version of Sturms theorem and the extension of

Counting the Number of Real Roots

where sign is the known signus function:

Alkiviadis G. Akritas and Panagiotis S. Vigklas

The two versions of Vincents theorem

Vincents theorem, in its original form, can be stated as follows:

Counting the Number of Real Roots

Alkiviadis G. Akritas and Panagiotis S. Vigklas

is possible to determine a positive quantity so that for every pair of positive

has exactly 0 or 1 variations in the sequence of its coefficients. The second

Figure 1: Obreschkoffs cone or sector.

Counting the Number of Real Roots

Algorithm to count the number of real roots in any open interval:

Alkiviadis G. Akritas and Panagiotis S. Vigklas

Counting the Number of Real Roots

[16] Alexandre Joseph Hidulphe Vincent, Sur la resolution des equations

Note on Vincent's Theorem

will be seen in the next section,

when the first

number is less than the second,

= CO + clx + . . . + cnxn = 0, then