In chapter 5 we zoomed out and looked at a differential
equation as a whole. Its solutions form a solution set S, and
its initial conditions form an initial-condition space I.
Evaluating at 𝑡0 matches each solution with its initial condition,
one-to-one. So a whole function is pinned down by finitely many numbers.
That picture tells us how many numbers to prescribe, but not how to find
the solution they pick out. In general, each initial condition is a new
problem. Solving from one initial condition doesn't necessarily give us
the solution from another.
In some cases, though, things are beautifully simple. For the spring, once
we know cos𝑡 and sin𝑡, we know every solution: each one is
𝐴cos𝑡+𝐵sin𝑡, built from those two by scaling and adding. Two
solutions produce all the others, with no more calculus. Which equations
let us build every solution from a few by adding and scaling? And when
they do, how much of the work of solving can we hand over to linear
algebra?
6.1Linear Equations and Their Solution Sets
The two spaces from chapter 5 look quite different. The
initial-condition space is easy to describe. An initial condition is a
list of numbers, a vector in ℝ𝑁, and usually I is
all of ℝ𝑁. We can add initial conditions and multiply them by
numbers: I is a vector space.
The solution set is harder to see. It is a subset of a space of
functions, and a subset of a vector space need not be a vector space
itself. It is one exactly when it is a subspace: it contains zero,
and sums and multiples of its members stay inside it. For a solution
set, that means the zero function solves the equation, and sums and
multiples of solutions are solutions again.
The spring passes this test. If 𝑢 and 𝑣 solve 𝑦″=−𝑦, then
(𝑢+𝑣)″=𝑢″+𝑣″=−𝑢−𝑣=−(𝑢+𝑣),
so 𝑢+𝑣 is a solution too. Similarly, (𝑐𝑢)″=𝑐𝑢″=−𝑐𝑢, and the zero
function solves 𝑦″=−𝑦. So the spring's solution set is a subspace: the
functions 𝐴cos𝑡+𝐵sin𝑡 really do form a plane through the origin.
Notice that we never used that formula. The calculation works for any
solutions at all.
The logistic equation 𝑦′=𝑦(1−𝑦) fails. The constant function 𝑦=1 is
a solution, since both sides are zero. But 𝑦=2 is not: its derivative
is 0, while 2(1−2)=−2. To see what goes wrong in general, write the
equation as 𝑦′=𝑦−𝑦2 and add two solutions 𝑢 and 𝑣:
(𝑢+𝑣)′=𝑢′+𝑣′=(𝑢−𝑢2)+(𝑣−𝑣2).
For 𝑢+𝑣 to be a solution, the equation would require
(𝑢+𝑣)′=(𝑢+𝑣)−(𝑢+𝑣)2=(𝑢−𝑢2)+(𝑣−𝑣2)−2𝑢𝑣.
The extra term −2𝑢𝑣 vanishes only where one of the two solutions does.
Scaling breaks the same way: (𝑐𝑢)′=𝑐𝑢−𝑐𝑢2, while the equation demands
𝑐𝑢−𝑐2𝑢2.
So the difference lies in the equations themselves. The spring's equation
𝑦″+𝑦=0 is built from differentiating and adding, and both respect sums
and multiples. The logistic equation contains a square, which does not.
Differentiation already respects the arithmetic we need. For functions
𝑢,𝑣∈𝐶1(𝐼) and numbers 𝑎,𝑏,
(𝑎𝑢+𝑏𝑣)′=𝑎𝑢′+𝑏𝑣′.
So differentiation takes a function in 𝐶1(𝐼) to a function in 𝐶(𝐼),
and it respects sums and multiples. Maps like this deserve a name.
Definition 6.1 (Linear operator). An operator is a map that takes functions to functions. An operator
𝐿 is linear if
𝐿[𝑎𝑢+𝑏𝑣]=𝑎𝐿[𝑢]+𝑏𝐿[𝑣]
for all functions 𝑢,𝑣 in its domain and all real numbers 𝑎,𝑏.
A linear operator is just a linear map, in the sense of linear algebra,
between two spaces of functions. Differentiation is one. Multiplying by a
known function of time is another: for any coefficient 𝑝(𝑡), we have
𝑝(𝑎𝑢+𝑏𝑣)=𝑎𝑝𝑢+𝑏𝑝𝑣. Sums of linear operators are linear, and so are
compositions. For the spring, consider
𝐿:𝐶2(𝐼)⟶𝐶(𝐼),𝐿[𝑦]=𝑦″+𝑦.
It differentiates a function twice and adds the original function. Both
steps respect linear combinations, so
Definition 6.2 (Linear differential equation). A linear differential equation is a differential equation that can be
written
𝐿[𝑦]=𝑔,
where 𝐿 is a linear operator and 𝑔 is a known function, called the
forcing. The unknown 𝑦 may be a single function or, for a system, a
vector-valued function. The equation is homogeneous when 𝑔=0 and
forced otherwise. A differential equation that cannot be written this
way is nonlinear.
Two special cases cover every linear equation in this book. The first is a
scalar equation of order 𝑛,
𝑎𝑛(𝑡)𝑦(𝑛)+⋯+𝑎1(𝑡)𝑦′+𝑎0(𝑡)𝑦=𝑔(𝑡),
where the coefficients 𝑎𝑘 and the forcing 𝑔 are known functions of
𝑡. For 𝑛=1, this includes the first-order linear equations
𝑦′=𝑓(𝑡)𝑦+𝑔(𝑡) from chapter 3. The left side is a linear operator for
the same reason 𝑦″+𝑦 is: differentiating and multiplying by a known
function both distribute over sums and let constants pull out.
Example 6.3 (Linear and nonlinear equations).
The spring 𝑦″+𝑦=0 is linear and homogeneous, with constant
coefficients.
The equation 𝑦′−𝑡𝑦=0 is linear and homogeneous. Its coefficient −𝑡
changes with time, but it is a known function.
The equation 𝑦′=2𝑡 is linear and forced, with forcing 𝑔(𝑡)=2𝑡.
The logistic equation 𝑦′=𝑦−𝑦2 is nonlinear, because of the term
𝑦2.
The second is a first-order system in which each rate is a combination of
the unknowns with known coefficients, plus a known function:
𝑥′𝑘=𝑎𝑘1(𝑡)𝑥1+⋯+𝑎𝑘𝑁(𝑡)𝑥𝑁+𝑔𝑘(𝑡),𝑘=1,…,𝑁.
Its operator sends 𝑋 to the vector whose 𝑘th component is
𝑥′𝑘−𝑎𝑘1(𝑡)𝑥1−⋯−𝑎𝑘𝑁(𝑡)𝑥𝑁. This is linear for the same
reason as before.
Matrices let us write this much more compactly. Let 𝐴(𝑡) be the
𝑁×𝑁 matrix whose entries are the coefficients 𝑎𝑘𝑗(𝑡), and let
𝑔(𝑡) be the vector with components 𝑔𝑘(𝑡). The 𝑘th component of
𝐴(𝑡)𝑋 is exactly 𝑎𝑘1(𝑡)𝑥1+⋯+𝑎𝑘𝑁(𝑡)𝑥𝑁. We call 𝐴(𝑡) the
coefficient matrix.
Proposition 6.4 (Linear first-order systems). A first-order system 𝑋′=𝐹(𝑡,𝑋) is linear exactly when it can be written
𝑋′=𝐴(𝑡)𝑋+𝑔(𝑡),
where the coefficient matrix 𝐴(𝑡) and the vector 𝑔(𝑡) are known
functions of 𝑡. Its operator is 𝐿[𝑋]=𝑋′−𝐴(𝑡)𝑋, and its forcing is 𝑔.
Proof. If the system has this form, then 𝐿[𝑋]=𝑋′−𝐴(𝑡)𝑋 is linear, and the
system is 𝐿[𝑋]=𝑔. Conversely, suppose 𝑋′−𝐹(𝑡,𝑋)=𝐿[𝑋]−𝑔 for a linear
operator 𝐿. Plug in a constant vector 𝑐. Its derivative is zero, so
𝐹(𝑡,𝑐)=𝑔(𝑡)−𝐿[𝑐](𝑡). The map 𝑐↦−𝐿[𝑐](𝑡) is linear on
ℝ𝑁, so at each time it is multiplication by some matrix 𝐴(𝑡).
Thus 𝐹(𝑡,𝑋)=𝐴(𝑡)𝑋+𝑔(𝑡). ∎
Example 6.5 (A linear system). The system 𝑥′=2𝑦,𝑦′=−3𝑥 is 𝑋′=𝐴𝑋 with
𝐴=(02−30),
so it is linear and homogeneous. We can find its solutions using the
spring equation we already know. Differentiate 𝑥′=2𝑦 and substitute
𝑦′=−3𝑥:
𝑥″=2𝑦′=−6𝑥.
So 𝑥 oscillates with frequency 𝜔=√6, and we try
𝑥(𝑡)=𝑎cos(𝜔𝑡)+𝑏sin(𝜔𝑡).
The initial position gives 𝑎=𝑥0. The system also tells us that
𝑥′(0)=2𝑦0, so 𝑏𝜔=2𝑦0. Once we have 𝑥, the first equation gives
𝑦=𝑥′/2. Thus
These functions have the prescribed initial values. We constructed them so
that 𝑥′=2𝑦, and 𝑦′=𝑥″/2=−3𝑥, so they satisfy both equations.
Example 6.6 (Linear and nonlinear systems). The system 𝑥′=𝑡𝑥+𝑦,𝑦′=−𝑥+sin𝑡 is linear and forced, with
𝐴(𝑡)=(𝑡1−10),𝑔(𝑡)=(0sin𝑡).
The system 𝑥′=𝑦,𝑦′=−𝑥2 is nonlinear, because of the term 𝑥2.
The spring equation is the homogeneous equation 𝐿[𝑦]=0 with
𝐿[𝑦]=𝑦″+𝑦. Its solutions are the functions that 𝐿 sends to zero: the
kernel of 𝐿. Kernels of linear maps are always subspaces, and that is
what our calculation for the spring was really showing.
Theorem 6.7 (Homogeneous linear equations). If 𝐿 is a linear operator, then the solution set of the differential
equation 𝐿[𝑦]=0 is a vector space.
Proof. The solutions are the kernel of 𝐿. It contains the zero function, since
𝐿[0]=0. If 𝐿[𝑢]=𝐿[𝑣]=0, then 𝐿[𝑎𝑢+𝑏𝑣]=𝑎𝐿[𝑢]+𝑏𝐿[𝑣]=0. ∎
In other words, sums and multiples of solutions are solutions again. So the
spring's solution set is a subspace because the operator 𝑦″+𝑦 is linear.
The same goes for 𝑦′−𝑡𝑦=0, for the system 𝑥′=2𝑦,𝑦′=−3𝑥, and for every
homogeneous linear equation, whether or not we can find a single solution.
The word linear describes this arithmetic of functions, not the shapes of
their graphs: our solutions can grow exponentially or oscillate and still
belong to a vector space.
What about forced equations? We have been solving one since
antidifferentiation: the solutions of 𝑦′=2𝑡 are the functions 𝑦=𝑡2+𝐶.
This solution set is not a vector space. The zero function is not a
solution, and adding two solutions gives 2𝑡2+𝐶1+𝐶2, whose derivative
is 4𝑡. But subtracting two solutions gives the constant 𝐶1−𝐶2, which
solves the homogeneous equation 𝑦′=0. So every solution is 𝑡2 plus a
solution of 𝑦′=0. Geometrically, the solution set is the line of constant
functions, shifted away from the origin by 𝑡2.
Definition 6.8 (Affine space). An affine space in a vector space 𝑉 is a set of the form
𝑝+𝑊={𝑝+𝑤:𝑤∈𝑊},
where 𝑊 is a subspace of 𝑉 and 𝑝 is a point of 𝑉.
The point 𝑝 locates the affine space, and the subspace 𝑊 gives its
directions. The solutions of 𝑦′=2𝑡 form the affine space 𝑡2+𝑊, where
𝑊 is the space of constant functions. The constants of integration were
describing an affine space all along.
Theorem 6.9 (Forced linear equations). Let 𝐿 be a linear operator, and suppose 𝑢𝑝 solves 𝐿[𝑦]=𝑔. Then the
solution set of 𝐿[𝑦]=𝑔 is the affine space 𝑢𝑝+ker𝐿.
Proof. If 𝑢 is any solution, then 𝐿[𝑢−𝑢𝑝]=𝐿[𝑢]−𝐿[𝑢𝑝]=𝑔−𝑔=0, so 𝑢−𝑢𝑝
belongs to ker𝐿. Conversely, if ℎ∈ker𝐿, then
𝐿[𝑢𝑝+ℎ]=𝐿[𝑢𝑝]+𝐿[ℎ]=𝑔+0=𝑔, so 𝑢𝑝+ℎ is a solution. ∎
So every solution of a forced linear equation is one particular solution
plus a solution of the homogeneous equation. When 𝑔=0 we can take
𝑢𝑝=0, and we recover the vector space ker𝐿. The theorem assumes that
we already have one solution 𝑢𝑝. It does not say that one exists.
In chapter 3, existence and uniqueness for 𝑦′=𝐹(𝑡,𝑦) was local: it
promised a solution only near 𝑡0. For a linear equation
𝑦′=𝑓(𝑡)𝑦+𝑔(𝑡) it became global, and the solution exists on the whole
interval where 𝑓 and 𝑔 are continuous. The same thing happens for
systems. theorem 5.10 is local, and for linear systems it
becomes global.
Theorem 6.10 (Existence and uniqueness for linear systems). Suppose the coefficients 𝑎𝑘𝑗(𝑡) and the forcing terms 𝑔𝑘(𝑡) are
continuous on an open interval 𝐼, and let 𝑡0∈𝐼. For every
𝑋0∈ℝ𝑁, the initial value problem
𝑋′=𝐴(𝑡)𝑋+𝑔(𝑡),𝑋(𝑡0)=𝑋0
has exactly one solution, and it exists on all of 𝐼.
So every initial condition has a solution on all of 𝐼. In particular, a
forced linear system always has a particular solution, so its solution set
is 𝑢𝑝+ker𝐿.
The same holds for scalar equations, through their first-order form.
Corollary 6.11 (Existence and uniqueness for linear equations of order $n$). Suppose 𝑎0,…,𝑎𝑛 and 𝑔 are continuous on 𝐼, and 𝑎𝑛 is never
zero there. Then for every initial condition
(𝑦(𝑡0),…,𝑦(𝑛−1)(𝑡0)), the equation
𝑎𝑛(𝑡)𝑦(𝑛)+⋯+𝑎0(𝑡)𝑦=𝑔(𝑡)
has exactly one solution, and it exists on all of 𝐼.
Dividing by 𝑎𝑛 makes the first-order form a linear system with
continuous coefficients, so the theorem applies.
6.2Evaluation and Solution Maps
In chapter 5 we connected the solution set and the
initial-condition space by the evaluation map 𝐸𝑡0, which sends each
solution to its initial condition. For a homogeneous linear equation, both
of these are vector spaces. Does evaluation respect their arithmetic? For
solutions 𝑈,𝑉 and numbers 𝑎,𝑏,
For a scalar equation, evaluation also records derivatives at 𝑡0, and
differentiation respects sums and multiples too. So evaluation is a linear
map.
Theorem 6.12 (Evaluation is an isomorphism). Consider a homogeneous linear system whose coefficients are continuous on
𝐼, or a homogeneous linear equation of order 𝑛 whose coefficients are
continuous on 𝐼 with 𝑎𝑛 never zero there. Then evaluation
𝐸𝑡0:S→I is a linear isomorphism. In particular,
dimS=dimI.
Proof. Evaluation is linear, as we just checked. It is one-to-one by uniqueness,
corollary 5.11. It is onto because every initial condition has
a solution on all of 𝐼, by theorem 6.10 and
corollary 6.11. A linear map that is one-to-one and onto is
an isomorphism, and isomorphic vector spaces have the same dimension. ∎
So we know the dimension of the solution set without finding a single
solution. The ambient space is infinite-dimensional, but the solutions of a
linear equation form a subspace whose dimension is the number of values in
an initial condition. The solutions of 𝑦′=𝑦 form a line, and those of the
spring form a plane. Those of 𝑥‴=𝑥″ form a three-dimensional space, and
those of the system 𝑥′=2𝑦,𝑦′=−3𝑥 form a plane in 𝐶1(𝐼,ℝ2).
This turns finding every solution into a finite job. In an 𝑁-dimensional
space, any 𝑁 linearly independent vectors form a basis. So if we find
dimI independent solutions, by any means at all, then every
solution is a combination of them, and we have found them all.
The solution map from chapter 5 improves too. Every initial
condition has a solution on all of 𝐼, so the solution map is defined on
all of I. As the inverse of a linear isomorphism, it is linear
itself: adding initial conditions adds their solutions, and scaling an
initial condition scales its solution.
A forced equation's solutions do not form a vector space, but by
theorem 6.9 they form an affine space 𝑢𝑝+S, where
S is now the solution set of the homogeneous equation 𝐿[𝑦]=0.
Global existence guarantees that a particular solution 𝑢𝑝 exists. The
dimension of an affine space 𝑝+𝑊 means the dimension of its directions
𝑊, so the solutions of a forced linear equation form an affine space of
dimension dimI. Finding all of them is a finite job too: one
particular solution, plus dimI independent solutions of the
homogeneous equation.
For example, consider 𝑦″+𝑦=𝑡2. For the homogeneous equation 𝑦″+𝑦=0,
we already have two solutions, cos𝑡 and sin𝑡. We found them by
guessing: calculus gave us functions that return to their negatives after
two derivatives. They are independent, since neither is a multiple of the
other.
We can find a particular solution the same way. The forcing is a
polynomial, and differentiating a polynomial gives another polynomial, so
we try 𝑦=𝑡2+𝑐. Its second derivative is 2, so we need 2+𝑡2+𝑐=𝑡2,
which means 𝑐=−2. Check: 𝑡2−2 has second derivative 2, and
2+(𝑡2−2)=𝑡2.
We found these functions by recognizing how their derivatives behave. The
theorem tells us something our guesses alone could not: they are enough to
give every solution. Every solution of 𝑦″+𝑦=𝑡2 is
𝑡2−2+𝑎cos𝑡+𝑏sin𝑡
for some numbers 𝑎 and 𝑏. The solution set is a plane in
𝐶2(ℝ), shifted away from the origin by 𝑡2−2.
6.3Bases of Solutions
To compute with vectors, we choose a basis and work in coordinates. The
initial-condition space comes with coordinates already: an initial
condition is a list of numbers. To give the solution set coordinates too,
we need a basis of solutions, and evaluation tells us when we have one.
Corollary 6.13 (Bases of solutions). Under the hypotheses of theorem 6.12, solutions
𝑈1,…,𝑈𝑁 form a basis of S exactly when their initial
conditions 𝐸𝑡0(𝑈1),…,𝐸𝑡0(𝑈𝑁) form a basis of
I.
This holds because an isomorphism carries bases to bases, in both
directions.
So we can tell whether whole functions are independent by looking at a
single time. For arbitrary functions this would be hopeless: 𝑡 and 𝑡2
are independent, yet both are zero at 𝑡=0. For solutions, uniqueness
makes it work. A combination of solutions whose initial condition is zero
is the zero solution everywhere.
Now fix a basis 𝑈1,…,𝑈𝑁 of S, and use the standard
basis on I=ℝ𝑁. Every solution is
𝑈=𝑐1𝑈1+⋯+𝑐𝑁𝑈𝑁 for unique coefficients 𝑐1,…,𝑐𝑁, and
since evaluation is linear,
𝐸𝑡0(𝑈)=𝑐1𝐸𝑡0(𝑈1)+⋯+𝑐𝑁𝐸𝑡0(𝑈𝑁)=𝑀𝑐,
where 𝑐=(𝑐1,…,𝑐𝑁) and 𝑀 is the matrix whose columns are the
initial conditions of the basis solutions:
𝑀=(𝐸𝑡0(𝑈1)⋯𝐸𝑡0(𝑈𝑁)).
This 𝑀 is the matrix of evaluation in these bases. It is invertible,
because its columns form a basis of I. The solution map is the
inverse of evaluation, so its matrix is 𝑀−1. To solve from an initial
condition 𝑋0, compute
𝑐=𝑀−1𝑋0,
and the solution is 𝑐1𝑈1+⋯+𝑐𝑁𝑈𝑁. Once we have a basis, solving
any initial value problem is a matrix computation.
Example 6.14 (Solving the spring equation). For the spring 𝑦″=−𝑦, we already know two solutions: cos𝑡 and
sin𝑡. Their initial conditions at 𝑡=0 are (1,0) and (0,1), the
standard basis of I=ℝ2. So by
corollary 6.13, these two functions form a basis of S.
Their initial conditions are also the columns of 𝑀, so 𝑀=𝐼 and
𝑀−1=𝐼. The coordinates of a solution in this basis are exactly its
initial position and velocity! The solution from (𝑦0,𝑣0) is
𝑦0cos𝑡+𝑣0sin𝑡.
The same works for any homogeneous linear equation. Solve from each
standard basis vector of I.
Definition 6.15 (Standard solutions). The standard solutions at 𝑡0 are the solutions 𝑅1,…,𝑅𝑁
whose initial conditions are the standard basis vectors:
𝐸𝑡0(𝑅𝑗)=𝑒𝑗.
Their evaluation matrix is the identity, so the solution from
𝑋0=(𝜉1,…,𝜉𝑁) is 𝜉1𝑅1+⋯+𝜉𝑁𝑅𝑁. But we don't have
to aim at particular initial conditions. Any dimI solutions
will do, however we find them, as long as their initial conditions are
independent. Then 𝑀 need not be the identity, and we invert it.
Example 6.16 (Solving $y''=y$). The equation 𝑦″=𝑦 asks for functions that equal their own second
derivative. Calculus gives us two: 𝑒𝑡 and 𝑒−𝑡. Their initial
conditions at 𝑡=0 are (1,1) and (1,−1), which are independent, so
𝑒𝑡 and 𝑒−𝑡 form a basis. Directly: their ratio 𝑒𝑡/𝑒−𝑡=𝑒2𝑡
is not constant. Two guessed solutions, and we have every solution:
𝑎𝑒𝑡+𝑏𝑒−𝑡.
Evaluation sends 𝑎𝑒𝑡+𝑏𝑒−𝑡 to (𝑎+𝑏,𝑎−𝑏), so
𝑀=(111−1),𝑀−1=12(111−1).
To solve from 𝑦(0)=3, 𝑦′(0)=1, multiply:
𝑀−1(31)=(21),
so the solution is 2𝑒𝑡+𝑒−𝑡. Check: its value at 0 is 2+1=3, and
its derivative at 0 is 2−1=1.
Example 6.17 (Solving the forced spring). In section 6.2 we found every solution of 𝑦″+𝑦=𝑡2:
𝑡2−2+𝑎cos𝑡+𝑏sin𝑡. To solve from an initial condition (𝑦0,𝑣0),
split off the particular solution. Its initial condition is (−2,0),
since 𝑡2−2 has value −2 and derivative 0 at 0. So the part
𝑎cos𝑡+𝑏sin𝑡 must have initial condition (𝑦0+2,𝑣0). In the
spring's standard basis, its coordinates are exactly that initial
condition: 𝑎=𝑦0+2 and 𝑏=𝑣0. The solution is
𝑦=𝑡2−2+(𝑦0+2)cos𝑡+𝑣0sin𝑡.
For example, 𝑦(0)=0, 𝑦′(0)=1 gives 𝑡2−2+2cos𝑡+sin𝑡. Check: its
value at 0 is −2+2=0, and its derivative 2𝑡−2sin𝑡+cos𝑡 is 1 at
0. Three guessed solutions now solve every initial value problem for this
equation.
Starting from any basis, we can always change to the standard solutions.
Each standard solution is a solution, so it is some combination of the
basis solutions. For 𝑦″=𝑦, finding those combinations produces two new
functions, the hyperbolic cosine and sine, which you can discover in the
exercises.
6.4Example: The Damped Spring
In section 4.4 we added drag to the spring and got the model
𝑚𝑥″=−𝑘𝑥−𝑏𝑥′. Take the mass to be 1, so the equation is
𝑥″+𝑏𝑥′+𝑘𝑥=0,
with spring constant 𝑘>0 and drag coefficient 𝑏≥0. In Part I we had
no way to solve this equation, so we computed its motions numerically. Now
we can do better.
This is a homogeneous linear equation of order 2, so by
theorem 6.12 its solutions form a two-dimensional space.
Before solving anything, we know that two independent solutions will give
us every solution. How do we find two? Let's look at the motions first.
Figure 6.1 Solutions of 𝑥″+𝑏𝑥′+𝑥=0 starting from 𝑥(0)=1 and 𝑥′(0)=0. Move the
drag coefficient 𝑏. At 𝑏=0 the mass oscillates forever. For small 𝑏
it still oscillates, but its swings shrink. For large 𝑏 it doesn't
oscillate at all: it creeps back toward 0.
When 𝑏=0 there is no drag, and the equation is the spring 𝑥″+𝑘𝑥=0,
with solutions cos(√𝑘𝑡) and sin(√𝑘𝑡). For small 𝑏,
the motion still looks like a cosine, but one whose swings shrink
exponentially. So let's try
𝑥=𝑒−𝑎𝑡cos(𝜔𝑡),
with numbers 𝑎 and 𝜔 to be found. Differentiating,
For this to be zero at every time, both coefficients must vanish.
Evaluating at 𝑡=0 shows that the first one is zero, and then the second
must be too. With 𝜔≠0, the second gives 𝑎=𝑏/2, and then the
first gives
𝜔2=𝑘−𝑏24.
This is possible exactly when 𝑏2<4𝑘. So when 𝑏2<4𝑘, the guess works,
with decay rate 𝑏/2 and frequency 𝜔=√𝑘−𝑏2/4. The same
calculation, with cosine and sine trading places, shows that
𝑒−𝑏𝑡/2sin𝜔𝑡 is a solution too.
Their initial conditions are (1,−𝑏/2) and (0,𝜔), which are
independent because 𝜔≠0. So when 𝑏2<4𝑘, every solution is
𝑥(𝑡)=𝑒−𝑏𝑡/2(𝑐1cos𝜔𝑡+𝑐2sin𝜔𝑡).
This explains what we saw in Part I. The mass oscillates with frequency
𝜔, a little slower than the undamped √𝑘, while its swings
shrink at the rate 𝑏/2.
For large 𝑏, the mass doesn't oscillate. It creeps back like an
exponential decay. So try 𝑥=𝑒𝑟𝑡. Then 𝑥′=𝑟𝑒𝑟𝑡 and
𝑥″=𝑟2𝑒𝑟𝑡, so
𝑥″+𝑏𝑥′+𝑘𝑥=(𝑟2+𝑏𝑟+𝑘)𝑒𝑟𝑡.
Since 𝑒𝑟𝑡 is never zero, the guess works exactly when 𝑟2+𝑏𝑟+𝑘=0,
that is, when
𝑟=−𝑏±√𝑏2−4𝑘2.
When 𝑏2>4𝑘, these are two different real numbers 𝑟1 and 𝑟2. Both
are negative, since their sum is −𝑏<0 and their product is 𝑘>0. The
initial conditions of 𝑒𝑟1𝑡 and 𝑒𝑟2𝑡 are (1,𝑟1) and
(1,𝑟2), which are independent because 𝑟1≠𝑟2. So when 𝑏2>4𝑘,
every solution is
𝑥(𝑡)=𝑐1𝑒𝑟1𝑡+𝑐2𝑒𝑟2𝑡,
a combination of two decaying exponentials, with no oscillation.
The two calculations have found the boundary between the two kinds of
motion: 𝑏2=4𝑘. That leaves exactly one case, 𝑏2=4𝑘, called critical
damping. Here the quadratic has only one root, 𝑟=−𝑏/2, so the
exponential guess gives only one solution, 𝑒−𝑏𝑡/2. The oscillating
guess doesn't help either: it needs 𝜔2=𝑘−𝑏2/4=0, which turns the
cosine guess into 𝑒−𝑏𝑡/2 again and the sine guess into 0.
But the solution space is two-dimensional, so another solution must exist.
What could it be? Let's keep guessing. Try the exponential times the
simplest function of 𝑡 we can think of, 𝑡 itself:
𝑥=𝑡𝑒−𝑎𝑡,𝑎=𝑏2.
Then 𝑥′=𝑒−𝑎𝑡(1−𝑎𝑡) and 𝑥″=𝑒−𝑎𝑡(𝑎2𝑡−2𝑎), so
𝑥″+𝑏𝑥′+𝑘𝑥=𝑒−𝑎𝑡((𝑏−2𝑎)+(𝑎2−𝑎𝑏+𝑘)𝑡).
With 𝑏=2𝑎 and 𝑘=𝑏2/4=𝑎2, both coefficients are zero. It works! The
initial conditions of 𝑒−𝑎𝑡 and 𝑡𝑒−𝑎𝑡 are (1,−𝑎) and (0,1),
which are independent. So when 𝑏2=4𝑘, every solution is
𝑥(𝑡)=(𝑐1+𝑐2𝑡)𝑒−𝑏𝑡/2.
Altogether, every motion of the damped spring falls into one of three
cases:
underdamped, 𝑏2<4𝑘:
𝑥=𝑒−𝑏𝑡/2(𝑐1cos𝜔𝑡+𝑐2sin𝜔𝑡), with
𝜔=√𝑘−𝑏2/4;
critically damped, 𝑏2=4𝑘: 𝑥=(𝑐1+𝑐2𝑡)𝑒−𝑏𝑡/2;
overdamped, 𝑏2>4𝑘: 𝑥=𝑐1𝑒𝑟1𝑡+𝑐2𝑒𝑟2𝑡.
In each case we found two independent solutions by looking at pictures and
guessing, and the dimension count guarantees that they give every solution.
We found them case by case, though. Why exponentials, why cosines and
sines, and why a factor of 𝑡 in the critical case? We'll see where they
come from in chapter 8.
6.5The Flow of a Linear System
Let's zoom back out and put everything we've learned into one form. Every
equation can be written as a first-order system, so consider a homogeneous
linear system 𝑋′=𝐴(𝑡)𝑋 with 𝑁 unknowns. Suppose we have found its
standard solutions 𝑅1,…,𝑅𝑁 at 𝑡0. In this basis, evaluation and
the solution map are both the identity, so the solution from
𝑋0=(𝜉1,…,𝜉𝑁) is
𝑋(𝑡)=𝜉1𝑅1(𝑡)+⋯+𝜉𝑁𝑅𝑁(𝑡).
At each time 𝑡, this is a combination of the vectors
𝑅1(𝑡),…,𝑅𝑁(𝑡), with the entries of 𝑋0 as coefficients. That is
exactly matrix multiplication. Put the vectors in columns,
Φ(𝑡)=(𝑅1(𝑡)⋯𝑅𝑁(𝑡)),
and then
𝑋(𝑡)=Φ(𝑡)𝑋0.
Definition 6.18 (Flow matrix). The flow matrixΦ(𝑡) of a homogeneous linear system is the
𝑁×𝑁 matrix whose columns are the states 𝑅1(𝑡),…,𝑅𝑁(𝑡) of
its standard solutions at 𝑡0.
So solving is matrix multiplication. The solution map returns a whole
function, and for each 𝑡, the flow matrix gives that function's value at
time 𝑡. The entire solution map is packaged into one matrix-valued
function of 𝑡.
Now read 𝑋(𝑡)=Φ(𝑡)𝑋0 a different way. Fix a time 𝑡. The matrix
Φ(𝑡) takes every initial condition 𝑋0 and moves it to the state its
solution reaches at time 𝑡. As 𝑡 changes, every point of ℝ𝑁
moves at once. So a linear system is a flow on the initial-condition
space, and at each time, that flow is a matrix.
At the starting time nothing has moved yet: each 𝑅𝑗(𝑡0) is the standard
basis vector 𝑒𝑗, so Φ(𝑡0)=𝐼. And Φ(𝑡) is invertible at every
time. It first solves and then evaluates at time 𝑡, so it is the matrix of
𝐸𝑡∘𝐸−1𝑡0. Both maps are isomorphisms by
theorem 6.12, since evaluation is an isomorphism at every
time in 𝐼. In words, the state at any time determines the initial
condition, so evolving never loses information.
Example 6.19 (The flow of the spring). Write the spring 𝑦″=−𝑦 as the system 𝑥′=𝑣,𝑣′=−𝑥, and start at
𝑡0=0. The standard solutions come from example 6.14. The
solution from (1,0) is cos𝑡, with velocity −sin𝑡, and the
solution from (0,1) is sin𝑡, with velocity cos𝑡. So
This is a rotation matrix: it turns the plane clockwise by angle 𝑡. So
the spring's flow rotates the whole space of initial conditions, carrying
every initial position and velocity around a circle centered at the
origin. After time 2𝜋, every point is back where it started.
Example 6.20 (The flow of $x'=2y,\ y'=-3x$). In example 6.5 we found the solution from (𝑥0,𝑦0), with
𝜔=√6:
So this matrix is Φ(𝑡). Its columns are the standard solutions, found
by setting (𝑥0,𝑦0) to (1,0) and (0,1), with no new calculus.
Figure 6.2 The plane of initial conditions, carried by the flow matrix. The grid, the
unit square, and the initial condition 𝑋0 all move together: the
standard basis vectors become the columns 𝑅1(𝑡) and 𝑅2(𝑡), and 𝑋0
becomes 𝑋(𝑡)=Φ(𝑡)𝑋0. Drag 𝑋0, or move time. For the spring, the
whole plane rotates. For 𝑥′=2𝑦,𝑦′=−3𝑥, it turns while stretching along
ellipses.
Remark 6.21 (Time-dependent coefficients). When the coefficients depend on time, the flow matrix also depends on when
we start. The flow from one starting time over a stretch of time need not
match the flow from another starting time over the same stretch, because
the coefficients change in between. The exercises work out an example.
This answers the question we started with. For a linear system, solving
every initial value problem comes down to solving just 𝑁 of them. Once we
have the standard solutions, linear algebra does the rest: the solution
from any initial condition is Φ(𝑡)𝑋0.
But those 𝑁 problems are still calculus problems of the kind we met in
Part I, and we have no uniform way to solve them. We have to attack each
one separately, by guessing, by calculus, or numerically. And when the
equations are intertwined, even one of them is hard. The rate of each
unknown depends on the others, so we can't solve for the unknowns one at a
time.
What we'd love is a zoomed-out way to solve: one that treats the system as
a whole. A linear system 𝑋′=𝐴(𝑡)𝑋 is completely described by its
coefficient matrix. Can we find its flow directly from that matrix? That is
the question of chapter 7.
Problems
Check Your Understanding
1Determine the dimension
For each homogeneous linear problem below, assume continuous coefficients
and equations solved for their highest derivatives. State how many
independent pieces of initial data are needed and hence the dimension of
the solution space. How many solutions form a basis? List the initial
data at 𝑡=0 for the solutions corresponding to the standard basis
of the initial-condition space.
(a)
𝑦‴+𝑎2(𝑡)𝑦″+𝑎1(𝑡)𝑦′+𝑎0(𝑡)𝑦=0.
(b)
A homogeneous first-order system in four unknown functions.
(c)
Two coupled second-order equations for 𝑥(𝑡) and 𝑦(𝑡). First rewrite the
state in first-order form before counting.
2When the leading coefficient vanishes
Consider the equation 𝑡𝑦′=𝑦, whose leading coefficient 𝑡 is zero at
𝑡=0.
(a)
Show that 𝑦=𝐶𝑡 is a solution on ℝ for every number 𝐶.
(b)
Show that every solution on ℝ has this form. On 𝑡>0 and on
𝑡<0 separately, the equation can be solved for 𝑦′. What must happen at
𝑡=0 for the two pieces to join into one differentiable function?
(c)
How many solutions satisfy 𝑦(0)=0? How many satisfy 𝑦(0)=1? Which
hypothesis of the existence and uniqueness theorem for linear equations of
order 𝑛 fails?
3Discovering the hyperbolic functions
Consider the equation 𝑦″=𝑦, with the basis (𝑒𝑡,𝑒−𝑡) of its
solution space and the evaluation map
𝐸0(𝑎𝑒𝑡+𝑏𝑒−𝑡)=(𝑎+𝑏,𝑎−𝑏).
(a)
Find a basis (𝑅1,𝑅2) of the solution space in which the matrix of
𝐸0 is the identity. What must 𝐸0(𝑅1) and 𝐸0(𝑅2) be? Write 𝑅1
and 𝑅2 as combinations of 𝑒𝑡 and 𝑒−𝑡.
(b)
Explain why the matrix of the solution map is then the identity too, and
write down the solution with initial condition (𝑦0,𝑣0). Use it to
solve 𝑦″=𝑦 with 𝑦(0)=2, 𝑦′(0)=−1.
(c)
The functions 𝑅1 and 𝑅2 are the hyperbolic cosine and sine,
cosh𝑡 and sinh𝑡. Show that cosh′=sinh and sinh′=cosh. Then
show that cosh2𝑡−sinh2𝑡=1, by checking that its derivative is zero
and evaluating at 𝑡=0. Explain why the point (cosh𝑡,sinh𝑡) moves
along the hyperbola 𝑥2−𝑦2=1.
4A flow that depends on the starting time
Consider the system 𝑥′=𝑡𝑦,𝑦′=−𝑡𝑥, whose coefficients depend on time.
(a)
Fix a starting time 𝑡0, and let 𝜃(𝑡)=(𝑡2−𝑡20)/2. Check that
𝑥=𝑥0cos𝜃+𝑦0sin𝜃,𝑦=−𝑥0sin𝜃+𝑦0cos𝜃
solves the system with 𝑥(𝑡0)=𝑥0 and 𝑦(𝑡0)=𝑦0.
(b)
Write down the flow matrix with starting time 𝑡0. Describe what it does
to the plane.
(c)
Compare the flow from 𝑡0=0 to 𝑡=1 with the flow from 𝑡0=1 to
𝑡=2. Both cover one unit of time. Through what angle does each one
rotate? Explain the difference in terms of the coefficients.
5Linear maps on polynomials
Let 𝑃2 be the space of real polynomials of degree at most two, with
ordered basis (1,𝑡,𝑡2). Consider the map
𝑇:𝑃2⟶ℝ2,𝑇(𝑝)=(𝑝(0),𝑝′(0)).
(a)
Show that 𝑇 is linear. Find its matrix using the given basis of 𝑃2
and the standard basis of ℝ2.
(b)
Find the kernel and image of 𝑇, giving a basis and dimension for each.
Is 𝑇 an isomorphism? Explain what information about a polynomial is
lost when we record only its value and first derivative at zero.
(c)
Now record one more derivative:
𝐽:𝑃2⟶ℝ3,𝐽(𝑝)=(𝑝(0),𝑝′(0),𝑝″(0)).
Find the matrix of 𝐽 in the same polynomial basis and the standard
basis of ℝ3. Is this map an isomorphism? Justify your answer,
and explain how to recover the polynomial from these three numbers.
6Solution sets inside function spaces
Now consider the whole set of solutions, with no initial condition imposed.
In each part, work with functions defined on all of ℝ.
(a)
For 𝑢″+4𝑢=0, define
𝐿:𝐶2(ℝ)⟶𝐶(ℝ),𝐿[𝑢]=𝑢″+4𝑢.
Verify that 𝐿 is linear. Explain why the solution set is ker𝐿,
and prove that it is a vector subspace of 𝐶2(ℝ). What is
the zero vector in this space? Your argument should not require a formula
for the general solution.
(b)
For 𝑢′+𝑢=1, define 𝐿[𝑢]=𝑢′+𝑢 on 𝐶1(ℝ) and check
that 𝐿 is linear. If 𝑢 and 𝑣 are solutions, compute 𝐿[𝑢+𝑣]
and 𝐿[𝑢−𝑣]. Explain why the solution set is not a vector subspace,
even though the equation is linear.
Find a particular solution 𝑢𝑝 and show that the full solution set is
𝑢𝑝+ker𝐿: prove both that every solution has this form and that
every function of this form is a solution. Describe this affine set by
identifying its translation and its space of directions.
(c)
For 𝑢′=𝑢(1−𝑢10), compare (2𝑢)′ with
the right-hand side evaluated at 2𝑢. Show that the only solution 𝑢
for which 2𝑢 is also a solution is 𝑢=0. Use a nonzero equilibrium
to conclude that the solution set is not a vector subspace of
𝐶1(ℝ). Since it contains the zero function, explain why it
cannot be an affine translate of a vector subspace either.
7A basis of spring solutions
Let 𝐴 and 𝐵 be the solutions of
𝑥″+4𝑥=0
with initial data (𝐴(0),𝐴′(0))=(1,0) and (𝐵(0),𝐵′(0))=(0,1).
(a)
Find formulas for 𝐴 and 𝐵, and verify that each satisfies the equation
and its initial data.
(b)
Let S be the solution space of this equation on ℝ.
Consider the initial-data map
𝐸0:S⟶ℝ2,𝐸0(𝑥)=(𝑥(0),𝑥′(0)).
Rewrite the equation as a first-order system and check the hypotheses
of the global linear existence-and-uniqueness theorem. What does the
theorem guarantee for each initial position and velocity?
(c)
Show that 𝐸0 is linear: for solutions 𝑢,𝑣∈S and
real numbers 𝑎,𝑏, verify that
𝐸0(𝑎𝑢+𝑏𝑣)=𝑎𝐸0(𝑢)+𝑏𝐸0(𝑣).
(d)
Explain how existence and uniqueness give the other two properties we need:
Surjectivity: Why is every vector in ℝ2 the initial
state of some solution in S?
Injectivity: Why must two solutions with the same initial state
be the same function?
Conclude that 𝐸0 is a linear isomorphism.
(e)
Compute 𝐸0(𝐴) and 𝐸0(𝐵). Do these vectors form a basis of
ℝ2? Use the isomorphism 𝐸0 to explain why 𝐴,𝐵
form a basis of the solution space S.
(f)
Without solving another differential equation, write down the solution
𝑥(𝑡) with 𝑥(0)=−3 and 𝑥′(0)=5. Explain how expressing the initial
state in a basis of ℝ2 determines the whole function in the
basis 𝐴,𝐵.
(g)
The coefficient of 𝐵 in your answer is the initial velocity 5, but the
coefficient of sin2𝑡 is not. Explain.
8Solving from a basis we find
Consider the equation
𝑦″−3𝑦′+2𝑦=0.
(a)
Before looking for any solutions, explain why the solution space
S of this equation is two-dimensional. Then look for
solutions by guessing and checking: try an exponential 𝑦=𝑒𝑟𝑡.
Find two linearly independent solutions.
(b)
Use your two solutions from the previous part as an ordered basis
B of the solution space. Write down the evaluation map
𝐸0:S⟶ℝ2
onto the space of initial conditions. Using the basis B for
S and the standard basis (1,0),(0,1) for ℝ2,
express 𝐸0 as a matrix.
(c)
The inverse of the evaluation map is the solution map𝐸−10:ℝ2→S. It takes an initial condition
(𝑦0,𝑣0) to the solution starting from it. In these same bases,
write down the matrix of the solution map, and use it to write the
solution from (𝑦0,𝑣0).
(d)
Use your work to solve 𝑦″−3𝑦′+2𝑦=0 with each of the following
initial conditions:
𝑦(0)=3, 𝑦′(0)=4;
𝑦(0)=1, 𝑦′(0)=−1;
𝑦(0)=0, 𝑦′(0)=2.
9A missing solution
Consider the third-order equation
𝑥‴−𝑥″=0.
(a)
Before looking for any solutions, explain why the solution space
S is three-dimensional. Then try 𝑥=𝑒𝑟𝑡 and find every
value of 𝑟 that works. How many linearly independent solutions does
this give? Why does the dimension of S tell you that these
cannot be all of the solutions?
(b)
Look for a third solution by trying powers 𝑥=𝑡𝑛. Which powers work?
Then check that your three solutions are linearly independent. You do
not need to work with the functions directly. The evaluation map
𝐸0(𝑥)=(𝑥(0),𝑥′(0),𝑥″(0)) is a linear isomorphism, and a linear
isomorphism preserves linear independence in both directions. So
solutions are linearly independent exactly when their initial data are
linearly independent vectors in ℝ3. Compute the initial data
of your three solutions, and show that these vectors are independent.
(c)
Use your three solutions as the basis B of S,
and the standard basis for ℝ3. Express the evaluation map
𝐸0 as a matrix. Then invert it to find the matrix of the solution
map, which takes an initial condition (𝑥0,𝑣0,𝑤0) to its solution.
(d)
Use your work to solve 𝑥‴−𝑥″=0 with each of the following initial
conditions:
𝑥(0)=1, 𝑥′(0)=0, 𝑥″(0)=2;
𝑥(0)=2, 𝑥′(0)=1, 𝑥″(0)=1;
𝑥(0)=0, 𝑥′(0)=3, 𝑥″(0)=−1.
Python: Reusing Computed Solutions
10Combine computed solutions
Euler's method advances an approximation by 𝑦𝑘+1=𝑦𝑘+ℎ𝐹(𝑡𝑘,𝑦𝑘). For
the linear equation 𝑦′=𝑡𝑦, this update is
𝑦𝑘+1=𝑦𝑘+ℎ𝑡𝑘𝑦𝑘=(1+ℎ𝑡𝑘)𝑦𝑘.
Each step multiplies the current value by the number 1+ℎ𝑡𝑘, which
depends on the step but not on 𝑦.
Use the Euler function from Chapter 2 to compute solutions of 𝑦′=𝑡𝑦 from
initial values 2, −1, and 5 on the same time grid.
(a)
Form the array 2𝑌2−𝑌−1 from the first two computed solutions. Which
initial value does this combination have? Compare it point by point with the
separately computed solution from that initial value.
(b)
Suppose two Euler runs 𝑈𝑘 and 𝑉𝑘 use the same grid, and
𝑌𝑘=𝑎𝑈𝑘+𝑏𝑉𝑘 at some step. Use the update to show that
𝑌𝑘+1=𝑎𝑈𝑘+1+𝑏𝑉𝑘+1. Conclude that a combination of Euler runs on a
common grid is itself an Euler run. Does this make the combined array an
exact solution of the differential equation?
(c)
Plot the difference between the combined and directly computed arrays. Is
the difference exactly zero, close to zero, or visibly large? Explain using
the previous part.
(d)
Repeat the experiment with the logistic equation 𝑦′=𝑦(1−𝑦/10). Its Euler
update is 𝑦𝑘+1=𝑦𝑘+ℎ𝑦𝑘(1−𝑦𝑘/10), which does not multiply by a fixed
number. Plot the combined approximation and the directly computed
approximation together, then plot their difference. At what step does the
mismatch first appear?