parallel numerical simulation€¦ · derivation and analysis of models • of which type is the...
TRANSCRIPT
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 1 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Fifth SimLab Short Course on
Parallel Numerical Simulation
Belgrade, October 1-7, 2006
Mathematical Models – ODE and PDE
October 2, 2006
Hans-Joachim BungartzDepartment of Computer Science – Chair VTechnische Universität München, Germany
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 2 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
1.1. Principles of Mathematical Modelling
• describe a given problem with some mathematical formalism inorder to
– get a formal and precise description– see fundamental properties due to the abstraction– allow a systematic treatment and, thus, solution
• (mathematical) model: formal description (and usually simplifi-cation) of (some) reality
• bigger or smaller evidence:
– exact natural science and engineering: long tradition (basicconservation laws of continuum mechanics, e.g.)
– economics, game theory, climate modelling: many openquestions (Keynes or not, MinMax or not, chaos or not?)
• to do: both derivation and analysis of models
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 3 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Derivation of a Model 1
• What do you want to model?
– a catalyst’s function or the detailed reactions in it?
• Which are the important quantities for that?
– Example 1: Optimum trajectory of the Space Shuttle – grav-itation of Pluto, gravitation of the Earth?
– Example 2: Prediction of Dow Jones Index tomorrow –statements of Mr. Greenspan, statements of myself?
• How important are they?
– Think of consequences of a neglection!
• What are their relations and interactions?
– qualitative and quantitative aspects
• How can these be (mathematically) described?
– algebraic or differential equations, graphs, automata,. . .
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 4 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Derivation of a Model 2
• description of relations and interactions:
– position, speed, and acceleration of an oscillating pendu-lum?
⇒ ordinary differential equation!
– deformation of a membrane under some load?⇒ partial differential equation!
– initial or boundary conditions of some growth process?⇒ algebraic equations!
– non-negativity of some quantity?⇒ algebraic inequality!
– order of several steps?⇒ graphs!
– state transitions?⇒ automata!
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 5 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Derivation and Analysis of Models
• Of which type is the resulting task?
– Is there a solution (Hamiltonian way in a graph)?– Find a/the solution (flow field around an aircraft)!– Find a/the best solution (shape optimization)!
• What can be said about solution(s) concerning
– their existence?– their uniqueness?– their dependency on the input data?
(well-posed problems: Hadamard 1923; Tikhonov, John; cf.however inverse problems)
• Is the model well-suited for a numerical treatment?
• Is the model derived so far correct?
– validation (experiments)!
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 6 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
What to do with models?
• the analytical approach:
– prove existence and uniqueness formally– construct or find solution(s) formally/directly/analytically– desirable, but almost never possible
• the heuristic approach:
– trial and error, following some (hopefully smart) strategy– useful in discrete problems (travelling salesman etc.)
• the direct numerical approach:
– follow some numerical algorithm and end up with the exactsolution (Simplex algorithm for linear programming)
• the approximative numerical approach:
– approximate/discretize the model equations and end up withsome approximate solution
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 7 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
1.2. Classes of Mathematical Models
• Models can be discrete or continuous:
– discrete models use a discrete/combinatoric description (in-teger numbers, graphs,. . . )
– continuous models use real quantities (real numbers, phys-ical quantities, differential equations,. . . )
– primarily, but not necessarily: discrete models for discretephenomena, continuous models for continuous phenomena
– examples: lattice-gas-automata for fluid flow, continuum me-chanics for traffic flow
• Models can be deterministic or stochastic:
– again no general relation between phenomena and models– roll the die: random phenomenon, stochastic model– crash test: deterministic phenomenon, deterministic model– but what about weather or data traffic through the internet?
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 8 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Discrete Models 1: Scheduling
• n jobs to be done on m machines working in parallel:
– no job simultaneously on more than one machine– no machine working simultaneously on more than one job
• several model parameters:
– characterizing the jobs: subjobs, processing time, earlieststarting time, due time, weighting, cost of delay, . . .
– characterizing machines and processing: identical or dif-ferent machines, order of subjobs important or not, prece-dence relations of jobs
– criteria of optimality: time of completion, delay, idle times
• task: find the best schedule with respect to some objective func-tion (minimizing overall time, e.g.)
• model: graph with disjunctive and conjunctive edges
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 9 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Discrete Models 2: Elections
• m candidates, n voters; model: theory of relations
• task: derive a reasonable collective preference from the differ-ent reasonable individual preferences
• possible choices:
– external dictator: individual preferences do not matter– internal dictator: one individual preference always wins– majority wins: cycles (a > b > c > a) cannot be excluded
– something else democratic — but what is democratic?
* all reasonable individual preferences are allowed
* the collective preference must be reasonable
* everything is possible with unanimity
* no dictator, independence of irrelevant alternativesimpossible if m > 2 and n > 1 (Arrow, 1951)!!
• drawback of democracy or drawback of our model??
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 10 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Discrete Models 3: Event Simulation
• task: model data or job traffic through computer system, opti-mize flow, find and avoid bottlenecks
• model: discrete event simulation, stochastic processes
– elementary queueing system (service stations, waiting pool)– quantities: waiting time, service time, staying time; filling,
capacity, throughput– arrival and completion of jobs: stochastic processes (deter-
ministic D, Markovian M, or general G)– notation: M|M|1 (arrival and completion follow a negative
exponential distribution, one station)– service strategies: FCFS, LCFS, random, round robin, . . .– queueing nets: network of elementary queueing systems
• computer systems: discrete space of states, state transitionsdriven by stochastic processes, Markovian chains or systems,resp.
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 11 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Hierarchy and Multiscale Property
• often to be chosen: scale/level of observation:
– Which resolution is necessary (w.r.t. the model’s accuracy)?– Which resolution can be tackled numerically?
• flow through a cylinder – how many dimensions?
1D: neglect cross-section2D: exploit symmetries3D: full resolution
• electric circuit simulation – spatial resolution or not?
– standard system simulators (SPICE, TITAN): only time, ODEs– parasitic cross effects: take space into account!
• turbulence – which vortices can be neglected?
– significant transport of energy between different scales– direct simulation – Large Eddy Simulation – averaging mod-
els
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 12 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Averaging and Homogenization
• often: coarse-grain phenomena are of interest, but fine-grainphenomena must not be neglected
• try to do some averaging:
– in time: turbulence, molecular dynamics– in space: flow and transport through porous media (a cata-
lyst or soil)
• formal concept: homogenization
– representative elementary volume– scaled reproduction, translation, periodic continuation– limit process of scaling factor– new quantities (effective parameters: porosity, permeabil-
ity)– new equations (porous media: instead of transport equa-
tions now Darcy-Forchheimer equation)
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 13 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
1.3. Continuous Models 1: ODE
• in scientific computing: numerical simulations and, hence, typi-cally continuous models
• two big classes:
– problems with a treatment of space (involving partial differ-ential equations (PDE))
– problems without a treatment of space (involving ordinarydifferential equations (ODE))
• standard example for the latter: population dynamics
– development (growth) of populations,
* either isolated (without external influences)
* or in coexistence (peaceful or hostile) of different species– modelling has a long tradition– classical representative: model of Maltus (1798)
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 14 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Model of Maltus
• one species considered:
– constant birth rate γ per time unit and individual– constant death rate δ per time unit and individual– thus, constant growth rate λ = γ − δ
• development of p(t), the number of individuals:
p(t + ∆t) = p(t) + λp(t)∆t
(growth is proportional to size of population and time)
• this leads to the ODE
p(t) = λp(t)
with solutionp(t) = p0e
λt if p(0) = p0
• note:
– exponential growth or decrease– discrete reality, but continuous model!
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 15 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Model Refinement 1
• Is exponential growth realistic?
– population of the earth between 1700 and 1960: yes!
* growth rate of about 0.02, population doubles in 34.67years
– generally: no!
* limited capacity of the earth, limited resources
* increasing competiton for food, water, or air slows downgrowth
• refinement following Verhulst and others (19. century):
– population tends towards some saturation limit– linear birth and death rates (now per time unit only):
γ(t) = γ0 − γ1p(t) δ(t) = δ0 + δ1p(t) γ0 > δ0 > 0, γ1, δ1 > 0
big population decreases birth rate and increases death rate
– ODE: p(t) = γ(t)− δ(t) = −m · (p(t)− p∞)
– Limit exists if t tends towards infinity:
p(t) = p∞ + (p0 − p∞) · e−m·t
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 16 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Model Refinement 2
• Is Verhulst’s approach realistic?
– Second derivative of p(t) does not change its sign!– S-shape is widespread, however (cf. US population 1790-
1950, for example)
• further refinement in order to obtain S-shape:
p(t) = a · p(t)− b · p2(t), a > b > 0
with solution
p(t) =a · p0
b · p0 + (a− b · p0)e−at
• discussion:
– limit for t towards infinity is a/b
– S-shape for p0 < a/b
– example USA 1790-1950: a = 0.03134, b = 1.5587 · 10−10
– better than our starting point, but still no externalinfluences
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 17 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Model Refinement 3
• Typically, a is much bigger than b:
– The quadratic term gets influence only for really big p(t).– Why quadratic and not cubic? – That’s modelling!– justification: an individual is disturbed proportional to p(t)
• S-shape growth is called logistic
• For that, other ODEs can be used, too (widespread inmodelling of tumours growth, e.g.):
– p(t) = λ(t) · p(t) with some continuous, positive, anddecreasing function lambda (empirically, growth rates haveto decrease)
– or, even more general, p(t) = f(p(t)) · p(t) with somesuitable non-negative, decreasing, and vanishing (forincreasing t) f
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 18 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
More than one Species
• next step of refinement: consider two species P and Q
p(t) = f(p(t), q(t)) · p(t),q(t) = g(p(t), q(t)) · q(t)
f and g defined for positive p and q
• A pair p, q with f(p, q) = g(p, q) = 0 allows for the definition of a
stationary solution(
p(t)q(t)
)=
(pq
)which is called equilibrium, if p > 0, q > 0.
• Is there an equilibrium? If yes, is it attractive?
– theory: a sufficient condition for attractiveness arenegative real parts of the eigenvalues of the Jacobian ofF (p, q) = (f(p, q)p, g(p, q)q) in p, q
– We study two special cases: predator-prey andcompetition characteristics.
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 19 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Competition Characteristics
• The species P and Q do not “eat” each other, but both strugglefor the same resources:
fp(p, q), fq(p, q), gp(p, q), gq(p, g) < 0 for p, q > 0
• Sufficient condition for attractive equilibrium reads
fp(p, q) · gq(p, q)− fq(p, q) · gp(p, q) > 0
(P’s influence on P is bigger than on Q and vice versa)
• simple concrete choice for f and g:
f(p, q) = a1 + a2 · p + a3 · q, g(p, q) = a4 + a5 · p + a6 · q
with (due to our model assumptions)
a1, a4 > 0, a2, a3, a5, a6 < 0, a2 · a6 > a3 · a5
• Attractive equilibrium f(p, q) = g(p, q) = 0, p, q > 0 exists!
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 20 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Predator-Prey Characteristics
• Q is “food” of P, which leads to a different growth behaviour:
fp(p, q), gp(p, q), gq(p, q) < 0, fq(p, q) > 0 for p, q > 0
a2, a5, a6 < 0, a3 > 0
(P, of course, enjoys an increasing population Q)
• Sufficient condition for attractive equilibrium is always fulfilled(i.e., if there is an equilibrium, it is attractive).
• classical representative: model of Volterra and Lotka:
a2 = a6 = 0
– no influence of P on P or Q on Q– There is an attractive equilibrium (though our sufficient
condition is not valid)!
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 21 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
1.4. Continuous Models 2: PDE
• so far: only time as independent variable
• ODE-based population models sometimes too coarse:
– population in the USA during California gold rush in the1850s
– predictions of the UN concerning world population(industrialized countries versus third world)
• therefore: suppose p(x, t) or p(x, y, t) instead of p(t)
– California gold rush: 1D sufficient (east-west)– world population: perhaps 1D (north-south), perhaps 2D
• taking space into account makes models
– more accurate (spatial effects are no longer neglected)– more complicated (analytical solution becomes harder,
numerical solution means a lot of additional work)
• standard example: heat conduction
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 22 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Modelling with PDE
• taking space into account is typical for many problems orphenomena from physics or continuum mechanics:
– fluid mechanics: where will we get a tornado?– structural mechanics: where will be the crack?– process engineering: where is it how hot in the reactor?– electromagnetism: where is which electron density?– geology: where will the earthquake happen?
• more independent variables entail partial derivatives
• we distinguish:
– stationary problems: no time-dependence– unsteady problems: time-dependence (perhaps, but not
necessarily, with a stationary limit for increasing time)
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 23 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Heat Conduction
• central problem of thermodynamics
• let heat affect an object’s boundary – propagation?
– a wire, heated at one end– a metal plate, heated at one side– water cooling the reactor in a nuclear power plant– a room in winter: where to place the heating– a room in summer: effect of direct sunshine– boiling water in a pot on a ceramic hob
• central function of interest: temperature T
T (x, y) or T (x, y, t) or T (x, y, z, t)
• The values of T will depend on the material and its heatconductivity.
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 24 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Modelling Heat Conduction 1
• part 1 of the model: the PDE, indicating the relations ofchanges of T with respect to time and space (3D):
κ · (Txx + Tyy + Tzz) = κ ·(
∂2T
∂x2+
∂2T
∂y2+
∂2T
∂z2
)=
∂T
∂t= Tt
or shortly κ ·∆T = Tt with the Laplace operator ∆
• short derivation (excursion to physics):
– starting point is the basic principle of energy conservation– changes of heat in some part D of our domain are due to
flux in/out D’s surface and to external sources and drainsin D
∂
∂t
∫D
ρcTdV =
∫D
qdV +
∫∂D
k∇T · ~n dS
– density ρ, specific heat c, external term q, heat conductivityk, outer normal vector ~n, volume/surface element dV/dS
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 25 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Modelling Heat Conduction 2
• derivation of the heat equation (continued):
– transform the above equation according to Gauß’ theorem:∫D
(ρcTt − q − k∆T )dV = 0
– This holds for an arbitrary part D of our domain. Hence,the integrand must vanish:
Tt = κ∆T +q
ρc, κ =
k
ρc
– κ > 0 is called the thermal diffusion coefficient (since theLaplace operator stands for a (heat) diffusion process)
– For vanishing external influence q = 0, we get (and, thus,have derived) the famous heat equation:
Tt = κ∆T
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 26 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Modelling Heat Conduction 3
• part 2 of the model: the PDE needs boundary orinitial-boundary conditions to provide a unique solution:
– Dirichlet boundary conditions: fix T on (part of) theboundary
T (x, y, z) = ϕ(x, y, z)
– Neumann boundary conditions: fix T ’s normal derivativeon (part of) the boundary:
∂T
∂n(x, y, z) = ϕ(x, y, z)
– pure Dirichlet and mixtures are allowed, pure Neumannb.c. do not lead to a unique solution (with T solvesT + const. the PDE, too)
– in case of time-dependence: initial conditions for t = 0
• in case of no time-dependence: Laplace equation
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 27 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Modelling Heat Conduction 4
• meaning of boundary conditions:
– Dirichlet: the temperature T is prescribed itself along(part of) the boundary (some defined heating or cooling)
– Neumann: the temperature flux through (part of) theboundary is prescribed (if vanishing: complete isolation,no orthogonal transport of heat into or out of the domain
• analytical solutions:
– In simple (1D) configurations, solutions can be givenexplicitly via separation of variables (Fourier’s method).We will discuss these in the exercises.
– The heat equation is a simple case of a PDE, wheregeneral statements concerning existence and uniquenessof solutions are possible. Often, such theorems can not beproven.
Principles of . . .
Classes of . . .
Continuous Models 1: . . .
Continuous Models 2: . . .
Page 28 of 28
Mathematical Models – ODE and PDEHans-Joachim Bungartz
Types of PDE
• The heat equation is a linear PDE of second order:d∑
i,j=1
ai,j(~x) · uxi,xj(~x) +
d∑i=1
ai(~x) · uxi(~x) + a(~x) · u(~x) = f(~x)
• three types are distinguished:
– elliptic PDE: the matrix A of the ai,j is positive or negativedefinite
– parabolic PDE: one eigenvalue of A is zero, the othershave the same sign, and the rank of A together with thevector of the ai is full (d)
– hyperbolic PDE: A has 1 positive and d− 1 negativeeigenvalues or vice versa.
• examples:
– elliptic: Laplace equation ∆u = 0
– parabolic: heat equation ∆u = ut
– hyperbolic: wave equation ∆u = utt