Download presentation
Presentation is loading. Please wait.
Published byBenjamin McCoy Modified over 6 years ago
1
Vincent’s Theorem of 1836: Overview and Recent Developments
Alkiviadis G. Akritas Department of Computer & Communication Engineering University of Thessaly Volos, Greece
2
Outline of the talk: We will present Vincent’s theorem and review the various real root isolation methods derived from it (2 bisection methods and 1 Continued Fractions method). We will concentrate on the Continued Fractions method (the fastest of them all) and show how it was recently speeded up by 40% over its initial implementation. ACA 2008, Linz, Austria
3
Descartes’ Rule of Signs --- in the open interval ]0, +∞[
ACA 2008, Linz, Austria
4
Special Cases of Descartes’ Rule of Signs
0 sign variations positive roots 1 sign variation positive root ACA 2008, Linz, Austria
5
Vincent’s Theorem (1836) – the Continued Fractions version
ACA 2008, Linz, Austria
6
Vincent’s Theorem … was kept “alive” by Uspensky --- in his book “Theory of Equations” (1949), and was rediscovered by Akritas in 1976; was the subject of his Ph.D. Thesis. ACA 2008, Linz, Austria
7
Historical Note on Uspensky (1883-1947):
Uspensky never read Vincent’s original paper of Instead, Uspesky read Vincent’s theorem (only) in a Russian translation of Serret’s (French) book on Algebra (1866). ACA 2008, Linz, Austria
8
Vincent’s Theorem – the Bisection version by Alesina & Galuzzi (2000)
ACA 2008, Linz, Austria
9
Alesina-Galuzzi’s proof of Vincent’s theorem …
is the most recent one; it uses Obreschkoff’s theorem of which gives the necessary condition for a polynomial with one positive root to have one sign variation! ACA 2008, Linz, Austria
10
Obreschkoff’s Cone or Sector Theorem (V=1)
ACA 2008, Linz, Austria
11
Obreschkoff’s Cone & Circles in picture …
ACA 2008, Linz, Austria
12
Real Root Isolation Using Vincent’s Theorem
ACA 2008, Linz, Austria
13
Different Ways to Isolate the Real Roots:
ACA 2008, Linz, Austria
14
Why bother with these and not use Numeric Methods?
Numeric Methods CANNOT isolate just the positive roots! They isolate ALL roots (real and complex). They can give wrong answers as the following example demonstrates. ACA 2008, Linz, Austria
15
Consider the polynomial: 10999(x-1)50 – 1 (2 positive roots ≠1)
Numeric Method using 1010 digits takes 56 ms and fails (finds all 50 roots =1)! Numeric Method using 1020 digits successfully isolates the roots but takes ms! The VAS Continued Fractions method, discussed below, takes 4 ms! ACA 2008, Linz, Austria
16
Therefore … We DO need the real root isolation methods derived from Vincent’s Theorem of 1836. Especially so since the one developed by Vincent in 1836 is exponential! ACA 2008, Linz, Austria
17
Vincent’s Exponential Method of 1836 …
… uses continued fractions and is described in his original paper of 1836!, and … was erroneously and unintentionally attributed by Uspensky to himself (1947). ACA 2008, Linz, Austria
18
Methods derived from Vincent’s Theorem …
… were developed to overcome the exponential behavior of Vincent’s method. We begin with the bisection methods ACA 2008, Linz, Austria
19
The Bisection Methods Derived from Vincent’s Theorem differ in:
the termination test (criterion) used, and the open interval, ]a, b[, they bisect. ACA 2008, Linz, Austria
20
The first bisection method, VCA, …
was developed in 1976, uses Uspensky’s termination test, and bisects the open interval ]0, 1[. ACA 2008, Linz, Austria
21
The termination test used in VCA is named after Uspensky because …
Uspensky was the one to use it as a test, in an isolation method that he erroneously attributed to himself; Uspensky developed the test since he was not aware of Budan’s theorem of 1807 (eclipsed by Fourier’s theorem). ACA 2008, Linz, Austria
22
Budan’s theorem – to be found in Vincent’s paper of 1836 (and my work).
ACA 2008, Linz, Austria
23
Vincent vs Uspensky ACA 2008, Linz, Austria
24
Uspensky’s Termination Test, for the interval ]a, b[ = ]0,1[
ACA 2008, Linz, Austria
25
VCA method: The 1st Bisection Method Derived from Vincent’s Theorem (1976)
ACA 2008, Linz, Austria
26
VCA is implemented in maple and its other names were:
“Modified Uspensky’s method” ( ) 1986 : “There is no Uspensky’s method” “Descartes’ method” OR “Collins-Akritas” ( ) 2007 : “There is no Descartes’ method” ACA 2008, Linz, Austria
27
Method’s correct name:
Since it does NOT use Descrates’ test Methode de Vincent-Collins-Akritas Coined by Francois Boulier, University of Lille, France, 2007 ACA 2008, Linz, Austria
28
What’s in a name? It makes clear the existing relation with Vincent’s theorem. Can use the results by Alesina & Galuzzi (2000) to estimate the computing time (termination) of all methods derived from Vincent’s theorem. ACA 2008, Linz, Austria
29
The Computing Time of the VCA Bisection Method …
is O( n4 τ2 ), where n is the degree of the polynomial, and τ bounds the coefficient bitsize. ACA 2008, Linz, Austria
30
… and … the fastest implementation of the VCA method was developed by Rouillier & Zimmermann in 2004 (REL)! ACA 2008, Linz, Austria
31
The second bisection method …
… uses Vincent’s termination test, and … bisects the interval ]a, b[ = ]0, ub[. ACA 2008, Linz, Austria
32
Vincent’s Termination Test, for the interval ]a, b[ = ]0, ub[
ACA 2008, Linz, Austria
33
The 2nd Bisection Method Derived from Vincent’s Theorem (A+G:2000)
ACA 2007, Oakland University
34
Comparison of the Two Bisection Methods
VCA, the method using the simpler termination test (Uspensky’s) is faster than the one using Vincent’s (more complex) termination test! ACA 2008, Linz, Austria
35
The Continued Fractions Method Derived from Vincent’s Theorem (1978) …
… uses Descartes’ Rule of Signs as the termination test, and … relies, heavily, on the (REPEATED) estimation of lower bounds on the values of the positive roots of polynomials. ACA 2008, Linz, Austria
36
VAS – Continued Fractions Method (VAS-CF)
ACA 2008, Linz, Austria
37
Comment on the VAS-CF Real Root Isolation Method:
Without steps 4 and 5 it is simply Vincent’s original exponential method. Strzebonski’s contribution is ommitted for simplicity. ACA 2008, Linz, Austria
38
VAS-CF is implemented in Mma. Over the past 30 years …
it has been using Cauchy’s bound on the values of the positive roots, it has been several thousand times faster than the VCA bisection method (for random polys) --- even up to times faster than VCA, for Mignotte polys (and nobody believed it until 2005!), only in the case of very many (>50) very large roots (~ 10100) has it been up to 4 times slower than VCA. ACA 2008, Linz, Austria
39
Computing Time of the VAS Continued Fractions Method
Using a plausible hypothesis and the fast translation algorithm by von zu Gathen, its computing time is O( n4 τ2 ), where n is the degree of the polynomial, and τ bounds the coefficient bitsize. (Akritas 1978, Tsigaridas-Emiris 2005) Without any hypotheses its computing time is O( n8 τ3 ) (Sharma 2007). HOWEVER, this bound does NOT match its performance. ACA 2008, Linz, Austria
40
To improve the performance of VAS-CF even further …
… new bounds on the values of the positive roots of polynomials were needed. To understand the nature of these bounds we used Doru Stefanescu’s inspirational work! ACA 2008, Linz, Austria
41
Stefanescu’s Theorem (2005) : it matches +ive with –ive coefficients when the # of sign variations is even ACA 2008, Linz, Austria
42
Our Theorem (2006): it breakes up & matches +ive with –ive coefficients, (a).
ACA 2008, Linz, Austria
43
Our Theorem (2006): it breakes up & matches +ive with –ive coefficients , (b).
ACA 2008, Linz, Austria
44
On our Theorem (Akritas, Strzebonski and Vigklas, 2006):
It is a general theorem from which ALL methods for computing positive bounds on the values of positive roots are derived! This generality is achieved by breaking up and pairing (unmatched) positive coefficients with negative ones in various ways! ACA 2008, Linz, Austria
45
For the estimation of the bounds …
ACA 2008, Linz, Austria
46
Bounds in the Literature are, in general, of Linear Complexity!
That is, EACH negative coefficient of the polynomial is paired with ONE of the preceding (unmatched) positive coefficients and the maximum of all the computed values is taken as the estimate of the bound. ACA 2008, Linz, Austria
47
Linear Complexity Bounds discussed:
We present four of them. Of those four bounds, the last two were developed by Akritas, Strzebonski and Vigklas in 2007. ACA 2008, Linz, Austria
48
1st Linear Complexity Bound Derived from Our Theorem
ACA 2008, Linz, Austria
49
2nd Linear Complexity Bound Derived from Our Theorem
ACA 2008, Linz, Austria
50
3rd Linear Complexity Bound Derived from Our Theorem
ACA 2008, Linz, Austria
51
4th Linear Complexity Bound Derived from Our Theorem
ACA 2008, Linz, Austria
52
Problems with a single method of computing bounds
There does NOT exist one which ALWAYS works better. Therefore, to compute the bound use two methods & pick the minimum ACA 2008, Linz, Austria
53
Example: the single positive root of the poly = 1.
x x2 – x - 1 The bound computed by the first three methods is UNACCEPTABLE (>= 1050)! However, with LM, we pick the max from the coefficient pairs {10100 x2 / 2, x} and {{10100 x2 / 22, -1} and obtain : 2 ACA 2008, Linz, Austria
54
Min(FL, LM), the Best Linear Complexity bound!
Experimental results showed that min( FL, LM) gives the best estimates. Using min( FL, LM) we obtain a 15% average overall speed-up over the VAS-CF implementation using Cauchy’s rule. VAS-CF using min( FL, LM) is ALWAYS faster than the VCA-bisection method! ACA 2008, Linz, Austria
55
To Further Improve the Performance of VAS-CF …
… we decided to use quadratic complexity bounds because their improved estimates SHOULD compensate for the extra time needed to compute them. ACA 2008, Linz, Austria
56
GI: The General Idea of the Quadratic Complexity Bounds:
EACH negative coefficient of the polynomial is paired with ALL the preceding positive coefficients and the minimum of the computed values is taken; the maximum of all those minimums is taken as the estimate of the bound. ACA 2008, Linz, Austria
57
Quadratic Complexity Bounds discussed:
We will present four Quadratic Complexity Bounds derived by applying the general idea (GI) to the four linear complexity bounds we discussed before. Three out of those four bounds, were developed by Akritas, Argyris, Strzebonski and Vigklas in 2008. ACA 2008, Linz, Austria
58
1st Quadratic Complexity Bound Derived from Our Theorem
ACA 2008, Linz, Austria
59
2nd Quadratic Complexity Bound Derived from Our Theorem (developed by Hong)
ACA 2008, Linz, Austria
60
3rd Quadratic Complexity Bound Derived from Our Theorem, (a)
ACA 2008, Linz, Austria
61
3rd Quadratic Complexity Bound Derived from Our Theorem, (b)
ACA 2008, Linz, Austria
62
FLQ is the Fastest Quadratic Complexity method.
EACH negative coefficient of the polynomial is paired only with one of the first λ preceding positive coefficients and the minimum is taken. Each of the first λ preceding positive coefficients is used once. ACA 2008, Linz, Austria
63
4th Quadratic Complexity Bound Derived from Our Theorem
ACA 2008, Linz, Austria
64
Revisited Example: the single positive root of the poly = 1.
x x2 – x - 1 The bound computed by three quadratic complexity bounds is 2! However, with FLQ we pick the Max of the Min{{x3, x}, {10100 x2, x}}=1, & Min{{x3, -1} , {10100 x2, -1}}=1 to obtain : 1 ACA 2008, Linz, Austria
65
Linear vs Quadratic Complexity Bounds
The quadratic complexity bounds CANNOT perform worse than the linear complexity ones; most of the times they perform (a lot) better! ACA 2008, Linz, Austria
66
LMQ, the best quadratic complexity bound!
LMQ always gives a better (<=) estimate than KQ. Experimental results indicate that FLQ, LMQ and min( FLQ, LMQ ) behave equally well! Therefore, pick LMQ. ACA 2008, Linz, Austria
67
VAS-CF using Linear and Quadratic Complexity Bounds: a
ACA 2008, Linz, Austria
68
VAS-CF using Linear and Quadratic Complexity Bounds: b
ACA 2008, Linz, Austria
69
Conclusions: Linear Complexity Bounds: VAS-CF using min ( FL, LM ), the best linear bound, was speeded up 15% over VAS-CF using Cauchy (AND became always faster than VCA). Quadratic Complexity Bounds: VAS-CF using LMQ, one of the three best quadratic bounds, was speeded up 40% over VAS-CF using Cauchy. ACA 2008, Linz, Austria
70
References I ACA 2008, Linz, Austria
71
References II ACA 2008, Linz, Austria
72
References III ACA 2008, Linz, Austria
73
References IV ACA 2008, Linz, Austria
Similar presentations
© 2024 SlidePlayer.com Inc.
All rights reserved.