validated numerics a short introduction to rigorous …validated numerics a short introduction to...
TRANSCRIPT
![Page 1: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/1.jpg)
Validated Numericsa short introduction to rigorous computations
Warwick Tucker
The CAPA group
Department of Mathematics
Uppsala University, Sweden
![Page 2: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/2.jpg)
Are floating point computations reliable?
Computing with the C/C++ single format
![Page 3: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/3.jpg)
Are floating point computations reliable?
Computing with the C/C++ single format
Example 1: Repeated addition
103∑
i=1
〈10−3〉 = 0.999990701675415,
104∑
i=1
〈10−4〉 = 1.000053524971008.
![Page 4: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/4.jpg)
Are floating point computations reliable?
Computing with the C/C++ single format
Example 1: Repeated addition
103∑
i=1
〈10−3〉 = 0.999990701675415,
104∑
i=1
〈10−4〉 = 1.000053524971008.
Example 2: Order of summation
1 +1
2+
1
3+ · · ·+ 1
106= 14.357357,
1
106+ · · ·+ 1
3+
1
2+ 1 = 14.392651.
![Page 5: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/5.jpg)
Are floating point computations reliable?
Given the point (x, y) = (77617, 33096), evaluate the function
f(x, y) = 333.75y6 + x2(11x2y2 − y6 − 121y4 − 2) + 5.5y8 + x/(2y)
![Page 6: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/6.jpg)
Are floating point computations reliable?
Given the point (x, y) = (77617, 33096), evaluate the function
f(x, y) = 333.75y6 + x2(11x2y2 − y6 − 121y4 − 2) + 5.5y8 + x/(2y)
IBM S/370 (β = 16) with FORTRAN:
type p f(x, y)
REAL*4 24 1.172603 . . .
REAL*8 53 1.1726039400531 . . .
REAL*10 64 1.172603940053178 . . .
![Page 7: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/7.jpg)
Are floating point computations reliable?
Given the point (x, y) = (77617, 33096), evaluate the function
f(x, y) = 333.75y6 + x2(11x2y2 − y6 − 121y4 − 2) + 5.5y8 + x/(2y)
IBM S/370 (β = 16) with FORTRAN:
type p f(x, y)
REAL*4 24 1.172603 . . .
REAL*8 53 1.1726039400531 . . .
REAL*10 64 1.172603940053178 . . .
Pentium III (β = 2) with C/C++ (gcc/g++):
type p f(x, y)
float 24 178702833214061281280
double 53 178702833214061281280
long double 64 178702833214061281280
![Page 8: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/8.jpg)
Are floating point computations reliable?
Given the point (x, y) = (77617, 33096), evaluate the function
f(x, y) = 333.75y6 + x2(11x2y2 − y6 − 121y4 − 2) + 5.5y8 + x/(2y)
IBM S/370 (β = 16) with FORTRAN:
type p f(x, y)
REAL*4 24 1.172603 . . .
REAL*8 53 1.1726039400531 . . .
REAL*10 64 1.172603940053178 . . .
Pentium III (β = 2) with C/C++ (gcc/g++):
type p f(x, y)
float 24 178702833214061281280
double 53 178702833214061281280
long double 64 178702833214061281280
Correct answer: −0.8273960599 . . .
![Page 9: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/9.jpg)
How do we control rounding errors?
Round each partial result both ways
If x, y ∈ F and ⋆ ∈ {+,−,×,÷}, we can enclose the exact resultin an interval:
x ⋆ y ∈ [▽(x ⋆ y),△(x ⋆ y)].
![Page 10: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/10.jpg)
How do we control rounding errors?
Round each partial result both ways
If x, y ∈ F and ⋆ ∈ {+,−,×,÷}, we can enclose the exact resultin an interval:
x ⋆ y ∈ [▽(x ⋆ y),△(x ⋆ y)].
Since all (modern) computers round with maximal quality, theinterval is the smallest one that contains the exact result.
![Page 11: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/11.jpg)
How do we control rounding errors?
Round each partial result both ways
If x, y ∈ F and ⋆ ∈ {+,−,×,÷}, we can enclose the exact resultin an interval:
x ⋆ y ∈ [▽(x ⋆ y),△(x ⋆ y)].
Since all (modern) computers round with maximal quality, theinterval is the smallest one that contains the exact result.
Question
How do we compute with intervals? And why, really?
![Page 12: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/12.jpg)
Arithmetic over IR
Definition
If ⋆ is one of the operators +,−,×,÷, and if a, b ∈ IR, then
a ⋆ b = {a ⋆ b : a ∈ a, b ∈ b},
except that a÷ b is undefined if 0 ∈ b.
![Page 13: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/13.jpg)
Arithmetic over IR
Definition
If ⋆ is one of the operators +,−,×,÷, and if a, b ∈ IR, then
a ⋆ b = {a ⋆ b : a ∈ a, b ∈ b},
except that a÷ b is undefined if 0 ∈ b.
Simple arithmetic
a+ b = [a+ b,a+ b]
a− b = [a− b,a− b]
a× b = [min{ab,ab,ab,ab},max{ab,ab,ab,ab}]a÷ b = a× [1/b, 1/b], if 0 /∈ b.
![Page 14: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/14.jpg)
Arithmetic over IR
Definition
If ⋆ is one of the operators +,−,×,÷, and if a, b ∈ IR, then
a ⋆ b = {a ⋆ b : a ∈ a, b ∈ b},
except that a÷ b is undefined if 0 ∈ b.
Simple arithmetic
a+ b = [a+ b,a+ b]
a− b = [a− b,a− b]
a× b = [min{ab,ab,ab,ab},max{ab,ab,ab,ab}]a÷ b = a× [1/b, 1/b], if 0 /∈ b.
On a computer we use directed rounding, e.g.a+ b = [▽(a⊕ b),△(a ⊕ b)].
![Page 15: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/15.jpg)
Interval-valued functions
Range enclosure
Extend a real-valued function f to an interval-valued F :
R(f ;x) = {f(x) : x ∈ x} ⊆ F (x)
![Page 16: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/16.jpg)
Interval-valued functions
Range enclosure
Extend a real-valued function f to an interval-valued F :
R(f ;x) = {f(x) : x ∈ x} ⊆ F (x)
x
f(x)
x
F (x)
![Page 17: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/17.jpg)
Interval-valued functions
Range enclosure
Extend a real-valued function f to an interval-valued F :
R(f ;x) = {f(x) : x ∈ x} ⊆ F (x)
x
f(x)
x
F (x)
y /∈ F (x) implies that f(x) 6= y for all x ∈ x.
![Page 18: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/18.jpg)
Interval-valued functions
Some explicit formulas are given below:
![Page 19: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/19.jpg)
Interval-valued functions
Some explicit formulas are given below:
ex = [ex, ex]√x = [
√x,
√x] if 0 ≤ x
logx = [logx, logx] if 0 < x
arctanx = [arctanx, arctanx] .
![Page 20: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/20.jpg)
Interval-valued functions
Some explicit formulas are given below:
ex = [ex, ex]√x = [
√x,
√x] if 0 ≤ x
logx = [logx, logx] if 0 < x
arctanx = [arctanx, arctanx] .
Set S+ = {2kπ + π/2: k ∈ Z} and S− = {2kπ − π/2: k ∈ Z}.Then sinx is given by
[−1, 1] : if x ∩ S− 6= ∅ and x ∩ S+ 6= ∅,[−1,max{sinx, sinx}] : if x ∩ S− 6= ∅ and x ∩ S+ = ∅,[min{sinx, sinx}, 1] : if x ∩ S− = ∅ and x ∩ S+ 6= ∅,[min{sinx, sinx},max{sinx, sinx}] : if x ∩ S− = ∅ and x ∩ S+ = ∅.
![Page 21: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/21.jpg)
Graph Enclosures
A controlled discretization
We can now select and adapt the level of discretization
![Page 22: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/22.jpg)
Graph Enclosures
A controlled discretization
We can now select and adapt the level of discretization
−5 −4 −3 −2 −1 0 1 2 3 4 5−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
−5
0
5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
0
1
2
3
4
x
f(x)
f(x) = cos(x)3 + sin(x)
Various levels of discretization of f(x) = cos3 x+ sinx.
![Page 23: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/23.jpg)
Solving non-linear equations
![Page 24: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/24.jpg)
Solving non-linear equations
Consider everything. Keep what is good.
Avoid evil whenever you recognize it.
St. Paul, ca. 50 A.D. (The Bible, 1 Thess. 5:21-22)
![Page 25: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/25.jpg)
Solving non-linear equations
Consider everything. Keep what is good.
Avoid evil whenever you recognize it.
St. Paul, ca. 50 A.D. (The Bible, 1 Thess. 5:21-22)
No solutions can be missed!
![Page 26: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/26.jpg)
Solving non-linear equation
The code is transparent and natural
01 function bisect(fcnName, X, tol)
02 f = inline(fcnName);
03 if ( 0 <= f(X) ) % If f(X) contains zero...
04 if Diam(X) < tol % and the tolerance is met...
05 X % print the interval X.
06 else % Otherwise, divide and conquer.
07 bisect(fcnName, interval(Inf(X), Mid(X)), tol);
08 bisect(fcnName, interval(Mid(X), Sup(X)), tol);
09 end
10 end
![Page 27: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/27.jpg)
Solving non-linear equation
The code is transparent and natural
01 function bisect(fcnName, X, tol)
02 f = inline(fcnName);
03 if ( 0 <= f(X) ) % If f(X) contains zero...
04 if Diam(X) < tol % and the tolerance is met...
05 X % print the interval X.
06 else % Otherwise, divide and conquer.
07 bisect(fcnName, interval(Inf(X), Mid(X)), tol);
08 bisect(fcnName, interval(Mid(X), Sup(X)), tol);
09 end
10 end
Nice property
If F is well-defined on the domain, the algorithm produces anenclosure of all zeros of f . [No existence is established, however.]
![Page 28: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/28.jpg)
Existence and uniqueness
Existence and uniqueness require fixed point theorems.
![Page 29: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/29.jpg)
Existence and uniqueness
Existence and uniqueness require fixed point theorems.
Brouwer’s fixed point theorem
Let B be homeomorhpic to the closed unit ball in Rn. Then given
any continuous mapping f : B → B there exists x ∈ B such thatf(x) = x.
![Page 30: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/30.jpg)
Existence and uniqueness
Existence and uniqueness require fixed point theorems.
Brouwer’s fixed point theorem
Let B be homeomorhpic to the closed unit ball in Rn. Then given
any continuous mapping f : B → B there exists x ∈ B such thatf(x) = x.
Schauder’s fixed point theorem
Let X be a normed vector space, and let K ⊂ X be a non-empty,compact, and convex set. Then given any continuous mappingf : K → K there exists x ∈ K such that f(x) = x.
![Page 31: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/31.jpg)
Existence and uniqueness
Existence and uniqueness require fixed point theorems.
Brouwer’s fixed point theorem
Let B be homeomorhpic to the closed unit ball in Rn. Then given
any continuous mapping f : B → B there exists x ∈ B such thatf(x) = x.
Schauder’s fixed point theorem
Let X be a normed vector space, and let K ⊂ X be a non-empty,compact, and convex set. Then given any continuous mappingf : K → K there exists x ∈ K such that f(x) = x.
Banach’s fixed point theorem
If f is a contraction defined on a complete metric space X, thenthere exists a unique x ∈ X such that f(x) = x.
![Page 32: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/32.jpg)
Newton’s method in IR
Theorem
Let f ∈ C1(R,R), and set x̌ = mid(x). We define
Nf (x)def= Nf (x, x̌) = x̌− [DF (x)]−1f(x̌).
If Nf (x) is well-defined, then the following statements hold:
![Page 33: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/33.jpg)
Newton’s method in IR
Theorem
Let f ∈ C1(R,R), and set x̌ = mid(x). We define
Nf (x)def= Nf (x, x̌) = x̌− [DF (x)]−1f(x̌).
If Nf (x) is well-defined, then the following statements hold:
(1) if x contains a zero x∗ of f , then so does Nf (x) ∩ x;
![Page 34: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/34.jpg)
Newton’s method in IR
Theorem
Let f ∈ C1(R,R), and set x̌ = mid(x). We define
Nf (x)def= Nf (x, x̌) = x̌− [DF (x)]−1f(x̌).
If Nf (x) is well-defined, then the following statements hold:
(1) if x contains a zero x∗ of f , then so does Nf (x) ∩ x;
(2) if Nf (x) ∩ x = ∅, then x contains no zeros of f ;
![Page 35: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/35.jpg)
Newton’s method in IR
Theorem
Let f ∈ C1(R,R), and set x̌ = mid(x). We define
Nf (x)def= Nf (x, x̌) = x̌− [DF (x)]−1f(x̌).
If Nf (x) is well-defined, then the following statements hold:
(1) if x contains a zero x∗ of f , then so does Nf (x) ∩ x;
(2) if Nf (x) ∩ x = ∅, then x contains no zeros of f ;
(3) if Nf (x) ⊆ x, then x contains a unique zero of f .
![Page 36: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/36.jpg)
Newton’s method in IR
Theorem
Let f ∈ C1(R,R), and set x̌ = mid(x). We define
Nf (x)def= Nf (x, x̌) = x̌− [DF (x)]−1f(x̌).
If Nf (x) is well-defined, then the following statements hold:
(1) if x contains a zero x∗ of f , then so does Nf (x) ∩ x;
(2) if Nf (x) ∩ x = ∅, then x contains no zeros of f ;
(3) if Nf (x) ⊆ x, then x contains a unique zero of f .
Similar statements hold for the Krawczyk operator
Kf (x)def= x̌− [Df(x̌)]−1f(x̌)−
(1− [Df(x̌)]−1F ′(x)
)[−r, r],
where we use the notation r = rad(x).
![Page 37: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/37.jpg)
Newton’s method in IR
Algorithm
Starting from an initial search region x0, we form the sequence
xi+1 = Nf (xi) ∩ xi i = 0, 1, . . . .
![Page 38: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/38.jpg)
Newton’s method in IR
Algorithm
Starting from an initial search region x0, we form the sequence
xi+1 = Nf (xi) ∩ xi i = 0, 1, . . . .
xixi+1
Nf (xi)
x̌i
![Page 39: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/39.jpg)
Newton’s method in IR
Algorithm
Starting from an initial search region x0, we form the sequence
xi+1 = Nf (xi) ∩ xi i = 0, 1, . . . .
xixi+1
Nf (xi)
x̌i
Performance
If well-defined, this method is never worse than bisection, and itconverges quadratically fast under mild conditions.
![Page 40: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/40.jpg)
Newton’s method in IR
Example
Take f(x) = −2.001 + 3x− x3 and start with x0 = [−3,−3/2].
![Page 41: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/41.jpg)
Newton’s method in IR
Example
Take f(x) = −2.001 + 3x− x3 and start with x0 = [−3,−3/2].
X(0) = [-3.000000000000000,-1.500000000000000]; rad = 7.50000e-01
X(1) = [-2.140015625000001,-1.546099999999996]; rad = 2.96958e-01
X(2) = [-2.140015625000001,-1.961277398284108]; rad = 8.93691e-02
X(3) = [-2.006849239640351,-1.995570580247208]; rad = 5.63933e-03
X(4) = [-2.000120104486270,-2.000103608530276]; rad = 8.24798e-06
X(5) = [-2.000111102890393,-2.000111102873815]; rad = 8.28893e-12
X(6) = [-2.000111102881727,-2.000111102881724]; rad = 1.55431e-15
X(7) = [-2.000111102881727,-2.000111102881724]; rad = 1.55431e-15
Finite convergence!
Unique root in -2.00011110288172 +- 1.555e-15
![Page 42: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/42.jpg)
Newton’s method in IR
Example
Take f(x) = −2.001 + 3x− x3 and start with x0 = [−3,−3/2].
X(0) = [-3.000000000000000,-1.500000000000000]; rad = 7.50000e-01
X(1) = [-2.140015625000001,-1.546099999999996]; rad = 2.96958e-01
X(2) = [-2.140015625000001,-1.961277398284108]; rad = 8.93691e-02
X(3) = [-2.006849239640351,-1.995570580247208]; rad = 5.63933e-03
X(4) = [-2.000120104486270,-2.000103608530276]; rad = 8.24798e-06
X(5) = [-2.000111102890393,-2.000111102873815]; rad = 8.28893e-12
X(6) = [-2.000111102881727,-2.000111102881724]; rad = 1.55431e-15
X(7) = [-2.000111102881727,-2.000111102881724]; rad = 1.55431e-15
Finite convergence!
Unique root in -2.00011110288172 +- 1.555e-15
Stopping condition
Stop when no further improvement takes place.
![Page 43: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/43.jpg)
The Krawczyk method with bisection
When we have several zeros, we must bisect to isolate the zeros.
Example
Take f(x) = sin (ex + 1) and start with x0 = [0, 3].
![Page 44: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/44.jpg)
The Krawczyk method with bisection
When we have several zeros, we must bisect to isolate the zeros.
Example
Take f(x) = sin (ex + 1) and start with x0 = [0, 3].
Domain : [0, 3]
Tolerance : 1e-10
Function calls : 71
Unique zero in the interval 0.761549782880[8890,9006]
Unique zero in the interval 1.664529193[6825445,7060436]
Unique zero in the interval 2.131177121086[2673,3558]
Unique zero in the interval 2.4481018026567[773,801]
Unique zero in the interval 2.68838906601606[36,68]
Unique zero in the interval 2.8819786295709[728,1555]
![Page 45: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/45.jpg)
The Krawczyk method with bisection
When we have several zeros, we must bisect to isolate the zeros.
Example
Take f(x) = sin (ex + 1) and start with x0 = [0, 3].
Domain : [0, 3]
Tolerance : 1e-10
Function calls : 71
Unique zero in the interval 0.761549782880[8890,9006]
Unique zero in the interval 1.664529193[6825445,7060436]
Unique zero in the interval 2.131177121086[2673,3558]
Unique zero in the interval 2.4481018026567[773,801]
Unique zero in the interval 2.68838906601606[36,68]
Unique zero in the interval 2.8819786295709[728,1555]
Applications
Counting short periodic orbits for ODEs [Z. Galias]Measuring the stable regions of the quadratic map [D. Wilczak]
![Page 46: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/46.jpg)
Implicit curves (level sets)
Example
Draw the level-set defined by
f(x, y) = sin (cos x2 + 10 sin y2)− y cos x = 0;
restricted to the domain [−5, 5] × [−5, 5].
![Page 47: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/47.jpg)
Implicit curves (level sets)
Example
Draw the level-set defined by
f(x, y) = sin (cos x2 + 10 sin y2)− y cos x = 0;
restricted to the domain [−5, 5] × [−5, 5].
MATLAB produces the following picture:
−5 0 5−5
0
5
![Page 48: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/48.jpg)
Implicit curves (level sets)
Example
Draw the level-set defined by
f(x, y) = sin (cos x2 + 10 sin y2)− y cos x = 0;
restricted to the domain [−5, 5] × [−5, 5].
MATLAB produces the following picture:
−5 0 5−5
0
5
According to the same m-file, thelevel set defined by |f(x, y)| = 0,however, appears to be empty.
But this is the same set!!!
![Page 49: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/49.jpg)
Implicit curves
The (increasingly tight) set-valued enclosures in both cases are
![Page 50: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/50.jpg)
Implicit curves
The (increasingly tight) set-valued enclosures in both cases are
![Page 51: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/51.jpg)
Quadrature
Example (A bonus problem)
Compute the integral∫ 80 sin
(x+ ex
)dx.
![Page 52: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/52.jpg)
Quadrature
Example (A bonus problem)
Compute the integral∫ 80 sin
(x+ ex
)dx.
A regular MATLAB session:
>> q = quad(’sin(x + exp(x))’, 0, 8)
q =
0.251102722027180
![Page 53: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/53.jpg)
Quadrature
Example (A bonus problem)
Compute the integral∫ 80 sin
(x+ ex
)dx.
A regular MATLAB session:
>> q = quad(’sin(x + exp(x))’, 0, 8)
q =
0.251102722027180
Using an adaptive validated integrator:
$$ ./adQuad 0 8 4 1e-4
Partitions: 8542
CPU time : 0.52 seconds
Integral : 0.347[3863144222905,4140198005782]
![Page 54: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/54.jpg)
Quadrature
Example (A bonus problem)
Compute the integral∫ 80 sin
(x+ ex
)dx.
A regular MATLAB session:
>> q = quad(’sin(x + exp(x))’, 0, 8)
q =
0.251102722027180
Using an adaptive validated integrator:
$$ ./adQuad 0 8 4 1e-4
Partitions: 8542
CPU time : 0.52 seconds
Integral : 0.347[3863144222905,4140198005782]
$$ ./adQuad 0 8 20 1e-10
Partitions: 874
CPU time : 0.45 seconds
Integral : 0.3474001726[492276,652638]
![Page 55: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/55.jpg)
Parameter estimation
Problem formulation
Given a finitely parametrized model function together with some(noisy) data, and a search region P in parameter space:
y = f(x; p)︸ ︷︷ ︸
model
{(xi, yi)}Ni=1︸ ︷︷ ︸
data
p ∈ P︸ ︷︷ ︸
space
try to find parameters that give a good agreement between thedata and the model. [A classic inverse problem]
![Page 56: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/56.jpg)
Parameter estimation
Problem formulation
Given a finitely parametrized model function together with some(noisy) data, and a search region P in parameter space:
y = f(x; p)︸ ︷︷ ︸
model
{(xi, yi)}Ni=1︸ ︷︷ ︸
data
p ∈ P︸ ︷︷ ︸
space
try to find parameters that give a good agreement between thedata and the model. [A classic inverse problem]
Existence: with noisy data, or with an incorrect model, thereis usually no parameter that produces a perfect fit.
![Page 57: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/57.jpg)
Parameter estimation
Problem formulation
Given a finitely parametrized model function together with some(noisy) data, and a search region P in parameter space:
y = f(x; p)︸ ︷︷ ︸
model
{(xi, yi)}Ni=1︸ ︷︷ ︸
data
p ∈ P︸ ︷︷ ︸
space
try to find parameters that give a good agreement between thedata and the model. [A classic inverse problem]
Existence: with noisy data, or with an incorrect model, thereis usually no parameter that produces a perfect fit.
Uniqueness: even with unlimited amounts of exact data,there might not exist a unique solution p♯ ∈ P such that
f(xi; p♯) = yi i = 1, . . . , N.
![Page 58: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/58.jpg)
Parameter estimation
Problem formulation
Given a finitely parametrized model function together with some(noisy) data, and a search region P in parameter space:
y = f(x; p)︸ ︷︷ ︸
model
{(xi, yi)}Ni=1︸ ︷︷ ︸
data
p ∈ P︸ ︷︷ ︸
space
try to find parameters that give a good agreement between thedata and the model. [A classic inverse problem]
Existence: with noisy data, or with an incorrect model, thereis usually no parameter that produces a perfect fit.
Uniqueness: even with unlimited amounts of exact data,there might not exist a unique solution p♯ ∈ P such that
f(xi; p♯) = yi i = 1, . . . , N.
Instability: inverse problems can be extremely ill-conditioned.
![Page 59: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/59.jpg)
Introduction
A statistical approach
Use a (weighted) least-squares approach to find the bestparameter:
argminp∈P
N∑
i=1
wi|f(xi; p)− yi|2.
![Page 60: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/60.jpg)
Introduction
A statistical approach
Use a (weighted) least-squares approach to find the bestparameter:
argminp∈P
N∑
i=1
wi|f(xi; p)− yi|2.
If the parameters enter f linearly, this is “straight-forward”.
![Page 61: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/61.jpg)
Introduction
A statistical approach
Use a (weighted) least-squares approach to find the bestparameter:
argminp∈P
N∑
i=1
wi|f(xi; p)− yi|2.
If the parameters enter f linearly, this is “straight-forward”.
Otherwise, we have moved the problem to global optimization.
![Page 62: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/62.jpg)
Introduction
A statistical approach
Use a (weighted) least-squares approach to find the bestparameter:
argminp∈P
N∑
i=1
wi|f(xi; p)− yi|2.
If the parameters enter f linearly, this is “straight-forward”.
Otherwise, we have moved the problem to global optimization.
The selection of weights is almost always a delicate issue.
![Page 63: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/63.jpg)
Introduction
A statistical approach
Use a (weighted) least-squares approach to find the bestparameter:
argminp∈P
N∑
i=1
wi|f(xi; p)− yi|2.
If the parameters enter f linearly, this is “straight-forward”.
Otherwise, we have moved the problem to global optimization.
The selection of weights is almost always a delicate issue.
A set-valued approach
Locate nearby models that are consistent with nearby data:
f(x; p) −→ f(x;p) (xi, yi) −→ (xi,yi).
![Page 64: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/64.jpg)
Set-valued computations
Points versus sets in parameter space
We move from the point-valued model function f(x; p) to theset-valued version f(x;p).
![Page 65: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/65.jpg)
Set-valued computations
Points versus sets in parameter space
We move from the point-valued model function f(x; p) to theset-valued version f(x;p).
0 5 10 15 20 25 300
0.5
1
1.5
2
2.5
3
0 5 10 15 20 25 300
0.5
1
1.5
2
2.5
3
Figure: (a) p = 0.15, a point in P . (b) p = [0.14, 0.16], a subset of P .The model function is f(x; p) = xe−px, and 10 samples are shown.
![Page 66: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/66.jpg)
Parameter estimation
Strategy
Adaptively bisect the parameter space into sub-boxes: P = ∪Kj=1pj
and examine each pj separately. [Good for parallelisation]
![Page 67: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/67.jpg)
Parameter estimation
Strategy
Adaptively bisect the parameter space into sub-boxes: P = ∪Kj=1pj
and examine each pj separately. [Good for parallelisation]
Each sub-box p of the parameter space falls into one of threecategories:
![Page 68: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/68.jpg)
Parameter estimation
Strategy
Adaptively bisect the parameter space into sub-boxes: P = ∪Kj=1pj
and examine each pj separately. [Good for parallelisation]
Each sub-box p of the parameter space falls into one of threecategories:
(1) consistent
if f(xi;p) ⊂ yi for all i = 0, . . . , N . SAVE
![Page 69: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/69.jpg)
Parameter estimation
Strategy
Adaptively bisect the parameter space into sub-boxes: P = ∪Kj=1pj
and examine each pj separately. [Good for parallelisation]
Each sub-box p of the parameter space falls into one of threecategories:
(1) consistent
if f(xi;p) ⊂ yi for all i = 0, . . . , N . SAVE
(2) inconsistent
if f(xi;p) ∩ yi = ∅ for at least one i. DROP
![Page 70: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/70.jpg)
Parameter estimation
Strategy
Adaptively bisect the parameter space into sub-boxes: P = ∪Kj=1pj
and examine each pj separately. [Good for parallelisation]
Each sub-box p of the parameter space falls into one of threecategories:
(1) consistent
if f(xi;p) ⊂ yi for all i = 0, . . . , N . SAVE
(2) inconsistent
if f(xi;p) ∩ yi = ∅ for at least one i. DROP
(3) undetermined
not (1), but f(xi;p) ∩ yi 6= ∅ for all i = 0, . . . , N . SPLIT
![Page 71: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/71.jpg)
Parameter estimation
Example
Consider the model function
f(x; p1, p2) = 5e−p1x − 4× 10−6e−p2x
with samples taken at x = 0, 5 . . . , 40 using p♯ = (0.11,−0.32).Accepting a relative noise level of 90%, we get the following set ofconsistent parameters:
![Page 72: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/72.jpg)
Parameter estimation
Example
Consider the model function
f(x; p1, p2) = 5e−p1x − 4× 10−6e−p2x
with samples taken at x = 0, 5 . . . , 40 using p♯ = (0.11,−0.32).Accepting a relative noise level of 90%, we get the following set ofconsistent parameters:
![Page 73: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/73.jpg)
Parameter estimation
Varying the relative noise levels between 10, 20 . . . , 90%, we getthe following indeterminate sets.
![Page 74: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/74.jpg)
Constraint propagation
Constraining the parameter/data space
We use set-valued constraint propagation to quickly discardinconsistent regions in the data and the parameter space.
This is done without bisection!
![Page 75: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/75.jpg)
Constraint propagation
Constraining the parameter/data space
We use set-valued constraint propagation to quickly discardinconsistent regions in the data and the parameter space.
This is done without bisection!
Example
Let f(x; p) = xe−px, and consider the situation p = [0, 1] and(x,y) = (2, [1, 3]).
![Page 76: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/76.jpg)
Constraint propagation
Constraining the parameter/data space
We use set-valued constraint propagation to quickly discardinconsistent regions in the data and the parameter space.
This is done without bisection!
Example
Let f(x; p) = xe−px, and consider the situation p = [0, 1] and(x,y) = (2, [1, 3]). By a forward (interval) evaluation, we have
f(2; [0, 1]) = 2e−2[0,1] = 2e[−2,0] = 2[e−2, 1] = [2e−2, 2].
![Page 77: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/77.jpg)
Constraint propagation
Constraining the parameter/data space
We use set-valued constraint propagation to quickly discardinconsistent regions in the data and the parameter space.
This is done without bisection!
Example
Let f(x; p) = xe−px, and consider the situation p = [0, 1] and(x,y) = (2, [1, 3]). By a forward (interval) evaluation, we have
f(2; [0, 1]) = 2e−2[0,1] = 2e[−2,0] = 2[e−2, 1] = [2e−2, 2].
This allows us to contract the data range according to
y 7→ y ∩ f(x;p) = [1, 3] ∩ [2e−2, 2] = [1, 2].
![Page 78: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/78.jpg)
Constraint propagation
Directed Acyclic Graphs (DAGs)
We use a DAG representation of the model function to automateconstraint propagations.
![Page 79: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/79.jpg)
Constraint propagation
Directed Acyclic Graphs (DAGs)
We use a DAG representation of the model function to automateconstraint propagations.
x
p
×
× neg exp
n1
n2
n3 n4 n5
n6 = y
![Page 80: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/80.jpg)
Constraint propagation
Directed Acyclic Graphs (DAGs)
We use a DAG representation of the model function to automateconstraint propagations.
x
p
×
× neg exp
n1
n2
n3 n4 n5
n6 = y
n1 = x
n2 = p
n3 = n1 × n2
n4 = −n3
n5 = en4
n6 = n1 × n5.
Figure: The DAG representation of a forward sweep of y = xe−px,together with the corresponding code list.
![Page 81: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/81.jpg)
Constraint propagation
Directed Acyclic Graphs (DAGs)
We can propagate constraints from data to the parameter bymoving backwards in the code list.
![Page 82: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/82.jpg)
Constraint propagation
Directed Acyclic Graphs (DAGs)
We can propagate constraints from data to the parameter bymoving backwards in the code list.
x
y
÷
÷neg log
n1p = n2
n3 n4 n5
n6
![Page 83: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/83.jpg)
Constraint propagation
Directed Acyclic Graphs (DAGs)
We can propagate constraints from data to the parameter bymoving backwards in the code list.
x
y
÷
÷neg log
n1p = n2
n3 n4 n5
n6
n5 = n6 ÷ n1
n4 = log n5
n3 = −n4
n2 = n3 ÷ n1.
Figure: The DAG representation of a backward sweep ofy = xe−px, together with the corresponding code list.
![Page 84: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/84.jpg)
Constraint propagation
Example
Again, we work on the model function y = f(x; p) = xe−px, butnow with the data (x,y) = (2, [1, 3]), together with the parameterdomain p = [0, 1].
![Page 85: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/85.jpg)
Constraint propagation
Example
Again, we work on the model function y = f(x; p) = xe−px, butnow with the data (x,y) = (2, [1, 3]), together with the parameterdomain p = [0, 1]. The forward sweep, performed in Example 7,contracts the interval data to y = [1, 2].
![Page 86: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/86.jpg)
Constraint propagation
Example
Again, we work on the model function y = f(x; p) = xe−px, butnow with the data (x,y) = (2, [1, 3]), together with the parameterdomain p = [0, 1]. The forward sweep, performed in Example 7,contracts the interval data to y = [1, 2]. Performing a backwardsweep contracts the interval parameter to p = [0, 12 log 2]:
n5 = n6 ÷ n1 = [1, 2] ÷ 2 = [12 , 1]n4 = log n5 = log [12 , 1] = [− log 2, 0]n3 = −n4 = [0, log 2]n2 = n3 ÷ n1 = [0, log 2]÷ 2 ≈ [0, 0.34657359].
![Page 87: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/87.jpg)
Constraint propagation
Example
Again, we work on the model function y = f(x; p) = xe−px, butnow with the data (x,y) = (2, [1, 3]), together with the parameterdomain p = [0, 1]. The forward sweep, performed in Example 7,contracts the interval data to y = [1, 2]. Performing a backwardsweep contracts the interval parameter to p = [0, 12 log 2]:
n5 = n6 ÷ n1 = [1, 2] ÷ 2 = [12 , 1]n4 = log n5 = log [12 , 1] = [− log 2, 0]n3 = −n4 = [0, log 2]n2 = n3 ÷ n1 = [0, log 2]÷ 2 ≈ [0, 0.34657359].
Note that, in one forward/backward sweep, we managed to excludeover 65% of the parameter domain, at the same time reducing thedata uncertainty by 50%.
![Page 88: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/88.jpg)
Mixed-effects models
Mixed-effects models
We are given several data sets (trajectories) corresponding to kdifferent “individuals”:
individual1 : (x11, y11), (x12, y12), . . . , (x1N , y1N1)
individual2 : (x21, y21), (x22, y22), . . . , (x2N , y2N2)
......
...
individualk : (xk1, yk1), (xk2, yk2), . . . , (xkN , ykNk).
Some model parameters are equal (shared) for all individuals, andsome are distinct.
![Page 89: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/89.jpg)
Mixed-effects models
Mixed-effects models
We are given several data sets (trajectories) corresponding to kdifferent “individuals”:
individual1 : (x11, y11), (x12, y12), . . . , (x1N , y1N1)
individual2 : (x21, y21), (x22, y22), . . . , (x2N , y2N2)
......
...
individualk : (xk1, yk1), (xk2, yk2), . . . , (xkN , ykNk).
Some model parameters are equal (shared) for all individuals, andsome are distinct.
We need to consider all individuals simultaneously. Otherwisethe number of unknown parameters may be too large.
![Page 90: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/90.jpg)
A mixed-effects model for orange tree truncs
Example
We will apply our method to the following scenario:
![Page 91: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/91.jpg)
A mixed-effects model for orange tree truncs
Example
We will apply our method to the following scenario:
Model function: f(x; p) =p1
1 + p2ep3x
![Page 92: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/92.jpg)
A mixed-effects model for orange tree truncs
Example
We will apply our method to the following scenario:
Model function: f(x; p) =p1
1 + p2ep3x
Individual parameter: pi1 = p♯1 + ηi, ηi ∼ N(0, σ2)
![Page 93: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/93.jpg)
A mixed-effects model for orange tree truncs
Example
We will apply our method to the following scenario:
Model function: f(x; p) =p1
1 + p2ep3x
Individual parameter: pi1 = p♯1 + ηi, ηi ∼ N(0, σ2)
Data perturbation: yij = y♯ij(1 + θij), θij ∼ U(−ǫ,+ǫ)
![Page 94: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/94.jpg)
A mixed-effects model for orange tree truncs
Example
We will apply our method to the following scenario:
Model function: f(x; p) =p1
1 + p2ep3x
Individual parameter: pi1 = p♯1 + ηi, ηi ∼ N(0, σ2)
Data perturbation: yij = y♯ij(1 + θij), θij ∼ U(−ǫ,+ǫ)
For this specific example, we will use Np ∈ {1, 2, 5, 50} subjects,sampled at Nd = 10 data sites, evenly spaced within [100, 1600].
![Page 95: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/95.jpg)
A mixed-effects model for orange tree truncs
Example
We will apply our method to the following scenario:
Model function: f(x; p) =p1
1 + p2ep3x
Individual parameter: pi1 = p♯1 + ηi, ηi ∼ N(0, σ2)
Data perturbation: yij = y♯ij(1 + θij), θij ∼ U(−ǫ,+ǫ)
For this specific example, we will use Np ∈ {1, 2, 5, 50} subjects,sampled at Nd = 10 data sites, evenly spaced within [100, 1600].
Target parameters:p♯ = (191.84, 8.153,−0.0029), σ = 20, ǫ ∈ {0.01, 0.1, 0.2, 0.5}.
![Page 96: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/96.jpg)
A mixed-effects model for orange tree truncs
Example
We will apply our method to the following scenario:
Model function: f(x; p) =p1
1 + p2ep3x
Individual parameter: pi1 = p♯1 + ηi, ηi ∼ N(0, σ2)
Data perturbation: yij = y♯ij(1 + θij), θij ∼ U(−ǫ,+ǫ)
For this specific example, we will use Np ∈ {1, 2, 5, 50} subjects,sampled at Nd = 10 data sites, evenly spaced within [100, 1600].
Target parameters:p♯ = (191.84, 8.153,−0.0029), σ = 20, ǫ ∈ {0.01, 0.1, 0.2, 0.5}.
Search region:P = ([0, 300], [0, 9], [−1, 0]).
![Page 97: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/97.jpg)
A mixed-effects model for orange tree truncs
0 200 400 600 800 1000 1200 1400 16000
50
100
150
200
250
Figure: Data inflation and contraction for the example. The graph of themodel function for one subject (blue line). The data points are markedwith red dots. The inflated data sets are shown as striped bars, and there-contracted data as green bars.
![Page 98: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/98.jpg)
A mixed-effects model for orange tree truncs
Numerical results
Np = 1 Np = 2
ǫ = 0.01 190.639 (– –) (0.010) 193.141 (19.6) (0.013)ǫ = 0.1 194.139 (– –) (0.092) 195.233 (21.1) (0.097)ǫ = 0.2 189.139 (– –) (0.190) 193.437 (20.3) (0.192)ǫ = 0.5 167.226 (– –) (0.604) 167.770 (26.6) (0.589)
Np = 5 Np = 50
ǫ = 0.01 191.675 (20.1) (0.014) 191.239 (20.1) (0.012)ǫ = 0.1 192.954 (21.4) (0.099) 198.428 (22.2) (0.110)ǫ = 0.2 191.773 (20.3) (0.203) 197.580 (23.6) (0.214)ǫ = 0.5 164.656 (23.9) (0.620) 174.318 (27.1) (0.618)
Table: The results of four experiments for the example, each using100 trial runs with p1 = 191.184, and σ = 20.0. For each pair(ǫ,Np), we display the triple µ(p1), µ(σ), and µ(ǫ) – the averageestimates of the distribution parameters for p1, and the data error.
![Page 99: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/99.jpg)
A mixed-effects model for orange tree truncs
202204
206208
210212
8.1
8.15
8.2
−2.98
−2.97
−2.96
−2.95
−2.94
x 10−3
p1p2
p3
Figure: The set of consistent parameters for two subjects from theexample.
![Page 100: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/100.jpg)
Conclusions
![Page 101: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/101.jpg)
Conclusions
Standard numerics does not produce mathematics.
![Page 102: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/102.jpg)
Conclusions
Standard numerics does not produce mathematics.
Set-valued mathematics enables validated numerics.
![Page 103: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/103.jpg)
Conclusions
Standard numerics does not produce mathematics.
Set-valued mathematics enables validated numerics.
Existence (uniqueness) comes from fixed point theorems.
![Page 104: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/104.jpg)
Conclusions
Standard numerics does not produce mathematics.
Set-valued mathematics enables validated numerics.
Existence (uniqueness) comes from fixed point theorems.
Set-valued methods are suitable for inverse problems.
![Page 105: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/105.jpg)
Conclusions
Standard numerics does not produce mathematics.
Set-valued mathematics enables validated numerics.
Existence (uniqueness) comes from fixed point theorems.
Set-valued methods are suitable for inverse problems.
Parameter estimation is done via relaxation.
![Page 106: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/106.jpg)
Conclusions
Standard numerics does not produce mathematics.
Set-valued mathematics enables validated numerics.
Existence (uniqueness) comes from fixed point theorems.
Set-valued methods are suitable for inverse problems.
Parameter estimation is done via relaxation.
The relaxed problem is solved via set inversion.
![Page 107: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/107.jpg)
Further study...
Interval Computations Web Page
http://www.cs.utep.edu/interval-comp
![Page 108: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/108.jpg)
Further study...
Interval Computations Web Page
http://www.cs.utep.edu/interval-comp
INTLAB – INTerval LABoratory
http://www.ti3.tu-harburg.de/~rump/intlab/
![Page 109: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/109.jpg)
Further study...
Interval Computations Web Page
http://www.cs.utep.edu/interval-comp
INTLAB – INTerval LABoratory
http://www.ti3.tu-harburg.de/~rump/intlab/
CXSC – C eXtensions for Scientific Computation
http://www.xsc.de/
![Page 110: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/110.jpg)
Further study...
Interval Computations Web Page
http://www.cs.utep.edu/interval-comp
INTLAB – INTerval LABoratory
http://www.ti3.tu-harburg.de/~rump/intlab/
CXSC – C eXtensions for Scientific Computation
http://www.xsc.de/
CAPA – Computer–Aided Proofs in Analysis
http://www.math.uu.se/~warwick/CAPA/
![Page 111: Validated Numerics a short introduction to rigorous …Validated Numerics a short introduction to rigorous computations Warwick Tucker TheCAPAgroup Department ofMathematics UppsalaUniversity,Sweden](https://reader035.vdocuments.site/reader035/viewer/2022071019/5fd2a88433ce490fe11123aa/html5/thumbnails/111.jpg)
A short message from your sponsors...
Validated Numerics:
A Short Introduction to Rigorous
Computations
Warwick Tucker
Princeton University Press, 2011
ISBN: 9780691147819
152 pp.|6 x 9|41 illus.|12 tables.
USD 45.00/GBP 30.95
http://press.princeton.edu/titles/9488.html