![Page 1: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/1.jpg)
IntroductionComputing the probability of collision
Conclusion
A new method to compute the probability of collision for short-termspace encounters
R. Serra, D. Arzelier, M. Joldes, J-B. Lasserre, A. Rondepierre and B. Salvy
ANR FastRelax MeetingMay, 2015
![Page 2: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/2.jpg)
IntroductionComputing the probability of collision
Conclusion
Fiction...
Credit Gravity (2013)
![Page 3: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/3.jpg)
IntroductionComputing the probability of collision
Conclusion
General ModelShort-term encounter probability of collision
On-orbit collision
Figure: Cerise hit by a debris in 1996 (source : CNES/D. Ducros)
2 / 28
![Page 4: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/4.jpg)
IntroductionComputing the probability of collision
Conclusion
General ModelShort-term encounter probability of collision
On-orbit collision
Figure: Space debris population model (source : ESA)
3 / 28
![Page 5: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/5.jpg)
IntroductionComputing the probability of collision
Conclusion
General ModelShort-term encounter probability of collision
On-orbit collision
Context
Two objects: primary P (operational satellite) and secondary S (space debris)
Information about their geometry, position, velocity at a given time Affected by uncertainty
Needs:Risk assessmentDesign of a collision avoidance strategy
compute the probability of collision
Classical assumptions
Spherical objects
Gaussian probability density functions
Independent probability distribution laws
Rp
Rs
Rc
P
S
Figure: Combined spherical object
Probability of collision
Generally: 12-dimensional, Gaussian integrand, Complex integration domain
Computation: Monte-Carlo trials and/or simplified models
4 / 28
![Page 6: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/6.jpg)
IntroductionComputing the probability of collision
Conclusion
General ModelShort-term encounter probability of collision
On-orbit collision
Context
Two objects: primary P (operational satellite) and secondary S (space debris)
Information about their geometry, position, velocity at a given time Affected by uncertainty
Needs:Risk assessmentDesign of a collision avoidance strategy
compute the probability of collision
Classical assumptions
Spherical objects
Gaussian probability density functions
Independent probability distribution laws
Rp
Rs
Rc
P
S
Figure: Combined spherical object
Probability of collision
Generally: 12-dimensional, Gaussian integrand, Complex integration domain
Computation: Monte-Carlo trials and/or simplified models
4 / 28
![Page 7: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/7.jpg)
IntroductionComputing the probability of collision
Conclusion
General ModelShort-term encounter probability of collision
On-orbit collision
Context
Two objects: primary P (operational satellite) and secondary S (space debris)
Information about their geometry, position, velocity at a given time Affected by uncertainty
Needs:Risk assessmentDesign of a collision avoidance strategy
compute the probability of collision
Classical assumptions
Spherical objects
Gaussian probability density functions
Independent probability distribution laws
Rp
Rs
Rc
P
S
Figure: Combined spherical object
Probability of collision
Generally: 12-dimensional, Gaussian integrand, Complex integration domain
Computation: Monte-Carlo trials and/or simplified models4 / 28
![Page 8: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/8.jpg)
IntroductionComputing the probability of collision
Conclusion
General ModelShort-term encounter probability of collision
Short-term encounter model
Framework: High relative velocity
Assumptions:
Rectilinear relative motion
No velocity uncertainty
Infinite encounter time horizon
Encounter frame of reference
p
s
v = vs − vp
x
y
z
xm
zm
Figure: Encounter plane
5 / 28
![Page 9: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/9.jpg)
IntroductionComputing the probability of collision
Conclusion
General ModelShort-term encounter probability of collision
Short-term encounter model and probability of collision
Probability of collision: bivariate normal law
2-D integral over a disk B(0, R).
Pc =1
2π√|Σxy |
∫B((0,0),R)
exp
(−
1
2[x− xm y] Σ−1
xy [x− xm y]T)
dxdy.
covariance matrix
Σxy =
[σ2x ρxyσxσy
ρxyσxσy σ2y
],
standard deviations of relativecoordinates σx, σy ∈ R+
∗ ,
correlation coefficient ρxy ∈ (−1, 1).
p
s
v = vs − vp
x
y
z
xm
zm
Figure: Encounter plane
6 / 28
![Page 10: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/10.jpg)
IntroductionComputing the probability of collision
Conclusion
General ModelShort-term encounter probability of collision
Short-term encounter model and probability of collision
Probability of collision: bivariate normal law 2-D integral over a disk B(0, R).
Pc =1
2π√|Σxy |
∫B((0,0),R)
exp
(−
1
2[x− xm y] Σ−1
xy [x− xm y]T)
dxdy.
covariance matrix
Σxy =
[σ2x ρxyσxσy
ρxyσxσy σ2y
],
standard deviations of relativecoordinates σx, σy ∈ R+
∗ ,
correlation coefficient ρxy ∈ (−1, 1).
p
s
v = vs − vp
x
y
z
xm
zm
Figure: Encounter plane
6 / 28
![Page 11: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/11.jpg)
IntroductionComputing the probability of collision
Conclusion
General ModelShort-term encounter probability of collision
Short-term encounter model and probability of collision
Probability of collision: bivariate normal law 2-D integral over a disk B(0, R).
P =1
2πσxσy
∫B((0,0),R)
exp
(−
(x− xm)2
2σx2−
(y − ym)2
2σy2
)dxdy,
x
y
ym
combined xm
level sets
θ
R
xm
x
y
disc
θ
Figure: Rotation in encounter plane
where
R: radius of combined object
xm, ym: mean relative coordinates
σx, σy : standard deviations of relativecoordinates
7 / 28
![Page 12: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/12.jpg)
IntroductionComputing the probability of collision
Conclusion
General ModelShort-term encounter probability of collision
Short-term encounter model and probability of collision
Probability of collision: bivariate normal law 2-D integral over a disk B(0, R).
P =1
2πσxσy
∫B((0,0),R)
exp
(−
(x− xm)2
2σx2−
(y − ym)2
2σy2
)dxdy,
x
y
ym
combined xm
level sets
θ
R
xm
x
y
disc
θ
Figure: Rotation in encounter plane
where
R: radius of combined object
xm, ym: mean relative coordinates
σx, σy : standard deviations of relativecoordinates
7 / 28
![Page 13: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/13.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Existing methods
Methods based on numerical integration schemes:Foster ’92, Patera ’01, Alfano ’05.
Analytic methods: Chan ’97 uses some simplifying assumptions (˜σx = ˜σy)
provides truncation error bounds
Pro’s and Con’s
use truncated power series, but no rigorous proof about convergence rate
truncation orders fixed by trial and error and by comparing with other existing software:Approximately 60,000 test cases were used to evaluate the numerical expression [...] Thereference (”truth”) probability was computed with MATHCAD 11 [...] - Alfano’05
Fast and already used in practice
Our purpose
Give a ”simple”, ”analytic” formula, suitable for double-precision evaluationand effective error bounds.
8 / 28
![Page 14: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/14.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Existing methods
Methods based on numerical integration schemes:Foster ’92, Patera ’01, Alfano ’05.
Analytic methods: Chan ’97 uses some simplifying assumptions (˜σx = ˜σy)
provides truncation error bounds
Pro’s and Con’s
use truncated power series, but no rigorous proof about convergence rate
truncation orders fixed by trial and error and by comparing with other existing software:Approximately 60,000 test cases were used to evaluate the numerical expression [...] Thereference (”truth”) probability was computed with MATHCAD 11 [...] - Alfano’05
Fast and already used in practice
Our purpose
Give a ”simple”, ”analytic” formula, suitable for double-precision evaluationand effective error bounds.
8 / 28
![Page 15: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/15.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Existing methods
Methods based on numerical integration schemes:Foster ’92, Patera ’01, Alfano ’05.
Analytic methods: Chan ’97 uses some simplifying assumptions (˜σx = ˜σy)
provides truncation error bounds
Pro’s and Con’s
use truncated power series, but no rigorous proof about convergence rate
truncation orders fixed by trial and error and by comparing with other existing software:Approximately 60,000 test cases were used to evaluate the numerical expression [...] Thereference (”truth”) probability was computed with MATHCAD 11 [...] - Alfano’05
Fast and already used in practice
Our purpose
Give a ”simple”, ”analytic” formula, suitable for double-precision evaluationand effective error bounds.
8 / 28
![Page 16: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/16.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Existing methods
Methods based on numerical integration schemes:Foster ’92, Patera ’01, Alfano ’05.
Analytic methods: Chan ’97 uses some simplifying assumptions (˜σx = ˜σy)
provides truncation error bounds
Pro’s and Con’s
use truncated power series, but no rigorous proof about convergence rate
truncation orders fixed by trial and error and by comparing with other existing software:Approximately 60,000 test cases were used to evaluate the numerical expression [...] Thereference (”truth”) probability was computed with MATHCAD 11 [...] - Alfano’05
Fast and already used in practice
Our purpose
Give a ”simple”, ”analytic” formula, suitable for double-precision evaluationand effective error bounds.
8 / 28
![Page 17: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/17.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Existing methods
Methods based on numerical integration schemes:Foster ’92, Patera ’01, Alfano ’05.
Analytic methods: Chan ’97 uses some simplifying assumptions (˜σx = ˜σy) provides truncation error bounds
Pro’s and Con’s
use truncated power series, but no rigorous proof about convergence rate
truncation orders fixed by trial and error and by comparing with other existing software:Approximately 60,000 test cases were used to evaluate the numerical expression [...] Thereference (”truth”) probability was computed with MATHCAD 11 [...] - Alfano’05
Fast and already used in practice
Our purpose
Give a ”simple”, ”analytic” formula, suitable for double-precision evaluationand effective error bounds.
8 / 28
![Page 18: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/18.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Our method - Underlying techniques
1 Laplace transform: Lasserre and Zeron, Solving a Class of Multivariate Integration Problems via Laplace Techniques, Applicationes Mathematicae, 2001.
2 D-finite functions Zeilberger, A holonomic systems approach to special functions identities. Journal of Computational and Applied Mathematics, 1990 Salvy, D-finiteness: Algorithms and applications. In Manuel Kauers, editor, ISSAC 2005
solution of linear differential equation with polynomial coefficients power series coefficients satisfy a linear recurrence relation with polynomial coefficients
Example:
f(x) = exp(x) =∞∑i=0
fnxn ↔ {f ′ − f = 0, f(0) = 1} ↔ {(n + 1)fn+1 = fn, f0 = 1}
cos, arccos, Airy functions, Bessel functions, ...
3 Finite-precision evaluation of power series prone to cancellation Gawronski, Muller, Reinhard, Reduced Cancellation in the Evaluation of Entire Functions and Applications to the Error Function, SIAMJournal on Numerical Analysis, 2007.
Chevillard, Mezzarobba, Multiple-Precision Evaluation of the Airy Ai Function with Reduced Cancellation, 21st IEEE SYMPOSIUM on
Computer Arithmetic, 2013
9 / 28
![Page 19: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/19.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Our method - Underlying techniques
1 Laplace transform: Lasserre and Zeron, Solving a Class of Multivariate Integration Problems via Laplace Techniques, Applicationes Mathematicae, 2001.
2 D-finite functions Zeilberger, A holonomic systems approach to special functions identities. Journal of Computational and Applied Mathematics, 1990 Salvy, D-finiteness: Algorithms and applications. In Manuel Kauers, editor, ISSAC 2005
solution of linear differential equation with polynomial coefficients power series coefficients satisfy a linear recurrence relation with polynomial coefficients
Example:
f(x) = exp(x) =
∞∑i=0
fnxn ↔ {f ′ − f = 0, f(0) = 1} ↔ {(n + 1)fn+1 = fn, f0 = 1}
cos, arccos, Airy functions, Bessel functions, ...
3 Finite-precision evaluation of power series prone to cancellation Gawronski, Muller, Reinhard, Reduced Cancellation in the Evaluation of Entire Functions and Applications to the Error Function, SIAMJournal on Numerical Analysis, 2007.
Chevillard, Mezzarobba, Multiple-Precision Evaluation of the Airy Ai Function with Reduced Cancellation, 21st IEEE SYMPOSIUM on
Computer Arithmetic, 2013
9 / 28
![Page 20: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/20.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Our method - Underlying techniques
1 Laplace transform: Lasserre and Zeron, Solving a Class of Multivariate Integration Problems via Laplace Techniques, Applicationes Mathematicae, 2001.
2 D-finite functions Zeilberger, A holonomic systems approach to special functions identities. Journal of Computational and Applied Mathematics, 1990 Salvy, D-finiteness: Algorithms and applications. In Manuel Kauers, editor, ISSAC 2005
solution of linear differential equation with polynomial coefficients power series coefficients satisfy a linear recurrence relation with polynomial coefficients
Example:
f(x) = exp(x) =
∞∑i=0
fnxn ↔ {f ′ − f = 0, f(0) = 1} ↔ {(n + 1)fn+1 = fn, f0 = 1}
cos, arccos, Airy functions, Bessel functions, ...
3 Finite-precision evaluation of power series prone to cancellation Gawronski, Muller, Reinhard, Reduced Cancellation in the Evaluation of Entire Functions and Applications to the Error Function, SIAMJournal on Numerical Analysis, 2007.
Chevillard, Mezzarobba, Multiple-Precision Evaluation of the Airy Ai Function with Reduced Cancellation, 21st IEEE SYMPOSIUM on
Computer Arithmetic, 2013
9 / 28
![Page 21: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/21.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Our method - Underlying techniques
1 Laplace transform: Lasserre and Zeron, Solving a Class of Multivariate Integration Problems via Laplace Techniques, Applicationes Mathematicae, 2001.
2 D-finite functions Zeilberger, A holonomic systems approach to special functions identities. Journal of Computational and Applied Mathematics, 1990 Salvy, D-finiteness: Algorithms and applications. In Manuel Kauers, editor, ISSAC 2005
solution of linear differential equation with polynomial coefficients power series coefficients satisfy a linear recurrence relation with polynomial coefficients
Example:
f(x) = exp(x) =
∞∑i=0
fnxn ↔ {f ′ − f = 0, f(0) = 1} ↔ {(n + 1)fn+1 = fn, f0 = 1}
cos, arccos, Airy functions, Bessel functions, ...
3 Finite-precision evaluation of power series prone to cancellation Gawronski, Muller, Reinhard, Reduced Cancellation in the Evaluation of Entire Functions and Applications to the Error Function, SIAMJournal on Numerical Analysis, 2007.
Chevillard, Mezzarobba, Multiple-Precision Evaluation of the Airy Ai Function with Reduced Cancellation, 21st IEEE SYMPOSIUM on
Computer Arithmetic, 2013
9 / 28
![Page 22: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/22.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Laplace Transform
∀z ∈ R+ :
g(z) := P(√z) =
1
2πσxσy
∫B((0,0),
√z)
exp
(−
(x− xm)2
2σx2−
(y − ym)2
2σy2
)dxdy, (1)
L(g)(t) =
∫ +∞
0g(z) exp(−tz)dz (2)
. . . =
exp
(−σx
2ym2+σy
2xm2
2σx2σy2 + ym2
2σy2(2tσy2+1)+ xm
2
2σx2(2tσx2+1)
)t√
(2tσx2 + 1)(2tσy2 + 1)
(3)
L(g) is D-finite !
10 / 28
![Page 23: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/23.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Laplace Transform
∀z ∈ R+ :
g(z) := P(√z) =
1
2πσxσy
∫B((0,0),
√z)
exp
(−
(x− xm)2
2σx2−
(y − ym)2
2σy2
)dxdy, (1)
L(g)(t) =
∫ +∞
0g(z) exp(−tz)dz (2)
. . . =
exp
(−σx
2ym2+σy
2xm2
2σx2σy2 + ym2
2σy2(2tσy2+1)+ xm
2
2σx2(2tσx2+1)
)t√
(2tσx2 + 1)(2tσy2 + 1)
(3)
L(g) is D-finite !
10 / 28
![Page 24: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/24.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Laplace Transform
∀z ∈ R+ :
g(z) := P(√z) =
1
2πσxσy
∫B((0,0),
√z)
exp
(−
(x− xm)2
2σx2−
(y − ym)2
2σy2
)dxdy, (1)
L(g)(t) =
∫ +∞
0g(z) exp(−tz)dz (2)
. . . =
exp
(−σx
2ym2+σy
2xm2
2σx2σy2 + ym2
2σy2(2tσy2+1)+ xm
2
2σx2(2tσx2+1)
)t√
(2tσx2 + 1)(2tσy2 + 1)(3)
L(g) is D-finite !
10 / 28
![Page 25: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/25.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Laplace Transform
∀z ∈ R+ :
g(z) := P(√z) =
1
2πσxσy
∫B((0,0),
√z)
exp
(−
(x− xm)2
2σx2−
(y − ym)2
2σy2
)dxdy, (1)
L(g)(t) =
∫ +∞
0g(z) exp(−tz)dz (2)
. . . =
exp
(−σx
2ym2+σy
2xm2
2σx2σy2 + ym2
2σy2(2tσy2+1)+ xm
2
2σx2(2tσx2+1)
)t√
(2tσx2 + 1)(2tσy2 + 1)(3)
L(g) is D-finite !
10 / 28
![Page 26: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/26.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Borel-Laplace
g(z) L(g)(t) =exp
(−σx
2ym2+σy
2xm2
2σx2σy2 + ym2
2σy2(2tσy2+1)+ xm
2
2σx2(2tσx2+1)
)t√
(2tσx2+1)(2tσy2+1)
g(z) =∞∑i=0
li(i+1)!
zi+1 L(g)(t) := t2L(g)(1t
)=∞∑i=0
li(1t
)i
g(z) is:
D-finite
entire function of exponential type
type σ = 12σ2y
sum prone to cancellation
L(g)(t) is:
D-finite
Finite radius of convergence 2σ2y
li satisfy a linear recurrence with polynomial coefficients. Compute everything with gfun
Laplace Transform
expansion at ∞
Borel Transform
11 / 28
![Page 27: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/27.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Borel-Laplace
g(z) L(g)(t) =exp
(−σx
2ym2+σy
2xm2
2σx2σy2 + ym2
2σy2(2tσy2+1)+ xm
2
2σx2(2tσx2+1)
)t√
(2tσx2+1)(2tσy2+1)
g(z) =∞∑i=0
li(i+1)!
zi+1 L(g)(t) := t2L(g)(1t
)=∞∑i=0
li(1t
)i
g(z) is:
D-finite
entire function of exponential type
type σ = 12σ2y
sum prone to cancellation
L(g)(t) is:
D-finite
Finite radius of convergence 2σ2y
li satisfy a linear recurrence with polynomial coefficients. Compute everything with gfun
Laplace Transform
expansion at ∞
Borel Transform
11 / 28
![Page 28: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/28.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Borel-Laplace
g(z) L(g)(t) =exp
(−σx
2ym2+σy
2xm2
2σx2σy2 + ym2
2σy2(2tσy2+1)+ xm
2
2σx2(2tσx2+1)
)t√
(2tσx2+1)(2tσy2+1)
g(z) =∞∑i=0
li(i+1)!
zi+1 L(g)(t) := t2L(g)(1t
)=∞∑i=0
li(1t
)i
g(z) is:
D-finite
entire function of exponential type
type σ = 12σ2y
sum prone to cancellation
L(g)(t) is:
D-finite
Finite radius of convergence 2σ2y
li satisfy a linear recurrence with polynomial coefficients. Compute everything with gfun
Laplace Transform
expansion at ∞
Borel Transform
11 / 28
![Page 29: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/29.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Borel-Laplace
g(z) L(g)(t) =exp
(−σx
2ym2+σy
2xm2
2σx2σy2 + ym2
2σy2(2tσy2+1)+ xm
2
2σx2(2tσx2+1)
)t√
(2tσx2+1)(2tσy2+1)
g(z) =∞∑i=0
li(i+1)!
zi+1 L(g)(t) := t2L(g)(1t
)=∞∑i=0
li(1t
)i
g(z) is:
D-finite
entire function of exponential type
type σ = 12σ2y
sum prone to cancellation
L(g)(t) is:
D-finite
Finite radius of convergence 2σ2y
li satisfy a linear recurrence with polynomial coefficients. Compute everything with gfun
Laplace Transform
expansion at ∞
Borel Transform
11 / 28
![Page 30: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/30.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Cancellation in finite precision power series evaluation
Example: σx = 115, σy = 1.41, xm = 0.15, ym = 3.88,√z = 15
g(z) =∞∑i=0
li(i+1)!
zi+1
g(225) = 0.16 · 10−1 + 1.5 + 16.1−250 . . .+2.2 · 1019−2.6 · 1019− . . .+4.3−0.14− 0.60 . . .
Values of∣∣∣ li225i+1
(i+1)!
∣∣∣, compared to g(225) ' 0.1004:
Lost Digits: dg(z) ' logmaxi
∣∣gizi∣∣|g(z)|
12 / 28
![Page 31: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/31.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Cancellation in finite precision power series evaluation
Example: σx = 115, σy = 1.41, xm = 0.15, ym = 3.88,√z = 15
g(z) =∞∑i=0
li(i+1)!
zi+1
g(225) = 0.16 · 10−1 + 1.5 + 16.1−250 . . .+2.2 · 1019−2.6 · 1019− . . .+4.3−0.14− 0.60 . . .
Values of∣∣∣ li225i+1
(i+1)!
∣∣∣, compared to g(225) ' 0.1004:
Lost Digits: dg(z) ' logmaxi
∣∣gizi∣∣|g(z)|
12 / 28
![Page 32: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/32.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Cancellation in finite precision power series evaluation
Example: σx = 115, σy = 1.41, xm = 0.15, ym = 3.88,√z = 15
g(z) =∞∑i=0
li(i+1)!
zi+1
g(225) = 0.16 · 10−1 + 1.5 + 16.1−250 . . .+2.2 · 1019−2.6 · 1019− . . .+4.3−0.14− 0.60 . . .
Values of∣∣∣ li225i+1
(i+1)!
∣∣∣, compared to g(225) ' 0.1004:
Lost Digits: dg(z) ' logmaxi
∣∣gizi∣∣|g(z)|
12 / 28
![Page 33: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/33.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Cancellation in finite precision power series evaluation
Example: exp(−x) =∞∑i=0
(−1)ixi
i!
exp(−20) = 1−20 . . .+1.66 · 107−1.23 · 107+ . . .+1.19 · 10−8−3.45 · 10−9 . . .
Values of∣∣∣ (−1)i20i
i!
∣∣∣, compared to exp(−20) ' 2.06 · 10−9:
Lost Digits: dg(z) ' logmaxi
∣∣gizi∣∣|g(z)|
BUT...
exp(−x) =1
exp(x)No cancellation!
13 / 28
![Page 34: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/34.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Cancellation in finite precision power series evaluation
Example: exp(−x) =∞∑i=0
(−1)ixi
i!
exp(−20) = 1−20 . . .+1.66 · 107−1.23 · 107+ . . .+1.19 · 10−8−3.45 · 10−9 . . .
Values of∣∣∣ (−1)i20i
i!
∣∣∣, compared to exp(−20) ' 2.06 · 10−9:
Lost Digits: dg(z) ' logmaxi
∣∣gizi∣∣|g(z)|
BUT...
exp(−x) =1
exp(x)No cancellation!
13 / 28
![Page 35: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/35.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Cancellation in finite precision power series evaluation
Example: exp(−x) =∞∑i=0
(−1)ixi
i!
exp(−20) = 1−20 . . .+1.66 · 107−1.23 · 107+ . . .+1.19 · 10−8−3.45 · 10−9 . . .
Values of∣∣∣ (−1)i20i
i!
∣∣∣, compared to exp(−20) ' 2.06 · 10−9:
Lost Digits: dg(z) ' logmaxi
∣∣gizi∣∣|g(z)|
BUT...
exp(−x) =1
exp(x)No cancellation!
13 / 28
![Page 36: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/36.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Cancellation in finite precision power series evaluation
Example: σx = 115, σy = 1.41, xm = 0.15, ym = 3.88,√z = 15
g(z) =∞∑i=0
li(i+1)!
zi+1
g(225) = 0.16 · 10−1 + 1.5 + 16.1−250 . . .+2.2 · 1019−2.6 · 1019− . . .+4.3−0.14− 0.60 . . .
Values of∣∣∣ li225i+1
(i+1)!
∣∣∣, compared to g(225) ' 0.1004:
g(z) =F (z)
G(z)No cancellation!
Gawronski, Muller, Reinhard (2007) Method:
Find F (z) and G(z) using ”complex analysistricks” indicator function.
14 / 28
![Page 37: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/37.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Cancellation in finite precision power series evaluation
Example: σx = 115, σy = 1.41, xm = 0.15, ym = 3.88,√z = 15
g(z) =∞∑i=0
li(i+1)!
zi+1
g(225) = 0.16 · 10−1 + 1.5 + 16.1−250 . . .+2.2 · 1019−2.6 · 1019− . . .+4.3−0.14− 0.60 . . .
Values of∣∣∣ li225i+1
(i+1)!
∣∣∣, compared to g(225) ' 0.1004:
g(z) =F (z)
G(z)No cancellation!
Gawronski, Muller, Reinhard (2007) Method:
Find F (z) and G(z) using ”complex analysistricks” indicator function.
14 / 28
![Page 38: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/38.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Cancellation in finite precision power series evaluation
Example: σx = 115, σy = 1.41, xm = 0.15, ym = 3.88,√z = 15
g(z) =∞∑i=0
li(i+1)!
zi+1
g(225) = 0.16 · 10−1 + 1.5 + 16.1−250 . . .+2.2 · 1019−2.6 · 1019− . . .+4.3−0.14− 0.60 . . .
Values of∣∣∣ li225i+1
(i+1)!
∣∣∣, compared to g(225) ' 0.1004:
g(z) =F (z)
G(z)No cancellation!
Gawronski, Muller, Reinhard (2007) Method:
Find F (z) and G(z) using ”complex analysistricks” indicator function.
14 / 28
![Page 39: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/39.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
GMR Method in a Nutshell
Lost Digits:
dg(reiθ) ' logmaxi|gi| ri∣∣g(reiθ)
∣∣' rρ(σ − h(θ))
where
g has order ρ and type σ:max|z|=r
|g(z)| ∼ exp (σrρ) for large r
indicator function h:∣∣g(reiθ)∣∣ ∼ exp (h(θ)r) for large r
Goal of GMR:
To efficiently evaluate g on a sector Sθ, find suitable G and F , s.t. g =F
G, and
σF − hF (θ) and σG − hG(θ) are minimized on Sθ.
15 / 28
![Page 40: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/40.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
GMR Method in a Nutshell
Lost Digits:
dg(reiθ) ' logmaxi|gi| ri∣∣g(reiθ)
∣∣' rρ(σ − h(θ))
where
g has order ρ and type σ:max|z|=r
|g(z)| ∼ exp (σrρ) for large r
indicator function h:∣∣g(reiθ)∣∣ ∼ exp (h(θ)r) for large r
Goal of GMR:
To efficiently evaluate g on a sector Sθ, find suitable G and F , s.t. g =F
G, and
σF − hF (θ) and σG − hG(θ) are minimized on Sθ.
15 / 28
![Page 41: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/41.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Borel-Laplace + GMR Method
g(z) =F (z)
G(z)L(g)(t) =
exp
(−σx
2ym2+σy
2xm2
2σx2σy2 + ym2
2σy2(2tσy2+1)+ xm
2
2σx2(2tσx2+1)
)t√
(2tσx2+1)(2tσy2+1)
g(z) =∞∑i=0
li(i+1)!
zi+1 L(g)(t) := t2L(g)(1t
)=∞∑i=0
li(1t
)i
g(z) is:
D-finite
entire function of exponential type
type σ = 12σ2y
indicator function:∣∣g(reiθ)∣∣ ∼ exp (h(θ)r) for large r
k(θ) = supz∈K{<(ze−iθ)}, θ ∈ [0, 2π]
Polya’s Theorem: h(θ) = k(−θ)
h(θ) =
{− cos θ2σy2 if θ ∈ [π
2, 3π
2]
0 if θ ∈ [0, π2
) ∪ ( 3π2, 2π].
Laplace
expansion at ∞
Borel Transform
16 / 28
![Page 42: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/42.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Borel-Laplace + GMR Method
g(z) =F (z)
G(z)L(g)(t) =
exp
(−σx
2ym2+σy
2xm2
2σx2σy2 + ym2
2σy2(2tσy2+1)+ xm
2
2σx2(2tσx2+1)
)t√
(2tσx2+1)(2tσy2+1)
g(z) =∞∑i=0
li(i+1)!
zi+1 L(g)(t) := t2L(g)(1t
)=∞∑i=0
li(1t
)i
g(z) is:
D-finite
entire function of exponential type
type σ = 12σ2y
indicator function:∣∣g(reiθ)∣∣ ∼ exp (h(θ)r) for large r
k(θ) = supz∈K{<(ze−iθ)}, θ ∈ [0, 2π]
Polya’s Theorem: h(θ) = k(−θ)
h(θ) =
{− cos θ2σy2 if θ ∈ [π
2, 3π
2]
0 if θ ∈ [0, π2
) ∪ ( 3π2, 2π].
Laplace
expansion at ∞
Borel Transform
16 / 28
![Page 43: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/43.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Borel-Laplace + GMR Method
g(z) =F (z)
G(z)L(g)(t) =
exp
(−σx
2ym2+σy
2xm2
2σx2σy2 + ym2
2σy2(2tσy2+1)+ xm
2
2σx2(2tσx2+1)
)t√
(2tσx2+1)(2tσy2+1)
g(z) =∞∑i=0
li(i+1)!
zi+1 L(g)(t) := t2L(g)(1t
)=∞∑i=0
li(1t
)i
g(z) is:
D-finite
entire function of exponential type
type σ = 12σ2y
indicator function:∣∣g(reiθ)∣∣ ∼ exp (h(θ)r) for large r
k(θ) = supz∈K{<(ze−iθ)}, θ ∈ [0, 2π]
Polya’s Theorem: h(θ) = k(−θ)
h(θ) =
{− cos θ2σy2 if θ ∈ [π
2, 3π
2]
0 if θ ∈ [0, π2
) ∪ ( 3π2, 2π].
Laplace
expansion at ∞
Borel Transform
16 / 28
![Page 44: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/44.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Zoom on Indicator functions
g(z) =F (z)
G(z)∣∣g(reiθ)∣∣ ∼ exp (h(θ)r) for large r
indicator of g:
h(θ) =
{− cos θ2σy2 if θ ∈ [π
2, 3π
2]
0 otherwise.
lost precision ∼ |σ − h(0)|
G(z) = exp(σz)indicator of G:
indicator of F :
17 / 28
![Page 45: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/45.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Zoom on Indicator functions
g(z) =F (z)
G(z)∣∣g(reiθ)∣∣ ∼ exp (h(θ)r) for large r
indicator of g:
h(θ) =
{− cos θ2σy2 if θ ∈ [π
2, 3π
2]
0 otherwise.
lost precision ∼ |σ − h(0)|
G(z) = exp(σz)indicator of G:
indicator of F :
17 / 28
![Page 46: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/46.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Zoom on Indicator functions
g(z) =F (z)
G(z)∣∣g(reiθ)∣∣ ∼ exp (h(θ)r) for large r
indicator of g:
h(θ) =
{− cos θ2σy2 if θ ∈ [π
2, 3π
2]
0 otherwise.
lost precision ∼ |σ − h(0)|
G(z) = exp(σz)indicator of G:
indicator of F :
17 / 28
![Page 47: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/47.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Zoom on Indicator functions
G(z)g(z) = F (z)∣∣g(reiθ)∣∣ ∼ exp (h(θ)r) for large r
indicator of g:
h(θ) =
{− cos θ2σy2 if θ ∈ [π
2, 3π
2]
0 otherwise.
lost precision ∼ |σ − h(0)|
G(z) = exp(σz)indicator of G:
indicator of F :
17 / 28
![Page 48: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/48.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Borel-Laplace + GMR Method
G(z)g(z) = F (z) L(g)(t− σ)
G(z)g(z) =∞∑k=0
αk(k+1)!
zk+1∞∑k=0
αk(1t
)k
G(z) = exp(σz)
F is D-finite
reduced cancellation for evaluating F,G on positive real line
recurrence for αk:
− (32kσ4xσ
10y + 128σ
4xσ
10y )αk+4 = (σ
4xy
2m − 2σ
2xσ
2yy
2m + σ
4yy
2m)αk
+((−4σ4xσ
4y + 8σ
2xσ
6y − 4σ
8y)k − 10σ
4xσ
4y − 6σ
4xσ
2yy
2m + 20σ
2xσ
6y + 8σ
2xσ
4yy
2m − 10σ
8y − 2σ
6yy
2m)αk+1
+((24σ4xσ
6y − 32σ
2xσ
8y + 8σ
10y )k + 72σ
4xσ
6y + 12σ
4xσ
4yy
2m − 92σ
2xσ
8y − 8σ
2xσ
6yy
2m + 20σ
10y + 4σ
8yx
2m)αk+2
+((−48σ4xσ
8y + 32σ
2xσ
10y )k − 168σ
4xσ
8y − 8σ
4xσ
6yy
2m + 104σ
2xσ
10y − 8σ
10y x
2m)αk+3,
Laplace
expansion at ∞
Borel Transform
18 / 28
![Page 49: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/49.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Sketch of the proof - Borel-Laplace + GMR Method
G(z)g(z) = F (z) L(g)(t− σ)
G(z)g(z) =∞∑k=0
αk(k+1)!
zk+1∞∑k=0
αk(1t
)k
G(z) = exp(σz)
F is D-finite
reduced cancellation for evaluating F,G on positive real line
recurrence for αk:
− (32kσ4xσ
10y + 128σ
4xσ
10y )αk+4 = (σ
4xy
2m − 2σ
2xσ
2yy
2m + σ
4yy
2m)αk
+((−4σ4xσ
4y + 8σ
2xσ
6y − 4σ
8y)k − 10σ
4xσ
4y − 6σ
4xσ
2yy
2m + 20σ
2xσ
6y + 8σ
2xσ
4yy
2m − 10σ
8y − 2σ
6yy
2m)αk+1
+((24σ4xσ
6y − 32σ
2xσ
8y + 8σ
10y )k + 72σ
4xσ
6y + 12σ
4xσ
4yy
2m − 92σ
2xσ
8y − 8σ
2xσ
6yy
2m + 20σ
10y + 4σ
8yx
2m)αk+2
+((−48σ4xσ
8y + 32σ
2xσ
10y )k − 168σ
4xσ
8y − 8σ
4xσ
6yy
2m + 104σ
2xσ
10y − 8σ
10y x
2m)αk+3,
Laplace
expansion at ∞
Borel Transform
18 / 28
![Page 50: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/50.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Prop. Coefficients αk are positive
Suppose σx > σy .
L(x) =∞∑k=0
αkxk satisfies L′(x) = F (x)L(x), L(0) =
exp
(−σ2xy
2m + σ2
yx2m
2σ2xσ
2y
)2σxσy
, where
F (x) =ym2
4σy4+
σy4xm2
(x (σx2 − σy2)− 2σx2σy2)2−
1
−2σy2 + x
+−σx2 + σy2
2x (σx2 − σy2)− 4σx2σy2.
0 ≤
fk =1 +
(1−
σ2y
σ2x
)k ((k + 1)
(x2mσ
2y
σ4x
)+ 1−
σ2y
σ2x
)2
(2σ2y)k+1
+
0, k > 0
ym2
4σ4y
, k = 0,
0 ≤ (n+ 1)αn+1 =n∑i=0
fi · αn−i, by induction.
19 / 28
![Page 51: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/51.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Prop. Coefficients αk are positive
Suppose σx > σy .
L(x) =∞∑k=0
αkxk satisfies L′(x) = F (x)L(x), L(0) =
exp
(−σ2xy
2m + σ2
yx2m
2σ2xσ
2y
)2σxσy
, where
F (x) =ym2
4σy4+
σy4xm2
(x (σx2 − σy2)− 2σx2σy2)2−
1
−2σy2 + x
+−σx2 + σy2
2x (σx2 − σy2)− 4σx2σy2.
0 ≤
fk =1 +
(1−
σ2y
σ2x
)k ((k + 1)
(x2mσ
2y
σ4x
)+ 1−
σ2y
σ2x
)2
(2σ2y)k+1
+
0, k > 0
ym2
4σ4y
, k = 0,
0 ≤ (n+ 1)αn+1 =n∑i=0
fi · αn−i, by induction.
19 / 28
![Page 52: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/52.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Prop. Coefficients αk are positive
Suppose σx > σy .
L(x) =∞∑k=0
αkxk satisfies L′(x) = F (x)L(x), L(0) =
exp
(−σ2xy
2m + σ2
yx2m
2σ2xσ
2y
)2σxσy
, where
F (x) =ym2
4σy4+
σy4xm2
(x (σx2 − σy2)− 2σx2σy2)2−
1
−2σy2 + x
+−σx2 + σy2
2x (σx2 − σy2)− 4σx2σy2.
0 ≤ fk =1 +
(1−
σ2y
σ2x
)k ((k + 1)
(x2mσ
2y
σ4x
)+ 1−
σ2y
σ2x
)2
(2σ2y)k+1
+
0, k > 0
ym2
4σ4y
, k = 0,
0 ≤ (n+ 1)αn+1 =n∑i=0
fi · αn−i, by induction.
19 / 28
![Page 53: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/53.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Prop. Coefficients αk are positive
Suppose σx > σy .
L(x) =∞∑k=0
αkxk satisfies L′(x) = F (x)L(x), L(0) =
exp
(−σ2xy
2m + σ2
yx2m
2σ2xσ
2y
)2σxσy
, where
F (x) =ym2
4σy4+
σy4xm2
(x (σx2 − σy2)− 2σx2σy2)2−
1
−2σy2 + x
+−σx2 + σy2
2x (σx2 − σy2)− 4σx2σy2.
0 ≤ fk =1 +
(1−
σ2y
σ2x
)k ((k + 1)
(x2mσ
2y
σ4x
)+ 1−
σ2y
σ2x
)2
(2σ2y)k+1
+
0, k > 0
ym2
4σ4y
, k = 0,
0 ≤ (n+ 1)αn+1 =n∑i=0
fi · αn−i, by induction.
19 / 28
![Page 54: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/54.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Bounds using majorant series
Let γ :=
1 + 12
(1−
σ2y
σ2x
+x2mσ
2y
σ4x
+ ym2
σy2
)2σy2
, αk := α0γk and αk := α0
(1
2σy2
)k.
Then αk ≤ αk ≤ αk, ∀k ∈ N.
Let PN (z) :=N−1∑k=0
αkzk
(k + 1)!. Then we have the following error bounds:
εN (z) ≤ g(z)− PN (z) ≤ εN (z), where
εN (z) := 2α0σ2ye− z
2σ2y
(z
2σ2y
)N+1
(N + 1)!,
εN (z) :=α0
γe− z
2σ2y(zγ)N+1
(N + 1)!.
20 / 28
![Page 55: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/55.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Bounds using majorant series
Let γ :=
1 + 12
(1−
σ2y
σ2x
+x2mσ
2y
σ4x
+ ym2
σy2
)2σy2
, αk := α0γk and αk := α0
(1
2σy2
)k.
Then αk ≤ αk ≤ αk, ∀k ∈ N.
Let PN (z) :=N−1∑k=0
αkzk
(k + 1)!. Then we have the following error bounds:
εN (z) ≤ g(z)− PN (z) ≤ εN (z), where
εN (z) := 2α0σ2ye− z
2σ2y
(z
2σ2y
)N+1
(N + 1)!,
εN (z) :=α0
γe− z
2σ2y(zγ)N+1
(N + 1)!.
20 / 28
![Page 56: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/56.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
ExamplesSample 1
Case Input parameters (km)# σx σy R xm ym1 0.05 0.025 0.005 0.01 02 0.05 0.025 0.005 0 0.013 0.075 0.025 0.005 0.01 04 0.075 0.025 0.005 0 0.015 3 1 0.01 1 06 3 1 0.01 0 17 3 1 0.01 10 08 3 1 0.01 0 109 10 1 0.01 10 0
10 10 1 0.01 0 1011 3 1 0.05 5 012 3 1 0.05 0 5
21 / 28
![Page 57: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/57.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Examples: quality η =ε10(z)− ε10(z)
ε10(z) + P10(z)and plot log
(piz
i)
Case 1: η = 23 Case 2: η = 22 Case 4: η = 22
Case 6: η = 47 Case 8: η = 33 Case 11: η = 34
22 / 28
![Page 58: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/58.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Numerical studySample 1
Case Probability of Collision (-)# Alfano Patera Chan New method1 9.742× 10−3 9.741× 10−3 9.754× 10−3 9.742× 10−3
2 9.181× 10−3 9.181× 10−3 9.189× 10−3 9.181× 10−3
3 6.571× 10−3 6.571× 10−3 6.586× 10−3 6.571× 10−3
4 6.125× 10−3 6.125× 10−3 6.135× 10−3 6.125× 10−3
5 1.577× 10−5 1.577× 10−5 1.577× 10−5 1.577× 10−5
6 1.011× 10−5 1.011× 10−5 1.011× 10−5 1.011× 10−5
7 6.443× 10−8 6.443× 10−8 6.443× 10−8 6.443× 10−8
8 0 3.219× 10−27 3.216× 10−27 3.219× 10−27
9 3.033× 10−6 3.033× 10−6 3.033× 10−6 3.033× 10−6
10 0 9.656× 10−28 9.645× 10−28 9.656× 10−28
11 1.039× 10−4 1.039× 10−4 1.039× 10−4 1.039× 10−4
12 1.564× 10−9 1.564× 10−9 1.556× 10−9 1.564× 10−9
− equal to reference value (from [Chan 2008])
23 / 28
![Page 59: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/59.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Examples: quality ηN =εN (z)− εN (z)εN (z) + PN (z)
and plot log(piz
i)
Case 3 Alfano: η800 = 30 Case 5 Alfano: η121000 = 20
Sample 2 (from [Alfano 2009])
Case Input parameters (m)# σx σy R xm ym3 114.25 1.41 15 0.15 3.885 177.8 0.038 10 2.12 -1.22
24 / 28
![Page 60: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/60.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
ExamplesSample 2 (from [Alfano 2009])
Case Input parameters (m)# σx σy R xm ym3 114.25 1.41 15 0.15 3.885 177.8 0.038 10 2.12 -1.22
Alfano’s test Probability of collision (-)case number Alfano New method Reference (MC)
3 0.10038 0.10038 0.100855 0.044712 0.045509 0.044499
25 / 28
![Page 61: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/61.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Saddle-point method for ”degenerate” cases
Mellin’s Inverse Formula
g(z) =1
2iπ
∫γL(g)(t)eztdt,
where γ is any contour that encloses the conjugate indicator diagram of L(g)(t).
Saddle-point intuition: choose γ such that
g(z) = L(g)(t0)ezt01
2iπ
∫γe−
M2u2
du,
Search for an approximate saddle-point t0 root of∂(L(g)(t)ezt)
∂t= 0.
Example Alfano 3:
Asymptotic behavior: g(z) ∼ L(g)(t0)ezt0
√2πM
26 / 28
![Page 62: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/62.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Saddle-point method for ”degenerate” cases
Mellin’s Inverse Formula
g(z) =1
2iπ
∫γL(g)(t)eztdt,
where γ is any contour that encloses the conjugate indicator diagram of L(g)(t).
Saddle-point intuition: choose γ such that
g(z) = L(g)(t0)ezt01
2iπ
∫γe−
M2u2
du,
Search for an approximate saddle-point t0 root of∂(L(g)(t)ezt)
∂t= 0.
Example Alfano 3:
Asymptotic behavior: g(z) ∼ L(g)(t0)ezt0
√2πM
26 / 28
![Page 63: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/63.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Saddle-point method for ”degenerate” cases
Mellin’s Inverse Formula
g(z) =1
2iπ
∫γL(g)(t)eztdt,
where γ is any contour that encloses the conjugate indicator diagram of L(g)(t).
Saddle-point intuition: choose γ such that
g(z) = L(g)(t0)ezt01
2iπ
∫γe−
M2u2
du,
Search for an approximate saddle-point t0 root of∂(L(g)(t)ezt)
∂t= 0.
Example Alfano 3:
Asymptotic behavior: g(z) ∼ L(g)(t0)ezt0
√2πM
26 / 28
![Page 64: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/64.jpg)
IntroductionComputing the probability of collision
Conclusion
Existing methodsOur methodExamples
Examples: Relative error using saddle-point approximation
Case Numbers of terms in the inverted series# 1 2 31 0.06 −0.8 · 10−3 0.1 · 10−4
2 0.05 −0.3 · 10−3 0.4 · 10−4
3 0.05 −0.4 · 10−3 0.4 · 10−4
4 0.05 −0.4 · 10−3 0.3 · 10−4
. . . . . . . . .
Alfano3 0.06 −0.73 · 10−3 0.2 · 10−3
Alfano5 0.06 −0.72 · 10−3 0.1 · 10−3
27 / 28
![Page 65: A new method to compute the probability of collision for short …fastrelax.gforge.inria.fr/files/fastrelax2015Joldes.pdf · 2015-06-11 · Introduction Computing the probability](https://reader030.vdocuments.us/reader030/viewer/2022040905/5e78dc9b34e84459315c03a8/html5/thumbnails/65.jpg)
IntroductionComputing the probability of collision
Conclusion
Conclusion
New method
Analytical formula
Reduced cancellation evaluation
Error bounds
No simplifying assumption
Current and future work
Saddle-point method for ”degenerate” cases
Long-term 3D encounter model
Extension to polygonal cross-sections
28 / 28