continuous dynamical systems and computation
TRANSCRIPT
![Page 1: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/1.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Continuous dynamical systems andcomputation
Computing with polynomial ODE’s
Manuel Campagnolo
DM/ISA, Technical University of Lisbonand
SQIG/Instituto de Telecomunicações, Lisbon
CUNY Logic Workshop, New York City, March 12, 2010
![Page 2: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/2.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Outline
1 IntroductionContinuous time computationDynamical systems
2 Polynomial IVPsPreliminariesSimulation of Turing machinesApplications
3 Polynomial IVPs and CAComputable analysisApproximation of computable real functions with PIVPs
4 References
![Page 3: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/3.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Continuous time computation
Motivation
How can we link computation and continuous dynamicalsystems?
What does this tell us about continuous dynamicalsystems?Is there a canonical continuous dynamical system withrespect to computability?
![Page 4: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/4.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Continuous time computation
Motivation
How can we link computation and continuous dynamicalsystems?What does this tell us about continuous dynamicalsystems?
Is there a canonical continuous dynamical system withrespect to computability?
![Page 5: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/5.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Continuous time computation
Motivation
How can we link computation and continuous dynamicalsystems?What does this tell us about continuous dynamicalsystems?Is there a canonical continuous dynamical system withrespect to computability?
![Page 6: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/6.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Continuous time computation
The differential analyser
Computing in continuous time with an analog computer:
The differential analyser, which concept dates to LordKelvin and his brother James Thomson in 1876, andwas constructed at MIT under the supervision ofVannevar Bush in 1932.
![Page 7: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/7.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Continuous time computation
The differential analyser
Figure: Hartree and Porter Meccano differential analyzer originallybuilt in 1934
![Page 8: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/8.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Continuous time computation
Analog circuits
- ---
- -
−1
t
∫ ×∫cos tsin t
− sin t
Figure: A circuit that calculates sin and cos. Its initial conditions aresin(0) = 0 and cos(0) = 1. The output w of the integrator unit
∫obeys dw = u dv where u and v are its upper and lower inputsrespectively.
The circuit above is represented by the system of equations
y ′1 = y2y ′2 = −y1
which solution is y1 = sin t and y2 = cos t given the initialconditions y1(0) = 0 and y2(0) = 1.
![Page 9: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/9.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Continuous time computation
Analog circuits
- ---
- -
−1
t
∫ ×∫cos tsin t
− sin t
Figure: A circuit that calculates sin and cos. Its initial conditions aresin(0) = 0 and cos(0) = 1. The output w of the integrator unit
∫obeys dw = u dv where u and v are its upper and lower inputsrespectively.
The circuit above is represented by the system of equations
y ′1 = y2y ′2 = −y1
which solution is y1 = sin t and y2 = cos t given the initialconditions y1(0) = 0 and y2(0) = 1.
![Page 10: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/10.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Continuous time computation
Shannon’s GPAC
In 1941 Claude Shannon proposed a mathematical model forthe Differential Analyser.
Shannon proved that, given a sufficient number of integrators,any differentially algebraic function, i.e., a solution of
p(t , y , y ′, . . . , y (n)) = 0
could be generated.
Shannon’s model was refined by Graça and Costa (2003) whoshowed that all non-degenerate GPAC functions are preciselyPIVP functions, i.e. solutions of polynomial initial valueproblems
y ′ = p(y , t) , y(0) = x .
![Page 11: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/11.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Continuous time computation
Shannon’s GPAC
In 1941 Claude Shannon proposed a mathematical model forthe Differential Analyser.
Shannon proved that, given a sufficient number of integrators,any differentially algebraic function, i.e., a solution of
p(t , y , y ′, . . . , y (n)) = 0
could be generated.
Shannon’s model was refined by Graça and Costa (2003) whoshowed that all non-degenerate GPAC functions are preciselyPIVP functions, i.e. solutions of polynomial initial valueproblems
y ′ = p(y , t) , y(0) = x .
![Page 12: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/12.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Dynamical systems
Dynamical systems
A discrete dynamical system [X ,N, φ] defined on thetopological space X over N is a function φ : X ×N→ X with theproperties:
1 initial condition: φ(x ,0) = x for all x ∈ X ;2 semigroup property: φ(φ(x , t1), t2) = φ(x , t1 + t2) for all
t1, t2 ∈ N (for φ(., .) ∈ X ).
A continuous dynamical system [X ,R+0 , φ] is defined
analogously, where N is replaced by R+0 .
![Page 13: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/13.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Dynamical systems
Dynamical systems
Maps and flows:
A discrete dynamical system can be written asyt+1 = f (yt ), with y0 = x , where f (.) = φ(.,1) is called thetransition function.
If φ is continuously differentiable with respect to t then, acontinuous dynamical system gives rise to an initial valueproblem y ′ = f (t , y), y(0) = x , where f is called a vectorfield.
![Page 14: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/14.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Dynamical systems
Dynamical systems
Maps and flows:
A discrete dynamical system can be written asyt+1 = f (yt ), with y0 = x , where f (.) = φ(.,1) is called thetransition function.
If φ is continuously differentiable with respect to t then, acontinuous dynamical system gives rise to an initial valueproblem y ′ = f (t , y), y(0) = x , where f is called a vectorfield.
![Page 15: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/15.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Dynamical systems
On [0,1]× [0,1] there is a piecewise linear function thatsimulates the transition function of an arbritrary TM (Moore1991)
On R there is an analytic closed form function thatsimulates the transition function of an arbritrary TM (Mooreand Koiran 1996)3-dimensional piecewise constant differential flows onbounded domains simulate the dynamics of an arbritraryTM (Asarin, Maler and Pnueli 1995)On R3 there is a continuous flow y ′ = f (t , y) that simulatesthe dynamics of an arbritrary TM (Branicky 1996)It is conjectured that no analytic map on a compact,finite-dimensional space can simulate a Turing machinethrough a reasonable encoding (Moore 1998).
![Page 16: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/16.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Dynamical systems
On [0,1]× [0,1] there is a piecewise linear function thatsimulates the transition function of an arbritrary TM (Moore1991)On R there is an analytic closed form function thatsimulates the transition function of an arbritrary TM (Mooreand Koiran 1996)
3-dimensional piecewise constant differential flows onbounded domains simulate the dynamics of an arbritraryTM (Asarin, Maler and Pnueli 1995)On R3 there is a continuous flow y ′ = f (t , y) that simulatesthe dynamics of an arbritrary TM (Branicky 1996)It is conjectured that no analytic map on a compact,finite-dimensional space can simulate a Turing machinethrough a reasonable encoding (Moore 1998).
![Page 17: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/17.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Dynamical systems
On [0,1]× [0,1] there is a piecewise linear function thatsimulates the transition function of an arbritrary TM (Moore1991)On R there is an analytic closed form function thatsimulates the transition function of an arbritrary TM (Mooreand Koiran 1996)3-dimensional piecewise constant differential flows onbounded domains simulate the dynamics of an arbritraryTM (Asarin, Maler and Pnueli 1995)
On R3 there is a continuous flow y ′ = f (t , y) that simulatesthe dynamics of an arbritrary TM (Branicky 1996)It is conjectured that no analytic map on a compact,finite-dimensional space can simulate a Turing machinethrough a reasonable encoding (Moore 1998).
![Page 18: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/18.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Dynamical systems
On [0,1]× [0,1] there is a piecewise linear function thatsimulates the transition function of an arbritrary TM (Moore1991)On R there is an analytic closed form function thatsimulates the transition function of an arbritrary TM (Mooreand Koiran 1996)3-dimensional piecewise constant differential flows onbounded domains simulate the dynamics of an arbritraryTM (Asarin, Maler and Pnueli 1995)On R3 there is a continuous flow y ′ = f (t , y) that simulatesthe dynamics of an arbritrary TM (Branicky 1996)
It is conjectured that no analytic map on a compact,finite-dimensional space can simulate a Turing machinethrough a reasonable encoding (Moore 1998).
![Page 19: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/19.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Dynamical systems
On [0,1]× [0,1] there is a piecewise linear function thatsimulates the transition function of an arbritrary TM (Moore1991)On R there is an analytic closed form function thatsimulates the transition function of an arbritrary TM (Mooreand Koiran 1996)3-dimensional piecewise constant differential flows onbounded domains simulate the dynamics of an arbritraryTM (Asarin, Maler and Pnueli 1995)On R3 there is a continuous flow y ′ = f (t , y) that simulatesthe dynamics of an arbritrary TM (Branicky 1996)It is conjectured that no analytic map on a compact,finite-dimensional space can simulate a Turing machinethrough a reasonable encoding (Moore 1998).
![Page 20: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/20.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Polynomial initial value problems
Analytic solutions;
Common: they define most of the usual mathematicalfunctions, in particular the “elementary functions” ofAnalysis;Widely used: e.g. Lorenz, Lokta-Volterra, or Van der Polequations;Challenging: many open questions;They satisfy an elimination result: Given y ′ = f (y , t) andy(0) = x , with the components of f being compositions ofpolynomials and PIVP functions, y is given by the firstcomponents of some PIVP function (Graça, Buescu andCampagnolo 2009).
![Page 21: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/21.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Polynomial initial value problems
Analytic solutions;Common: they define most of the usual mathematicalfunctions, in particular the “elementary functions” ofAnalysis;
Widely used: e.g. Lorenz, Lokta-Volterra, or Van der Polequations;Challenging: many open questions;They satisfy an elimination result: Given y ′ = f (y , t) andy(0) = x , with the components of f being compositions ofpolynomials and PIVP functions, y is given by the firstcomponents of some PIVP function (Graça, Buescu andCampagnolo 2009).
![Page 22: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/22.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Polynomial initial value problems
Analytic solutions;Common: they define most of the usual mathematicalfunctions, in particular the “elementary functions” ofAnalysis;Widely used: e.g. Lorenz, Lokta-Volterra, or Van der Polequations;
Challenging: many open questions;They satisfy an elimination result: Given y ′ = f (y , t) andy(0) = x , with the components of f being compositions ofpolynomials and PIVP functions, y is given by the firstcomponents of some PIVP function (Graça, Buescu andCampagnolo 2009).
![Page 23: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/23.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Polynomial initial value problems
Analytic solutions;Common: they define most of the usual mathematicalfunctions, in particular the “elementary functions” ofAnalysis;Widely used: e.g. Lorenz, Lokta-Volterra, or Van der Polequations;Challenging: many open questions;
They satisfy an elimination result: Given y ′ = f (y , t) andy(0) = x , with the components of f being compositions ofpolynomials and PIVP functions, y is given by the firstcomponents of some PIVP function (Graça, Buescu andCampagnolo 2009).
![Page 24: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/24.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Polynomial initial value problems
Analytic solutions;Common: they define most of the usual mathematicalfunctions, in particular the “elementary functions” ofAnalysis;Widely used: e.g. Lorenz, Lokta-Volterra, or Van der Polequations;Challenging: many open questions;They satisfy an elimination result: Given y ′ = f (y , t) andy(0) = x , with the components of f being compositions ofpolynomials and PIVP functions, y is given by the firstcomponents of some PIVP function (Graça, Buescu andCampagnolo 2009).
![Page 25: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/25.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Polynomial initial value problems
Some particular questions for PIVPs:
Are PIVP functions computable?
Is the domain of the solution computable?Is it even decidable if the domain is bounded?Is the reachability problem for PIVPs decidable?
![Page 26: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/26.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Polynomial initial value problems
Some particular questions for PIVPs:
Are PIVP functions computable?Is the domain of the solution computable?
Is it even decidable if the domain is bounded?Is the reachability problem for PIVPs decidable?
![Page 27: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/27.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Polynomial initial value problems
Some particular questions for PIVPs:
Are PIVP functions computable?Is the domain of the solution computable?Is it even decidable if the domain is bounded?
Is the reachability problem for PIVPs decidable?
![Page 28: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/28.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Polynomial initial value problems
Some particular questions for PIVPs:
Are PIVP functions computable?Is the domain of the solution computable?Is it even decidable if the domain is bounded?Is the reachability problem for PIVPs decidable?
![Page 29: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/29.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Overview
Unbounded domain Suspensionin Rn
yt+1 = f (yt ) −→ y ′ = p(t , y)
Encoding computationin Nn of y
TM
TM: Turing machineDiscrete dynamical system:y0 = x , yt+1 = f (yt ), x ∈ Nn, f : Nn → Nn, t ∈ NContinuous dynamical system:y(0) = x , y ′ = p(t , y), x ∈ Rn, p : Rn+1 → Rn, t ∈ R+
0
![Page 30: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/30.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Computability of solutions
Definition (PIVP function with parameters in S)y (or yi ) is a PIVP function with parameters in S if y is thesolution of a polynomial initial value problem y ′ = p(y , t),y(0) = y0 where the coefficients of p and the components of y0are in S.
Theorem (follows from Graça, Zhong and Buescu, 2007)
Let y : (α, β) ⊂ R→ Rk be a PIVP function with computableparameters. Then, y is computable on (α, β).
![Page 31: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/31.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Preliminaries
Computability of solutions
Definition (PIVP function with parameters in S)y (or yi ) is a PIVP function with parameters in S if y is thesolution of a polynomial initial value problem y ′ = p(y , t),y(0) = y0 where the coefficients of p and the components of y0are in S.
Theorem (follows from Graça, Zhong and Buescu, 2007)
Let y : (α, β) ⊂ R→ Rk be a PIVP function with computableparameters. Then, y is computable on (α, β).
![Page 32: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/32.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Encoding TMs
Encoding Turing machine as a discrete dynamical system onN3:
(ω,q,h)ψ−→ x ∈ N3
δ ↓ ↓ f
(ω′,q′,h′)ψ−→ f (x) ∈ N3
If the tape contents is
. . . 000a−p . . . a−1a0a1 . . . an000 . . .
ai ∈ 1, . . . ,9, q ∈ 1, . . . ,m, then
ψ :
x1 = a0 + a1 10 + · · ·+ an 10n
x2 = a−1 + a−2 10 + · · ·+ a−p 10p−1
x3 = q
![Page 33: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/33.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Extension
DefinitionThe map Ω : Rm → Rm is a robust extension of the mapω : Nm → Nm if there are δin, δev , δout ∈ (0, 1
2) such that
||n0 − x0||∞ ≤ δin, ||Ω− Ω||∞ ≤ δev
implies that||ω[k ](n0)− Ω
[k ](x0)||∞ ≤ δout
for all k ∈ N
![Page 34: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/34.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Let Q[π] := bnπn + · · ·+ b1π + b0 ∈ R |b0, . . . ,bn ∈ Q.
Theorem (Graça, Campagnolo and Buescu, 2008)
The transition function ω : N3 → N3 of a Turing machine (underthe encoding ψ) admits a robust extension Ω : R3 → R3.Ω can be chosen to be a composition of polynomials withcoefficients in Q[π] and PIVP functions with parameters in Q[π].
Sketch of the proof.Use a trigonometric interpolation of x mod 10 to obtain a0from x1. Use polynomial interpolations to represent the nextstate, the next symbol, the direction of the move of the headand update x1 and x2. Use error contracting functions(compositions of polynomials and trigonometric functions) tokeep the error within, say, 1
4 of the exact encoding.
![Page 35: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/35.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Let Q[π] := bnπn + · · ·+ b1π + b0 ∈ R |b0, . . . ,bn ∈ Q.
Theorem (Graça, Campagnolo and Buescu, 2008)
The transition function ω : N3 → N3 of a Turing machine (underthe encoding ψ) admits a robust extension Ω : R3 → R3.Ω can be chosen to be a composition of polynomials withcoefficients in Q[π] and PIVP functions with parameters in Q[π].
Sketch of the proof.Use a trigonometric interpolation of x mod 10 to obtain a0from x1. Use polynomial interpolations to represent the nextstate, the next symbol, the direction of the move of the headand update x1 and x2. Use error contracting functions(compositions of polynomials and trigonometric functions) tokeep the error within, say, 1
4 of the exact encoding.
![Page 36: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/36.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Suspension
Definition (Robust suspension)The solution φ of y ′ = f (t , y), y(0) = x ∈ Nm
is a robust suspension of the map ω : Nm → Nm if there areδin, δev , δout , δtime ∈ (0, 1
2) such that
||x − y0||∞ ≤ δin, ||f − f ||∞ ≤ δev , ||t − k ||∞ ≤ δtime
implies that the solution φ of y ′ = f (t , y), y(0) = y0
satisfies||ω[k ](x)− φ(t)||∞ ≤ δout for all k ∈ N
.
![Page 37: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/37.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Construction of the suspension: Branicky’s clocks
The solution ofy ′ = c (b − y)3φ(t),
where φ(t) > 0, approaches at t = 1 the “target” b witharbitrary precision depending on c, irrespectively of theinitial condition at t = 0.
The following system of polynomial ODEs allows to iteratethe map Ω
z ′1 = c1(Ω(r(z2))− z1)3 θ(sin 2πt)z ′2 = c2(r(z1)− z2)3 θ(− sin 2πt)
z1(0) = x0z2(0) = x0
where r is a rounding function and θ = H(x) ∗ xk .
![Page 38: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/38.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Construction of the suspension: Branicky’s clocks
The solution ofy ′ = c (b − y)3φ(t),
where φ(t) > 0, approaches at t = 1 the “target” b witharbitrary precision depending on c, irrespectively of theinitial condition at t = 0.The following system of polynomial ODEs allows to iteratethe map Ω
z ′1 = c1(Ω(r(z2))− z1)3 θ(sin 2πt)z ′2 = c2(r(z1)− z2)3 θ(− sin 2πt)
z1(0) = x0z2(0) = x0
where r is a rounding function and θ = H(x) ∗ xk .
![Page 39: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/39.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Construction of the suspension: example
Figure: Suspension of the iteration of f (n) = 2n with ODEs
![Page 40: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/40.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Construction of the suspension: removing the non-PIVP functions
Consider the perturbed version of the ODE y ′ = c (b − y)3φ(t):
y ′ = c (b(t)− y)3φ(t) + e(t)
where |b(t)− b| ≤ ρ and |e(t)| ≤ δ.
LemmaIf |y(1)− b| ≤ γ then |y(1)− b| ≤ γ + ρ+ δ/2.
Again, this allows to replace the rounding function r and thecontrol functions θ(sin 2πt) by appropriate PIVP functions andre-write the suspension of Ω as a PIVP.
![Page 41: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/41.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Construction of the suspension: removing the non-PIVP functions
Consider the perturbed version of the ODE y ′ = c (b − y)3φ(t):
y ′ = c (b(t)− y)3φ(t) + e(t)
where |b(t)− b| ≤ ρ and |e(t)| ≤ δ.
LemmaIf |y(1)− b| ≤ γ then |y(1)− b| ≤ γ + ρ+ δ/2.
Again, this allows to replace the rounding function r and thecontrol functions θ(sin 2πt) by appropriate PIVP functions andre-write the suspension of Ω as a PIVP.
![Page 42: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/42.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Construction of the suspension: removing the non-PIVP functions
Consider the perturbed version of the ODE y ′ = c (b − y)3φ(t):
y ′ = c (b(t)− y)3φ(t) + e(t)
where |b(t)− b| ≤ ρ and |e(t)| ≤ δ.
LemmaIf |y(1)− b| ≤ γ then |y(1)− b| ≤ γ + ρ+ δ/2.
Again, this allows to replace the rounding function r and thecontrol functions θ(sin 2πt) by appropriate PIVP functions andre-write the suspension of Ω as a PIVP.
![Page 43: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/43.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Suspension for discrete dynamical systems with PIVP functions
Theorem (Graça, Buescu and Campagnolo, 2009)If the map ω : Nm → Nm admits a robust extensionΩ : Rm → Rm whose components are compositions ofpolynomials and PIVP functions with parameters in Q[π], thenω admits a robust suspension φ which a PIVP function withparameters in Q[π].
Corollary
The transition function ω : N3 → N3 of a Turing machine (underthe encoding ψ) admits a robust suspension φ. Moreover φ is aPIVP function with parameters in Q[π].
![Page 44: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/44.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Simulation of Turing machines
Suspension for discrete dynamical systems with PIVP functions
Theorem (Graça, Buescu and Campagnolo, 2009)If the map ω : Nm → Nm admits a robust extensionΩ : Rm → Rm whose components are compositions ofpolynomials and PIVP functions with parameters in Q[π], thenω admits a robust suspension φ which a PIVP function withparameters in Q[π].
Corollary
The transition function ω : N3 → N3 of a Turing machine (underthe encoding ψ) admits a robust suspension φ. Moreover φ is aPIVP function with parameters in Q[π].
![Page 45: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/45.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Applications
Reachability
CorollaryThe following problem is undecidable:
Given a vector of polynomial p : Rn+1 → Rn withcoefficients in Q[π], y0 ∈ Q×Qn, and an open set A inRn decide if the solution of
y ′ = p(t , y), y(0) = y0
crosses A.
![Page 46: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/46.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Applications
Boundedness of the maximal interval of existence
Note: the maximal interval of the PIVP
y ′ = α(y2 − 1)t , y(0) = 3
is bounded for α > 0 and unbounded for α ≤ 0.
Unless one can decide α 6= 0 this gives rise to trivialundecidability results.
Let’s restrict the parameters of the PIVP to Q[π], which is acomparable set, i.e., given α, β ∈ Q[π] we can decide if α = βand α < β.
![Page 47: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/47.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Applications
Boundedness of the maximal interval of existence
Note: the maximal interval of the PIVP
y ′ = α(y2 − 1)t , y(0) = 3
is bounded for α > 0 and unbounded for α ≤ 0.
Unless one can decide α 6= 0 this gives rise to trivialundecidability results.
Let’s restrict the parameters of the PIVP to Q[π], which is acomparable set, i.e., given α, β ∈ Q[π] we can decide if α = βand α < β.
![Page 48: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/48.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Applications
Theorem (Graça, Buescu and Campagnolo, 2009)The following problem is undecidable:
Given a vector of polynomial p : Rn+1 → Rn withcoefficients in Q[π] and y0 ∈ Qn, decide if the maximalinterval of existence of the solution ofy ′ = p(t , y), y(0) = y0 is bounded.
Sketch of the proof.Let xq be the component of the PIVP that encodes the state inthe suspension of a TM M. We set M s.t. it halts iff xq reachesm. Consider the system (equivalent to a PIVP)
z ′1 = xq − (m − 12
), z2 =1z1, z1(0) = z2(0) = −1.
If M halts then z ′1 becomes eventually larger than, say, 1/8 andz2 blows up. Otherwise, z1 always decreases and the solutionis defined everywhere.
![Page 49: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/49.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Applications
Theorem (Graça, Buescu and Campagnolo, 2009)The following problem is undecidable:
Given a vector of polynomial p : Rn+1 → Rn withcoefficients in Q[π] and y0 ∈ Qn, decide if the maximalinterval of existence of the solution ofy ′ = p(t , y), y(0) = y0 is bounded.
Sketch of the proof.Let xq be the component of the PIVP that encodes the state inthe suspension of a TM M. We set M s.t. it halts iff xq reachesm. Consider the system (equivalent to a PIVP)
z ′1 = xq − (m − 12
), z2 =1z1, z1(0) = z2(0) = −1.
If M halts then z ′1 becomes eventually larger than, say, 1/8 andz2 blows up. Otherwise, z1 always decreases and the solutionis defined everywhere.
![Page 50: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/50.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Summary and questions (part 1)
There are PIVPs able to simulate arbitrary Turingmachines (on unbounded domains);
Although the solutions of PIVPs with computableparameters are computable, several properties of thosedynamical systems are undecidable, even if theparameters of the system are in Q[π].PIVPs, which are a well known model of physicalphenomena, are also a robust yet powerful model ofcontinuous time computation.Can we get rid of π and obtain the result aboutsuspensions for PIVPs with parameters in Q?Can we do a more genuine suspension of discretedynamical systems, which doesn’t rely on “clocks”?
![Page 51: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/51.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Summary and questions (part 1)
There are PIVPs able to simulate arbitrary Turingmachines (on unbounded domains);Although the solutions of PIVPs with computableparameters are computable, several properties of thosedynamical systems are undecidable, even if theparameters of the system are in Q[π].
PIVPs, which are a well known model of physicalphenomena, are also a robust yet powerful model ofcontinuous time computation.Can we get rid of π and obtain the result aboutsuspensions for PIVPs with parameters in Q?Can we do a more genuine suspension of discretedynamical systems, which doesn’t rely on “clocks”?
![Page 52: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/52.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Summary and questions (part 1)
There are PIVPs able to simulate arbitrary Turingmachines (on unbounded domains);Although the solutions of PIVPs with computableparameters are computable, several properties of thosedynamical systems are undecidable, even if theparameters of the system are in Q[π].PIVPs, which are a well known model of physicalphenomena, are also a robust yet powerful model ofcontinuous time computation.
Can we get rid of π and obtain the result aboutsuspensions for PIVPs with parameters in Q?Can we do a more genuine suspension of discretedynamical systems, which doesn’t rely on “clocks”?
![Page 53: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/53.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Summary and questions (part 1)
There are PIVPs able to simulate arbitrary Turingmachines (on unbounded domains);Although the solutions of PIVPs with computableparameters are computable, several properties of thosedynamical systems are undecidable, even if theparameters of the system are in Q[π].PIVPs, which are a well known model of physicalphenomena, are also a robust yet powerful model ofcontinuous time computation.Can we get rid of π and obtain the result aboutsuspensions for PIVPs with parameters in Q?
Can we do a more genuine suspension of discretedynamical systems, which doesn’t rely on “clocks”?
![Page 54: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/54.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Summary and questions (part 1)
There are PIVPs able to simulate arbitrary Turingmachines (on unbounded domains);Although the solutions of PIVPs with computableparameters are computable, several properties of thosedynamical systems are undecidable, even if theparameters of the system are in Q[π].PIVPs, which are a well known model of physicalphenomena, are also a robust yet powerful model ofcontinuous time computation.Can we get rid of π and obtain the result aboutsuspensions for PIVPs with parameters in Q?Can we do a more genuine suspension of discretedynamical systems, which doesn’t rely on “clocks”?
![Page 55: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/55.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Motivation
Is f :⊂ R→ R computable?
Several notions of computability for real functions:
Turing machine approach: Computable AnalysisGPAC functionsBSS machines...
![Page 56: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/56.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Computable analysis
Computable analysis
Since real numbers and many other objects studied inanalysis are “infinite” objects containing an “infiniteamount of information”, one has to approximate themby “finite” objects containing only a “finite amount ofinformation” and to perform the actual computationson these finite objects.(Brattka et al., A Tutorial on Computable Analysis,2008)
For instance, a real number is computable if there is a Turingmachine with no input that outputs a binary expansion of thatnumber (note: the output tape is one-way).
![Page 57: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/57.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Computable analysis
Computable analysis
Since real numbers and many other objects studied inanalysis are “infinite” objects containing an “infiniteamount of information”, one has to approximate themby “finite” objects containing only a “finite amount ofinformation” and to perform the actual computationson these finite objects.(Brattka et al., A Tutorial on Computable Analysis,2008)
For instance, a real number is computable if there is a Turingmachine with no input that outputs a binary expansion of thatnumber (note: the output tape is one-way).
![Page 58: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/58.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Computable analysis
Computable reals
Definition (Cauchy representation)A sequence rn of rationals is a ρ-name of a real number x ifthere exists three functions a,b, c from N to N such that for alln ∈ N
rn = (−1)a(n) b(n)
c(n) + 1and |rn − x | ≤ 2−n.
Definition (Computable real number)x ∈ R is computable if it has a computable ρ-name, i.e., if a, band c are computable.
![Page 59: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/59.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Computable analysis
Computable reals
Definition (Cauchy representation)A sequence rn of rationals is a ρ-name of a real number x ifthere exists three functions a,b, c from N to N such that for alln ∈ N
rn = (−1)a(n) b(n)
c(n) + 1and |rn − x | ≤ 2−n.
Definition (Computable real number)x ∈ R is computable if it has a computable ρ-name, i.e., if a, band c are computable.
![Page 60: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/60.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Computable analysis
Computable real functions
M is an oracle Turing machine if, at any step of the computationof M using oracle φ : N→ Nk , M is allowed to query the valueof φ(n) for any n. (Below, φ is a ρ-name for x .)
Definition (Computable function)A function f : D ⊂ Rm → Rp is computable if there is an oracleTuring machine such that for any accuracy n and any ρ-namefor x ∈ D given as oracle, computes a rational vector rsatisfying ||r − f (x)|| ≤ 2−n.
In other words, the machine produces a rapidly convergingrational sequence with limit f (x) from a rapidly convergingrational sequence with limit x .
DefinitionC(R) denotes the set of computable functions.
![Page 61: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/61.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Computable analysis
Computable real functions
M is an oracle Turing machine if, at any step of the computationof M using oracle φ : N→ Nk , M is allowed to query the valueof φ(n) for any n. (Below, φ is a ρ-name for x .)
Definition (Computable function)A function f : D ⊂ Rm → Rp is computable if there is an oracleTuring machine such that for any accuracy n and any ρ-namefor x ∈ D given as oracle, computes a rational vector rsatisfying ||r − f (x)|| ≤ 2−n.
In other words, the machine produces a rapidly convergingrational sequence with limit f (x) from a rapidly convergingrational sequence with limit x .
DefinitionC(R) denotes the set of computable functions.
![Page 62: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/62.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Computable analysis
Computable real functions
M is an oracle Turing machine if, at any step of the computationof M using oracle φ : N→ Nk , M is allowed to query the valueof φ(n) for any n. (Below, φ is a ρ-name for x .)
Definition (Computable function)A function f : D ⊂ Rm → Rp is computable if there is an oracleTuring machine such that for any accuracy n and any ρ-namefor x ∈ D given as oracle, computes a rational vector rsatisfying ||r − f (x)|| ≤ 2−n.
In other words, the machine produces a rapidly convergingrational sequence with limit f (x) from a rapidly convergingrational sequence with limit x .
DefinitionC(R) denotes the set of computable functions.
![Page 63: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/63.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Computable analysis
Equivalent formulation of CA: modulus of continuity
Theorem (see Ko (1991), Corollary 2.14)f : [0,1]→ R is in C(R) iff there exist three computablefunctions m : N→ N, sgn, abs : N3 → N such that:
1 m is a modulus of continuity for f , i.e. for all n ∈ N and allx , y ∈ [0,1],
|x − y | ≤ 2−m(n) =⇒ |f (x)− f (y)| ≤ 2−n
2 For all (j , k) ∈ N2 such that j2k ∈ [0,1], and all n ∈ N,∣∣∣∣(−1)sgn(j,k ,n) abs(j , k ,n)
2n − f(
j2k
)∣∣∣∣ ≤ 2−n.
![Page 64: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/64.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
C(R)?= PIVP functions with parameters in X (X ⊂ R)
⊇ TRUE, if X is a subset of the computable reals(follows from Graça, Zhong and Buescu, 2007).
⊆ FALSE, even if X = R. This follows from the factthat, for instance, Euler’s gamma function
Γ(x) =
∫ +∞
0tx−1e−tdt
is not differentially algebraic (Hölder, 1887).
![Page 65: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/65.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
C(R)?= PIVP functions with parameters in X (X ⊂ R)
⊇ TRUE, if X is a subset of the computable reals(follows from Graça, Zhong and Buescu, 2007).
⊆ FALSE, even if X = R. This follows from the factthat, for instance, Euler’s gamma function
Γ(x) =
∫ +∞
0tx−1e−tdt
is not differentially algebraic (Hölder, 1887).
![Page 66: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/66.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
C(R)?= PIVP functions with parameters in X (X ⊂ R)
⊇ TRUE, if X is a subset of the computable reals(follows from Graça, Zhong and Buescu, 2007).
⊆ FALSE, even if X = R. This follows from the factthat, for instance, Euler’s gamma function
Γ(x) =
∫ +∞
0tx−1e−tdt
is not differentially algebraic (Hölder, 1887).
![Page 67: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/67.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
Approximation of computable real functions with PIVPs
Definitiony(t ; x1, . . . , xk ) is a PIVP function with parameters in S if thereare polynomials p : Rn → Rn with coefficients in S and PIVPfunctions q1, . . . ,qn with parameters in S such that y is thesolution of y ′ = p(y , t) and y(0) = (q1(x), . . . ,qn(x)), where xis some x1, . . . , xk ∈ R.
![Page 68: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/68.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
(A) approximating C(R): two initial conditions x and η ≈ n
Lemma (Bournez, Campagnolo, Graça and Hainry, 2007)Let f : [0,1]→ R be a computable function. Then there exists aPIVP function y(t ; x , η) with parameters in Q[π], wherex ∈ [0,1] and |η − n| ≤ ε < 1/2 (n ∈ N), and some T ≥ 0 s.t.
|y1(t ; x , η)− f (x)| ≤ 2−n for all t ≥ T .
Sketch of the proof (i): a “cascade” of ODEsODE initial conditions depend on value after Ti
y ′ = p1(y , t) x , η y1(t ; x , η) ≈ 2n
y2(t ; x , η) ≈ x 2m(n)
y ′ = p2(y , t) x1 ≈ 2n, x2 ≈ x 2m(n) y3(t ; x1, x2) ≈abs(x 2m(n),m(n),n)
y ′ = p3(y , t) x3 ≈ abs(x 2m(n),m(n),n) y4 ≈ f (x)x4 ≈ 2n
![Page 69: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/69.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
(A) approximating C(R): two initial conditions x and η ≈ n
Lemma (Bournez, Campagnolo, Graça and Hainry, 2007)Let f : [0,1]→ R be a computable function. Then there exists aPIVP function y(t ; x , η) with parameters in Q[π], wherex ∈ [0,1] and |η − n| ≤ ε < 1/2 (n ∈ N), and some T ≥ 0 s.t.
|y1(t ; x , η)− f (x)| ≤ 2−n for all t ≥ T .
Sketch of the proof (i): a “cascade” of ODEsODE initial conditions depend on value after Ti
y ′ = p1(y , t) x , η y1(t ; x , η) ≈ 2n
y2(t ; x , η) ≈ x 2m(n)
y ′ = p2(y , t) x1 ≈ 2n, x2 ≈ x 2m(n) y3(t ; x1, x2) ≈abs(x 2m(n),m(n),n)
y ′ = p3(y , t) x3 ≈ abs(x 2m(n),m(n),n) y4 ≈ f (x)x4 ≈ 2n
![Page 70: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/70.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
Switching dynamics
Sketch of the proof (ii): define y ′ = f (y , t) where:
f (y , t) =
p1(y , t) , t < T1p2(y , t) , T1 + δ < t < T1 + T2p3(y , t) , T1 + T2 + δ < t < T1 + T2 + T3.
Given f1 and f2 (dotted lines) definesome f = φ1
y f1 + φ1y f2 (solid line)
such that
|f (t)− f1(t)| ≤ ε if y(t) ≤ α− 1/4|f (t)− f2(t)| ≤ ε if y(t) ≥ α+ 1/4.
y(t) is the control function:
![Page 71: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/71.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
Switching dynamics
Sketch of the proof (ii): define y ′ = f (y , t) where:
f (y , t) =
p1(y , t) , t < T1p2(y , t) , T1 + δ < t < T1 + T2p3(y , t) , T1 + T2 + δ < t < T1 + T2 + T3.
Given f1 and f2 (dotted lines) definesome f = φ1
y f1 + φ1y f2 (solid line)
such that
|f (t)− f1(t)| ≤ ε if y(t) ≤ α− 1/4|f (t)− f2(t)| ≤ ε if y(t) ≥ α+ 1/4.
y(t) is the control function:
![Page 72: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/72.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
(B) approximating C(R): one initial condition x .
Theorem (Bournez, Campagnolo, Graça and Hainry, 2007)If f : [0,1]→ R is computable, then there is a PIVP functiony(t ; x) with parameters in Q[π] such that:
1 limt→∞ y2(t ; x) = 0;2 for x ∈ [0,1], and t ∈ [0,+∞), |y1(t ; x)− f (x)| ≤ y2(t ; x).
Sketch of the proof.Use the PIVP of the previous lemma with initial conditionx ∈ [0,1] and replace initial condition η by a component y3 with:
y3 ≈ 1 for 0 < t < T1, y3 ≈ 2 for T1 + δ < t < T2 , . . . ,depending on y4;y4 ≈ the state of a TM that indicates that|y1(t ; x)− f (x)| ≤ 2−n.
![Page 73: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/73.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
(B) approximating C(R): one initial condition x .
Theorem (Bournez, Campagnolo, Graça and Hainry, 2007)If f : [0,1]→ R is computable, then there is a PIVP functiony(t ; x) with parameters in Q[π] such that:
1 limt→∞ y2(t ; x) = 0;2 for x ∈ [0,1], and t ∈ [0,+∞), |y1(t ; x)− f (x)| ≤ y2(t ; x).
Sketch of the proof.Use the PIVP of the previous lemma with initial conditionx ∈ [0,1] and replace initial condition η by a component y3 with:
y3 ≈ 1 for 0 < t < T1, y3 ≈ 2 for T1 + δ < t < T2 , . . . ,depending on y4;y4 ≈ the state of a TM that indicates that|y1(t ; x)− f (x)| ≤ 2−n.
![Page 74: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/74.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
Characterization of C(R)
Definition (LIM∗)Let C be a class of functions over R. LIM∗ is an operationwhich takes f1, f2 ∈ C, such that limt→∞ f2(t , x) = 0, and returnsf (x) = limt→+∞ f1(t , x) if |f1(t , x)− f (x)| ≤ f2(t , x) for positive t .C(LIM∗) is the closure of C under LIM∗.
Lemma (Campagnolo and Ojakian 2007)C(R) = C(R)(LIM∗).
Theorem (Reformulation of the main result of Bournez,Campagnolo, Graça and Hainry 2007)On [0,1], C(R) = PIVPQ[π](LIM∗).
![Page 75: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/75.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Approximation of computable real functions with PIVPs
Characterization of C(R)
Definition (LIM∗)Let C be a class of functions over R. LIM∗ is an operationwhich takes f1, f2 ∈ C, such that limt→∞ f2(t , x) = 0, and returnsf (x) = limt→+∞ f1(t , x) if |f1(t , x)− f (x)| ≤ f2(t , x) for positive t .C(LIM∗) is the closure of C under LIM∗.
Lemma (Campagnolo and Ojakian 2007)C(R) = C(R)(LIM∗).
Theorem (Reformulation of the main result of Bournez,Campagnolo, Graça and Hainry 2007)On [0,1], C(R) = PIVPQ[π](LIM∗).
![Page 76: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/76.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Summary and questions (part 2)
Real computable functions can be approximated by PIVPfunctions on compact intervals, which is a further evidenceof their computational completeness.
Is the result true for unbounded intervals?Can we replace LIM∗ by a more natural limit with theapproximation requirement |f1(t , x)− f (x)| ≤ 1
t ?Can we replace the set of parameters Q[π] by Q?
![Page 77: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/77.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Summary and questions (part 2)
Real computable functions can be approximated by PIVPfunctions on compact intervals, which is a further evidenceof their computational completeness.Is the result true for unbounded intervals?
Can we replace LIM∗ by a more natural limit with theapproximation requirement |f1(t , x)− f (x)| ≤ 1
t ?Can we replace the set of parameters Q[π] by Q?
![Page 78: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/78.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Summary and questions (part 2)
Real computable functions can be approximated by PIVPfunctions on compact intervals, which is a further evidenceof their computational completeness.Is the result true for unbounded intervals?Can we replace LIM∗ by a more natural limit with theapproximation requirement |f1(t , x)− f (x)| ≤ 1
t ?
Can we replace the set of parameters Q[π] by Q?
![Page 79: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/79.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Summary and questions (part 2)
Real computable functions can be approximated by PIVPfunctions on compact intervals, which is a further evidenceof their computational completeness.Is the result true for unbounded intervals?Can we replace LIM∗ by a more natural limit with theapproximation requirement |f1(t , x)− f (x)| ≤ 1
t ?Can we replace the set of parameters Q[π] by Q?
![Page 80: Continuous dynamical systems and computation](https://reader031.vdocument.in/reader031/viewer/2022020703/61fb40362e268c58cd5bf35f/html5/thumbnails/80.jpg)
Introduction Polynomial IVPs Polynomial IVPs and CA References
Some references
D.S. Graça, J. Buescu, and M.L. Campagnolo. Computational boundson polynomial differential equations, Applied Mathematics andComputation, 215 (2009) 1375:1385.
D.S. Graça, M.L. Campagnolo, and J. Buescu. Computability withpolynomial differential equations, Advances in Applied Mathematics, 40,(2008) 330:349.
O. Bournez, M. L. Campagnolo, D. S. Graça, and E. Hainry. Polynomialdifferential equations compute all real computable functions oncomputable compact intervals. Journal of Complexity, 23:317–335,2007.
M. L. Campagnolo and K. Ojakian. Using Approximation to RelateComputational Classes over the Reals. In J. Durand-Lose and M.Margenstern (Eds.): MCU 2007, Lecture Notes in Computer Science,4664, (2007) 39:61.