Distributed Parameter Control of Thermal Fluids

Diana Rubio

Dissertation submitted to the Faculty of the

Virginia Polytechnic Institute and State University
in partial fulfillment of the requirements for the degree of

Doctor of Philosophy
in
Mathematics

John A. Burns, Chair

Terry L. Herdman
Eugene M. Cliff
Tao Lin
Jeffrey T. Borggaard

April 21, 1997

Blacksburg, Virginia

Keywords: Boundary Control, Dynamical Systems, Semigroups, Boussinesq

approximation
Copyright 1997, Aurora Diana Rubio
 Distributed Parameter Control of Thermal Fluids

Diana Rubio

(ABSTRACT)

We consider the problem of controlling a thermal convection flow by feedback. The system
is governed by the Boussinesq approximation of the coupled set of Navier-Stokes and heat
equations. The control is applied through Dirichlet boundary conditions. We concentrate
on a two-dimensional model and use a semidiscrete Galerkin scheme for numerical compu-
tations. We construct both a linear control and a non-linear quadratic control and apply
them to the full non-linear model. First, we test these controllers on a one-mode approx-
imation. The convergence of the numerical scheme is analyzed. We also consider LQR
control for a two-dimensional heat equation.

This work received support from Air Force Office of Scientific Research under grants
F49620-93-1-0280 and F49620-96-1-0329.
Acknowledgments

I would like to express my sincere appreciation to all the people who encouraged and
supported me throughout my academic endeavor. I wish to thank those with whom I have
discussed my work as well as those who were beside me with their friendship during these
years. In particular, my deepest gratitude to my advisor, John A. Burns for his constant
support, contagious enthusiasm and guidance. It was gratifying to work with him. He was
not only my advisor but also a good friend, always there to listen.
I would also like to thank Terry Herdman, Eugene Cliff, Jeff Borggaard and Tao Lin for
serving in my committee and all the people at the Interdisciplinary Center for Applied
Mathematics (ICAM) for their friendship and kindness.
Finally, I would like to thank my family for their encouragement, in particular my father
who would have be happy to share this moment with me. Thanks also to my friends, those
from Argentina who follow each of my steps and those who I met at Virginia Tech. They
gave me the indispensable support to accomplish this work.

iii
Contents

1 Introduction 1

1.1 Review of Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Outline of the Dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 The Thermal Convection Control Problem 4

2.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.2 Equations of Motion and Navier-Stokes Equation . . . . . . . . . . . . . . 6

2.2.1 The Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . 6

2.2.2 The Continuity Equation . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2.3 Euler’s Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2.4 The Viscosity Tensor . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.3 The Boussinesq approximation . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.4 The Equations for the Convection Loop . . . . . . . . . . . . . . . . . . . . 11

2.4.1 The Divergence Free Condition . . . . . . . . . . . . . . . . . . . . 12

2.4.2 An Integro-Differential Equation . . . . . . . . . . . . . . . . . . . . 12

2.4.3 The Heat Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

iv
 2.4.4 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.4.5 The PDE Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3 Abstract and Weak Formulations 17

3.1 The Linear State Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.2 The Linear Input Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3.3 The Nonlinear State Operator . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.4 Abstract Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.5 Well Posedness of the Open Loop Problem . . . . . . . . . . . . . . . . . . 25

3.6 A Variational Form of the Problem . . . . . . . . . . . . . . . . . . . . . . 26

4 Control of the Boussinesq Equations 32

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4.2 The Linear Quadratic Regulator Problem . . . . . . . . . . . . . . . . . . . 33

4.3 Krener’s Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4.4 Application to a One Mode Approximation . . . . . . . . . . . . . . . . . . 38

4.4.1 Derivation of the Equations . . . . . . . . . . . . . . . . . . . . . . 38

4.4.2 The Controlled System . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.4.3 LQR Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.4.4 Krener’s Nonlinear Control . . . . . . . . . . . . . . . . . . . . . . . 40

4.4.5 A Modification of Krener’s Controller . . . . . . . . . . . . . . . . . 42

4.4.6 An Ad Hoc Controller . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.4.7 A Comparison of the Four Controllers . . . . . . . . . . . . . . . . 42

v
 4.4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

5 A Finite Element Approximation for the Boussinesq Model 47

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

5.2 Discretization in Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

5.2.1 Approximating the Velocity . . . . . . . . . . . . . . . . . . . . . . 48

5.2.2 Approximating the Temperature . . . . . . . . . . . . . . . . . . . 49

5.2.3 A Galerkin Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

5.2.4 The Reduced System . . . . . . . . . . . . . . . . . . . . . . . . . . 55

5.2.5 The Finite Element Model . . . . . . . . . . . . . . . . . . . . . . . 63

6 Numerical Results for the Finite Element Approximation 65

6.1 LQR problem for the Finite Element Approximation . . . . . . . . . . . . 65

6.2 Krener’s Control for the Finite Element Approximation . . . . . . . . . . . 66

6.3 Comparison of the LQR and Krener Controls . . . . . . . . . . . . . . . . 67

6.3.1 Dimension of the Pipe and Fluid Properties . . . . . . . . . . . . . 67

6.3.2 State and Control Weight Operators for LQR Problem . . . . . . . 67

6.3.3 Mesh Dimensions and Plots Interpretation . . . . . . . . . . . . . . 68

6.3.4 Approximating Solutions and Controls . . . . . . . . . . . . . . . . 69

6.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

7 Convergence Issues 89

7.1 Comments on Convergence of the Finite Element Scheme . . . . . . . . . . 90

7.2 The One-Dimensional Heat Equation Scheme . . . . . . . . . . . . . . . . 92

vi
 7.3 Functional Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

7.4 Heat Equation Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

7.4.1 Description of the Problem . . . . . . . . . . . . . . . . . . . . . . . 104

7.4.2 Functional Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

Conclusions 112

Bibliography 113

vii
List of Figures

2.1 Description of the thermal convection loop . . . . . . . . . . . . . . . . . . 5

2.2 Decomposition of the gravity acceleration in radial and angular directions

at a point (r, ϕ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4.1 Lorenz Attractor. Parameters: P = 4, R = 50. Equilibrium points (7, 7, −1),

(−7, −7, −1) and (0, 0, −50). Initial Point x0 = (−3, 9, −8). . . . . . . . . . 44

4.2 Solutions of the nonlinear closed-loop system feeding by four feedback con-
trollers. Parameter values: P = 4, R = 50. Equilibrium Point xe =
(7, 7, −1). Initial Point x0 = (−3, 9, −8). . . . . . . . . . . . . . . . . . . . 45

4.3 Components of the feedback controllers. Parameter values: P = 4, R =

50. Equilibrium Point to be Stabilize xe = (7, 7, −1). Initial Point x0 =
(−3, 9, −8). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

5.1 Polar Coordinates Triangulation Mapping . . . . . . . . . . . . . . . . . . 48

5.2 Elements and Nodes in a Rectangular Grid. . . . . . . . . . . . . . . . . . 51

5.3 Map Gk : Qref → Qk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

6.1 Mesh for nr = 3, nϕ = 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

6.2 Example 1. Open Loop Velocity. Final time=600 sec. . . . . . . . . . . . . 69

6.3 Example 1. Closed Loop Velocity with LQR control. Final time=600 sec. . 70

6.4 Example 1. Closed Loop Velocity with Krener’s control. Final time=600 sec. 70

viii
6.5 Example 1. Fixed radius r = 1.23031in. Final time=600 sec. . . . . . . . . 72

6.6 Example 1. Fixed radius r = 1.26312in. Final time=600 sec. . . . . . . . . 73

6.7 Example 1. Fluid temperatures at r = 1.23031in. Final time=60 sec. . . . 74

6.8 Example 1. Fluid temperatures at r = 1.26312in. Final time=60 sec. . . . 75

6.9 Example 1. Linear and Non-linear Controllers applied on the inner wall.
Final time=60 sec. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

6.10 Example 1. Linear and Non-linear Controllers applied on the outer wall.
Final time=60 sec. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

6.11 Example 2. Open Loop Velocity. Final time=600 sec. . . . . . . . . . . . 78

6.12 Example 2 (Controlling Outer Wall Temperature). Closed Loop Velocity

using LQR control. Final time=600 sec. . . . . . . . . . . . . . . . . . . . 78

6.13 Example 2 (Controlling Outer Wall Temperature). Closed Loop Velocity

using quadratic control. Final time=600 sec. . . . . . . . . . . . . . . . . 79

6.14 Example 2 (Controlling Outer Wall Temperature). Fluid temperatures at

r = 1.23031in. Final time=120 sec. . . . . . . . . . . . . . . . . . . . . . . 80

6.15 Example 2 (Controlling Outer Wall Temperature). Fluid temperatures at

r = 1.26312in. Final time=120 sec. . . . . . . . . . . . . . . . . . . . . . . 81

6.16 Example 2. Linear and Non-linear Controls on the Outer Wall. Final
time=120 sec. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

6.17 Example 3. Open Loop Velocity. Final time=120 sec. . . . . . . . . . . . . 83

6.18 Example 3 (Controlling Outer Wall Temperature). Closed Loop Velocity

using LQR control. Final time=120 sec. . . . . . . . . . . . . . . . . . . . 83

6.19 Example 3 (Controlling Outer Wall Temperature). Closed Loop Velocity

using non-linear control. Final time=120 sec. . . . . . . . . . . . . . . . . . 84

6.20 Example 3 (Controlling Outer Wall Temperature). Fluid temperatures at

r = 1.23031in. Final time=120 sec. . . . . . . . . . . . . . . . . . . . . . . 85

ix
6.21 Example 3 (Controlling Outer Wall Temperature). Fluid temperatures at
r = 1.26312in. Final time=120 sec. . . . . . . . . . . . . . . . . . . . . . . 86

6.22 Example 3. Linear and Non-linear Controls on the Outer Wall. Final
time=120 sec. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

7.1 Functional Gains hhv (π, r), r ∈ (R1 , R2 ), nr = 6, 8, 10, 12, nϕ = 5nr . . . . . 98

7.2 Functional Gains hhv (π, r), r ∈ (R1 , R2 ), nr = 6, 10, 14, 18, 22, nϕ = 30. . . 99

7.3 Functional Gains hhv (π, r), r ∈ (R1 , R2 ), nr = 15, nϕ =10,15,20,25,30,35,40. 99

7.4 Functional Gains hhT (π, r, π), r ∈ (R1 , R2 ), nr = 6, 8, 10, 12, nϕ = 5nr . . . . 100

7.5 Functional Gains hhT (π, r, π), r ∈ (R1 , R2 ), nr = 6, 10, 14, 18, 22, nϕ = 30. . 100

7.6 Functional Gains hhT (π, r, π), r ∈ (R1 , R2 ), nr = 15, nϕ =10,15,20,25,30,35,

40. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

7.7 Functional Gains hhv (ξ, r), r ∈ (R1 , R2 ), nr = 6, 8, 10, 12, nϕ = 5nr . . . . . 102

7.8 Functional Gains hhT (π, r, ϕ), r ∈ (R1 , R2 ), ϕ ∈ (0, 2π), nr = 6, 8, 10, 12,
nϕ = 5nr . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

7.9 Center weighted. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

7.10 Functional Gain hh (0.5, x, y), 0 < x, y, < 1, nx = ny = 12, 16, 20, 24. . . . . 108

7.11 Functional Gain hh (0.5, 0.5, y), 0 < y < 1, nx = ny = 12, 16, 20, 24. . . . . 109

7.12 Functional Gain hh (0.5, x, y), nx = ny = 12, 16, 20, 24. . . . . . . . . . . . 110

7.13 Side weighted. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

x
Chapter 1

Introduction

1.1 Review of Literature

Fluid flow with thermal convection is often used as a starting point for the study of chaotic
motions. In thermal convection processes, fluid flow is created by a buoyancy force due
to thermal differences in the fluid, since changes in temperature are related to changes in
density. When heated the fluid becomes “lighter” and tends to move upward. On the other
hand, under cooling, the fluid tends to move downward. Thermal convection processes
have been used as models for solar energy systems and in the cooling of nuclear reactor
cores. An extensive review is presented in Mertol and Grief [27].

Stability properties of thermal convection systems have been widely discussed. In 1967
Welander [37] presented the first work on stability analysis of a fluid contained in a closed
loop that is heated from below and cooled from above. Also, it was assumed that the
fluid has a uniform temperature over a tube cross-section. In 1975, Creveling, De Paz,
Baladi and Schoenhals [15] presented analytical and experimental stability results for water
flowing inside a toroidal glass loop. They assumed a constant heat flux on the lower half
and a constant wall temperature on the upper half. For the analysis, they considered a one-
dimensional flow where the axial heat conduction and viscous heating are neglected. Mertol
and Grief [27] also presented analytical and experimental results for fluid flow in a toroidal
loop. Conditions similar to Creveling et al. [15] were assumed, where they analyzed one-,
two- and three-dimensional models. Numerical and experimental results have also been
obtained by Singer, Wang and Bau [31]. Joseph [17] considered the problem of stability
in domain of arbitrary shape. Bau and Torrance [4] considered an open convection loop
which is symmetrically heated.

1
 2

Perhaps the most commonly studied problem of thermal convection is the stability of a
Bénard flow, where the fluid is contained between two rigid horizontal plates. For exam-
ple, see the papers by Krishnamurti [21], Foias, Manley and Temam [11], Curry, Herring,
Loncaric and Orszag [8], Goldhirsch, Pelz and Orszag [13], and more recently by Yanagita
and Kaneko [39] where map lattices are used to model the system.

One common model is based on the Boussinesq approximation of the Navier-Stokes and
heat equations. In this model it is assumed that density variations are sufficiently small
and can be neglected except in the buoyancy force. Thus, the incompressible Navier-
Stokes equations govern the fluid motion. Also, a linear relation between the density and
the temperature is assumed.

The Lorenz equations may be viewed as a one mode approximation to the full Boussinesq
system (see [14], [30], [36] and [40]). Non-linear feedback control laws based on the Lorenz
model have been studied in [30], [35] and [36]. In [30] and [35], various linear and non-
linear feedback laws were constructed and compared. In both cases it was shown that a
non-linear controller greatly enhances the performance when applied to the Lorenz system.
However, it is interesting to note that while the behavior of finite-mode approximations
of the Boussinesq equations indicates the existence of two dimensional chaotic solutions,
chaos can also be the product of inadequate spatial resolution in the model (see [8]).
This observation raises interesting control theory problems for the distributed parameter
version of the Boussinesq system as well as for high order finite element approximations of
the corresponding partial differential equations.

In their work, Wang, Singer and Bau [36] promote and prevent chaotic flow by heating and
cooling the pipe. Based on their observations, they proposed a non-linear feedback law
that was tested on the one-mode model. However, the control was constructed based on
experimental observations and not on any control theoretical method. In particular, there
has not been any work that applies specific algorithms (LQR, feedback linearization, etc.)
to this problem. This is one objective of this work.

The fundamental mathematical control problem may be described as a Dirichlet boundary

control problem for a non-linear convection-diffusion system. Most of the current literature
is restricted to linear problems and almost all existing theory is directed towards standard
PDE problems (wave equations, heat equations, etc.). The work by Banks, Smith and
Wang [3] is an exception to this statement. However, their work concentrates on smart
structures and fluid/structures interactions. Here, we are concerned with a heat driven
flow.

The lecture notes by Lasiecka and Triggiani [24] provide a good starting point for the basic
“theory” (see also [5] I & II). However, these references do not deal with the hybrid system
 3

(thermal/fluid) of concern here. Also, non-linearity is seldom considered. We use these

references for results on linear problems and to guide our work. However, we also use some
ideas found in the work of Krener ([19], [20]) to construct non-linear controllers for finite
dimensional approximated systems.

1.2 Outline of the Dissertation

The description of the problem is presented in Chapter 2. The non-linear PDE system that
governs the system is derived by applying the Boussinesq approximation and an initial
boundary control problem is formulated. In Chapter 3, we define a state space model by
defining linear and non-linear state operators and the unbounded input operator. We show
that the non-linear uncontrolled system is well-posed and provide a variational formulation.

In Chapter 4 we consider an LQR control problem and a non-linear control based on

Krener’s method [19], [20]. As a first approach, we consider a one-mode approximation,
where the Lorenz equations govern the system. We apply and compare the controllers
mentioned above with a modification of the Krener controller and an ad hoc controller
introduced by Wang, Singer and Bau [36].

In Chapter 5, a finite element scheme is constructed. Quadratic basis functions are used
for velocity approximations and bilinear basis functions are used for the temperature ap-
proximation.

The approximating non-linear system is coupled with LQR and Krener’s controllers. In
Chapter 6 we present numerical results of the finite element open loop and closed loop
systems for different initial conditions. Also, we consider the case of controlling the tem-
perature of only one wall. An example of this case is also illustrated.

In Chapter 7, we discuss the convergence of the approximating problems. We give a proof

of convergence of the approximating scheme applied to the 1D heat equation. For the
2D thermal convection problem the proof is more complicated. We investigate this issue
numerically by considering the convergence of the approximating gain operator of the LQR
control. Also, the scheme is applied to a 2D heat equation in order to obtain smaller and
more accurate approximating systems.
Chapter 2

The Thermal Convection Control

Problem

2.1 Problem Description

A thermal convection loop consists of a viscous fluid contained in a circular pipe standing
in a vertical plane (see Figure 2.1). When there is a temperature difference between the top
and the bottom of the pipe, the thermal expansion causes a temperature gradient opposite
gravity and the fluid tends to flow. On the other hand, viscosity and thermal diffusivity
resist this motion. When the dissipative terms are overcome by the buoyancy force, fluid
motion is created. This effect is called buoyancy driven free convection.

Experiments shows that, when the difference in temperature between the top and the
bottom of the pipe is large enough, the fluid presents unstable motions which may also be
chaotic (see [15], [27], [31] and [36]).

The balance between the buoyancy force and the dissipative terms is measured by the
dimensionless Rayleigh number Ra , defined by

gβd3δT
Ra = , (2.1)
χν
where g is the gravity acceleration magnitude, β is the thermal expansion coefficient, d is
the diameter of the pipe, δT is the difference between the fluid temperature at the bottom
and the top of the pipe, χ is the fluid’s thermal diffusivity and ν is the kinematic viscosity.
Fluid motion occurs when the Rayleigh number Ra exceed a threshold value Ram , and

4
 5

g
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
R2
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
ϕ
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
R1
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////
/////////////////////////////////////

Figure 2.1: Description of the thermal convection loop

chaotic flow is obtained when Ra > Rac for certain critical value Rac (see [31]).

Here we consider the two-dimensional problem. In particular, we apply temperature control

at the boundary of the loop.

The interior radius of the pipe is denoted by R1 and the exterior radius by R2 . The radial
position of a fluid particle is denoted by r ∈ [R1 , R2 ] and the angle is measured by ϕ
counterclockwise from the horizontal position as shown in Figure 2.1.

We restrict ourselves to the case of a thin pipe where the width of the pipe is small compared
with the interior radius, i.e., R2 − R1 << R1 . In this case, the fluid may be considered as
flowing in a straight pipe of width R2 − R1 , so that the velocity depends only on the radial
coordinate. Moreover, we assume that the fluid flow has circular streamlines so that each
fluid particle flows at a fixed distance from the center of the pipe. Therefore, by denoting
the vector velocity by v, we have

v(t, r, ϕ) = v(t, r)eϕ (2.2)

where eϕ is a unit vector in the direction of increasing ϕ.

 6

2.2 Equations of Motion and Navier-Stokes Equation

The system consists of coupled nonlinear equations for velocity and temperature. We
first consider the equations of motion for a three-dimensional viscous flow in cartesian
coordinates. Then, we consider the Boussinesq approximation. The resulting system,
known as Boussinesq equations, is also valid for two-dimensional flows. Hence, we use
these equations to describe our problem. We rewrite the system in polar coordinates and
integrate along a circular path at a fixed radius to eliminate the pressure. Finally, we
obtain a nonlinear initial-boundary control problem given in polar coordinates.

2.2.1 The Navier-Stokes Equations

The Navier-Stokes equations are obtained from the equations of conservation of mass and
momentum with viscous effects. Usually, in thermal convection problems, a body force is
introduced in the equation in order to include the effect of the gravity. Here, we denote the
body force per unit mass by Fg . Thus, for a Newtonian viscous fluid, the Navier-Stokes
equations are given by
!
∂v 1 1 µ+λ
+ (v·∇)v = Fg − ∇p + ν∆v + ∇(divv) (2.3)
∂t ρ ρ ρ

∂ρ
+ div (ρv) = 0, (2.4)
∂t
where v is the velocity vector, ρ is the density of the fluid, ν = µ/ρ is the kinematic
viscosity, µ is the viscosity, λ is a second viscosity coefficient and p is the pressure.

The equations of motion, (2.3) and (2.4), are derived from the conservation laws of mo-
mentum and mass, respectively. In the following, we present a basic derivation of these
equations. The approach presented here is standard, see for example [10], [22] and [34].

2.2.2 The Continuity Equation

Conservation of mass implies that the mass variation at a fixed volume V is due only to
the mass of fluid flowing through the surface of V . In other words, the increase of mass
inside a volume V must be equal to the fluid flow coming into the volume minus the fluid
flowing out of the volume.
 7

The mass of fluid occupying a volume V is given by

Z
m(t) = ρ(t, x1 , x2 , x3 )dV, (2.5)
V

where ρ is the fluid density. Thus, the rate of increase of mass in a volume V is
∂ ∂ Z
m(t) = ρ(t, x1 , x2 , x3 )dV. (2.6)
∂t ∂t V
Assume that V does not change with time and let η denote the unit outward normal to the
boundary surface ∂V . The total outward mass flux (outflow minus inflow) through ∂V is
given by the integral I Z
ρvη dA = div(ρv) dV. (2.7)
∂V V

Thus, from equations (refdmdt) and (2.7), the conservation of mass is expressed by the
equation
∂ Z Z
ρ(t, x1 , x2 , x3 )dV = − div(ρv) dV. (2.8)
∂t V V

Equation (2.8) must hold for any volume. Consider an infinitesimal element of volume δV ,
conservation of mass implies
∂ρ
δV = −div (ρv) δV
∂t
or equivalently,
∂ρ
+ div (ρv) = 0. (2.9)
∂t
Equation (2.9) is referred to as the continuity equation.

Observe that when ρ is constant, equation (2.9) reduces to

divv = 0. (2.10)

2.2.3 Euler’s Equation

The balance of momentum requires that the rate of change of momentum of a volume of
fluid equals the force applied to it. By Newton’s second law, force is the product of mass
and acceleration, that is,
!
dv ∂v
m(t) = m(t) + (v·∇)v . (2.11)
dt ∂t
 8

If the force exerted across a surface per unit area at time t is given by the pressure
p(t, x1 , x2 , x3 ) in the direction of the outward normal to the surface, the fluid is called
ideal. In that case the force is assumed to be orthogonal to the surface, so that there are
no tangential forces. Thus, for ideal fluids the total force acting on a volume is equal to
the integral of the pressure over the boundary of that volume
Z Z
pη dA = ∇p dV, (2.12)
∂V V

and the conservation of momentum may be expressed by

Z !
∂v
m(t) + (v·∇)v = − ∇p dV. (2.13)
∂t V

Also, (2.13) must hold for any volume V . If we consider an infinitesimal volume element
δV , we have δm = ρδV and the integral in (2.13) is equivalent to −∇p δV yielding
!
∂v
ρδV + (v·∇)v = −∇p δV. (2.14)
∂t

Dividing the above equation by ρδV , leads to the equation

∂v 1
+ (v·∇)v = − ∇p. (2.15)
∂t ρ

Equation (2.15) was derived by Euler in 1755 and is known as Euler’s equation.

A body force can be added to the right hand side of (2.15) producing the balance of
momentum equation
∂v 1 1
+ (v·∇)v = − ∇p + Fg , (2.16)
∂t ρ ρ
where Fg is an external force per unit mass.

2.2.4 The Viscosity Tensor

The continuity equation

∂ρ
+ div (ρv) = 0, (2.17)
∂t
and momentum equation
∂v 1
+ (v·∇)v = − ∇p, (2.18)
∂t ρ
 9

hold for ideal fluids. However, the concept of ideal fluids is very restrictive and an important
physical phenomenon is omitted. During the motion of a real fluid, energy dissipation
occurs which is not accounted for in the above equations. This dissipation is caused by
internal friction between neighboring particles due to relative motion. This internal friction
leads to a viscous force. Although the continuity equation still holds, Euler’s equation needs
to be modified.

The stress vector is defined as the force acting on a unit area. It is obtained by adding the
viscous force to the pressure, that is,

τ (t, x1 , x2 , x3 ) = −p(t, x1 , x2 , x3 )η + σ(t, x1 , x2 , x3 )·η. (2.19)

The first term in the equation (2.19) corresponds to the pressure in the direction of the
outward normal to the surface. The matrix σ represents the viscous stress, which opposes
relative movements among fluid particles trying to prevent deformations and it is not
necessarily normal to the surface. Viscous effects appear when different fluid particles have
different velocities. Thus, σ must vanish when the velocity is constant and therefore all
terms in σ depend on spatial velocity variations of the form ∂vi /∂xj .

When the stress is directly proportional to the velocity gradient, the fluid is said to be
Newtonian. We say that the fluid is Newtonian if there is a linear relation between the
stress tensor and the rate of strain tensor (or deviation of the fluid motion from a rigid
body motion). The rate of strain tensor is a matrix having entries
!
1 ∂vi ∂vj
Rij = + .
2 ∂xj ∂xi

In 2D flows (two space dimensions) it has the form

  
∂v1 1 ∂v1 ∂v2
+ ∂x
R =  1  ∂v1 ∂x1
∂v2
 2 ∂x2
∂v2
1 .
2 ∂x2
+ ∂x1 ∂x2

Assuming that the fluid is Newtonian, the viscous tensor is a matrix of the form
X ∂vl
σ = 2µR + λI , (2.20)
l ∂xl
or
σ = 2µR + λI div v, (2.21)
where µ and λ are considered constants (known as Lamé constants) and I denotes the
identity matrix. The constant µ is called the shear viscosity or simply viscosity of the fluid
 10

and λ is called a second viscosity coefficient. Now, the total force acting on a volume V is
given by the following integral
Z Z Z
(−pη + σ·η) dA = − ∇p dV + div σ dV. (2.22)
∂V V V

Here, div σ denotes the vector obtained by applying the divergence operator to each row
of σ and is given by
div σ = µ∆v + (λ + µ)∇(div v).
Hence, when the viscous effect is included, the momentum equation
∂v 1 1
+ (v·∇)v = − ∇p + div σ
∂t ρ ρ
1 µ λ+µ
= − ∇p + ∆v + ∇(div v),
ρ ρ ρ
is obtained. Therefore, after adding the body force per unit mass Fg , we obtain the Navier-
Stokes equation (2.3).

2.3 The Boussinesq approximation

Although the Navier-Stokes equations are valid for the general problem, the complexity
of these equations can be reduced for certain flows by making additional assumptions
and introducing approximations. We shall consider the Boussinesq approximation which
neglects any variation in fluid properties, so they are regarded as constants with the only
exception being the density in the buoyancy term. This is a reasonable assumption since
changes in temperature produce changes in density and this change causes fluid motion.
In addition, it is assumed that the relation between the density ρ and the temperature T
is linear and has the form
ρ = ρ0 (1 − β(T − T 0 )), (2.23)
where T 0 is a reference temperature and ρ0 and β are the reference fluid density and the
thermal expansion coefficient, respectively. Usually, the reference temperature T 0 is taken
to be the average temperature of the fluid.

In the convection loop problem under consideration, the body force Fg is due to the gravity
acceleration g and buoyancy force, denoted by Fb . The buoyancy force is the vertical force
produced by changes in density and for a unit mass it is given by Fb = (ρ0 − ρ)(−g). The
relation (2.23) yields Fb = ρ0 β(T − T 0 )(−g). Hence, the external force per unit mass is
given by
Fg = ρg + Fb = ρg + ρ0 gβ(T 0 − T ). (2.24)
 11

Since variations in density are only considered in the buoyancy force (all other changes in
density are neglected), the continuity equation (2.9) is reduced to div v = 0. Thus, we
have the incompressible Navier-Stokes equations
∂v 1 1
+ (v·∇)v = Fg − ∇p + ν ∆v, (2.25)
∂t ρ ρ
div v = 0, (2.26)

where ν = µρ is called the kinematic viscosity. A complete description of the fluid behavior
is obtained when heat transfer is included. Here, the heat equation
∂T
+ v·∇T = χ∆T (2.27)
∂t
describes this transfer. In particular, heat is transferred by convection (represented by the
term v·∇T ) and diffusion χ∆T . The constant χ introduced in (2.27) is the coefficient of
thermal diffusivity. Finally, combining these equations we have the Boussinesq equations
∂v 1
+ (v·∇)v = g (1 + β(T 0 − T )) − ∇p + ν ∆v, (2.28)
∂t ρ
div v = 0, (2.29)
∂T
+ v·∇T = χ∆T. (2.30)
∂t
1 ρ0
Observe that in the body force term F ,
ρ g
we take ρ
= 1, i.e., density variations are
neglected.

2.4 The Equations for the Convection Loop

The geometry of the convection loop also allows us to make additional simplifications. In
particular, if we integrate equation (2.28) along a circular path at a fixed r = r̃, important
simplifications are obtained. Also, is convenient to use polar coordinates instead of cartesian
coordinates.

Let er , eϕ denote the reference system of unit vectors in the directions of increasing r and
ϕ, respectively. Let wr , wϕ be the components of the vector w with respect to the unit
vectors er , eϕ , then we write
w = wr er + wϕ eϕ .
 12

2.4.1 The Divergence Free Condition

The divergence operator in polar coordinates is given by

1 ∂(r wr ) 1 ∂wϕ
div w = + , (2.31)
r ∂r r ∂ϕ

for any vector w of the form w = wr er + wϕ eϕ . If v(t, r, ϕ) = v(t, r)eϕ (i.e. if (2.2) is
assumed), vr = 0, vϕ = v(t, r) and hence,

1 ∂v(t, r)
div v = = 0.
r ∂ϕ

Therefore, the divergence free condition (2.29) always holds and all flows of the form (2.2)
are divergence free. Hence, the Boussinesq equations in this case, reduce to
∂v 1
+ (v·∇)v = g (1 + β(T 0 − T )) − ∇p + ν ∆v, (2.32)
∂t ρ
∂T
+ v·∇T = χ∆T. (2.33)
∂t

2.4.2 An Integro-Differential Equation

Let C denote a circular path at radius r = r̃, R1 < r̃ < R2 . Integrating both sides of the
Boussinesq equation (2.32) around this path, leads to the identity
I ! I !
∂v 1
+ (v·∇)v ds = g (1 + β(T 0 − T )) − ∇p + ν ∆v ds. (2.34)
C ∂t C ρ

For v(t, r, ϕ) = v(t, r)eϕ we have

I I
∂v ∂v ∂v
(t, r̃, ϕ) ds = (t, r̃)eϕ ds = (t, r̃) 2πr̃. (2.35)
C ∂t C ∂t ∂t

The second term on the left of (2.34) vanishes. To see this, we write (v·∇)v in polar
coordinates. The gradient operator ∇ in polar coordinates has the form
∂ 1 ∂
∇= er + eϕ . (2.36)
∂r r ∂ϕ
 13

For vectors w = wr er + wϕ eϕ and z = zr er + zϕ eϕ , we have

! !
∂zr wϕ ∂zr wϕ zϕ ∂zϕ wϕ ∂zϕ wϕ zr
(w·∇)z = wr + − er + wr + + eϕ .
∂r r ∂ϕ r ∂r r ∂ϕ r
Since the polar components of v are vr = 0 and vϕ = v(t, r) all terms in (v·∇)v vanish
2
except [v(t,r)]
r
er . Thus, we have
I I
[v(t, r)]2
(v·∇)v ds = − er ds = 0 (2.37)
C C r
since er is orthogonal to the path. Thus, the left side of (2.34) reduces to
I !
∂v ∂v
+ (v·∇)v ds = (t, r̃) 2πr̃. (2.38)
C ∂t ∂t

Now consider the right hand side of (2.34). The first term corresponds to the body force.
The radial and angular components of the gravity acceleration at a point (r, ϕ) may be
decomposed as in Figure 2.2. Thus, we write g = g sin ϕ er − g cos ϕ eϕ . Hence,

ϕ)
n(
si
g

ϕ
g
co

g
s(
ϕ)

ϕ
r

Figure 2.2: Decomposition of the gravity acceleration in radial and angular directions at a
point (r, ϕ).

I   Z 2π  
g 1 + β(T 0 − T (t, r̃, ϕ) ) ds = g sin ϕ er 1 + β(T 0 − T (t, r̃, ϕ) ) ds
C 0
Z 2π  
+ (−g cos ϕ) eϕ 1 + β(T 0 − T (t, r̃, ϕ) ) ds
0
Z 2π  
= (−g cos ϕ) 1 + β(T 0 − T (t, r̃, ϕ) ) r̃ dϕ
0
Z 2π Z 2π
= −g(1 + βT 0 )r̃ cos ϕdϕ + gβr̃ cos ϕT (t, r̃, ϕ) dϕ
0 0
Z 2π
= gβr̃ T (t, r̃, ϕ) cos(ϕ)dϕ. (2.39)
0
 14

Turning to the pressure term it follows that

I ! Z Z
1 1 2π r̃ 2π ∂p
− ∇p ds = − ∇p·eϕ r̃dϕ = − dϕ (2.40)
C ρ ρ 0 ρr 0 ∂ϕ

since ∇p = ∂p
e
∂r r
+ 1 ∂p
e .
r ∂ϕ ϕ
Thus, equation (2.40) vanishes by the periodicity condition.

The last term on the right hand side of (2.34) reduces to

I I
ν ∆v(t, r̃) eϕ ds = ν∆v(t, r̃) eϕ ds = 2πr̃ν∆v(t, r̃), (2.41)
C C

and hence, equation (2.34) becomes

Z 2π
∂v
(t, r̃) 2πr̃ = gβr̃ T (t, r̃, ϕ) cos(ϕ)dϕ + 2πr̃ν∆v(t, r̃). (2.42)
∂t 0

The Laplacian operator in polar coordinates is given by

∂2 1 ∂ 1 ∂2
∆= + + . (2.43)
∂r2 r ∂r r2 ∂ϕ2
Applying (2.43) to v(t, r), the last term vanishes and the Laplacian ∆ reduces to ∆ r defined
by !
∂2 1 ∂
∆v(t, r) = ∆r v(t, r) ≡ + v(t, r).
∂r2 r ∂r
This equation is now divided by 2πr̃ yielding the Navier-Stokes equation
∂v gβ Z 2π
(t, r̃) = T (t, r̃, ϕ) cos ϕ dϕ + ν ∆r v(t, r̃), (2.44)
∂t 2π 0
for R1 < r̃ < R2 .

2.4.3 The Heat Equation

Finally, we write the heat equation in polar coordinates. Since

∂T 1 ∂T
∇T = er + eϕ , (2.45)
∂r r ∂ϕ
 15

and v = v(t, r)eϕ , it follows that

v(t, r) ∂T
v·∇T =
r ∂ϕ

and the heat equation (2.33) can be written as

∂T v(t, r) ∂T
(t, r, ϕ) + (t, r, ϕ) = χ ∆T (t, r, ϕ). (2.46)
∂t r ∂ϕ

2.4.4 Boundary Conditions

At the walls of the pipe we assume the “no-slip” condition, that is, the velocity of the fluid
adjacent to the walls of the pipe has the same velocity as the wall. This implies that the
fluid’s velocity at the wall is zero, i.e.

v(t, Ri ) = 0, t≥0 i = 1, 2.

We also assume “perfect” conductivity at the walls of the pipe. This assumption implies
that the fluid temperature at the wall equals the wall temperature and hence we have the
Dirichlet boundary conditions

T (t, R1 , ϕ) = TW1 (t, ϕ), t > 0, ϕ ∈ (0, 2π)

T (t, R2 , ϕ) = TW2 (t, ϕ), t > 0, ϕ ∈ (0, 2π).

2.4.5 The PDE Model

At this point we have a complete set of equations for v(t, r) and T (t, r, ϕ). In particular, we
have the controlled system governed by the nonlinear integro-partial differential equations
Z
∂v gβ 2π
(t, r) = T (t, r, ϕ) cos ϕ dϕ + ν ∆r v(t, r) (2.47)
∂t 2π 0
∂T v(t, r) ∂T
(t, r, ϕ) = − (t, r, ϕ) + χ ∆T (t, r, ϕ) (2.48)
∂t r ∂ϕ
with Dirichlet boundary conditions

v(t, R1 ) = v(t, R2 ) = 0, t≥0 (2.49)

 16

T (t, R1 , ϕ) = TW1 (t, ϕ) + uW1 (t, ϕ) = u1 (t, ϕ), t > 0, ϕ ∈ (0, 2π) (2.50)
T (t, R2 , ϕ) = TW2 (t, ϕ) + uW2 (t, ϕ) = u2 (t, ϕ), t > 0, ϕ ∈ (0, 2π), (2.51)

where uW1 and uW2 are the applied temperature controls. In order to define the dynamical
system we specify initial conditions of the form

v(0, r) = v0 (r), R1 < r < R2 (2.52)

T (0, r, ϕ) = T0 (r, ϕ) R1 < r < R2 , 0 < ϕ < 2π (2.53)
Chapter 3

Abstract and Weak Formulations

In this chapter we present an abstract form of the problem described in Chapter 2 and we
show that the linearized system is stable. We first define a linear state operator in Section
3.1 and show its stability property. Then, in Section 3.2, the Dirichlet map is introduced in
order to define the unbounded input operator. Finally, a nonlinear state operator is defined
in Section 3.3 and the abstract form is given in Section 3.4. In Section 3.5, semigroup theory
is used to show the well-posedness of the uncontrolled system. Finally, in Section 3.6, we
discuss the variational formulation of the PDE system that will be used in the finite element
model.

3.1 The Linear State Operator

Let Ω1 = (R1 , R2 ), Ω2 = (0, 2π) and Ω = Ω1 × Ω2 = (R1 , R2 ) × (0, 2π). Let Γ1 =

{R1 } × (0, 2π), Γ2 = {R2 } × (0, 2π) and Γ = Γ1 ∪ Γ2 . We define the inner products on
L2 (Ω1 ) and L2 (Ω) by
Z R2
< v, φ >L2 (Ω1 ) = v(r)φ(r) dr (3.1)
R1

and Z Z
2π R2
< T, ψ >L2 (Ω) = T (r, ϕ)ψ(r, ϕ) r dr dϕ, (3.2)
0 R1

respectively. Let H = L2 (Ω1 ) × L2 (Ω) be the Hilbert space with inner product
   
v1 v
, 2 = hv1 , v2 iL2 (Ω1 ) + hT1 , T2 iL2 (Ω) . (3.3)
T1 T2

17
 18

In order to define a state space model, we define two operators A0 and A1 . Let
h i h i
D(A0 ) = H 2 (Ω1 ) ∩ H01 (Ω1 ) × H 2 (Ω) ∩ H01 (Ω) (3.4)

and for  
v
z= ∈ D(A0 ),
T
define A0 by       
v ν ∆r 0 v ν ∆r v
A0 = = . (3.5)
T 0 χ∆ T χ ∆T

Let L : L2 (Ω) → L2 (Ω1 ) be the integral operator defined by

gβ Z 2π
[Lh] (r) = cos ϕ h(r, ϕ) dϕ. (3.6)
2π 0
Observe that L is a compact operator (Hilbert-Schmidt) from L2 (Ω) to L2 (Ω1 ).
Let A1 : H → H be the compact linear operator defined by
 
0 L
A1 = . (3.7)
0 0

We define the linear operator

A = A0 + A1 (3.8)
on D(A) = D(A0 ).

It is well known that the Laplacian operator ∆ is dissipative and has only point spectrum
σ(∆) on the real axis. This observation leads to the following results.

Lemma 3.1 The spectrum σ(A0 ) of A0 is only point spectrum and it is given by

σ(A0 ) = σ(ν∆r ) ∪ σ(χ∆).

Proof: Let λ ∈ σ(A0 ), then

      
ν∆r 0 v v v
=λ , ∈ D(A0 ). (3.9)
0 χ∆ T T T
 19

From the second equation we have χ∆T = λT. Hence, if λ 6∈ σ(χ∆), T ≡ 0 is the only
solution and the first equation reduces to ν∆r v = λv. Thus, λ ∈ σ(ν∆r ), otherwise v ≡ 0
and λ is not an eigenvalue. Therefore, σ(A0 ) ⊂ σ(ν∆r ) ∪ σ(χ∆).

The reverse inclusion follows easily by taking either v = 0 or T = 0. 2

Lemma 3.2 The operator A0 generates an analytic semigroup of contractions S0 (t) defined
on H. Moreover, there is a constant µ0 > 0 such that
kS0 (t)k ≤ e−µ0 t . (3.10)

This result follows directly from the previous Lemma and the fact that ∆ is a dissipative
operator.

Lemma 3.3 Let ρ(A0 ) be the resolvent set of the operator A0 and ρ(A) be the resolvent
set of A. Then, ρ(A0 ) ⊂ ρ(A) and σ(A) ⊂ σ(A0 ).

Proof: Let λ ∈ ρ(A0 ), then (A − λI)−1 exists and is given by

 −1  
ν∆r − λI L (ν∆r − λI)−1 − (ν∆r − λI)−1 L (χ∆ − λI)−1
= . (3.11)
0 χ∆ − λI 0 (χ∆ − λI)−1
The continuity of the operator (A − λI)−1 follows from the continuity of the operators
(ν∆r − λI)−1 , (χ∆ − λI)−1 and L. Moreover, since ∆r and ∆ are maximal, A is maximal
so that λ ∈ ρ(A).

Since σ(A) = Cl − ρ(A), σ(A0 ) = Cl − ρ(A0 ), it follows that σ(A) = Cl − ρ(A) ⊆ Cl − ρ(A0 ) =
σ(A0 ). 2

Theorem 3.1 The operator A generates an analytic semigroup S(t) on H. Moreover,

there are constants M, γ > 0 such that
kS(t)k ≤ Me−γt . (3.12)

Proof: Observe that A is a bounded (compact) perturbation of A0 , thus A generates an

analytic semigroup S(t) on H (see [29], p.81). Also, since σ(A) ⊂ σ(A0 ), it follows that
µ = sup{Reλ : λ ∈ σ(A)} ≤ sup{Reλ : λ ∈ σ(A0 )} = µ0 < 0.
Consequently, (3.12) follows from Theorem 4.3, Chapter 4, in [29]. 2
 20

3.2 The Linear Input Operator

Consider the Dirichlet problem

∆y = 0 in Ω (3.13)
y = u on Γ. (3.14)
Following Lions and Magenes [26], we say that y ∈ L2 (Ω) is a weak solution of the Dirichlet
problem (3.13)-(3.14) if y satisfies
Z
∂ω
hy, ∆∗ωiL2 (Ω) = hy, ∆ωiL2 (Ω) = u(ϕ) (r, ϕ) r dϕ (3.15)
Γ ∂η
for all ω ∈ D(∆∗ ) = D(∆) = H 2 (Ω) ∩ H01 (Ω). The integral over Γ on the right hand side
of (3.15) denotes the integral over Γ1 ∪ Γ2 and should be interpreted as follows
Z Z Z
∂ω ∂ω ∂ω
u(ϕ) (r, ϕ) r dϕ = R1 u1 (ϕ) (R1 , ϕ) dϕ + R2 u2(ϕ) (R2 , ϕ) dϕ. (3.16)
Γ ∂η Γ1 ∂η Γ2 ∂η
∂
Here, ∂η
denotes the (outward) normal derivative on Γ.

Observe that since ω ∈ D(∆) we can apply Green’s formula twice to get (3.15) for y ∈ L2 (Ω)
with y(r, ϕ)|Γ = u(ϕ). The following result is a consequence of Lions and Magenes ([26],
p.170).

Lemma 3.4 If u ∈ U = L2 (Γ), then there is a unique function y ∈ L2 (Ω) that satisfies
(3.13)-(3.14) in the weak sense, i.e. (3.15). Moreover, the Dirichlet map D : L2 (Γ) → L2 (Ω)
defined by
Du = y (3.17)
is bounded from L2 (Γ) into H 1/2 (Ω).

Formally, we say that the Dirichlet Map D : L2 (Γ) → L2 (Ω) satisfies

Z
∂ω
hDu, ∆ωi = u(ϕ) (r, ϕ) r dϕ. (3.18)
Γ ∂η
for all ω ∈ H 2 (Ω) ∩ H01 (Ω). Consequently, if ω ∈ L2 (Ω), it follows that
D  E
hDu, ωi = Du, ∆ ∆−1 ω
Z
∂  −1 
= u(ϕ) ∆ ω (r, ϕ) r dϕ. (3.19)
Γ ∂η
 21

Equations (3.18) and (3.19) will be used to define the input operator B.

Let A∗ denote the Hilbert adjoint of A = A0 +A1 . Since A generates an analytic semigroup
on H, A∗ is a densely defined closed operator.

Theorem 3.2 The operator A∗ has domain

D(A∗ ) = D(A0 )

and A∗ = A∗0 + A∗1 = A0 + A∗1 . Moreover, A∗1 is the bounded linear operator defined by
      
v 0 0 0 v
A∗1 = ∗ =
T Lv L∗ 0 T

where L∗ : L2 (Ω1 ) → L2 (Ω) is given by

gβ
[L∗ v] (r, ϕ) = v(r) cos ϕ. (3.20)
2π

Proof: The operator A0 is self-adjoint, thus A∗ = A∗0 + A∗1 = A0 + A∗1 . If z1 = [v1 T1 ]T ,

z2 = [v2 T2 ]T ∈ D(A) = D(A0 ), then
Z R2 Z 2π
gβ
hA1 z1 , z2 i = cos ϕ v2 (r) T1 (r, ϕ) r dr dϕ
R
*1
0 2π +
gβ
= T1 , cos ϕ v2
2π L2 (Ω)
    
v1 0 0 v2
= , ∗ = hz1 , A∗1 z2 i , (3.21)
T1 L 0 T2

where L∗ : L2 (Ω1 ) → L2 (Ω) is given by (3.20). 2

Let W denote the space D(A) = D(A0 ) with graph norm

     
v v
= A v + . (3.22)
T T T
W H H
 22

The dual space of W is denoted by W 0 . Also, let V1 = H01 (Ω1 ) and V2 = H01 (Ω) denote the
energy spaces with inner products
* +
∂v ∂ṽ
< v, ṽ >V1 =< v, ṽ >L2 (Ω1 ) + , (3.23)
∂r ∂r L2 (Ω1 )

and
* + * +
∂T ∂T 1 ∂T 1 ∂T
< T, T̃ >V2 = < T, T̃ >L2 (Ω) + , + ,
∂r ∂r L2 (Ω)
r ∂ϕ r ∂ϕ L2 (Ω)
D E
= < T, T̃ >L2 (Ω) + ∇T, ∇T̃ , (3.24)
L2 (Ω)

respectively. Finally, set V = V1 × V2 . If H is identified with H 0 through the Riesz Map,

then we have continuous dense injections

W ,→ V ,→ H = H 0 ,→ V 0 ,→ W 0 . (3.25)

The Dirichlet map D : L2 (Γ) → L2 (Ω) is extended to H = L2 (Ω1 ) × L2 (Ω) by

 
0
D̂ = . (3.26)
D

To complete the state space formulation of the boundary control problem, we need to
define the input operator “B” and the non-linear operator f . First we need to lift A to
 : H → W 0 . This is done by defining
ˆ = ψz ∈ W 0
Az (3.27)

if and only if
< z, A∗ w >= ψz (w) =< ψz , w >W 0 ×W (3.28)
for all w ∈ W.
This uniquely defines  as an operator from H to W 0 . Define the operator B : L2 (Γ) → W 0
by  
−LD
B = −ÂD̂ = . (3.29)
−∆D

Now we will show that

Z R2 Z 2π Z
gβR2 ∂ω2
hBu, ωi = hu, B ∗ωi = − u(ϕ) cos ϕ ω1 (r) drdϕ − u(ϕ) (r, ϕ) r dϕ,
2π R1 0 Γ ∂η
(3.30)
for u ∈ U and ω = (ω1 , ω2 )T ∈ [H 2 (Ω1 ) ∩ H01 (Ω1 )] × [H 2 (Ω) ∩ H01 (Ω)] .
 23

Note that  
∗ D ∗ L∗
∗ ∗
B = −D Â = − , (3.31)
D∗ ∆
and recall that for ω ∈ L2 (Ω)
Z
∂  −1 
hDu, ωi = u(ϕ) ∆ ω (r, ϕ) r dϕ.
Γ ∂η

Thus, if u ∈ U and ω1 ∈ H01(Ω1 ), we have

− hu, D∗L∗ ω1 i = − hDu, L∗ω1 i

Z
gβ ∂  −1 
= − u(ϕ) ∆ cos ϕ ω1 (r) r dϕ
2π Γ ∂η
Z  
gβ
= − u(ϕ) ∇ ∆−1 cos ϕ ω1 (r) · η r dϕ, (3.32)
2π Γ
where η denotes the outward normal unit vector to Γ.

Set (Φ1 , Φ2 )T = ∇ (∆−1 cos ϕ ω1 (r)). We note that (Φ1 , Φ2 )T must satisfy the equation
   
Φ1
div = ∆ ∆−1 cos ϕ ω1 (r) = cos ϕ ω1 (r), (3.33)
Φ2

where  
Φ1 ∂Φ1 1 ∂Φ2
div = + .
Φ2 ∂r r ∂ϕ
Observe that  Z 
  r
Φ1 cos ϕ ω1 (ξ) dξ 
= R1
Φ2 0
satisfies equation (3.33). Therefore,
 
gβ Z Z r
− hDu, L∗ω1 i = − u(ϕ) cos ϕ ω1 (ξ) dξ ηr r dϕ. (3.34)
2π Γ R1

Here ηr denotes the first component (in the polar coordinate system) of the normal (out-
ward) vector to Γ. Thus, ηr = 1 at Γ2 (r = R2 ) and ηr = −1 at Γ1 (r = R1 ). Therefore,
we have
Z Z
∗ ∗ gβ R2
− hu, D L ω1 i = − u(ϕ) cos ϕ ω1 (r)dr R2 dϕ
2π Γ2 R1
gβR2 Z 2π Z R2
= − u(ϕ) cos ϕ ω1 (r) drdϕ (3.35)
2π 0 R1
 24

for u ∈ U and ω1 ∈ H 2 (Ω1 ) ∩ H01 (Ω1 ).

On the other hand, if ω2 ∈ H 2 (Ω) ∩ H01 (Ω)

Z
∗ ∂ω2
h−u, D ∆ω2 i = − hDu, ∆ω2i = − u(ϕ) (r, ϕ) r dϕ. (3.36)
Γ ∂η

Equation (3.35) together with (3.36) gives

hBu, ωi = − hu, D∗L∗ ωi − hu, D∗∆ωi

Z Z Z
gβR2 2π R2 ∂ω2
= − u(ϕ) cos ϕ ω1 (r) drdϕ − u(ϕ) (r, ϕ) r dϕ,
2π 0 R1 Γ ∂η

which is exactly (3.30). We shall return to these identities to formulate a weak form of the
PDE system.

Note that the linear part of the model can be written in W 0 as

ż = Az + Bu, t > 0, (3.37)

z(0) = z0 . (3.38)

However, to make sense of (3.37)-(3.38) we need to formulate the system in weak form.

3.3 The Nonlinear State Operator

To complete the model we define the non-linear operator f : V → H by

" #
0
f (z) = (3.39)
− vr ∂T
∂ϕ
(., .)
 
v
Observe that f is well defined since ∈ V = H01 (Ω1 ) × H01 (Ω) implies that v(.) is
T
continuous and ∂ϕ∂
T (., .) belongs to L2 (Ω). Thus, v(.)
r ∂ϕ
∂
T (., .) belongs to L2 (Ω) because it
is the product of a bounded function with an L2 function.
 25

3.4 Abstract Model

The controlled Boussinesq equations defined by the system (2.47)-(2.48) can now be for-
mulated as a state space system in W 0 given by

ż = Az + f (z) + Bu, t > 0, (3.40)

with initial data

z(0) = z0 ∈ H ⊂ W 0 , (3.41)
where A, B and f (z) are defined in (3.8), (3.29) and (3.39), respectively.

3.5 Well Posedness of the Open Loop Problem

Since A0 generates an analytic semigroup with 0 ∈ ρ(A), (−A0 )1/2 is well defined (see
[29], p. 69). Let V be the Hilbert space H01 (Ω1 ) × H01 (Ω) equipped with the norm kzkV =
k(−A0 )1/2 zkH . We have the following lemma.

Lemma 3.5 Consider the non-linear operator f : V → H defined by

" #
0
f (z) = . (3.42)
− vr ∂T
∂ϕ
(., .)

For any z ∈ V there is a neighborhood N and a constant C such that

kf (u) − f (w)k ≤ Cku − wkV (3.43)

for all u, w ∈ N .

Proof: If u, w ∈ V , then f is well-defined and we have

∂u ∂w2
2
kf (u) − f (w)kH = u1 − w1
∂ϕ ∂ϕ H

∂u ∂u ∂w2
2
+ kw1 kH
2
≤ ku1 − w1 kH −
∂ϕ ∂ϕ ∂ϕ H
H
≤ kukV ku − wkV + kwkV ku − wkV . (3.44)
 26

Let z0 ∈ V and assume that N (z0 , δ) is a neighborhood of z0 . There is a constant C1 such

that if z ∈ N (z0 , δ), then kzkV ≤ C1 . Hence, if u, w ∈ N (z0 , δ), then

kf (u) − f (w)kH ≤ 2C1 ku − wkV ,

and (3.43) holds for C = 2C1 . 2

Theorem 3.3 If u = 0 and z0 ∈ V , then there exists a time tf > 0 such that (3.40) − (3.41)
has a unique local solution z(t) on [0, tf ).

Proof: We consider the uncontrolled system

ż = Az + f (z), t > 0,
z(0) = z0 .

We have shown that A generates an analytic semigroup with 0 ∈ ρ(A) and f satisfies
(3.43). Then, the existence and uniqueness of a local solution to (3.40)-(3.41) with u = 0
follows from Theorem 3.1, Chapter 6, in Pazy [29]. 2

Of course (3.40)-(3.41) requires that one defines what is meant by a “solution” for u ∈
L2 ((0, tf ); L2 (Γ)). In order to do this, we need to define weak solutions. This is the subject
of the next section.

3.6 A Variational Form of the Problem

The abstract form of the PDE model in W 0 is

ż = Az + f (z) + Bu, t > 0,

z(0) = z0 ∈ H.

Therefore, for ω ∈ W = D(A0 ), it follows that

hż, ωi = hAz, ωi + hf (z), ωi + hBu, ωi

= hA0 z, ωi + hA1 z, ωi + hf (z), ωi + hBu, ωi

hż, ωi = hz, A0 ωi + hA1 z, ωi + hf (z), ωi + hu, B ∗ ωi . (3.45)

 27

Note that if a finite element scheme is based on the above weak problem, then one needs
test functions ω ∈ D(A0 ) = [H 2 (Ω1 ) ∩ H01 (Ω1 )] × [H 2(Ω) ∩ H01(Ω)] .

Thus, for ω ∈ W , we obtain

D E
hż, ωi = (−A0 )1/2 z, (−A0 )1/2 ω + hA1 z, ωi + hf (z), ωi + hu, B ∗ωi . (3.46)

The last term in this equation is the only difficult term. In order to relax the smoothness
on ω one must define < u, B ∗ ω > for ω ∈ H01 (Ω1 ) × H01 (Ω). This is best done by appealing
directly to the integro-partial differential equation.

First observe that

Z Z Z
∗ gβR2 R2 2π ∂ω2
hu, B ωi = − u(ϕ) cos ϕ ω1 (r) drdϕ − u(ϕ) (r, ϕ) r dϕ,
2π R1 0 Γ ∂η

holds for ω = (ω1 , ω2 )T with ω1 ∈ H01 (Ω1 ), ω2 ∈ H 2 (Ω) ∩ H01 (Ω). Moreover, if u ∈ H 1/2 (Γ)
and ω2 ∈ H 1 (Ω), then the trace theorem implies that, ∂ω ∂η
2
∈ H −1/2 (Γ) ([26], p.188) and
Z * +
∂ω2 ∂ω2
u(ϕ) (r, ϕ) r dϕ ≡ u,
Γ ∂η ∂η H 1/2 (Γ)×H −1/2 (Γ)

is well defined. Thus, if u ∈ H 1/2 (Γ) and ω2 ∈ H 1 (Ω) we can define hDu, ∆ω2i by defining
Z * +
∂ω2 ∂ω2
hDu, ∆ω2i ≡ u(ϕ) (r, ϕ) r dϕ = u, (3.47)
Γ ∂η ∂η H 1/2 (Γ)×H −1/2 (Γ)

This will be important in our finite element approximations. For example, if u ∈ H 1/2 (Γ),
then Bu ∈ V 0 and

∗ gβR2 Z R2 Z 2π Z
∂ω2
hBu, ωi = hu, B ωi = − u(ϕ) cos ϕ ω1 (r) drdϕ − u(ϕ) (r, ϕ) r dϕ,
2π R1 0 Γ ∂η
(3.48)
for ω ∈ H01 (Ω1 ) × H01 (Ω). Consequently, (3.48) is well defined for any u ∈ H 1/2 (Γ) and
ω ∈ H01 (Ω1 ) × H01 (Ω). Thus, in this case, piecewise linear functions that vanish on the
boundary may be considered in the finite element scheme. Standard finite elements schemes
can be developed based on (3.46) (see for example Lasiecka and Triggiani [25], Banks and
Ito [2]).

Our approach differs from these schemes in that we first consider piecewise bilinear functions
for the approximate temperature that do not necessarily vanish on the boundary and
 28

construct a global finite element model. This model is then reduced to obtain
amodel in
T
H01 (Ω1 ) × H01 (Ω). We start by considering a weak form of the PDE. For ẑ = v T̂ ∈ V̂ =
H01 (Ω1 ) × H 1 (Ω) and Φ ∈ H01 (Ω1 ), Ψ ∈ H 1 (Ω) we have
* +
∂
v, Φ = hν∆r v, ΦiL2 (Ω1 ) + hLT, ΦiL2 (Ω1 ) (3.49)
∂t L2 (Ω1 )
* + * +
∂ v ∂T
T, Ψ = hχ∆T, ΨiL2 (Ω) − ,Ψ . (3.50)
∂t L2 (Ω)
r ∂ϕ L2 (Ω)

We apply Green’s formula to obtain

* +
∂
v, Φ = −ν h∇r v, ∇r ΦiL2 (Ω1 ) + hLT, ΦiL2 (Ω1 ) (3.51)
∂t L2 (Ω1 )
* + * +
∂ ∂T
T, Ψ = −χ h∇T, ∇ΨiL2 (Ω) + χ ,Ψ
∂t L2 (Ω)
∂η L2 (Γ)
* +
v ∂T
− ,Ψ . (3.52)
r ∂ϕ L2 (Ω)

Here ∇r = ∂
∂r
.

Our scheme is based on a variational form given by the equations (3.51)-(3.52). We consider
v h ∈ H01 (Ω1 ) as a linear combination of piecewise linear functions on Ω1 , and T̂ h ∈ H 1 (Ω)
as a linear combination of bilinear functions on Ω. Thus, the control uh (t) = T̂ h |Γ will be
approximated by piecewise linear functions on Γ and hence uh ∈ H 1 (Γ) ⊆ H 1/2 (Γ). This
observation provides the basis for the Galerkin approximations below.

We note that if V̂ h ⊆ H01(Ω1 ) × H 1 (Ω) and U h ⊆ H 1 (Γ) ⊆ L2 (Γ), then restricting (3.51)-
(3.52) to V̂ h × U h leads to

* +
∂ h h D E D E
v ,Φ = −ν ∇r v h , ∇r Φh + LT̂ h , Φh (3.53)
∂t L2 (Ω1 )
L2 (Ω1 ) L2 (Ω1 )
* + * +
∂ h h D ∂ T̂ h h E
T̂ , Ψ = −χ ∇T̂ , ∇Ψ 2 + χ
h h
,Ψ
∂t L2 (Ω)
L (Ω) ∂η L2 (Γ)
* +
v h ∂ T̂ h h
− ,Ψ . (3.54)
r ∂ϕ L2 (Ω)
 29

 T
Observe that (3.53)-(3.54) is well defined for ẑh = v h T̂ h ∈ V̂ and if T̂ h |Γ = uh , then
* + Z
∂ T̂ h h ∂uh
,Ψ = (ϕ)Ψh (r, ϕ) rdϕ.
∂η L2 (Γ) Γ ∂η

Let T h = T̂ h − Duh and observe that T h |Γ = T̂ h |Γ − Duh |Γ = 0, thus T h ∈ V h ⊆ H01 (Ω).

Hence, T̂ h = T h + Duh and from equation (3.35), the second term of the equation (3.53)
can be written as
D E D E D E D E
LT̂ h , Φh = T̂ h , L∗ Φh = T h , L∗ Φh + Duh , L∗ Φh
D E Z Z
∗ gβR2 2π R2
= T ,L Φ
h h
+ uh (ϕ) cos ϕ Φh (r) drdϕ. (3.55)
2π 0 R1

 T
This observation leads to a model for zh (t) = v h (t) T h (t) of the form
* +
∂ h h D E D E
v ,Φ = −ν ∇r v h , ∇r Φh + T h , L∗ Φh
∂t L2 (Ω1 )
L2 (Ω1 ) L2 (Ω1 )

gβR2 Z 2π Z R2 h
+ u (ϕ) cos ϕ Φh (r) drdϕ
2π 0 R1
(3.56)

* + * +
∂ h D ∂uh h E
(T + Duh ), Ψh = −χ ∇T , ∇Ψ 2 + χ ,Ψ h h
∂t L2 (Ω)
L (Ω) ∂η L2 (Γ)
* +
vh ∂
− (T h + Duh ), Ψ . (3.57)
r ∂ϕ L2 (Ω)

The last term in equation (3.57) reduces to the equation

* + * + * +
vh ∂ v h ∂T h h vh ∂
− (T h + Duh ), Ψh =− ,Ψ − (Duh ), Ψh . (3.58)
r ∂ϕ L2 (Ω)
r ∂ϕ L2 (Ω)
r ∂ϕ L2 (Ω)

Observe that the first term on the right in equation (3.58) becomes
* +
v h ∂T h h
− ,Ψ =< f (z h ), Ψh >L2 (Ω) . (3.59)
r ∂ϕ L2 (Ω)
 30

We will show that the second term in (3.58) vanishes. Integrating by parts we obtain
* + * +
vh ∂ v h ∂Ψh
(Duh ), Ψh = − Du , h
r ∂ϕ L2 (Ω)
r ∂ϕ L2 (Ω)
Z R2 v h h
h i
+ Du (2π)Ψh (r, 2π) − Duh(0)Ψh (r, 0) dr.
R1 r
(3.60)

The periodic boundary condition forces the last term to vanish, thus,
* + * +
v h ∂(Duh ) h v h ∂Ψh
,Ψ = − Du , h
r ∂ϕ L2 (Ω)
r ∂ϕ L2 (Ω)
Z " #!
h h
−1 v (r) ∂Ψ
= − h
u (ϕ)∇ ∆ (r, ϕ) · η r dϕ.
Γ r ∂ϕ
(3.61)

 h i
Set (Φ1 , Φ2 )T = ∇ ∆−1 vh ∂Ψh
r ∂ϕ
and observe that (Φ1 , Φ2 )T satisfies
 
Φ1 v h (r) ∂Ψh
div = (r, ϕ), (3.62)
Φ2 r ∂ϕ
or, equivalently,
∂Φ1 1 ∂Φ2 v h (r) ∂Ψh
(r, ϕ) + (r, ϕ) = (r, ϕ). (3.63)
∂r r ∂ϕ r ∂ϕ
If we define Φ1 = 0 and Φ2 = v h (r)Ψh (r, ϕ), then equation (3.62) is satisfied and equation
(3.61) reduces to
* + Z " #!
vh ∂ −1 v h (r) ∂Ψh
(Duh ), Ψh = − h
u (ϕ)∇ ∆ (r, ϕ) · η r dϕ
r ∂ϕ L2 (Ω) Γ r ∂ϕ
Z
= − uh (ϕ)v h (r)Ψh (r, ϕ)ηϕ r dϕ. (3.64)
Γ

Observe that ηϕ is the second component (in polar coordinates) of the vector normal to Γ.
Hence, ηϕ = 0 and * +
vh ∂ h h
(Du ), Ψ = 0. (3.65)
r ∂ϕ L2 (Ω)

This observation combined with equations (3.58)-(3.59) yields the identity

* +
vh ∂
(T h + Duh ), Ψh =< f (z h ), Ψh >L2 (Ω) . (3.66)
r ∂ϕ L2 (Ω)
 31

Therefore, equations (3.56) and (3.57) can be “reduced” to a system for

 T
zh (t) = v h (t) T h (t) ∈ V h ⊆ H01 (Ω1 ) × H01 (Ω). In particular, we have
* +
∂ h h D E D E
v ,Φ = −ν ∇r v h , ∇r Φh + T h , L∗Φh
∂t L2 (Ω1 )
L2 (Ω1 ) L2 (Ω1 )

gβR2 Z 2π Z R2 h
+ u (ϕ) cos ϕ Φh (r) drdϕ (3.67)
* +
2π 0 R1 * +
∂ h D E ∂u h
(T + Duh ), Ψh = −χ ∇T h , ∇Ψh 2 + χ , Ψh
∂t L2 (Ω)
L (Ω) ∂η L2 (Γ)
* +
v h ∂T h
− ,Ψ . (3.68)
r ∂ϕ L2 (Ω)

∂
Finally, we note that equation (3.68) contains terms involving ∂t
Duh . However, as we see
later, this term can also be eliminated.
Chapter 4

Control of the Boussinesq Equations

4.1 Introduction

We recall that the dynamics of the flow can be described by the following quasilinear
infinite-dimensional distributed parameter system
Z
∂v gβ 2π
(t, r) = T (t, r, ϕ) cos ϕ dϕ + ν ∆r v(t, r) (4.1)
∂t 2π 0
∂T v(t, r) ∂T
(t, r, ϕ) = − (t, r, ϕ) + χ ∆T (t, r, ϕ) (4.2)
∂t r ∂ϕ
with boundary conditions
v(t, r1 ) = v(t, r2 ) = 0

T (t, R1 , ϕ) = u1 (t, ϕ), t > 0, ϕ ∈ (0, 2π) (4.3)

T (t, R2 , ϕ) = u2 (t, ϕ), t > 0, ϕ ∈ (0, 2π) (4.4)

where g denotes the gravity acceleration magnitude, β is the thermal expansion, ν is the
kinematic viscosity of the fluid and χ is the fluid’s thermal conductivity.

As has been shown in Chapter 2, the control problem (4.1)-(4.2), may be written as

ż = Az + f (z) + Bu, t > 0, (4.5)

z(0) = z0 ∈ H,

where the state space is H = L2 (Ω1 ) × L2 (Ω) and B is an unbounded operator.

32
 33

Observe that the system (4.1)-(4.2) is invariant under translations of T (t, r, ϕ). Thus,
we may consider T as a difference in temperature from a given constant. We take T =
T emp−60◦ F . We consider the problem of driving the system to v = 0, T = 0 by controlling
the temperature at the walls of the loop. The control is of the form
 
u (t, ϕ)
u(t, ϕ) = 1 ,
u2 (t, ϕ)
where u1 (t, ϕ) is applied to the inner wall and u2 (t, ϕ) is applied to the outer wall of the
loop. We also consider the problem with either u1 = 0 or u2 = 0, so that only one wall is
heated/cooled.

We consider both linear and nonlinear feedback controllers. Section 4.2 is concerned with
the linear quadratic optimal control problem. The system is linearized about an equilibrium
point and the LQR problem is solved. Once the LQR linear control is obtained, it is applied
to the original nonlinear problem. We also consider nonlinear controllers. In particular, we
apply Krener’s approach [20] to finite element approximations. However, in order to provide
some indications that a Krener type feedback control might be useful for the Boussinesq
problem, we first investigate its application to a one-mode approximation. In particular, in
Section 4.4 we consider the Lorenz equations where the LQR optimal control, a quadratic
control based on Krener’s method and a nonlinear control designed by Wang, Singer and
Bau are applied and compared.

4.2 The Linear Quadratic Regulator Problem

We linearize the system about the equilibrium point v = 0, T = 0. Since f (0, 0) = (0, 0)T ,
the linearized system becomes

ż(t) = Az(t) + Bu(t), t > 0, z(0) = z0 , (4.6)

where z0 ∈ H, u ∈ U = L2 (Γ) and the operators A and B are defined as in Chapter 3.

The LQR problem is to minimize the quadratic cost defined by

Z ∞
J(z0 , u) = [hQz(t), z(t)iH + hRu(t), u(t)iU ] dt (4.7)
0

over all controls u ∈ L2 ((0, ∞); U), subject to the linear system (4.6). The state weighting
operator for the LQR problem is
 
Qv 0
Q=
0 QT
 34

with Qv = qv IL2 (Ω1 ) , QT = qT IL2 (Ω) and qv , qT positive constants. The operators IL2 (Ω1 ) , IL2 (Ω)
denote the identity operators in L2 (Ω1 ) and L2 (Ω), respectively. The control weighting op-
erator is given by R = qu IU , IU denotes the identity operator on U and qu is a positive
constant.

Recall that H and U are Hilbert spaces and A : D(A) → H generates an analytic semigroup
on H (Theorem 3.1). The dense continuous injection D(A) ,→ H allowed us to lift A to H.
The input operator B defined by −ÂD̂ is not bounded as an operator into H. However,
we have the following result.

Lemma 4.1 The operator B : U → W 0 = (D(A))0 satisfies

−(3/4+)
 B ∈ L(L2 (Γ); L2 (Ω1 ) × L2 (Ω)). (4.8)

Proof: The input operator B is given by

hBu, ωi = hu, B ∗ ωi
gβR2 Z R2 Z 2π Z
∂ω2
= − u(ϕ) cos ϕ ω1 (r) drdϕ − u(ϕ) (r, ϕ) r dϕ. (4.9)
2π R1 0 Γ ∂η
We define the operators Bv and BT as follows
gβR2 Z R2 Z 2π
hBv u, ω1 i = hu, Bv∗ ω1 i =− u(ϕ) cos ϕ ω1 (r) drdϕ (4.10)
2π R1 0
Z
∂ω2
hBT u, ω2 i = hu, BT∗ ω2 i = − u(ϕ) (r, ϕ) r dϕ. (4.11)
Γ ∂η
For each  > 0, the trace operator
∂ω2
ω2 → : H 3/2+2 (Ω) → H 2 (Γ) ⊂ L2 (Γ)
∂η
is continuous (see [1], [26], [38]). Thus, we have
 h i0 
BT ∈ L U, H 3/2+2 (Ω) ∀ > 0.

Also, if ω1 ∈ H 3/2+2 (Ω1 ) we have Bv∗ ω1 ∈ L2 (Γ). Thus, for ω ∈ H 3/2+2 (Ω1 ) × H 3/2+2 (Ω),
we have B ∗ ω ∈ L2 (Γ). Hence,
 h i0 
B ∈ L U, H 3/2+2 (Ω1 ) × H 3/2+2 (Ω) .
 35

On the other hand,  

H 3/2+2
(Ω1 ) × H 3/2+2
(Ω) = D b 3/4+
A ,

thus,   
B ∈ L U, D b −(3/4+)
A ,

or equivalently,
b −(3/4+)
A B ∈ L(U; H). (4.12)
2

In order to obtain existence of the LQR problem we recall that A generates an exponentially
stable semigroup and hence we have the finite cost condition.

Finite Cost Condition: For every z0 ∈ H, there exists u ∈ U such that the cost function
defined in (4.7) is finite.

Theorem 4.1 Let A and B be the operators defined in Chapter 3. There exist a self-
adjoint, non-negative definite operator P ∈ L(H) that satisfies the Algebraic Riccati Equa-
tion (ARE)

hP z, AωiH + hAz, P ωiH − hR−1 B ∗ P z, B ∗ P ωiU + hQz, ωiH = 0. (4.13)

Moreover,

1. (A∗ )(1−) P ∈ L(H), ∀ > 0,

2. R−1 B ∗ P ∈ L(H, U),

3. J(z0 , uopt ) = hP z0 , z0 iX .

The LQR problem has a solution of the form

uopt (t, z0 ) = −R−1 B ∗ P zopt (t, z0 ), (4.14)

where zopt is the corresponding solution to (4.6) with u = uopt.

Proof: We observe that our problem satisfies the assumption (1.5) found on page 12 in
Lasiecka and Triggiani [24] and the Finite Cost Condition. The result follows from Theorem
2.1 in [24]. 2
 36

Thus, for a given initial condition z0 ∈ H, we solve the LQR problem (4.6)-(4.7) for the
optimal control uopt. Note that the nonlinear operator f (z) is not taken into account in
the LQR problem. However, in order to see how well this linear feedback performs we feed
it into the full non-linear system (4.5) to obtain the closed-loop nonlinear system

ż = (A − BK)z + f (z), t > 0, (4.15)

z(0) = z0 .

Here K is the bounded linear operator defined by K = R−1 B ∗ P .

In practice, we must use some type of approximation. We consider finite element and one-
mode approximations and use these models to construct the feedback controllers. There-
fore, we can use existing finite dimensional algorithms. We now summarize Krener’s algo-
rithm.

4.3 Krener’s Algorithm

Once the PDE has been approximated we can apply finite dimensional design. Here we
review an approach due to Krener which consists of choosing a transformation and a state
feedback that linearize the system. We emphasize that Krener’s algorithm applies only to
finite dimensional systems and has not yet been extended to infinite dimensional systems.
We shall apply Krener algorithm to finite dimensional approximations of the Boussinesq
equations. The Boussinesq equations lead to approximating systems with a special form so
we limit our discussion to these systems. For more details or more general cases we refer
the reader to [19], [20].

Consider an autonomous nonlinear system

ẋ = f (x) + Bu (4.16)

where x ∈ IRn , u ∈ IRm , f (x) is a nonlinear vector function in x, with equilibrium point
x = 0 and B ∈ IRnxm .

In order to apply the method, the system is linearized and written as follows

ẋ = Ax + f (k) (x) + Bω + O(xk+1), (4.17)

where A and B are matrices, A : IRn → IRn , B : IRm → IRn , and f (k) (x) is a nonlinear vector
of dimension n, having entries of degree k ≥ 2 in x, for example, if x = (x1 , x2 ) ∈ IR2 ,
f (3) (x) contains terms which are linear combinations of x31 , x21 x2 , x1 x22 , x32 . Terms of order
 37

greater than k are contained in O(xk+1 ). Thus, the system is assumed to have nonlinear
terms of order greater or equal k.

This can be accomplished by taking A as the Jacobian of f at 0. Thus f (k) (x) represents
the difference.

One then seeks a change of coordinates and a nonlinear control of order k in x,

zk = x − φ(k) (x) (4.18)

ωk = α(k) (x) + β (k−1) (x)v + v, (4.19)

where φ(k) (x) is a vector of degree k in x and α(k) and β (k−1) are polynomials in x of degree
k and k − 1, respectively, such that the terms of order k cancel.

Let [., .] denote the Lie bracket defined for two vector fields f, g by
∂g ∂f
[f, g] = f− g.
∂x ∂x
The transformation φ(k) (x) and the nonlinear vectors α(k) and β (k−1) have to satisfy the
homological equations

f (k) (zk ) = −Bα(k) (zk ) + [Azk , φ(k) (zk )] (4.20)

Bβ (k−1) (zk )v = [Bv, φ(k) (zk )]. (4.21)

Necessary and sufficient conditions for the existence of a solution to (4.20)-(4.21) are given
in [19]. Assuming that a solution to (4.20)-(4.21) can be found, the resulting system is
given by
żk = Azk + Bvk + O(zk+1k ), (4.22)
and can be written as

żk = Azk + f (k+1) (zk ) + Bvk + O(zk+2

k ). (4.23)

This produces a system of the form (4.17). The procedure may be applied repeatedly to
obtain a system of the desired order of linearization, given by

ż = Az + Bv + O(zm+1 ), (4.24)

for any m > 0. Here v is the control, which is calculated by solving an LQR problem. In
particular, we minimize a given cost function J(z0 , v) of the form (4.7) subject to the linear
system
ż = Az + Bv. (4.25)
 38

If the system is controllable, the LQR problem has a unique solution

vopt = −Kz = R−1 B T P z,

where P is the unique, symmetric, non-negative matrix satisfying the algebraic Riccati
equation
AT P + P AT − P BR−1 BT P + Q = 0.
The control v is linear in z but is nonlinear in the original state x. The transformations
φ(j) and the polynomials α(j) and β (j−1) , j = 2, . . . , m are used to rewrite the control v
in terms of x. In Section 4.4.2 we illustrate all the steps in this algorithm and apply the
results to the Lorenz equations.

4.4 Application to a One Mode Approximation

Lorenz equations are typically used to describe chaotic systems. In particular, the Lorenz
equations may be viewed as a one mode approximation to the full Boussinesq system for
the thermal convection loop (see [36] and [40]). Nonlinear feedback control laws based
on the Lorenz model have been studied in [35] and [36]. In [36], Wang, Singer and Bau
present experimental and numerical results by applying a nonlinear feedback law based
on Lorenz model for the problem under consideration. We shall use this model to test
and compare the LQR/linearization based control with the non-linear control generated by
Krener’s method.

4.4.1 Derivation of the Equations

Yorke, Yorke and Mallet-Paret [40], Wang, Singer and Bau [36], assume a Fourier series
expansion for the difference between the fluid temperature and the temperature at the wall,
as follows ∞ X
T (t, r, ϕ) − TW (t, ϕ) = cn (t, r) cos(nϕ) + sn (t, r) sin(nϕ). (4.26)
n=0

This expansion is then substituted into the partial differential equations (2.47)-(2.48) ob-
taining differential equations for v(t, r) and the coefficients of the Fourier series. The
equations for v(t, r), c1 (t, r), and s1 (t, r) decoupled from the rest of the system.

In [36], Wang, Singer and Bau assumed that the velocity and the temperature are inde-
pendent of r. In [40], York, York and Mallet-Paret expanded v(t, r), c1 (t, r) and s1 (t, r)
 39

in a series of Bessel functions of order zero. Both approaches produce an infinite set of
ordinary differential equations. The first three equations, which correspond to the first
mode, decoupled from the others. They are similar to the Lorenz equations and are given
by

ẋ1 (t) = P (−x1 (t) + x2 (t))

ẋ2 (t) = −x2 (t) − x1 (t)x3 (t) (4.27)
ẋ3 (t) = x1 (t)x2 (t) − x3 (t) − Rũ,

where, in both cases, P and R are related to the loop’s Prandtl number and the loop’s
Rayleigh number, respectively. Also , R and ũ are related to the temperature at the
wall (see [40], [36]). These relationships depend on the approximation used to derive the
equations and are obtained after considering the dimensionless numbers P and R.

4.4.2 The Controlled System

As in [36], we set ũ(t) = 1 − R1 u(t) leading to the lumped parameter control system

ẋ1 (t) = P (−x1 (t) + x2 (t))

ẋ2 (t) = −x2 (t) − x1 (t)x3 (t) (4.28)
ẋ3 (t) = x1 (t)x2 (t) − x3 (t) − R + u(t).

For ũ = 1 (i.e. u(t) = 0), the three equilibrium points of the system (4.28) are given by
√ √
x1 = (0, 0, −R)T , x2 = ( R − 1, R − 1, −1)T

√ √
and x3 = (− R − 1, − R − 1, −1)T .

The stability of each fixed point depends on the parameter values (see [32]).

4.4.3 LQR Optimal Control

We consider now the LQR problem associated to equations (4.28). First, we linearize the
system about an equilibrium xe = x = (x1 , x2 , x3 )T . Then, we have

x̃˙ = AL x̃ + BL u (4.29)
 40

where x̃ = x − x,  
−P P 0
AL =  
 −x3 −1 −x1  (4.30)
x2 x1 −1
and  
0
 
BL = 0. (4.31)
1

The system (4.27) is equivalent to the system in the new variable x̃, given by

x̃˙ = AL x̃ + fL x̃ + BL u (4.32)

where fL is defined by  
0
 
fL (x̃) =  −x̃1 x̃3  . (4.33)
x̃1 x̃2

The LQR problem is to minimizes the quadratic functional

Z ∞
J= (x̃T (t)QL x̃(t) + RL [u(t)]2 dt, (4.34)
0

where QL = QTL ≥ 0 and RL > 0, subject to the linear system (4.29)-(4.31).

The optimal control

ũopt (t) = −KL x̃(t) (4.35)
is calculated solving the above stationary linear quadratic regulator problem.

We then feed the system (4.32) to get the nonlinear closed-loop system

x̃˙ = (AL − BL KL )x̃ + fL x̃, (4.36)

4.4.4 Krener’s Nonlinear Control

Here we show how the nonlinear control proposed by Krener in [19] is applied to (4.28).
The system (4.28) is written in the form

x̃˙ = AL x̃ + fL (x̃) + BL u, (4.37)

 41

where A,B and fL are defined in (4.30),(4.31) and (4.33), respectively. Introduce now a
quadratic change of coordinates
z = x̃ − φ(2) (x̃) (4.38)
and a quadratic control of the form

u = α(2) (x̃) + β(x̃)v + v, (4.39)

in order to obtain a higher order approximate system. We used a Matlab package provided
by Krener in order to accomplish this numerically.

Observe that the given system has a nonlinearity of degree 2, thus it is completely linearized
in one step. The resulting linear system, in the new variable z, is

ż = AL z + BL v, (4.40)

where AL and BL are defined in (4.30) and (4.31), respectively. The matrices AL and BL
are the same as in the LQR problem considered previously (see Section 4.4.3). Thus, the
optimal control is v = −KL z, where KL is the same gain matrix as in equation (4.35).

The input u to the equation (4.37) in the x̃-coordinates can be obtained form (4.38)-(4.39)
as follows

u(x̃) = α(2) (x̃) + (1 + β(x̃)))v

= α(2) (x̃) + (1 + β(x̃))(−KL x̃ + KL φ(2) (x̃))
h i
= −KL x̃ + α(2) (x̃) − KL β(x)x̃ + KL φ(2) (x̃)
+ KL φ(2) (x̃)β(x̃). (4.41)

The resulting nonlinear closed-loop system in x̃ is given by

x̃˙ = (AL − BL KL )x̃ + fL (x̃). (4.42)

Although the controller (4.41) has a cubic term, this term is relative small and is neglected.
Consequently, we are lead to the non-linear feedback law

u(x̃) = −KL x̃ + q(x̃) (4.43)

where q(x) is the quadratic function given by

q(x̃) = α(2) (x̃) − KL β(x̃)x̃ + KL φ(2) (x̃). (4.44)

 42

4.4.5 A Modification of Krener’s Controller

While experimenting with the control (4.43) we observed that we could improve perfor-
mance by scaling the non-linear term. In particular, we weighted the quadratic term by an
scalar κ so that a modification of nonlinear control (4.43) becomes
u(x̃) = −KL x̃ + κq(x̃). (4.45)

4.4.6 An Ad Hoc Controller

Wang, Singer and Bau [36] propose a nonlinear control of the form
u(x) = C(sgn(x2 )x2 − x2 ),
= C(abs(x2 ) − x2 ),
where C is a constant and x2 is the second component of the fixed point to be stabilized.
Although this controller is nonlinear, it is based on “engineering insight” and not on any
given algorithm.

4.4.7 A Comparison of the Four Controllers

Numerical experiments were conducted to test the controllers.

For P = 4 the critical value of R is Rc = 16. Here, P = 4 and R = 50 were chosen. Hence
all three equilibrium points,
{(0, 0, −50), (7, 7, −1), (−7, −7, −1)}
are unstable. We set Q = I, N = 1 and controlled to the equilibrium xe = (7, 7, −1).

The Lorenz attractor obtained by setting P = 4 and R = 50 is shown in Figure 4.1. The
marks 0 ∗0 in the figure indicate the location of the equilibrium points for the parameter
values {(0, 0, −50); (7, 7, −1); (−7, −7, −1)}.

In Figure 4.2 we show the trajectories of the closed loop system for each of the control laws
described above. Here Q = I, N = 1, the initial point is (−3, 9, −8) and the equilibrium
point to be stabilized is xe = (7, 7, −1).

In Figure 4.3 the components of each of the controls used are plotted.
 43

4.4.8 Conclusions

Similar results are obtained for different initial points. Here we present a particular case
as an example. We conducted several simulations with different initial data and similar
results were observed. From the numerical results we see that all of the feedback controllers
stabilize the chaotic system. The linear LQR optimal control and the nonlinear controller
suggest by Wang, Singer and Bau have slow response. Similar results are obtained for the
nonlinear control (4.43). However, the scaled controller (4.45) improved performance as
illustrated in Figure 4.2.

Observe that the non-linear controller (4.43) shows some improvement over the LQR con-
troller and greatly reduces the oscillations found in the nonlinear controller proposed by
Wang, Singer and Bau. Hence, it is worthwhile to investigate the effectiveness of these
approach on a more detailed finite element model.
 44

Lorenz Attractor

20

10

−10
*
−20

−30
*
−40

−50

20

10 *

0
15 20
5 10
−10 −5 0
−15 −10
−20

Figure 4.1: Lorenz Attractor. Parameters: P = 4, R = 50. Equilibrium points (7, 7, −1),
(−7, −7, −1) and (0, 0, −50). Initial Point x0 = (−3, 9, −8).
 45

Closed Loop, u = −K(x−xe) u = [0 35 0] (sign(x(2)) x−xe)

10 10

0 0
x(3)

x(3)
−10 −10

−20 10 −20 10
−5 0 5 0 −5 0 5 0
10 −10 10 −10
x(2) x(2)
x(1) x(1)
Closed Loop, u = −K (x−xe) + q(x−xe) u = −K (x−xe) + 20 q(x−xe)

10 10

0 0
x(3)

x(3)

−10 −10

−20 10 −20 10
−5 0 5 0 −5 0 5 0
10 −10 10 −10
x(2) x(2)
x(1) x(1)

Figure 4.2: Solutions of the nonlinear closed-loop system feeding by four feedback con-
trollers. Parameter values: P = 4, R = 50. Equilibrium Point xe = (7, 7, −1). Initial Point
x0 = (−3, 9, −8).
 46

u=−K(x−xe) u=[0 35 0] (sign(x(2)) x −xe)

20 20

10 10

0 0

−10 −10

−20 −20
0 5 10 0 5 10
t t
u=−K(x−xe) + q(x−xe) u=−K(x−xe) + 20q(x−xe)
20 20

10 10

0 0

−10 −10

−20 −20
0 5 10 0 5 10
t t

Figure 4.3: Components of the feedback controllers. Parameter values: P = 4, R = 50.

Equilibrium Point to be Stabilize xe = (7, 7, −1). Initial Point x0 = (−3, 9, −8).
Chapter 5

A Finite Element Approximation for

the Boussinesq Model

In this chapter we construct approximating optimal control problems which are based on
finite element approximations of the partial differential equations (2.47)-(2.48). We assume
that the reader is familiar with the finite element approximations and refer the reader to
[6], [7], [9], [12], [16], [28] and [41] for specifics of the method.

Galerkin’s method is applied to the Boussinesq model of the thermal convection problem.
The discretization in space (only) yields a semidiscrete scheme and an approximate solution
is obtained by solving a finite dimensional ordinary differential equation.

The finite element spaces for the velocity and temperature are different. They are described
in Sections 5.2.1 and 5.2.2, respectively. The specific finite element scheme is detailed in
Sections 5.2.3 - 5.2.4. In Section 5.2.5, we summarize the model.

5.1 Introduction

Basically, the Galerkin’s method produces an approximate problem, by projecting onto a

finite dimensional space. This is best done by using a variational form of the problem.
An approximate solution is written as linear combination of basis functions for a finite
element space. Substituting this expression into the weak formulation, a system of algebraic
equations for the coefficients is obtained. There are several approaches to this process. We
detail the particular scheme used in the calculations below.

47
 48

5.2 Discretization in Space

The first step is to define the triangulation, i.e., to choose the geometrical objects and
their sizes to cover Ω. For the given domain, sector elements are used since the partition

hr
ϕ =2π////

}
////////////////// ////
////////////////// ////
////
////////////////// ////
////////////////// ////
//////////////////
R
//////////////////
2

////
//////////////////
R 1
////
////////////////// ////
//////////////////
// ////
////
////////////////// ϕ =0, 2π ////
//////////////////
////////////////// ////
////////////////// ////
////////////////// ////
////////////////// ////
////
//////////////////
////////////////// ////
////////////////// ////
////////////////// ////
////
////////////////// ////
Sector Elements
////
////} hϕ
ϕ =0////
////R
R 1 2

Quadrilateral Elements

Figure 5.1: Polar Coordinates Triangulation Mapping

is natural and it provides a perfect fit at the boundary (see Figure 5.1). Also, since the
equations are written in polar coordinates, the elements and the basis functions of the
approximate space for V , will be also given in polar coordinates. Thus, sector elements are
regarded as rectangles in the polar coordinate system. We consider a uniform triangulation
in the polar plane. Let nr and nϕ be the number of subdivisions in the radial and angular
direction, respectively, and let hr = (R2 − R1 )/nr and hϕ = 2π/nϕ be step sizes and
h = max{hr , hϕ }. The finite dimensional space is given by a product space of the form
V h = V1h × V2h .

5.2.1 Approximating the Velocity

We take V1h to be the space of quadratic splines defined on Ω1 = [R1 , R2 ]. The approximate
velocity is obtained by using Lagrangian quadratic elements in IR.
 49

The domain Ω1 = [R1 , R2 ] is divided into nr equal intervals,

Sih = [ri−1 , ri ], ri = i.hr + R1 , 1 ≤ i ≤ nr ,
where hr denote the step size, hr = (R2 − R1 )/nr . Let S h denote the partition of [R1 , R2 ]
given by S h = {Sih , 1 ≤ i ≤ nr }. Each element Sih has three nodes, two located at the
endpoints, ri−1 and ri , and the third one at the middle, (ri−1 +ri )/2. For each node ri there
is a basis function which takes the value 1 at r = ri and 0 at any other node. The number
of nodes in [R1 , R2 ] corresponding to the triangulation S h is 2nr + 1 but the boundary
condition imposed on the velocity, v(t, R1 ) = v(t, R2 ) = 0, implies that no basis functions
are needed at R1 or R2 , and therefore, the number of basis functions in V1h ⊂ H01 (Ω1 ) is
Nv = 2nr − 1.

Let {Φhk , 1 ≤ k ≤ Nv } be a basis of V1h , the approximate velocity v h (t, r) can be expressed
as a linear combination of {Φhk (r)},
X
Nv
v h (t, r) = vkh (t)Φhk (r) (5.1)
k=1

with vkh (t) ∈ IR, 1 ≤ k ≤ Nv .

For our computations, we selected a reference element of side 2, Sref = [−1, 1], with local
(or natural) coordinate ξ. Its nodes are located at ξ1 = −1, ξ2 = 1 and ξ3 = 0. Introduce
the local basis functions qk (ξ), 1 ≤ k ≤ 3 given by


 2
− ξ)
1 2
(ξ if k = 1
1 2
qk (ξ) =  2
+ ξ)
(ξ if k = 2

1 − ξ2 if k = 3
which have the property that qi (ξj ) = δij , 1 ≤ i, j ≤ 3.

Also, define a linear map Fk : Sref → Skh , 1 ≤ k ≤ nr by

!
hr hr
Fk (ξ) = ξ + rk + hr = rk − rk−1.
2 2
This transformation allow us to do all computations in the reference element Sref .

5.2.2 Approximating the Temperature

In order to approximate the temperature, we used the space V̂2h of piecewise bilinear func-
tions defined on Ω. As we mentioned before, the domain is divided into sector elements.
 50

We shall always take nr ≥ 3, nϕ ≥ 7 to be the number of subdivisions in the radial and

angular direction, respectively. Uniform rectangles in the polar-coordinate system are used,
with sides of length hr = (R2 − R1 )/nr and hϕ = 2π/nϕ . The partition has Ne = nr .nϕ
elements defined by

Qij = [ri , ri+1 ] × [ϕj , ϕj+1], 0 ≤ i ≤ nr − 1, 1 ≤ j ≤ nϕ

having four nodes, one at each vertex. For each node (ri , ϕj ) there exists a bilinear basis
function wich takes the value 1 at (ri , ϕj ) and 0 at all other nodes. The number of nodes
in the mesh is NT = (nr + 1).nϕ .

Denote by {Ψhk , 0 ≤ k ≤ NT − 1} a basis of V̂2h ⊂ H 1 (Ω). The approximate temperature

T̂ h (t, r, ϕ) can be expressed as a linear combination of {Ψhk (r, ϕ)},
T −1
NX
h
T̂ (t, r, ϕ) = T̂kh (t)Ψhk (r, ϕ), (5.2)
k=0

with T̂kh (t) ∈ IR, 0 ≤ k ≤ NT − 1 and periodic condition T̂ h (t, r, 0) = T̂ h (t, r, 2π). The
elements are numbered from bottom to top and from left to right as shown in Figure 5.2.
Therefore,
Q1 = [R1 , R1 + hr ] × [0, hϕ ],
Q2 = [R1 , R1 + hr ] × [hϕ , 2hϕ ]
..
.
QNe = [R2 − hr , R2 ] × [2π − hϕ , 2π].

Note that there are two nodes numbered as k.nϕ , for 0 ≤ k ≤ nr due to the periodic
condition ( since they corresponds to points with polar coordinates (rk , 0) and (rk , 2π)).

Consider a reference element, which in this case is Qref = [−1, 1] × [−1, 1], with local
coordinates (ξ, η) (see Figure 5.3). The nodes are numbered in counterclockwise direction
as follows

(ξ1 , η1 ) = (−1, −1), (ξ2 , η2 ) = (1, −1), (ξ3 , η3 ) = (1, 1), (ξ4, η4 ) = (−1, 1).

We introduce the local basis functions hk (ξ, η), 1 ≤ k ≤ 4, defined by

1
h1 (ξ, η) = (1 − ξ)(1 − η),
4
1
h2 (ξ, η) = (1 + ξ)(1 − η),
4
 51

ϕ
!!!!!!!!!!!!!!!!!!
0 n ϕ n n r. ϕ
2π
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
Qn Q nϕ QN
2 ϕ
!!!!!!!!!!!!!!!!!!
e

n −1
!!!!!!!!!!!!!!!!!!
ϕ N −1 T
2π−hϕ
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
i+1!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
i+1+n ϕ

!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
Qi+1
i!!!!!!!!!!!!!!!!!!
ϕ
!!!!!!!!!!!!!!!!!!
i+n

Γ1 !!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
Qi
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
i−1 i−1+n ϕ Γ2
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
Q
!!!!!!!!!!!!!!!!!!
3

2 hϕ
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
Q Qn
!!!!!!!!!!!!!!!!!!
2 ϕ+2

1
!!!!!!!!!!!!!!!!!!
ϕ
n +1 n n +1 r. ϕ
hϕ
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!!!!!!
Q Qn ϕ+1

!!!!!!!!!!!!!!!!!!
1

0 0
!!!!!!!!!!!!!!!!!!
n ϕ n n r. ϕ

R1 R1+hr R2−hr R2 r
Figure 5.2: Elements and Nodes in a Rectangular Grid.
 52

Gk
η 1
//////
//////
ϕj+1
//////
4 3 4 3

//////
ξ
//////
//////
//////
−1 1
//////
//////
//////
ϕ //////
//////
j
1 2 1 2
−1 //////
ri ri+1
Reference quadrilateral
Element Qk
Qref

Figure 5.3: Map Gk : Qref → Qk

1
h3 (ξ, η) = (1 + ξ)(1 + η),
4
1
h4 (ξ, η) = (1 − ξ)(1 + η).
4
We can write them in one expression (often found it in the literature) as follows
1
hk (ξ, η) = (1 + ξξk )(1 + ηηk ), k = 1, ..., 4.
4
Observe that hi (ξj , ηj ) = δij , 1 ≤ i, j ≤ 4.

Finally, we define a linear map Gk : Qref → Qk , k = 1, Ne , by

  " hr
#  !
ξ 2
0 ξ rk + h2r
Gk = hϕ + ,
η 0 2
η ϕk + h2ϕ

where hr and hϕ are defined above. All computations take place in the reference element
Qref by using the map Gk .

Remark: When we consider a basis for V1h , we threw away the functions associated to the
boundary points since V1h ⊂ H01 (Ω1 ). Here, we have included the basis functions associated
to the controlled boundary nodes. We will refer to the resulting system as the “extended
system.” Then, this extended system is reduced, eliminating the equations associated with
boundary values (which are input values). We now describe how the extended system is
obtained and the steps leading to the final reduced scheme.
 53

5.2.3 A Galerkin Scheme

We now construct a basis for V̂ h . Let N = Nv + NT and consider the set of functions
B = {Bi }N
i=1
 h   h   h       
Φ1 Φ2 ΦNv 0 0 0
= , ,..., , , ,..., ,
0 0 0 Ψh0 Ψh1 ΨhNT −1

where {Φhk , 1 ≤ k ≤ Nv } is a basis for V1h and {Ψhk , 0 ≤ k ≤ NT − 1} is a basis for V̂2h .
Clearly, B is a basis for V̂ h .
 T
The approximate solution ẑh (t) = v h (t), T̂ h (t) ∈ V̂ h to the abstract system (3.40)-
(3.41) is written as
X
N
ẑh (t) = ẑhj (t)Bj (r, ϕ),
j=1

for some ẑhj (t) ∈ IR, 1 ≤ j ≤ N. A variational formulation on V̂ h for the PDE system given
in (2.47)-(2.48) is constructed by first considering the formal equation
X
N
d h ZZ
ẑj (t) Bj (r, ϕ) ω(r, ϕ) r drdϕ =
j=1 dt Ω

X
N ZZ ZZ
ẑhj (t) ABj (r, ϕ) ω(r, ϕ) r drdϕ + f (ẑh (t)) ω(r, ϕ) r drdϕ (5.3)
j=1 Ω Ω

for ω ∈ V̂ h . Observe that ABj is not well defined since Bj ∈ / D(A). However, the weak
variational form of the equation allow us to make equation (5.3) well defined. In particular,
from equations (3.51)-(3.52), we have
* +
∂ h D E D E
v (t), ω1 = −ν ∇r v h , ∇r ω1 + LT̂ h , ω1 (5.4)
∂t L2 (Ω1 )
L2 (Ω1 ) L2 (Ω1 )
* + * +
∂ h D ∂ T̂ h E
T̂ , ω2 = −χ ∇T̂ , ∇ω2 2 + χ h
, ω2
∂t L2 (Ω)
L (Ω) ∂η L2 (Γ)
* +
h
v ∂ h
− T̂ , ω2 , (5.5)
r ∂ϕ L2 (Ω)

for ω = [ω1 ω2 ]T ∈ H01 (Ω1 )×H 1 (Ω). We now replace ω by Bk and construct a finite element
system.
* +
∂ h D E D E
v (t), Bk1 = −ν ∇r v h , ∇r Bk1 + LT̂ h , Bk1 (5.6)
∂t L2 (Ω1 )
L2 (Ω1 ) L2 (Ω1 )
 54

* + * +
∂ h D E ∂ T̂ h
T̂ , Bk2 = −χ ∇T̂ , ∇Bk2 h
+ χ , Bk2
∂t L2 (Ω)
L2 (Ω) ∂η L2 (Γ)
* +
h
v ∂ h
− T̂ , Bk2 , (5.7)
r ∂ϕ L2 (Ω)

for k = 1, . . . , N. Taking into account that the basis functions in B satisfy the following
orthogonality property
hBk , B` i = 0 1 ≤ k ≤ Nv , Nv + 1 ≤ ` ≤ N,
for 1 ≤ k ≤ Nv , Bk = Φhk , we obtain the equations
X
Nv
d h D h hE X
Nv D E
vj (t) Φj , Φk 2 = −ν vjh (t) (Φhj )0 , (Φhk )0 2
j=1 dt
L (Ω1 ) L (Ω1 )
j=1
T −1 Z 
βg NX 2π
+ T̂ h (t) cos ϕΨhj (., ϕ) dϕ, Φhk , (5.8)
2π j=0 j 0 L2 (Ω1 )

for k = 1, . . . , Nv . The above linear system of equations can be expressed as follows

d
M̂vh v h (t) = K̂vh v h (t) + K̂1h T̂ h (t) (5.9)
dt
where v h (t) ∈ V1h is represented by its coefficients in the bases of V1h , that is,
[v1h (t), . . . , vN
h
v
(t)]T , T h (t) ∈ V̂2h is represented by [T̂0h (t), . . . , T̂NhT −1 (t)]T , and the operators
M̂vh , K̂vh and K̂1h are represented by the following matrices
D E 
M̂vh = Φhi , Φhj 2 , i, j = 1, . . . , Nv , (5.10)
L (Ω1 )
"  #
 0  0
K̂vh = −ν Φhi , Φhj , i, j = 1, . . . , Nv , (5.11)
L2 (Ω1 )
and
" Z  #
βg 2π
K̂1h = Φhi , cosϕΨhj (., ϕ) dϕ , i = 1, . . . , Nv , j = 0, . . . , NT − 1, (5.12)
2π 0 L2 (Ω1 )

respectively.

For Nv + 1 ≤ k ≤ N, Bk = Ψhk−Nv −1 , setting i = k − Nv − 1 yields

T −1 T −1
NX
d h D h hE NX D E
T̂j (t) Ψj , Ψi 2 = −χ T̂jh (t) ∇Ψhj , ∇Ψhi 2
j=0 dt L (Ω)
j=0
L (Ω)

T −1
NX Z
∂Ψhj
+χ T̂jh (t) (r, ϕ)Ψhi (r, ϕ) r dϕ
j=0 Γ ∂η
D E
+ f (ẑh ), Ψhi , (5.13)
L2 (Ω)
 55

for k = 0, . . . , NT − 1, where η denotes the unit vector normal to the boundary Γ pointing
outward. Let us introduce the following matrices
D E 
M̂ =h
Ψhi , Ψhj 2 , i, j = 0, . . . , NT − 1, (5.14)
L (Ω)
D E 
K̂ = −χ
h
∇Ψhi , ∇Ψhj 2 , i, j = 0, . . . , NT − 1, (5.15)
L (Ω)
D E
f̂ h (ẑh ) = f (ẑh ), Ψhi , i = 0, . . . , NT − 1. (5.16)
L2 (Ω)

Then, equation (5.13) may be written as follows

 
NXT −1 Z
d ∂Ψhj
M̂ h T̂ h = K̂h T̂ h + f̂ h (ẑh ) + χ  T̂jh (t) (r, ϕ)Ψhi (r, ϕ) r dϕ . (5.17)
dt j=0 Γ ∂η
i=0,...,NT −1

Here T̂ h ∈ V̂2h is represented by the vector of coefficients [T̂0h , . . . , T̂NhT −1 ]T and the operators
M̂ h , K̂h , f̂ h (ẑh ) by the matrices given above. Equations (5.9) and (5.17) together, give the
system
d h
M̂vh v = K̂vh v h + K̂1h T̂ h , (5.18)
dt  
NXT −1 Z h
d ∂Ψ
M̂ h T̂ h = K̂h T̂ h + f̂ h (ẑh ) + χ  T̂jh (t) j
(r, ϕ)Ψhi (r, ϕ) r dϕ .
(5.19)
dt j=0 Γ ∂η
i=0,...,NT −1

Remark: From here on, we do not distinguish between an approximate operator and its
matrix representation. However, it must be clear that finite element methods generates an
algebraic system. Thus, the unknowns “v h ” and “T̂ h ” are actually the vector of coefficients
of v h ∈ V1h and T̂ h ∈ V̂2h with respect to the chosen basis.

5.2.4 The Reduced System

Because we allow the boundary nodes where the control is applied, we have an “extended”
vector T̂ h for the temperature which contains control values. The system of equations
for the velocity (5.18) is modified by separating the terms with boundary values of the
temperature. The reduction in the second set of the equations (5.19) (heat equation),
requires row elimination.
 56

Reducing the Velocity Equation

The equations for v h (5.18) corresponds to the Navier-Stokes equations and is given by

d h
M̂vh v = K̂vh v h + K̂1h T̂ h (5.20)
dt

where the matrices M̂vh , K̂vh , K̂1h were given in (5.10), (5.11) and (5.12), respectively.

Setting T̂0h = uh (t)|Γ1 and T̂NhT −1 = uh (t)|Γ2 , the matrix K̂1h splits in two matrices as follows
" Z  #
βg 2π
K1hv = Φhi , cosϕΨhj (., ϕ) dϕ , (5.21)
2π 0 L2 (Ω1 )

for i = 1, . . . , Nv , nϕ ≤ j < NT − nϕ and

" Z  #
βg 2π
B̂vh = Φhi , cosϕΨhj (., ϕ) dϕ , (5.22)
2π 0 L2 (Ω1 )

for i = 1, . . . , Nv , 0 ≤ j < nϕ or NT − nϕ ≤ j ≤ NT − 1. In particular, B̂vh are the columns

of K̂1h which correspond to values on the boundary of Ω.

Therefore, the approximate system for the incompressible Navier-Stokes is

d h
v = Ahv v h + Ah1v T h + Bvh uh (5.23)
dt
where  −1
Ahv = M̂vh K̂vh , (5.24)
 −1
Ah1v = M̂vh K̂1hv (5.25)
and  −1
Bvh = M̂vh B̂vh . (5.26)

Reducing the Heat Equation

Consider now the second set of equations given in (5.19), that is,
 
NXT −1 Z
h d h ∂Ψhj
M̂ T̂ = K̂ T̂ + f̂ (ẑ ) + χ
h h h h  h
T̂j (t) (r, ϕ)Ψhi (r, ϕ) r dϕ . (5.27)
dt j=0 Γ ∂η
i=0,...,NT −1
 57

The expresion inside brackets in the above equation, contains an integral on the boundary
Γ. Thus, all terms vanish except for those basis functions associated with a boundary node.
that is, for Ψhi where i = 0, . . . , nϕ − 1 or i = NT − nϕ , . . . , NT − 1. Also, we can write
Γ = Γ1 ∪ Γ2 where Γi = {(Ri , ϕ), ϕ ∈ [0, 2π)}, i = 1, 2}. Then, the outward normal unit
vector η = ηr er + ηϕ eϕ , has coordinates
(
−1 at Γ1
ηr = , ηϕ = 0.
1 at Γ2

∂Ψhj
Thus, the normal derivative ∂η
is given by
 ∂Ψh

 − j
at Γ1
∂Ψhj  ∂r
= ∇Ψhj · η =  .
∂η 
 ∂Ψhj
∂r
at Γ2

It follows that
Z
∂Ψhj
(r, ϕ)Ψhi (r, ϕ) r dϕ =
Γ ∂η
 R
 ∂Ψh



− j h
Γ1 ∂r (R1 , ϕ)Ψi (R1 , ϕ) R1 dϕ if 0 ≤ i ≤ nϕ − 1




R ∂Ψh (5.28)


j h
Γ2 ∂r (R2 , ϕ)Ψi (R2 , ϕ) R2 dϕ if NT − nϕ ≤ i ≤ NT − 1






0 otherwise

Also, the support of Ψhi is contained in the surrounding elements of the ith -node, thus for
a node i on the boundary, the support of Ψhi is contained in the two contiguous elements
that have the node i as a vertex.

Let us take 0 ≤ i ≤ nϕ − 1, i.e., a node on Γ1 , then the node i is a vertex of Qi and Qi+1
for i > 0 and for i = 0 it is a vertex of Q1 and Qnϕ (see Figure 5.2). Let us denote by
I1 the set of the 6 nodes that delimit the support of Ψhi . Let Ib1 be the subset of I1 with
nodes on Γ1 , i.e. Ib1 = I1 ∩ Γ1 . We now have for 0 ≤ i ≤ nϕ − 1

T −1
NX Z
∂Ψhj
χ T̂jh (t) (r, ϕ)Ψhi (r, ϕ) r dϕ =
j=0 Γ ∂η

X Z
∂Ψhj
−χ T̂jh (t) (R1 , ϕ)Ψhi (r, ϕ) R1 dϕ
j∈I1 Γ1 ∂r
 58

X Z 2π ∂Ψhj
= −χ T̂jh (t) (R1 , ϕ)Ψhi (R1 , ϕ) R1 dϕ
j∈Ib1 0 ∂r
X Z 2π ∂Ψhj
−χ T̂jh (t) (R1 , ϕ)Ψhi (R1 , ϕ) R1 dϕ. (5.29)
j∈I1 −Ib1 0 ∂r

From the boundary condition T (t, r, ϕ)|Γ = u(t) it follows that T̂jh (t) = uhj (t) for j ∈ Ib .
Hence, for 0 ≤ i ≤ nϕ − 1 the boundary integral term in (5.13) is
T −1
NX Z
∂Ψhj
χ T̂jh (t) (r, ϕ)Ψhi (r, ϕ) r dϕ =
j=0 Γ ∂η
X Z 2π ∂Ψhj
−χ uh1j (t) (R1 , ϕ)Ψhi (R1 , ϕ) R1 dϕ
j∈Ib1 0 ∂r
X Z 2π ∂Ψhj
−χ T̂jh (t) (R1 , ϕ)Ψhi (R1 , ϕ) R1 dϕ. (5.30)
j∈I−Ib1 0 ∂r

Analogously, for NT − nϕ ≤ i ≤ NT − 1, i.e., for i on the exterior boundary Γ2 , the integral

becomes
T −1
NX Z
∂Ψhj
χ T̂jh (t) (r, ϕ)Ψhi (r, ϕ) r dϕ =
j=0 Γ ∂η

X Z 2π ∂Ψhj
χ uh2j (t) (R2 , ϕ)Ψhi (R2 , ϕ) R2 dϕ
j∈Ib2 0 ∂r
X Z 2π ∂Ψhj
+χ T̂jh (t) (R2 , ϕ)Ψhi (R2 , ϕ) R2 dϕ, (5.31)
j∈I2 −Ib2 0 ∂r

where I2 represents the set of nodes delimiting the support of Ψhi , for NT −nϕ ≤ i ≤ NT −1.
The subset of I2 with nodes on Γ2 is Ib2 = I2 ∩ Γ2 .

Therefore, equation (5.27) can be written as

∂ h
M̂ h T̂ = K̂h T̂ h + f̂ h (ẑh ) + GhT uh (5.32)
∂t
where uh = [uh1 uh2 ]T and GhT is of the form
 
Gh1 0

 0 0 

 . 
GhT = 
 .
. ,
 (5.33)
 
 0 0 
0 Gh2
 59

where Ghi , i = 1, 2 are submatrices of dimension nϕ by nϕ .

The finite elements have the property that the basis functions have a small support. Thus,
all the matrices involved in the equations can be also written in diagonal block form, where
each block is a square matrix of dimension nϕ by nϕ . We then have nr + 1 systems of nϕ
equations. The vector of coefficients for the approximate temperature can be written as
 
T̂0h
 
 T̂1h 
 
 .. 
T̂ h =  . ,
 
 
 T̂nhr −1 
T̂nhr

where T̂kh , k = 0, . . . , nr , represent vectors of length nϕ corresponding to coefficients of T

at nodes (rk , ϕj ), j = 1, . . . , nϕ .

The matrices have the following forms

 h h

M̂00 M̂01
 h h h 
 M̂10 M̂11 M̂12 
 
 ... ... ... 
M̂ h =  , (5.34)
 
 h
M̂(n h
M̂(n h
M̂(n 
 r −1)(nr −2) r −1)(nr −1) r −1)nr 
M̂nhr (nr −1) M̂nhr nr

 
K̂00
h
K̂01
h
 h 
 K̂10 K̂11
h
K̂12
h

 
 .. .. .. 
K̂h = 

. . . ,

(5.35)
 K̂(n
h
K̂(n
h
K̂(n
h 
 r −1)(nr −2) r −1)(nr −1) r −1)nr 
K̂nhr (nr −1) K̂nhr nr
and  
Gh1 0

 0 0 

 . 
GhT = 
 .
. .
 (5.36)
 
 0 0 
0 Gh2
 60

Finally, the nonlinear term has the form

 
f̂0 (ẑh )
 
 f̂1 (ẑh ) 
 
 .. 
f̂ h (zh ) = 
 . .

(5.37)
 
 f̂(nr −1) (ẑh ) 
f̂nr (ẑh )

Note that the homogeneous Dirichlet boundary condition imposed to v(t, r) and the defi-
nition of f !T
v(t, r) ∂T
f (z(t)) = 0 − (t, r, ϕ) ,
r ∂r
implies that f̂ h does not depend on the values of T̂ h at the boundary. Therefore, there are
no nonlinear boundary terms.

The first nϕ equations in (5.32) are given by

d h h d h
h
M̂00 T̂0 + M̂01 T̂ = K̂00
h
T̂0h + K̂01
h
T̂1h + Gh1 uh1 + f̂0 (zh ). (5.38)
dt dt 1
d h
Solving for T̂
dt 0
one has
!
d h  h −1 h d h
T̂0 = M̂00 −M̂01 T̂ + K̂00
h
T̂0h + K̂01
h
T̂1h + Gh1 uh1 + f̂0 (zh ) . (5.39)
dt dt 1

The second set of equations is given by

d h h d h h d h
h
M̂10 T̂0 + M̂11 T̂1 + M̂12 T̂ = K̂10
h
T̂0h + K̂11
h
T̂1h + K̂12
h
T̂2h + f̂1 (zh ), (5.40)
dt dt dt 2
and when equation (5.39) is substituted into the above equation, it follows that
   
h −1 d h h d h
h
M̂11 − h
M̂10 M̂00 h
M̂01 T̂1 + M̂12 T̂ =

dt 
dt 2 
 −1  −1
K̂11
h
− M̂10
h h
M̂00 K̂01
h
T̂1h + K̂10
h
− M̂10
h h
M̂00 (K̂00
h
+ Gh1 ) uh1
 −1
+ f̂1 (zh ) − M̂10
h h
M̂00 f̂0 (zh ) (5.41)

where T̂0h has been replaced by uh1 .

 61

Orthogonality implies that the next equations do not depend on boundary values, then
they need not be modified. The last 2nϕ equations depend on T̂nhr which are the values of
T at Γ2 = R2 × (0, 2π). Thus, we apply a similar transformation to deal with the last two
blocks. These blocks are given by
h d h h d h h d h
M̂(n r −1)(nr −2)
T̂nr −2 + M̂(n r −1)(nr −1)
T̂nr −1 + M̂(n r −1)nr
T̂ =
dt dt dt nr
K̂(n
h
T̂ h + K̂(n
r −1)(nr −2) nr −2
h
T̂ h
r −1)(nr −1) nr −1

+ K̂(n
h
T̂ h + f̂(nr −1) (zh )
r −1)nr nr
(5.42)
and
d h d
M̂nhr (nr −1)T̂nr −1 + M̂nhr nr T̂nhr = K̂nhr (nr −1) T̂nhr −1 + K̂nhr nr T̂nhr + Gh2 uh2 + (f̂nr (zh )), (5.43)
dt dt
respectively. It follows from equation (5.43) that
d h  −1 d
T̂nr = − M̂nhr nr M̂nhr (nr −1) T̂nhr −1
dt dt
 −1  −1
+ M̂nhr nr K̂nhr (nr −1) T̂nhr −1 + M̂nhr nr K̂nhr nr T̂nhr
 −1  −1
+ M̂nhr nr Gh2 uh2 + M̂nhr nr f̂nr (zh ). (5.44)

Let  −1
h
M = M̂(n r −1)nr
M̂nhr nr .
Substituting equation (5.44) into system (5.42) and replacing T̂nhr by uh2 we obtain
d h h i d
h
M̂(n r −1)(nr −2)
T̂nr −2 + M̂(n h
r −1)(nr −1)
− M M̂ h
nr (nr −1) T̂ h =
dt dt nr −1
K̂(n
h
T̂ h + f̂(nr −1) (zh )
r −1)(nr −2) nr −2
h i
+ K̂(n
h
r −1)(nr −1)
− MK̂nhr (nr −1) T̂nhr −1
h  i
+ K̂(n
h
r −1)nr
− M K̂nhr nr + Gh2 uh2
+ f̂(nr −1) (zh ) − M f̂nr (zh ). (5.45)

System (5.32) has been reduced by 2nϕ equations and the new reduced matrices have the
form
 h h 
MT11 M̂12
 h h h 
 M̂21 M̂22 M̂23 
 
h  ... ... ... 
MT =  , (5.46)
 
 h h h
M̂(nr −2)(nr −3) M̂(nr −2)(nr −2) M̂(nr −2)(nr −1)  

h
M̂(n r −1)(nr −2)
MTh(nr −1)(nr −1)
 62

where  −1
MTh11 = M̂11
h
− M̂10
h h
M̂00 h
M̂01
and
MTh(nr −1)(nr −1) = M̂(n
h
r −1)(nr −1)
− MM̂nhr (nr −1) ,

 
KTh11 K̂12
h
 h 
 K̂21 K̂22
h
K̂23
h

 
 ... ... ... 
KTh =  , (5.47)
 
 K̂(n
h
K̂(n
h
K̂(nr −2)(nr −1) 
h
 r −2)(nr −3) r −2)(nr −2) 
K̂(n
h
r −1)(nr −2)
KTh(nr −1)(nr −1)
with  −1
KTh11 = K̂11
h
− M̂10
h h
M̂00 K̂01
h

and KTh(nr −1)(nr −1) = K̂(n

h
r −1)(nr −1)
− MK̂nhr (nr −1)

 
f̂1 (zh )
 
 f̂2 (zh ) 
 
 .. 
f̂Th (zh ) = 
 . ,

(5.48)
 h 
 f̂(nr −2) (z ) 
f̂(nr −1) (zh )
where  −1
f̂1 (zh ) = f̂1 (zh ) − M̂10
h h
M̂00 f̂0 (zh ),
and
f̂(nr −1) (zh ) = f̂(nr −1) (zh ) − M f̂nr (zh ),

 
B̂1h 0

 0 0 

 . 
B̂Th = 
 .
. ,
 (5.49)
 
 0 0 
0 B̂2h
 −1  
B̂1h = K̂10
h
− M̂10
h h
M̂00 K̂00
h
+ Gh1
where  
B̂2h = K̂(n
h
r −1)nr
− M K̂nhr nr + Gh2
 63

and    
T1h T̂1h
   
 T2h   T̂2h 
   
 ..   .. 
Th = 
 . 

= 
 . .

 Tnhr −2   
   T̂nhr −2 
Tnhr −1 T̂nhr −1

Here T h denote the approximate values of T at the interior nodes of the partition, thus
T h ∈ V2h ⊂ H01 (Ω) and the reduced system of equations for the temperature is
d h
T = AhT T h + fTh (zh ) + BTh uh , (5.50)
dt
with  −1
AhT = MTh KTh , (5.51)
 −1
BTh = Mvh B̂Th (5.52)
and  −1
fTh (zh ) = Mvh f̂Th (zh ), (5.53)

with MTh , KTh , B̂Th , f̂Th (zh ) defined above.

5.2.5 The Finite Element Model

The approximate state in V h = V1h × V2h ⊆ H01 (Ω1 ) × H01 (Ω) is

 
vh
zh = .
Th
From (5.23) and (5.50), we have
d h
v = Ahv v h + Ah1v T h + Bvh uh , (5.54)
dt
d h
T = AhT T h + fTh (zh ) + BTh uh (5.55)
dt
or equivalently,
żh (t) = Ah zh (t) + f h (zh (t)) + B h uh , (5.56)
with    
h Ahv 0 0 Ah1v
A = + ,
0 AhT 0 0
 64

" #
h h 0
f (z ) = h h ,
fT (z )

" #
h Bvh
B = .
BTh
Chapter 6

Numerical Results for the Finite

Element Approximation

In this chapter we present some results of the approximating control problem for the Ther-
mal Convection Loop. In 6.1 we describe the approximating LQR problem and the optimal
control. In Section 6.2, the Krener’s control is considered. Finally, in Section 6.3, we
present and compare the numerical results by using each method.

6.1 LQR problem for the Finite Element Approxima-

tion

The approximating system in V h is given by

ż h (t) = Ah z h (t) + f h (z h (t)) + B h uh (t), t > 0, (6.1)

z h (0) = z0h (6.2)

where z h ∈ V h , Ah , f h and B h are the approximate operators for A, B and f , respectively,

obtained by using the interpolating functions on V h .

The partial differential equations (2.47)-(2.48) are now approximated by ordinary differen-
tial equations in a finite dimensional space.

The Linear Quadratic Regulator Problem for the Boussinesq approximation was described
in Chapter 4, Section 4.2. Here we apply it to the finite element model.

65
 66

We linearize the system about the equilibrium v = 0, T = 0 obtaining

ż h (t) = Ah z h (t) + B h uh (t), t > 0, z h (0) = z0h . (6.3)
The approximate LQR problem is to minimize the quadratic cost defined by
Z ∞ Z Z 
h h h T h h h T h h
J (u ) = (z ) Q z r dr dϕ + (u ) R u dϕ dt (6.4)
0 Ω Γ

over all controls uh ∈ L2 ([0, ∞); U h ), subject to the finite dimensional linear system (6.3).
Let Ivh , ITh , IUh be appropriate matrices associated with the identity operators IL2 (Ω1 ) , IL2 (Ω)
and IL2 (Γ) , respectively. The state weighting operator for the approximate LQR problem is
 
h Qhv 0
Q = , Qhv = qv Ivh , QhT = qT ITh ,
0 QhT
and the control weighting operator is Rh = qu IUh with qv , qT and qu positive constants.

Let P h : V h → V h be the approximate Riccati operator and K h the corresponding ap-

proximate
 −1  ∗
feedback operator. The approximate feedback operator is given by K h =
Rh B h P h : V h → U h , where U h is the approximating finite dimensional control
space, U h ⊂ U = L2 (Γ). The approximate linear optimal control has the form
uh (t, ξ) = −[K h z h (t, ., .)](ξ). (6.5)

6.2 Krener’s Control for the Finite Element Approx-

imation

Once again, consider the approximate model in V h ,

ż h (t) = Ah z h (t) + f h (z h (t)) + B h uh (t), t > 0, (6.6)
z h (0) = z0h . (6.7)
The model is finite dimensional and has the same structure than the one-mode approxi-
mation defined in Chapter 4 (see equation 4.37). We apply Krener’s algorithm to (6.6).
The steps in this case are analogous to the one-mode approximation and we will not be
repeated here. The quadratic control has the form
uh (t) = −K h z h (t) + q h (z h (t)) (6.8)
where K h is the same gain matrix as in (6.5), and q h (z h ) is a vector polynomial of degree
two in z h .
 67

6.3 Comparison of the LQR and Krener Controls

In this section we present some numerical results to analyze the performance of the quadratic
control (6.8) and compare it to the optimal control given in (6.5).

6.3.1 Dimension of the Pipe and Fluid Properties

For simulation we consider a pipe with the same dimensions as the one used by Wang,
Singer and Bau in their experiments [36]. We assume that water is flowing in a pipe having
inner radius R1 = 1.1975in (36.5cm) and outer radius R2 = 1.2959in (39.5cm).

The fluid parameters are

Kinematic viscosity ν = 1.22e − 5 ft2 /s

Thermal expansion β = 8.0e − 5 /◦ F
Thermal diffusivity χ = 1.514e − 6 ft2 /s

As a initial condition for velocity and temperature, we consider a linear combination of two
of the first 4 modes.

6.3.2 State and Control Weight Operators for LQR Problem

The state and control weights play an important role in the performance of the closed loop
system. We solve the approximate LQR problem with state weight operator
 
h qv Ivh 0
Q = ,
0 qT ITh
and control weight operator Rh = qu IUh . We observed that when the identity operators are
considered as state and control weights, that is, qv = 1, qT = 1 and qu = 1, the feedback
controllers used here do not show important improvements in the behavior of the system.
On the other hand, we found out that penalizing the velocity over the temperature state,
a better performance is obtained. The results presented here correspond to state weight
operator  
1500Ivh 0
Qh = ,
0 50ITh
and control weight operator Rh = qu IUh with qu = 7.5 10−3 with different initial conditions.
 68

Temperature nodes

Radius values

R1=1.19751

1.23031

1.26312

R2=1.29593

Velocity nodes

R1 R2

Figure 6.1: Mesh for nr = 3, nϕ = 7.

6.3.3 Mesh Dimensions and Plots Interpretation

We consider nr = 3 and nϕ = 7 as the number of subdivisions in radial and angular

direction, respectively. In this case, our finite element system is of dimension 19, where
Nv = 5 are velocity unknown coefficients and NT = 14 are temperature coefficients. Figure
6.1 shows the mesh in the 2D circular pipe. The temperature nodes are marked with red
cruxes along the loop and the velocity nodes are marked with red stars in the segment
[R1 , R2 ].

To plot the velocity variations, we interpolate the resulting coefficient values vih (t), i =
1, . . . , Nv , and consider the boundary condition v(R1 ) = v(R2 ) = 0. Then we use a three
dimensional plot to show the variations of v along the time.

To plot the temperature, we first fixed the radius value to be one of the grid values. Then,
a three dimensional plot is used to show the temperature variations with respect to the
angular position along the time for the chosen radius value r.

The control is of the form u(t, ϕ) = (u1 (t, ϕ) u2 (t, ϕ))T , where u1 (t, ϕ) and u2 (t, ϕ) are the
controls applied to the inner wall (r = R1 ) and outer wall (r = R2 ), respectively. Plots of
 69

Open Loop Velocity

0.1

0.08

0.06

velocity (ft/s)
0.04

0.02

−0.02
1.2
1.25
0 100 200 300 400 500 600

time (sec)
radius (ft)

Figure 6.2: Example 1. Open Loop Velocity. Final time=600 sec.

u1 (t, ϕ) and u2 (t, ϕ) are also presented.

6.3.4 Approximating Solutions and Controls

By setting u1 = 0 the inner wall temperature is uncontrolled and we assume that is fixed
at 60◦ F . Similarly, if u2 = 0.
We considered three cases. Following we show the case when we control both walls, i.e.
u1 6= 0 6= u2 . Then we present the case when u1 (t, ϕ) = 0 (control on outer wall) since in
the other case (u2 = 0) the results are symmetric.

Control on Inner and Outer Walls

Linear Optimal Control and Krener’s Control were applied on Γ. The inner and outer wall
temperatures are given by the control components u1 (t, ϕ) and u2 (t, ϕ).

Figures 6.2, 6.3 and 6.4 show the variations of the velocity field with the time for the open-
loop system, closed-loop by using LQR control and closed loop by feeding with non-linear
control, respectively. In this example, we see that velocity of the open loop changes sign.
This indicates a change of the direction of the flow due to temperature differences along
 70

Closed Loop Velocity − LQR Control

0.1

0.08

0.06
velocity (ft/s)

0.04

0.02

−0.02
1.2
1.25
0 100 200 300 400 500 600

time (sec)
radius (ft)

Figure 6.3: Example 1. Closed Loop Velocity with LQR control. Final time=600 sec.

Closed Loop Velocity − Non−Linear Control

0.1

0.08

0.06
velocity (ft/s)

0.04

0.02

−0.02
1.2
1.25
0 100 200 300 400 500 600

time (sec)
radius (ft)

Figure 6.4: Example 1. Closed Loop Velocity with Krener’s control. Final time=600 sec.
 71

the pipe. It is interesting to see that in the controlled cases (both, LQR and Krener’s) this
effect does not appear. Also, we see that the quadratic control drives the velocity to zero
faster, then the LQR linear controller, eventhough the differences are not significant.

In Figure 6.5, we plot temperature variations along the loop and the time. Here, the radius
value is r = R1 + hr, and the plots corresponds to solutions of the open loop, closed loop
by using linear control and closed loop by using non-linear control. In Figure 6.6, the plots
correspond to r = R2 − hr. We recall that the temperature is interpreted as its deviation
from 60◦ F.
In this example, we see some improvement of the controlled systems with respect to the
open loop. If we compare a shorter time behavior, we can see bigger differences in the
temperature’s fluid, where the quadratic controller shows a better performance. This is
illustrated in Figure 6.7 for r = 1.23031in and Figure 6.8 for r = 1.26312in.

We also need to see the physical applicability of the controls. In this case, the temperature
should be maintained between 32◦ F and 212◦ F in order to maintain the properties of the
water. To check this and for comparison purpose, we plot the linear optimal control and
the non-linear control on each boundary. In Figure 6.9, we plot the controls on the inner
wall as functions of time and angular position. Figure 6.10 illustrates the controls on the
outer wall. Note that we have plotted actual temperatures for the controls. In both cases
we observe that they are kept in the desired range of temperature values. Also, we can see
that the non-linear control achieves the equilibrium faster than the LQR control.

The behavior of the control depend on the initial conditions as well as the weight operators
of the cost functional. However, in most cases, we have observed that for fixed initial condi-
tions, we can find weight operators for which the non-linear control have better performance
than the linear operator when both wall temperatures are controlled and viceversa.

Control on Outer Boundary

Observe that the pipe has a very small diameter (0.1016in). This fact motivates the
question of how good could be controlling the temperature of one wall only. We have
considered both cases, control the outer wall and the inner wall temperature. Best results
are often obtained when the outer wall temperature is controlled. That case is considered
in this section ( i.e. u1 (t, ϕ) = 0). Figures 6.11, 6.12 and 6.13 illustrate the velocity
variations for the open loop, closed loop feeded by LQR control and closed loop by using
non-linear control, respectively. As in the previous example, the open loop flow changes
direction, indicated by the sign change in the velocity. Again, the closed-loop flow remains
in the same direction achieving the stationary state faster.
 72

Open Loop Temperature

2
0
−2
0 2 400 600
4 6 200
0
time (s)
angle (rad)
Closed Loop Temperature − LQR Control

2
0
−2
0 2 400 600
4 6 200
0
time (s)
angle (rad)
Closed Loop Temperature − Non−Linear Control

2
0
−2
0 2 400 600
4 6 200
0
radius r=1.23031 time (s)
angle (rad)

Figure 6.5: Example 1. Fixed radius r = 1.23031in. Final time=600 sec.

 73

Open Loop Temperature

2
0
−2
−4
0 2 400 600
4 6 200
0
time (s)
angle (rad)
Closed Loop Temperature − LQR Control

2
0
−2
−4
0 2 400 600
4 6 200
0
time (s)
angle (rad)
Closed Loop Temperature − Non−Linear Control

2
0
−2
−4
0 2 400 600
4 6 200
0
radius r=1.26312 time (s)
angle (rad)

Figure 6.6: Example 1. Fixed radius r = 1.26312in. Final time=600 sec.

 74

Open Loop Temperature

2
0
−2
0 2 40 60
4 6 20
0
time (s)
angle (rad)
Closed Loop Temperature − LQR Control

2
0
−2
0 2 40 60
4 6 20
0
time (s)
angle (rad)
Closed Loop Temperature − Non−Linear Control

2
0
−2
0 2 40 60
4 6 20
0
radius r=1.23031 time (s)
angle (rad)

Figure 6.7: Example 1. Fluid temperatures at r = 1.23031in. Final time=60 sec.

 75

Open Loop Temperature

2
0
−2
−4
0 2 40 60
4 6 20
0
time (s)
angle (rad)
Closed Loop Temperature − LQR Control

2
0
−2
−4
0 2 40 60
4 6 20
0
time (s)
angle (rad)
Closed Loop Temperature − Non−Linear Control

2
0
−2
−4
0 2 40 60
4 6 20
0
radius r=1.26312 time (s)
angle (rad)

Figure 6.8: Example 1. Fluid temperatures at r = 1.26312in. Final time=60 sec.

 76

Control Applied at r=1.19751

Linear optimal Control
temperature (F)

70
60
50
40
30
0 60
2 40
4 20
6 0
time (sec)
angle (rad) Non−linear Control
temperature (F)

70
60
50
40
30
0 60
2 40
4 20
6 0
time (sec)
angle (rad)

Figure 6.9: Example 1. Linear and Non-linear Controllers applied on the inner wall. Final
time=60 sec.
 77

Control Applied at r=1.29593

Linear optimal Control
temperature (F)

70
60
50
40
30
0 60
2 40
4 20
6 0
time (sec)
angle (rad) Non−linear Control
temperature (F)

70
60
50
40
30
0 60
2 40
4 20
6 0
time (sec)
angle (rad)

Figure 6.10: Example 1. Linear and Non-linear Controllers applied on the outer wall. Final
time=60 sec.
 78

Open Loop Velocity

0.09

0.08

0.07

0.06
velocity (ft/s)
0.05

0.04

0.03

0.02

0.01

−0.01
1.2
1.25
0 100 200 300 400 500 600

radius (ft) time (sec)

Figure 6.11: Example 2. Open Loop Velocity. Final time=600 sec.

Closed Loop Velocity − LQR Control

0.09

0.08

0.07

0.06
velocity (ft/s)

0.05

0.04

0.03

0.02

0.01

−0.01
1.2
1.25
0 100 200 300 400 500 600

time (sec)
radius (ft)

Figure 6.12: Example 2 (Controlling Outer Wall Temperature). Closed Loop Velocity using
LQR control. Final time=600 sec.
 79

Closed Loop Velocity − Non−Linear Control

0.09

0.08

0.07

0.06

velocity (ft/s) 0.05

0.04

0.03

0.02

0.01

−0.01
1.2
1.25
0 100 200 300 400 500 600

radius (ft) time (sec)

Figure 6.13: Example 2 (Controlling Outer Wall Temperature). Closed Loop Velocity using
quadratic control. Final time=600 sec.

In Figure 6.14, we plot the temperatures as functions of time and angular position, for
the open loop and closed loops using both linear optimal control and non-linear control,
for r = 1.23031in. Although these temperatures correspond to fluid particles close to the
uncontrolled wall, we can see some improvement of the controlled systems response (linear
and non-linear) over the open loop. In Figure 6.15, we illustrate the fluid temperatures at
r = 1.26312, that is, close to the controlled wall. Here, the improvement of the closed loop
systems is even more clear.

Observe that even when the control is applied only on one wall, it acts rapidly all across
the pipe.

Again, the temperature is driven to zero faster by controlling with the quadratic feedback
control. In particular, for fluid particles close to the inner wall, r = 1.23031, we observe con-
siderable relative temperature differences with respect to the initial temperatures, between
the two controlled systems.

In Figure 6.16, we compare the LQR and the quadratic control applied on the outer wall.
Although the quadratic control shows high peaks of temperatures, these peaks are inside the
“allowed” range and they last only few seconds, after which linaer and non-linear controls
have similar values. For the given weights and initial conditions, we see that by controlling
either both walls or one wall of the pipe, the results suggest that the non-linear control
 80

Open Loop Temperature

2
0
−2
0 2 80 100
4 20 40 60
6 0
time (s)
angle (rad)
Closed Loop Temperature − LQR Control

2
0
−2
0 2 80 100
4 20 40 60
6 0
time (s)
angle (rad)
Closed Loop Temperature − Non−Linear Control

2
0
−2
0 2 80 100
4 20 40 60
6 0
radius r=1.23031 time (s)
angle (rad)

Figure 6.14: Example 2 (Controlling Outer Wall Temperature). Fluid temperatures at

r = 1.23031in. Final time=120 sec.
 81

Open Loop Temperature

2
0
−2
0 2 80 100
4 20 40 60
6 0
time (s)
angle (rad)
Closed Loop Temperature − LQR Control

2
0
−2
0 2 80 100
4 20 40 60
6 0
time (s)
angle (rad)
Closed Loop Temperature − Non−Linear Control

2
0
−2
0 2 80 100
4 20 40 60
6 0
radius r=1.26312 time (s)
angle (rad)

Figure 6.15: Example 2 (Controlling Outer Wall Temperature). Fluid temperatures at

r = 1.26312in. Final time=120 sec.
 82

Linear optimal Control

temperature (F)

70
60
50
40
30
0 1 80 100
2 3 4 40 60
5 6 0 20
time (sec)
angle (rad) Non−linear Control
temperature (F)

70
60
50
40
30
0 1 80 100
2 3 4 40 60
5 6 0 20
time (sec)
angle (rad)

Figure 6.16: Example 2. Linear and Non-linear Controls on the Outer Wall. Final
time=120 sec.
 83

Open Loop Velocity

0

−0.01

−0.02

−0.03

velocity (ft/s) −0.04

−0.05

−0.06

−0.07

−0.08

−0.09

−0.1
1.2
1.25
0 20 40 60 80 100 120

radius (ft) time (sec)

Figure 6.17: Example 3. Open Loop Velocity. Final time=120 sec.

Closed Loop Velocity − LQR Control

0

−0.01

−0.02

−0.03

−0.04
velocity (ft/s)

−0.05

−0.06

−0.07

−0.08

−0.09

−0.1
1.2
1.25
0 20 40 60 80 100 120

time (sec)
radius (ft)

Figure 6.18: Example 3 (Controlling Outer Wall Temperature). Closed Loop Velocity using
LQR control. Final time=120 sec.
 84

Closed Loop Velocity − Non−Linear Control

0

−0.01

−0.02

−0.03

−0.04
velocity (ft/s)
−0.05

−0.06

−0.07

−0.08

−0.09

−0.1
1.2
1.25
0 20 40 60 80 100 120

radius (ft) time (sec)

Figure 6.19: Example 3 (Controlling Outer Wall Temperature). Closed Loop Velocity using
non-linear control. Final time=120 sec.

drives the system faster to the desire final state than the LQR optimal control. However,
this depend not only on the weights of the cost functions but also on the initial condition.
To illustrate this, we consider the same state and control weights and change the sign of
the velocity, that is, we change the direction of the flow. Figures 6.17, 6.18 and 6.19 show
the velocity field variation in each case. Observe that the velocity remains negative and
goes to zero asymptotically in the three cases. Also, we see that the LQR control acts
faster on the velocity for this initial condition. The temperatures for each case, open loop,
closed loop by using LQR control and quadratic control, are plotted in Figures 6.20 and
6.21, for r = 1.23031in and r = 1.26312in, respectively. The controllers act rapidly on the
fluid temperature close to the controlled wall. However, the controllers do not respond as
well as before across the pipe.

In Figure 6.22, linear and non-linear controllers are plotted. Observe the similarity between
the two controllers.

For this example we consider the initial condition of the previous example and change the
sign of the velocity, that is, changing the direction of the flow. Same weights for the cost
function as before were used. However, the reponses are quite different.
 85

Open Loop Temperature

2
0
−2
0 2 80 100
4 20 40 60
6 0
time (s)
angle (rad)
Closed Loop Temperature − LQR Control

2
0
−2
0 2 80 100
4 20 40 60
6 0
time (s)
angle (rad)
Closed Loop Temperature − Non−Linear Control

2
0
−2
0 2 80 100
4 20 40 60
6 0
radius r=1.23031 time (s)
angle (rad)

Figure 6.20: Example 3 (Controlling Outer Wall Temperature). Fluid temperatures at

r = 1.23031in. Final time=120 sec.
 86

Open Loop Temperature

2
0
−2
0 2 80 100
4 20 40 60
6 0
time (s)
angle (rad)
Closed Loop Temperature − LQR Control

2
0
−2
0 2 80 100
4 20 40 60
6 0
time (s)
angle (rad)
Closed Loop Temperature − Non−Linear Control

2
0
−2
0 2 80 100
4 20 40 60
6 0
radius r=1.26312 time (s)
angle (rad)

Figure 6.21: Example 3 (Controlling Outer Wall Temperature). Fluid temperatures at

r = 1.26312in. Final time=120 sec.
 87

Linear optimal Control

70
temperature (F)

65

60

55

50
0 100
2 60 80
4 20 40
6 0
time (sec)
angle (rad) Non−linear Control

70
temperature (F)

65

60

55

50
0 100
2 60 80
4 20 40
6 0
time (sec)
angle (rad)

Figure 6.22: Example 3. Linear and Non-linear Controls on the Outer Wall. Final
time=120 sec.
 88

6.4 Conclusions

We have used the finite element model developed in Chapter 5 to compute and compare
the performance of LQR linear controllers with a non-linear controller constructed by using
Krener’s algorithm. The results are mixed. However, there are cases where significant
differences do occur. We have not addressed any of the convergence issues for the non-
linear controller. The LQR problem is discussed below.

Although there is much more to do before a theoretical resolution of these issues can be
found, the numerical evidence seems to suggest that, for the full Boussinesq equations,
non-linear controllers can enhance performance. There is some improvement over linear
controllers. However, it is not yet clear that this improvement is significant. One reason
for this may be that, if the approximate model sufficiently resolves the weakly excited
small-scale (spatial) modes, then the open-loop system can have considerable damping.
Hence, feedback controllers (neither linear nor non-linear) may not significantly enhance
the existing natural dissipation.
Chapter 7

Convergence Issues

Lasiecka and Triggiani [24], Banks and Ito [2], have shown convergence of different ap-
proximating scheme for the heat equation with boundary control when it is defined on a
finite dimensional space V h ⊂ V = [H01 (Ω1 ) × H01 (Ω)]. We used a scheme based on a space
V̂ h ⊆ [H01 (Ω1 ) × H 1 (Ω)] but which is not contain in V . In particular, the basis elements
for temperature did not vanish on Γ. A reduced finite element model was obtained by
eliminating the “extra” terms in the global finite element model. This leaves open the
question of convergence. In Section 7.2, we show the convergence of the scheme when it is
applied to the 1D heat equation.

Then, we analyze the numerical convergence of the approximating scheme to the infinite
dimensional problem. This is done by considering the gain operators K h obtained by
solving the approximating LQR problem. In finite dimensional systems, the gain operators
K h have an integral representation with a kernel hh ∈ L2 (Γ; Ω). That is, K h is a Hilbert-
Schmidt operator. Here we analyze the numerical convergence of hh . However, there is
no analytical result about the existence of a kernel for the gain operator in the infinite
dimensional case.

In Section 7.3, we consider functional gains of the approximating gain K h for the convection
loop problem. The experimental results suggest convergence of these kernels to a “limit
function” in L2 (Γ, Ω). However, the elements do not satisfy the requirement of an accurate
model (aspect ratio close to 1). Due to the dimension of the pipe, we need a big number of
elements to satisfy this condition and computer memory problems arises when solving LQR
problem. Thus, we consider the two-dimensional heat equation in a square and analyze
the convergence in that case. The description of this problem and numerical results are
presented in Section 7.4.2.

89
 90

7.1 Comments on Convergence of the Finite Element

Scheme

The fundamental issue here is the question of convergence of the solutions of approxi-
mate LQR problems to the solution of the infinite dimensional LQR problem. Sufficient
conditions may be found in [23], p.366 and in [24], p.110. We make use of the following
framework. Let Πh : V → V h be the projection operator. We assume that for z ∈ V , the
projection z h ∈ V h ⊆ V ⊆ H of z, z h = Πh z, satisfies

kz − z h kH ≤ C1 hs kzkW

for some s > 0 and constant C1 > 0. This condition satisfies most finite element schemes,
where the value of s depends on the chosen elements. The following conditions are also
needed:
There are constants ω, C2 and C3 such that
 α eωt
e−A t k ≤ C2
h
k Ah , t > 0, (7.1)
tα
and  −1
k Ah − A−1 kL(H) ≤ C3 hs . (7.2)

In [23], condition (7.1) is called the “uniform analyticity” assumption. Likewise, we assume
that there exist constants C4 and C5 such that

k(B h )∗ z h kU + kB ∗ z h kU ≤ C4 h−γs kz h kH (7.3)

and  
k B ∗ − (B h )∗ zkU + kB ∗ (I − P h )zkU ≤ C5 h(1−γ)s kzkW (7.4)
for all z ∈ W .

The following result can be found in [23],p 370.

Theorem 7.1 (Convergence of Riccati Operator) Let P h be the Riccati operator of

the approximating LQR problem, and let P be the Riccati operator of the LQR problem in
the infinite dimensional space. If conditions (7.1)-(7.4) are satisfied, then

kP h − P kL(H) → 0, as h → 0, (7.5)

kP h zkW 1− ≤ C̃1 kzkH , as h → 0 where W 1− = D(Ah∗)1− , (7.6)

 91

k(B h )∗ P h zkU ≤ C̃2 kzkH , as h → 0, (7.7)

k(B h )∗ P h − B ∗ P kL(H) → 0, as h → 0, (7.8)
where the constants C̃1 , C̃2 are uniform with respect to h.

Consider the approximating linear incompressible Navier-Stokes equation (5.23) and heat
equation (5.50) given by
v̇ h = Ahv v h + Ah1 T h + Bvh uh (7.9)
d h
T = AhT T h + BTh uh . (7.10)
dt

The approximating operators Ahv , Ah1 , Bvh , AhT and BTh are defined by
 −1
Ahv = M̂vh K̂vh , (7.11)
 −1
Ah1 = M̂vh K̂1hv , (7.12)
 −1
Bvh = M̂vh B̂vh , (7.13)
 −1
AhT = MTh KTh , (7.14)
 −1
BTh = Mvh B̂Th , (7.15)

with M̂vh , K̂vh , K̂1hv , MTh , KTh , B̂Th given in equations (5.10), (5.11), (5.21), (5.47), (5.46) and
h h h h h
(5.49), respectively. Let Av , A1 , B v , AT , B T be the operators obtained by restricting the
variational form (3.46) to V h ⊆ H01 (Ω1 ) × H01 (Ω).

Observe that the mass matrix Mvh , the stiffness matrix Kvh , the compact operator K1hv and
input operator B̂vh are derived from (3.46) by restricting v to H01 (Ω1 ). Thus, we have
h h h
Ahv = Av , Ah1 = A1 and Bvh = B v .

h h
On the other hand, the approximate operators AhT and BTh are perturbations of AT , B T .
Let Πh be the projection operator onto V h . Observe that Ah1 = A1 Πh , thus

kAh1 z h − A1 z h k ≤ kA1 Πh z − A1 zk = kA1 (Πh − I)zk → 0, as h → 0.

On the other hand, Lasiecka and Triggiani guarantee convergence of the Dirichlet operator
∆ and the associated approximating LQR problem, when the weak form is restricted to a
 92

finite element space with functions vanishing on the boundary (see [24], p.126). Banks and
Ito [2] also give a convergence proof in that case. We show that for boundary control of
the 1D heat equation, the approximation scheme based on reducing a global finite element
model converges. We rely on the fact that the corresponding scheme based on V h ⊆ H01 (Ω)
is known to converge ([2]), ([24]). The extension to the convection loop problem is somewhat
more complex and will not be considered here.

7.2 The One-Dimensional Heat Equation Scheme

The heat equation on a finite interval [0, 1] with Dirichlet boundary control is given by
∂ ∂2
T (t, x) = χ T (t, x), 0 < x < 1, t > 0, (7.16)
∂t ∂x2
T (t, 0) = u0 (t), t>0 (7.17)
T (t, 1) = u1 (t), t>0 (7.18)
T (0, x) = g(x), 0 < x < 1. (7.19)
A weak form of the above system is given by
Z Z 1 2 Z 1
1 ∂ ∂ ∂ ∂
T (t, x)φ(x)dx = χ T (t, x)φ(x)dx − χ T (t, x) φ(x)dx
0 ∂t 0 ∂x2 0 ∂x ∂x
∂ ∂
+χ T (t, 1)φ(1) − χ T (t, 0)φ(0) (7.20)
∂x ∂x

We divide the domain into equal spaced subintervals of length h = 1/(N + 1), given by
[xi , xi+1 ], xi = i/(N + 1), i = 0, . . . , N. The piecewise linear basis functions bhi (x) are
defined by
(
x1 −x
, if 0 ≤ x ≤ x1
bh0 (x) = h (7.21)
0 otherwise,



x−xi−1
h
, if xi−1 ≤ x ≤ xi
xi+1 −x
bhi (x) =  h
, if xi ≤ x ≤ xi+1 (7.22)

0 otherwise,
(
x−xN
, if xN ≤ x ≤ 1
bhN +1 (x) = h (7.23)
0 otherwise.

The space V̂ h is defined as the space of piecewise linear polynomials of the form
X
N +1
φ(x) = φk bhk (x), x ∈ [0, 1], (7.24)
k=0
 93

for real numbers φk , k = 0, . . . , N + 1. The approximate solution T̂ h (t, x) is written as a

linear combination of the basis functions. In particular,
X
N +1
T̂ h (t, x) = T̂kh (t)bhk (x), (7.25)
k=0

with T̂kh (t) ∈ IR, k = 0, . . . , N + 1. Formal differentiation of this solution yields,

∂ h X
N +1
d h
T̂ (t, x) = T̂k (t)bhk (x),
∂t k=0 dt
∂ h X
N +1
T̂ (t, x) = T̂kh (t)(bhk )0 (x),
∂x k=0

where the derivatives of bhk are given by

(
− h1 , if 0 ≤ x ≤ x1
(bh0 )0 (x) = (7.26)
0 otherwise,



1
h
, if xi−1 ≤ x ≤ xi
(bhi )0 (x) = − h1 , if xi ≤ x ≤ xi+1 (7.27)


0 otherwise,
(
1
, if xN ≤ x ≤ 1
(bhN +1 )0 (x) = h (7.28)
0 otherwise.

Also, note that the boundary conditions T (t, 0) = u0 (t), T (t, 1) = u1 (t) imply T̂0h (t) =
u0 (t), T̂Nh +1 (t) = u1 (t). Assuming for the moment that u̇0 (t) and u̇1(t) exist, then the weak
formulation (7.20) can be written in the approximating space V̂ h as

d h Z1 h
X
N +1 X
N +1 Z 1
T̂k (t) bk (x)bhj (x) = χ T̂kh (t) (bhk )0 (x)(bhj )0 (x) + χT̂Nh (t)(bhN )0 (1)bhN +1 (1)
k=0 dt 0 k=0 0

+χT̂Nh +1 (t)(bhN +1 )0 (1)bhN +1 (1) − χT̂0h (t)(bh0 )0 (0)bh0 (0)

+χT̂1h (t)(bh1 )0 (0)bh1 (0)
X
N +1 Z 1 1
= χ (bhk )0 (x)(bhj )0 (x) + χT̂N (t)(− )
T̂kh (t)
k=0 0 h
1 1 1
+χu1 (t) − χu0 (t)(− ) + χT̂1 (t) . (7.29)
h h h
Then, a system of ordinary differential equations is obtained
∂ h
M̂ h T̂ (t) = χK̂h T̂ h (t) + χĜh u(t) (7.30)
∂t
 94

h iT h iT
where T̂ h (t) = T̂0h (t), . . . , T̂Nh +1 (t) = ûh0 (t), . . . , ûh1 (t) . The matrices M̂ h , K̂h and Ĝh
are of dimension (N + 2)x(N + 2), (N + 2)x(N + 2) and (N + 2)x2, respectively, and they
are given by  
2 1 0 0 0 ... 0
1 4 1 0 0 ... 0
 
 
h  0 1 4 1 0 . . . 0 
h
M̂ =  ..  ... ... ... ..  , (7.31)
6. .
 
0 0 1 4 1
0 0 0 1 2
 
−1 0 0 0 0 ... 0
 1 −2 1 0 0 ... 0 
 
 
χ  0 1 −2 1 0 ... 0 
K̂h =  .. .. .. .. ..  , (7.32)
h . . . . . 

 
 0 0 1 −2 1 
0 0 0 0 −1
 
1 0
χ 0 0
 
Ĝh =  .. ..  . (7.33)
h. .
0 1

We apply the same procedure as for the Thermal Convection Loop. It is easy to see, that
we can eliminate the first and the last equation. The new system has no terms involving
u̇0 (t) nor u̇1 (t) and the reduced system matrices M h , Kh and Ĝh are of dimension NxN,
NxN and Nx2, respectively. In particular, we have
 
4 − 1/2 1 0 0 0 ... 0
 1 4 1 0 0 ... 0 
 
 
h  0 1 4 1 0 ... 0 
Mh =  .. ... ... ... .. ,
 (7.34)
6 . . 
 
 0 0 1 4 1 
0 0 0 1 4 − 1/2
 
−2 1 0 0 0 ... 0
 1 −2 1 0 0 ... 0 
 
 
χ  0 1 −2 1 0 ... 0 
Kh =  .. ... ... ... ..  , (7.35)
h . . 

 
 0 0 1 −2 1 
0 0 0 1 −2
 95

and  
1 0
χ 0 0

Gh =  
 .. ..  . (7.36)
h . .

0 1

Then, for T h ∈ V h ⊆ H01 (0, 1) the reduced system is given by

∂ h
T (t) = χAh T h (t) + χB h uh (t) (7.37)
∂t
h iT
where T h (t) = T̂1h (t), . . . , T̂Nh (t) , Ah = (M h )−1 Kh and B h = (M h )−1 Gh .

The matrices Kh and B̂ h are the same as for the approximating problem obtained by
restricting the variational form to V h ⊆ H01 (0, 1). Let M0h denote the mass matrix on
V h ⊆ H01 (0, 1), then it is the NxN matrix given by
 
4 1 0 0 0 ... 0
1 4 1 0 0 ... 0
 
 
h 0 1 4 1 0 ... 0
M0h =  . .. .. .. ..  . (7.38)
6 .. . . . .
 
0 0 1 4 1
0 0 0 1 4
Observe that
M h = M0h + MPh
where
h
MPh = diag(1, 0, . . . , 0, 1).
12

The finite element scheme defined by projecting onto V h is defined by

∂ h
T (t) = χAh0 T h (t) + χB0h uh (t) (7.39)
∂t
where Ah = (M0h )−1 Kh and B h = (M0h )−1 B̂ h .

Lemma 7.1 Let M h and M0h be the operators having representation matrices given in
(7.34) and (7.38), respectively. Then,
 −1
h
I
− I + h
(M0h )−1 MPh
→0 as h → 0. (7.40)

Here I h denotes the identity operator in V h .

 96

Proof: It is well known that the mass matrix M0h is invertible and approaches the identity
operator as h → 0. Thus, there exists a constant C and a constant δ1 > 0 such that
k(M0h )−1 k < C for h < δ1 . Hence, for any  > 0 there is a δ2 = δ2 () > 0 such that for
h < δ2
h
k(M0h )−1 MPh k ≤ k(M0h )−1 kkMPh k ≤ CkMPh k = C < .
12
Thus, by Theorem 1.5, Chapter IV, in Taylor and Lay [33], we have

−1
h k(M0h )−1 MPh k 
I
− I +h
(M0h )−1 MPh
≤ < , (7.41)
1 − k(M0 ) MP k
h −1 h
1−

for any  > 0 and h < δ = min{δ1 , δ2 }. 2

Lemma 7.2 Let Ah be given by Ah = (M h )−1 Kh and let Ah0 be given by Ah0 = (M0h )−1 Kh .
Then, for any z h ∈ V h we have

kAh z h − Ah0 z h k → 0, as h → 0. (7.42)

Proof: Since, M h = M0h + MPh by the previous lemma we have

 −1

h −1 h h
kA z −
h h
Ah0 z h k = Mh + Mh K z − (M0 ) K z
h h
0 P
 −1 

h −1 h h
= Mh + Mh M0 − I (M0 ) K z
h h
0 P
 −1 

= I h + (M h )−1 M h h −1 h h
− I (M0 ) K z
h
0 P
 
< k(M0h )−1 Kh z h k ≤ C2 kAzkH ,
1− 1−
for some constant C2 , since (M0h )−1 Kh Πh z converges to Az. Then, (7.42) follows. 2

By using similar arguments one can show the following lemma.

Lemma 7.3 Let B h be given by B h = (M h )−1 B̂ h and let B0h be given by B h = (M0h )−1 B̂ h .
Then, for any U h ∈ U h we have

kB h uh − B0h uh k → 0, as h → 0. (7.43)

Since equation (7.39) converges to (7.16), thus the reduced system (7.37) converges.
 97

7.3 Functional Gains

We have analyzed the numerical convergence of the control

uh (t, ξ) = −[K h z h (t, ., .)](ξ)

defined in equation (6.5) to the optimal control

uopt (t, ξ) = −[Kz(t, ., .)](ξ)

given in equation (4.1). In order to do this, we consider the gain operator K h . In a finite
dimensional system, the Riesz Representation Theorem yields an integral representation
for K h given by
Z Z
[K h z h (t, ., .)](ξ) = hh (ξ, r, ϕ)z h (t, r, ϕ) r dr dϕ
Z Ω Z Z
= hhv (ξ, r)v h(t, r) dr + hhT (ξ, r, ϕ)T h(t, r, ϕ)r dr dϕ
Ω1 Ω
= hh (ξ, ., .), z h (t, ., .)iH
h
∀z ∈ V h , ξ ∈ Γ
h

where hh (ξ, r, ϕ) ∈ L2 (Γ × Ω), i.e., K h is a Hilbert-Schmidt operator.

However, in the infinite dimensional case, it is uncertain if there exists a kernel h(ξ, r, ϕ)
such that
Z Z
[Kz(t, ., .)](ξ) = h(ξ, r, ϕ)z(t, r, ϕ) r dr dϕ
Z Ω Z Z
= hv (ξ, r)v(t, r) dr + hT (ξ, r, ϕ)T (t, r, ϕ) r dr dϕ
Ω1 Ω
= hh(ξ, ., .), z(t, ., .)iH ∀z ∈ V, ξ ∈ Γ

holds. From the numerical results shown below, we suspect that a limit function
limh→0 hh (ξ, r, ϕ) = h(ξ, r, ϕ) exists for ξ ∈ Γ, r ∈ (R1 , R2 ), ϕ ∈ (0, 2π).

The graphs presented here correspond to approximate functional gains for the control
applied on the boundary Γ2 = {R2 }×[0, 2π) for different grids. Similar results are obtained
when the interior boundary Γ1 = {R1 } × [0, 2π) is considered. Observe that the functional
gains hh are of the form  
h hhv (ξ, r)
h (ξ, r, ϕ) = h . (7.44)
hT (ξ, r, ϕ)
The kernel hhv (ξ, r) acts as a weight for the velocity, we refer to as the functional gain for
the velocity. The kernel hhT (ξ, r, ϕ) is referred to as the functional gain for the temperature.
Observe that they are defined on different domains, thus they are plotted separately.
 98

−50

−100

−150
nr=12, nphi=60

nr=10, nphi=50
−200
nr=8, nphi=40

nr=6, nphi=30
−250

−300
1.2 1.21 1.22 1.23 1.24 1.25 1.26 1.27 1.28 1.29
radius (ft)

Figure 7.1: Functional Gains hhv (π, r), r ∈ (R1 , R2 ), nr = 6, 8, 10, 12, nϕ = 5nr .

To analyze numerical converges we refine the mesh and compare the results. We do this by
refining in one or both directions. First we show the results corresponding to the functional
gain for the velocity. In Figure 7.1 the meshes have a fixed aspect radio. We considered
2π
nϕ = 5nr , or equivalently, hϕ = 5n r
= R22π hr
−R1 5
, with nr = 6, 8, 10, 12. In Figure 7.2,
we plot hv (π, r), r ∈ (R1 , R2 ) with a fixed number of angular subdivisions (nϕ = 30) and
h

refining in the radial direction (nr = 6, 10, 14, 18, 22). Observe that either fixing the aspect
ratio or the step in angular direction, similar results are obtained. However, in both cases
the aspect ratio is far away from 1. In order to see how this may affect the results, we
decrease the aspect ratio by fixing the radial step and refining in angular direction. In
particular, in Figure 7.3, the radial subdivision is fixed to be nr = 15 and the angular
subdivisions considered are nϕ = 10, 15, 20, 25, 30, 35, 40. Eventhough they are still far
from the desire aspect ratio, we see that the variations are very small as we refine the mesh
in the angular direction.

In order to see if this occurs also with the functional gain for the temperature hhT (ξ, r, ϕ),
we consider the same meshes as before. For plotting purpose we take ξ = π = ϕ and
let r ∈ (R1 , R2 ). In Figure 7.4, we plot hhT (π, r, π) at meshes of fixed aspect ratio hhϕr =
2π
5(R2 −R1 )
where the subdivisions in radial direction are nr = 6, 8, 10, 12. Figure 7.5 shows
the functional gain hhT at (π, r, π), r ∈ (R1 , R2 ), where nϕ = 30 and nr = 6, 10, 14, 18, 22.
We can see some similarity in these cases too. Figure 7.6 shows hhT (π, r, π) when nr = 15
and nϕ = 10, 15, 20, 25, 30, 35, 40.
 99

−50

−100

nr=22, nphi=30
−150
nr=18, nphi=30

nr=14, nphi=30
−200
nr=10, nphi=30

nr=6, nphi=30
−250

−300
1.2 1.21 1.22 1.23 1.24 1.25 1.26 1.27 1.28 1.29
radius (ft)

Figure 7.2: Functional Gains hhv (π, r), r ∈ (R1 , R2 ), nr = 6, 10, 14, 18, 22, nϕ = 30.

−10

−20

−30
nr=15, nphi=40

nr=15, nphi=35
−40
nr=15, nphi=30

nr=15, nphi=25
−50
nr=15, nphi=20

nr=15, nphi=15
−60
nr=15, nphi=10

−70
1.2 1.21 1.22 1.23 1.24 1.25 1.26 1.27 1.28 1.29
radius (ft)

Figure 7.3: Functional Gains hhv (π, r), r ∈ (R1 , R2 ), nr = 15, nϕ =10,15,20,25,30,35, 40.
 100

14000

12000

10000

8000
nr=12, nphi=60

nr=10, nphi=50
6000
nr=8, nphi=40

nr=6, nphi=30
4000

2000

0
1.2 1.21 1.22 1.23 1.24 1.25 1.26 1.27 1.28 1.29
radius (ft)

Figure 7.4: Functional Gains hhT (π, r, π), r ∈ (R1 , R2 ), nr = 6, 8, 10, 12, nϕ = 5nr .

9000

8000

nr=22, nphi=30
7000

nr=18, nphi=30
6000

nr=14, nphi=30
5000

nr=20, nphi=30
4000

nr=6, nphi=30
3000

2000

1000

0
1.2 1.21 1.22 1.23 1.24 1.25 1.26 1.27 1.28 1.29
radius (ft)

Figure 7.5: Functional Gains hhT (π, r, π), r ∈ (R1 , R2 ), nr = 6, 10, 14, 18, 22, nϕ = 30.
 101

10000

9000
nr=15, nphi=40
8000
nr=15, nphi=35
7000
nr=15, nphi=30
6000
nr=15, nphi=25
5000
nr=15, nphi=20
4000
nr=15, nphi=15
3000
nr=15, nphi=10
2000

1000

0
1.2 1.21 1.22 1.23 1.24 1.25 1.26 1.27 1.28 1.29
radius (ft)

Figure 7.6: Functional Gains hhT (π, r, π), r ∈ (R1 , R2 ), nr = 15, nϕ =10,15,20,25,30,35, 40.

Observe that as we consider meshes with aspect ratio closer to 1, the functional gains show
a clear non-zero support which is concentrated close to the boundary where the control
is applied. This observation indicates the possibility of the existence of a limit element
h(ξ, r, ϕ) ∈ L2 (Γ × Ω). However, we can not assure that such a limit exists.

Now we plot the functional gains for the velocity hhv (ξ, r) for ξ ∈ (0, 2π) and r ∈ (R1 , R2 ) for
different mesh sizes. In particular we consider meshes where the aspect ratio is fixed. Figure
7.7 shows hhv (ξ; r) for nϕ = 5nr with nr = 6, 8, 10, 12. Figure 7.8 shows the approximate
functional gains hhT (π; r, ϕ), with (r, ϕ) ∈ Ω on the same meshes.

Although we observe numerical convergence of the functional gains, in order to get an

accurate model, the finite element theory points out that the elements should have aspect
ratio close to 1. For the problem under consideration the radial step is hr = 0.098/nr ≈
0.1/nr . Choosing nr = 3 we have hr ≈ 0.033, thus the step in angular direction, hϕ should
be approximately 0.033 and then the number of subdivisions in angular direction should be
nϕ ≈ 190. The dimension of the system, in this case, is 385. Note that this is the smaller
case that we can consider with aspect ratio approximately 1. If we take nr = 5, nϕ = 380
the dimension of the system is about 1900. We can not study numerical convergence with
the desire aspect ratio because of the memory requirement. We thus consider the problem
of the heat equation in a square (0, 1) × (0, 1). For this case, we can consider reasonable
refinements of the mesh. Even when this is not a proof, it will give us some indication of
 102

nr=6, ns=30 nr=8, ns=40

200 200
100 100
0 0
−100 −100
−200 −200
1.2 1.2
0 1.25 0 1.25
5 radius (ft) 5 radius (ft)
angle (rad) angle (rad)
nr=10, ns=50 nr=12, ns=60

200 200
100 100
0 0
−100 −100
−200 −200
1.2 1.2
0 1.25 0 1.25
5 radius (ft) 5 radius (ft)
angle (rad) angle (rad)

Figure 7.7: Functional Gains hhv (ξ, r), r ∈ (R1 , R2 ), nr = 6, 8, 10, 12, nϕ = 5nr .
 103

nr=6, ns=30 nr=8, ns=40

10000 10000

5000 5000

0 0
1.2 1.2
5 1.25 5 1.25
0 radius (ft) 0 radius (ft)
angle (rad) angle (rad)
nr=10, ns=50 nr=12, ns=60

10000 10000

5000 5000

0 0
1.2 1.2
5 1.25 5 1.25
0 radius (ft) 0 radius (ft)
angle (rad) angle (rad)

Figure 7.8: Functional Gains hhT (π, r, ϕ), r ∈ (R1 , R2 ), ϕ ∈ (0, 2π), nr = 6, 8, 10, 12,
nϕ = 5nr .
 104

convergence.

7.4 Heat Equation Example

We consider now, the heat equation in a 2D domain. Observe that a similar abstract form
obtained for the Thermal Convection Loop applies here. We briefly describe the problem
in 7.4.1 and then show the numerical results for the approximate functional gains in 7.4.2.

7.4.1 Description of the Problem

Consider a plate occupying the region Ω = (0, 1) × (0, 1) ⊆ R2 with boundary Γ. We

shall control at the y = 0 boundary defined by Γc = {(x, y) | y = 0, 0 ≤ x ≤ 1} and let
Γu = Γ − Γc . The state space is X = L2 (Ω) and we define the operator A on the domain

D(A) = H 2 (Ω) ∩ H01 (Ω), (7.45)

by " #
2 2 ∂2w ∂2w
Aw = c ∆w = c + . (7.46)
∂x2 ∂y 2

Let  denote the lifting of A to L2 (Ω) defined from L2 (Ω) to D(A∗ )0 = D(A)0 by

Âw = ϕ if and only if ϕ(ψ) =< w, A∗ψ >L2 (Ω) for ψ ∈ D(A∗ ), (7.47)

and let D : U = L2 (Γc ) −→ L2 (Ω) denote the Dirichlet map. The controlled heat equation
is written as

" #
∂ ∂ 2 w(t, x, y) ∂ 2 w(t, x, y)
w(t, x, y) = c2 + , (7.48)
∂t ∂x2 ∂y 2
with boundary condition

w |Γu = 0, w |Γc = u(t, x) (7.49)

and initial data
 105

w(0, x, y) = w0 (x, y). (7.50)

If one defines B : U = L2 (Γc ) −→ D(A)0 by

B = −ÂD (7.51)
then the control problem may be written (in D(A)0) as

ẋ(t) = Ax(t) + Bu(t). (7.52)

Here, B is an unbounded operator when considered a a map from the control space to the
state space. We note that this problem falls within the framework considered in [24] and
[18].

Let Q = q(x, y)I with q(x, y) > 0 and R = wr IU where IU denotes the identity operator
on U. The state weight operator satisfies Q∗ = Q > 0 and wr is a positive constant. The
LQR problem is to minimize
Z ∞ Z Z 
J(u) = 2
Q(x, y)|w(t, x, y)| dxdy + wr |u(t, x)| dx dt
2
(7.53)
0 Ω Γc

over all controls u ∈ L2 ([0, ∞); U) subject to

ẇ(t) = Aw(t) + Bu(t). (7.54)

Also, for q(x, y) = 1, ∀(x, y) ∈ Ω we have Q = IL2 (Ω) , the identity operators on X. It can
be shown (see [24]) that the optimal control exists and it is given in feedback form
uopt (t) = −B ∗ P wopt(t) = −Kwopt (t), (7.55)
where P is the non-negative definite solution to the algebraic Riccati equation (ARE2)
hP w, AΦiX + hAw, P ΦiX − hB ∗ P w, B ∗P ΦiU + hw, ΦiX = 0, (7.56)
for all w, Φ in Dom(A).

7.4.2 Functional Gains

We use standard bilinear finite elements with uniform meshes on Ω with same step size in
both direction. We will denote by N either the subdivision in x-direction or y-direction,
 106

that is N = nx = ny . Let h = 1/N be the step size. Solving the approximate finite element
LQR problem produces a feedback operator
K h : L2 (Ω) → L2 (Γc ), (7.57)
with the representation
Z
h
[K w](ξ) = hh (ξ, x, y)w(x, y)dxdy, (7.58)
Ω

where hh (ξ, ., .) ∈ L2 (Ω) for each h. Moreover, standard results [24] imply that
K h w → Kw (7.59)
for each w ∈ L2 (Ω).

We shall see that hh (ξ, x, y) → h(ξ, x, y) on int(Ω) (pointwise for each fixed ξ ∈ Γc ) and
hence Z
[K h w](ξ) = hh (ξ, x, y)w(x, y)dxdy ≈ [Kw](ξ) (7.60)
Ω
provides an approximation to the optimal feedback operator.

To be precise, we consider a gold plate with thermal diffusivity c2 = .22167e−3 ft2 /s. We
consider two weighting functions. One that penalizes at the center of the plate and another
that penalizes in a non-symmetric local region.

We have consider a weight that penalizes the center of the plate (see Figure 7.9)

75, if 1/3 ≤ x, y ≤ 2/3;
q1 (x, y) = (7.61)
10, otherwise.

For each h, the finite dimensional problem has an approximate feedback operator K h with
a representation of the form

Z
h
[K φ](ξ) = hh (ξ, x, y)φ(x, y)dxdy, (7.62)
Ω

where hh (ξ, x, y) is a functional gain in L2 (Γc × Ω).

The functional gains for the approximating problem are plotted in the same manner as
in the case of the Thermal Convection Loop. Since hh (ξ, x, y) depends on 3 variables, we
fixed ξ = 0.5 for the plots. Similar figures are obtained for other values of ξ ∈ (0, 1).

In Figure 7.10, we plot hh (0.5, x, y), 0 < x, y, < 1, N = 12, 16, 20, 24. Observe that most of
ˆ x, y) is concentrated near the boundary Γc and hence the feedback law
the support of hh (ξ,
 107

Q(x,y) − minimum=10, maximum=75

80

70

60

50

40

30

20

10
1
0.8 1
0.6 0.8
0.4 0.6
0.4
0.2 0.2
0 0

Figure 7.9: Center weighted.

(7.62) is almost local in space. We can also see a peak near the boundary. This spike seems
to increase as the mesh is refined, but the limit function is integrable (at least numerically).

Observe that h(0.5, x, y) has non-zero support in int(Ω). We set x = 0.5 and plot hh (0.5, 0.5, y),
0 < y < 1 as a function of y (see Figure 7.11). Note that hh (0.5, 0.5, y) has non-zero sup-
port. This figure clearly shows that the “limit function” h(0.5, x, y) has a non-zero compact
support contained in Ω.

In Figure 7.12 we show the approximate functional gain hh (ξ, x, y), for h = 0.05 (N = 20)
and ξ = 0.1 ∗ i, i = 1, 9 when the weight q(x, y) in (7.53) is concentrated in one side of the
plate instead of the center. In particular we consider (see Figure 7.13)

2000, if 0.2 ≤ x ≤ 0.4, 0.2 ≤ y ≤ 0.8
q2 (x, y) = (7.63)
10, otherwise.

ˆ x, y) appears to have a singularity

Observe that for both choices of qi (x, y), i = 1, 2, h(ξ,
ˆ
at y = 0. Thus, it is not clear that h(ξ, x, y) is even integrable on Γc × Ω. The limited
numerical evidence obtained thus far however suggests that h(ξ, ˆ x, y) ∈ L1 (Γc × Ω).
 108

N = 12 N = 16

6 6

4 4
z

z
2 2

0 0
1 1 1 1
0.5 0.5 0.5 0.5
0 0 0 0
y x y x

N = 20 N = 24

6 6

4 4
z

2 2

0 0
1 1 1 1
0.5 0.5 0.5 0.5
0 0 0 0
y x y x

Figure 7.10: Functional Gain hh (0.5, x, y), 0 < x, y, < 1, nx = ny = 12, 16, 20, 24.
 109

hN(0.5,0.5,y)
6

3
N=24

2 N=20

N=16

1 N=12

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
y

Figure 7.11: Functional Gain hh (0.5, 0.5, y), 0 < y < 1, nx = ny = 12, 16, 20, 24.
 110

h(0.1,x,y) h(0.2,x,y) h(0.3,x,y)

15 12
15 10
10 10 8
z

z
6
5 5 4
2
1 0.5 1 1 1 1 1
0 0 x0.5
y h(0.4,x,y)
0.5 0 0 x0.5
y h(0.5,x,y)
0.5 0 0 x0.5
y h(0.6,x,y)

12 12 12
10 10 10
8 8 8
6 6
z

z
6
4 4 4
2 2 2
1 0.5 1 1 1 1 1
0 0 x0.5
y h(0.7,x,y)
0.5 0 0 x0.5
y h(0.8,x,y)
0.5 0 0 x0.5
y h(0.9,x,y)

12 12 12
10 10 10
8 8 8
6 6 6
z

4 4 4
2 2 2
1 0.5 0.5 1 1 0.5 0.5 1 1 0.5 0.5 1
y 0 0 x y 0 0 x y 0 0 x

Figure 7.12: Functional Gain hh (0.5, x, y), nx = ny = 12, 16, 20, 24.
 111

Q(x,y) − minimum=10, maximum=2000

2000

1500

1000

500

0.2

0.4

0.6

0.8 1
0.8
0.6
0.4
1 0.2
0

Figure 7.13: Side weighted.

7.5 Conclusions

We observe that the functional gains for the approximate heat equation (7.44) in Ω =
(R1 , R2 ) × (0, 2π), hhT (π, r, ϕ), π ∈ [0, 2π), (r, ϕ) ∈ Ω, are similar to the functional gains
ˆ x, y), x, y ∈ (0, 1) × (0, 1) of the approximating heat equation (7.62). In the case of
h(ξ,
the Thermal Convection Loop, we could not refine enough the mesh to assure numerical
convergence with an appropriate aspect ratio. However, in the case of the heat equation
in a square, that problem was overcome and the numerical results agree with the Thermal
Convection Loop, and convergence seems to be satisfied by the proposed scheme, although
need to be shown formally.
Conclusions

We have shown that LQR optimal control and non-linear control can be used to control a
non-linear PDE. In particular, a Boussinesq model is considered where the input operator is
unbounded. The well-posedness of the uncontrolled system is shown as well as the existence
of the optimal LQR control for the linearized system. The Lorenz equations, which are
frequently used to described the system, were also considered. The comparison of four
different controllers, showed the superiority of the nonlinear control , based on Krener’s
algorithm, over the LQR optimal control and a non-linear ad hoc control proposed by
Wang, Singer and Bau. Then, the LQR control and Krener’s non-linear control are applied
to a semidiscrete scheme obtained by using a Galerkin method. Numerical results are
mixed and more analysis are needed to decide the superiority of one over the other. The
convergence of the scheme is treated numerically by means of the functional Gains of
the finite dimensional model. We have shown the convergence of the scheme when it is
applied to a one-dimensional heat equation. However, the thermal convection case has been
analyzed only numerically.

112
Bibliography

[1] R.A. Adams. Sobolev Spaces. Academic Press, New York, 1975.
[2] H.T. Banks and K. Ito. Approximation in LQR Problems for Infinite Dimensional
Systems with Unbounded Input Operators. J. Math. Systems, Estimation and Control,
(to appear).
[3] H.T. Banks, R. Smith, and Y. Wang. Smart Material Structures. Modeling, Estimation
and Control, Masson, Paris, 1996.
[4] H.H. Bau and K.E. Torrance. Transient and Steady Behavior of an Open,
Symmetrically-Heated, Free Convection Loop. J. Heat and Mass Transfer, 24(4):597–
609, 1981.
[5] A. Bensoussan, G. Da Prato, M. Delfour, and S. Mitter. Representation and Control
of Infinite Dimensional Systems, volume I,II. Birkhäuser, Boston, 1992.
[6] P.G. Ciarlet. The Finite Element Method, volume Part 1. North-Holland, New York,
1980.
[7] P.G. Ciarlet. The Finite Element Method for Elliptic Problems, volume I,II. North-
Holland, New York, 1980.
[8] J.H. Curry, J. R. Herring, J. Loncaric, and S. A. Orszag. Order and Disorder in Two-
and Three-Dimensional Bénard Convection. J. Fluid Mechanics, 147:1–38, 1984.
[9] C. Cuvelier, A. Segal, and A.A. van Steenhoven. Finite Element Methods and Navier-
Stokes Equations. D. Reidel Pub. Co., Boston, 1986.
[10] C.R. Doering and J.D. Gibbon. Applied analysis of the Navier-Stokes Equations.
Cambridge University Press.
[11] C. Foias, O. Manley, and R. Temam. Attractors for the Bénard problem: Existence
and Physical Bounds on their Fractal Dimension. Nonlinear Analysis, Theory, Methods
and Applications, 11(8):939–967, 1987.

113
 114

[12] V. Girault and P.A. Raviart. Finite Element Methods for Navier-Stokes Equations,
Theory and Algorithms. Springer-Verlag, New York, 1986.

[13] I. Goldhirsch, R.B. Pelz, and S. A. Orszag. Numerical Simulation of Thermal Convec-
tion in a Two-Dimensional Finite Box. J. Fluid Mechanics, 199:1–28, 1989.

[14] M. Gorman and P.J. Widmann. Chaotic Flow Regimes in a Convection Loop. Physical
Review Letters, 52:2241–2244, 1984.

[15] H.F.Creveling, J.F.De Paz, J.Y. Baladi, and R.J. Schoenhals. Stability Characteristics
of a Single-Phase Free Convection Loop. J. Fluid Mech., 67:65–84, 1975.

[16] C. Johnson. Numerical Solutions of Partial Differential Equations by the Finite Ele-
ment Method. Cambridge University Press, Cambridge, 1987.

[17] D.D. Joseph. Stability of Convection in Containers of Arbitrary Shape. J. Fluid

Mechanics, 47:257–282, 1971.

[18] B.B. King. Representation of Feedback Operators for Parabolic Control Problems.

[19] A.J. Krener. Approximate Linearization by State Feedback and Coordinate Change.
Systems & Control Letters, 5:181–185, 1984.

[20] A.J. Krener, S. Karahan, M. Hubbard, and R. Frezza. Higher Order Linear Approxi-
mation to Nonlinear Control Systems. Proceedings of the 28 th Conference on Decision
and Control, pages 519–523, Dec. 1987.

[21] R. Krishnamurti. On the Transition to Turbulent Convection. Part 1. The Transition

from Two- to Three-Dimensional Flow. J. Fluid Mechanics, 42:295–307, 1970.

[22] L.D. Landau and E.M. Lifshitz. Fluid Mechanics. Pergamon Press, 1959.

[23] I. Lasiecka. Convergence Rates for the Approximations of the Solutions to Algebraic
Riccati Equations with Unbounded Coefficients: Case of Analytic Semigroups. Nu-
merische Mathematik, 63:357–390, 1992.

[24] I. Lasiecka and R. Triggiani. Differential and Algebraic Riccati Equations with Appli-
cation to Boundary/Point Control Problems: Continuous Theory and Approximation
Theory. Lecture Notes in Control and Information Sciences, volume 164. Springer-
Verlag, New York, 1991.

[25] I. Lasiecka and R. Triggiani. The Regulator Problem for Parabolic Equations with
Dirichlet Boundary Control. part ii: Galerkin Approximation. Applied Mathematics
and Optimization, 16:187–216, 1987.
 115

[26] L.J. Lions and E. Magenes. Non-homogeneous Boundary Value Problems and Appli-
cacions, volume 1. Springer-Verlag, New York, 1972.
[27] A. Mertol and R. Greig. Natural Convection: Fundamentals and Applications, 1985.
[28] A.R. Mitchel and R. Wait. The Finite Element Method in Partial Differential Equa-
tions. John Wiley and Sons, New York, 1977.
[29] A. Pazy. Semigroup of Linear Operator and Applications to Partial Differential Equa-
tions. Springer-Verlag, New York, 1983.
[30] I.G. Rosen. On Hilbert-Schmidt Norm Convergence of Galerkin Approximation for
Operator Riccati Equations. International Series of Numerical Mathematics, 91:335–
349, 1989.
[31] J. Singer, Y. Wang, and H. H. Bau. Controlling a Chaotic System. Phys. Rev. Lett.,
(66):1123–1125, 1991.
[32] C. Sparrow. The Lorenz Equations: Bifurcations, Chaos and Strange Attractors.
Springer, Berlin, 1982.
[33] A.E. Taylor and D.C. Lay. Introduction to Functional Analysis. Krieger Publ.Co.,
Malabar, Florida, 1986.
[34] D.J. Tritton. Physical Fluid Dynamics. Van Nostrand Publ, 1979.
[35] T.L. Vincent and J. Yu. Control of a Chaotic System. PJ. Dynamics and Control,
1:35–52, 1991.
[36] Y. Wang, J. Singer, and H. H. Bau. Controlling Chaos in a Thermal Convection Loop.
J. Fluid Mech., 237:479–498, 1992.
[37] P. Welander. On the Oscillatory Instability of a Differentially Heated Fluid Loop. J.
Fluid Mech., 29:7–30, 1967.
[38] J. Wloka. Partial Differential Equations. Cambridge University Press, Cambridge,
1987.
[39] T. Yanagita and K. Kaneko. Rayleigh-Bénard Convection. Patterns, Chaos, Spa-
tiotemporal Chaos and Turbulence. Physica D, 82:288–313, 1995.
[40] J.A. Yorke, E. D. Yorke, and J. Mallet-Paret. Lorenz-like Chaos in a Partial Differential
Equation for a Heated Fluid Loop. Physica, 24D:279–291, 1987.
[41] O.C. Zienkiewicz and R.L. Taylor. The Finite Element Method, volume 1. Mc Graw-
Hill, England, 1989.
Vita

Aurora Diana Rubio was born in Buenos Aires, Argentina on August 29, 1959. Diana grad-
uated with a B.S./M.S. in Applied Mathematics in May 1988 from Universidad de Buenos
Aires, Argentina, where she was a teaching assistant from 1984 to 1992 and a research
fellow in 1987. In 1988, she began working as a research assistant for the (CNEA) National
Commission of Atomic Energy, Buenos Aires, Argentina. She would leave Argentina in
August, 1992 to continue graduate studies in the United States of America. She received
a M.S. and a Ph.D. degree in Mathematics from Virginia Polytechnic Institute and State
University in 1994 and 1997, respectively.

116