5 · Equations and Solution Spaces
Chapter 5

Equations and Solution Spaces

In the preceding chapters, many of our solutions arrived with constants attached. Antidifferentiating 𝑦′ =2𝑡 gave the whole family 𝑦 =𝑡2 +𝐶. Guessing and checking gave 𝑦 =𝐶𝑒𝑡 for 𝑦′ =𝑦, and 𝑥 =𝐴cos⁡𝑡 +𝐵sin⁡𝑡 for the spring 𝑥″ = −𝑥. Even the system 𝑓′ =𝑔, 𝑔′ = −𝑓, which the very first page of the book asked you to solve just by thinking about it, has a whole family of solutions:

𝑓(𝑡)=𝐴cos⁡𝑡+𝐵sin⁡𝑡,𝑔(𝑡)=−𝐴sin⁡𝑡+𝐵cos⁡𝑡.

Often, our next move was to use an initial condition to pick out the constants and study that one solution. Sometimes the formulas were already written in terms of initial data: the logistic solution in terms of the initial population 𝑁0, and the spring's motion in terms of its initial position and velocity. We could see the whole family, but the problem at hand often asked us to select one member.

Let's not do that this time. For the spring, keep every function 𝐴cos⁡𝑡 +𝐵sin⁡𝑡, and regard the whole collection as one mathematical object. What kind of object is it, and how can we describe it?

Each member is an entire function, yet two numbers pick it out. Somehow a collection of whole functions is being described by a handful of numbers. We would like to understand how this works even for equations whose solutions we cannot write down. To begin, we need a way to think about a function itself as a single point.

5.1Spaces of Functions and the Solution Set

Linear algebra already gives us a way to do this. In ℝ𝑛, we add vectors and multiply them by numbers, and use these operations to describe lines, planes, and higher-dimensional subspaces. The same ideas apply to matrices, polynomials, and functions. It may have been a while since you thought of a function as a vector, but that is exactly the viewpoint we need.

A solution is always a solution on an interval. On an interval 𝐼, we need a function with enough derivatives to make every expression in the equation meaningful, satisfying the equation at every time in 𝐼. The interval matters. The function 𝑦(𝑡) =1/(1 −𝑡) solves 𝑦′ =𝑦2 on ( −∞,1), but cannot provide a solution through 𝑡 =1. A formula and the interval on which it satisfies the equation belong together.

Fix an open interval 𝐼. If 𝑓 and 𝑔 are real-valued functions on 𝐼, their sum and a scalar multiple are defined by

(𝑓+𝑔)(𝑡)=𝑓(𝑡)+𝑔(𝑡),(𝑐𝑓)(𝑡)=𝑐𝑓(𝑡)for every 𝑡∈𝐼.

These are pointwise operations: perform the arithmetic separately at each input to obtain a new function. The familiar algebraic rules hold because they hold at every input. For instance, 𝑓 +𝑔 =𝑔 +𝑓 because 𝑓(𝑡) +𝑔(𝑡) =𝑔(𝑡) +𝑓(𝑡) for every 𝑡. The zero vector is the function which is zero everywhere, and the additive inverse of 𝑓 is the function −𝑓. Thus real-valued functions on a fixed interval form a vector space.

For differential equations, we need functions we can differentiate. Write 𝐶𝑘(𝐼) for the space of real-valued functions on 𝐼 with continuous derivatives through order 𝑘. In particular, 𝐶1(𝐼) consists of continuously differentiable functions, while 𝐶2(𝐼) requires two continuous derivatives. These are vector spaces too: adding or scaling functions preserves the required differentiability. We use 𝐶(𝐼) for the space of continuous functions. The linear algebra appendix reviews these spaces alongside the familiar spaces of numerical vectors.

These spaces are enormous! Even 𝐶𝑘(𝐼) for a fixed 𝑘 contains all polynomials. The functions 1,𝑡,𝑡2,… are linearly independent: a finite linear combination that vanishes throughout an interval is a zero polynomial, so all its coefficients vanish. Consequently, 𝐶𝑘(𝐼) contains arbitrarily large linearly independent collections. It is infinite-dimensional; no finite list of functions can form a basis for the whole space.

This enormous space is simply the place where solutions live. From all of these functions, the equation picks out the ones that solve it. For example, the solutions of 𝑦′ =𝑦 on 𝐼 form the subset

S={𝑦∈𝐶1(𝐼):𝑦′(𝑡)=𝑦(𝑡) for every 𝑡∈𝐼}.

This set S is exactly the object we were after: every solution of 𝑦′ =𝑦, bundled together in one place. We call it the solution set of the equation. The larger space it lives in is its ambient space. The equation 𝑦′ =𝑦 involves one derivative, so each solution must be continuously differentiable. Thus all of its solutions live in 𝐶1(𝐼). For the spring, the equation 𝑥″ = −𝑥 involves two derivatives, so each solution must be twice continuously differentiable, and its solutions live in 𝐶2(𝐼).

Systems fit into the same picture. A solution of the system 𝑓′ =𝑔, 𝑔′ = −𝑓 from the introduction is a pair of functions, and neither is an answer on its own: the derivative of the first must equal the second, and the derivative of the second must equal the negative of the first. We can regard the pair as a single vector-valued function,

𝑋(𝑡)=(𝑓(𝑡)𝑔(𝑡)).

At each time, 𝑋(𝑡) is a vector in ℝ2. The whole function 𝑋 records how that vector changes with time. Its derivative is computed component by component, so the system can be written

𝑋′(𝑡)=(𝑓′(𝑡)𝑔′(𝑡))=(𝑔(𝑡)−𝑓(𝑡)).

This is still the same pair of equations. Collecting the functions into a vector lets us speak of one solution while keeping track of all its components.

Write 𝐶1(𝐼,ℝ2) for the space of functions from 𝐼 to ℝ2 whose components are continuously differentiable. We add and scale these functions component by component, so this is again a vector space. Each solution of the system has continuously differentiable components, so the solutions form a subset of 𝐶1(𝐼,ℝ2).

The same goes for a system with 𝑁 unknown functions. Collecting them into 𝑋 =(𝑥1,…,𝑥𝑁) gives a function whose value at each time is a vector in ℝ𝑁, and its solutions live in 𝐶𝑘(𝐼,ℝ𝑁), the space of functions from 𝐼 to ℝ𝑁 whose components belong to 𝐶𝑘(𝐼). Notice that 𝑁 counts the components of 𝑋(𝑡) at one time; it is not the dimension of this function space. Even one component can be any of infinitely many linearly independent functions.

Definition 5.1 (Solution set). Fix an open interval 𝐼. The solution set S of a differential equation on 𝐼 is the set of all functions on 𝐼 that satisfy the equation at every 𝑡 ∈𝐼. It is a subset of an ambient space, which consists of functions with enough continuous derivatives for the equation to make sense: 𝐶𝑛(𝐼) for a scalar equation of order 𝑛, and for a system, vector-valued functions each of whose components has as many continuous derivatives as its equation requires.

Example 5.2 (The solution set of $y'=y$). On 𝐼 =ℝ, every solution of 𝑦′ =𝑦 has the form 𝐶𝑒𝑡 for some number 𝐶, and every such function is a solution. So the solution set is

S={𝐶𝑒𝑡:𝐶∈ℝ}.

Each number 𝐶 gives one solution, and different numbers give different solutions, since evaluating at 𝑡 =0 recovers 𝐶. Geometrically, S is a line through the origin in 𝐶1(ℝ): all scalar multiples of the single function 𝑒𝑡. Each entire exponential graph is just one point on this line.

Example 5.3 (Solutions of the spring equation). For the spring equation 𝑥″ = −𝑥 on ℝ, every function

𝑥𝐴,𝐵(𝑡)=𝐴cos⁡𝑡+𝐵sin⁡𝑡

is a solution, so the solution set contains one function for each pair of numbers (𝐴,𝐵). Different pairs give different functions, since 𝑥𝐴,𝐵(0) =𝐴 and 𝑥′𝐴,𝐵(0) =𝐵. These functions form a plane in 𝐶2(ℝ): all combinations of cos⁡𝑡 and sin⁡𝑡. Choosing a point (𝐴,𝐵) chooses one point of this plane.

Figure 5.1 Choose a point (𝐴,𝐵) on the left to select the entire function 𝑥(𝑡) =𝐴cos⁡𝑡 +𝐵sin⁡𝑡 on the right. Moving the time marker along that graph leaves (𝐴,𝐵) fixed; moving (𝐴,𝐵) chooses a different function. The left panel gives coordinates for functions, while the right panel shows the values of one function over time.

Example 5.4 (Solutions of a system). For the system 𝑓′ =𝑔, 𝑔′ = −𝑓 on ℝ, every vector-valued function

𝑋𝐴,𝐵(𝑡)=(𝐴cos⁡𝑡+𝐵sin⁡𝑡−𝐴sin⁡𝑡+𝐵cos⁡𝑡)

is a solution: differentiating the first component gives the second, and differentiating the second gives the negative of the first. So the solution set, a subset of 𝐶1(ℝ,ℝ2), contains one solution for each pair of numbers (𝐴,𝐵). Different pairs give different solutions, since 𝑋𝐴,𝐵(0) =(𝐴,𝐵).

One warning before the next example. A solution set lives on a fixed interval 𝐼, but some equations have solutions that blow up, and different solutions can blow up at different times. A function can then solve the equation near 𝑡0 and still not belong to the solution set on 𝐼, simply because it doesn't last that long.

Example 5.5 (A less familiar family). The functions

𝑥𝑎,𝑏(𝑡)=𝑎+log⁡(cosh⁡𝑡+𝑏sinh⁡𝑡)

solve 𝑥″ =1 −(𝑥′)2 wherever they are defined. To check this, write ℎ(𝑡) =cosh⁡𝑡 +𝑏sinh⁡𝑡, so that ℎ″ =ℎ. Then

𝑥′𝑎,𝑏=ℎ′ℎ,𝑥″𝑎,𝑏=ℎ″ℎ−(ℎ′ℎ)2=1−(𝑥′𝑎,𝑏)2.

Since ℎ(0) =1 and ℎ′(0) =𝑏, we have 𝑥𝑎,𝑏(0) =𝑎 and 𝑥′𝑎,𝑏(0) =𝑏. If |𝑏| ≤1, then ℎ(𝑡) =1+𝑏2𝑒𝑡 +1−𝑏2𝑒−𝑡 is positive everywhere, so 𝑥𝑎,𝑏 is a solution on all of ℝ. If |𝑏| >1, then ℎ reaches zero at a finite time on one side of 0. There 𝑥𝑎,𝑏 blows up, so it does not belong to the solution set on ℝ.

We will see in the next chapter that this never happens for linear equations: their solutions last across the entire interval where the coefficients are continuous, so working on a fixed interval loses nothing.

In each of our examples, a solution was picked out by a few numbers that we could read off a formula: 𝐶, or (𝐴,𝐵), or (𝑎,𝑏). What about equations where we can't find a formula?

For a single first-order equation 𝑦′ =𝐹(𝑡,𝑦), we already know the answer, and it doesn't need a formula at all. The existence and uniqueness theorem from chapter 3 says that, under its hypotheses, each initial condition 𝑦(𝑡0) =𝑦0 has exactly one solution, at least near 𝑡0. So there is one solution for every initial condition. And what is an initial condition? Just a choice of the starting value 𝑦0: a single real number. The solutions form a one-parameter family, one for each real number, whether or not we can ever write them down.

Can we say the same for every equation we've met? For the spring, or for a system, what plays the role of that single number 𝑦0?

5.2A Common Formulation: First-Order Systems

So far we have zoomed out once: from a single solution to the set of all solutions of one equation. Now we want to go further, and build a single language for talking about the solutions of any differential equation. The trouble is that differential equations come in many shapes. Some are first order, like 𝑦′ =𝑦. Some are second order, like the spring. Some are systems, like 𝑓′ =𝑔, 𝑔′ = −𝑓, and a system can even mix equations of different orders. To speak about all of them in one way, we first need a single form that every equation can be written in.

We have already seen the key idea, when we modeled the spring. Its equation 𝑥″ = −𝑥 tells us the acceleration from the position, but position alone does not determine the motion. Two masses can pass through the same position with different velocities. To describe how either one moves, we need to keep track of both quantities.

Give velocity its own name, 𝑣 =𝑥′. We can now ask for the first derivative of each quantity we are tracking. The derivative of position is velocity, and the derivative of velocity is the acceleration supplied by the equation:

𝑥′=𝑣,𝑣′=−𝑥.

We have replaced one second-order equation with two first-order equations. The new variable records information the original problem already needed. The initial conditions 𝑥(𝑡0) =𝑥0 and 𝑥′(𝑡0) =𝑣0 become the single vector condition 𝑋(𝑡0) =(𝑥0,𝑣0) for 𝑋 =(𝑥,𝑣).

Nothing about this construction depended on the particular spring equation. Apply it to the equation 𝑥″ =1 −(𝑥′)2 from our examples. With the same choice 𝑣 =𝑥′, we obtain

𝑥′=𝑣,𝑣′=1−𝑣2.

The second equation now involves only 𝑣, so we can solve it by separation and then integrate 𝑥′ =𝑣. Two of its solutions are 𝑥 =𝑡 and 𝑥 = −𝑡, the members of our family with 𝑎 =0 and 𝑏 = ±1. They correspond to the system solutions (𝑥,𝑣) =(𝑡,1) and (𝑥,𝑣) =( −𝑡, −1). In the system, we can see why those motions work: the velocities 1 and −1 make 𝑣′ =0 and remain constant.

Figure 5.2 Follow one solution in its two descriptions. The graphs of 𝑥(𝑡) and 𝑣(𝑡) =𝑥′(𝑡) share a time marker; their values give the gold point (𝑥(𝑡),𝑣(𝑡)) in the state plane. Move time to trace the same solution, or change either initial value to choose another. Switching equations shows the same correspondence for 𝑥″ = −𝑥 and 𝑥″ =1 −(𝑥′)2. Here a point in the state plane records a solution at one instant; the entire state curve records its motion over the displayed interval.

Higher-order equations work the same way, except that we must record more derivatives. Take the third-order equation

𝑥‴=𝑥″.

Recording 𝑥 and 𝑣 =𝑥′ is no longer enough: the derivative of 𝑣 is 𝑥″, which we have not recorded. So give it a name too, 𝑤 =𝑥″. Now the derivative of every recorded quantity is something we know:

𝑥′=𝑣,𝑣′=𝑤,𝑤′=𝑤.

The first two equations say what 𝑣 and 𝑤 mean, and the last one is the original equation 𝑥‴ =𝑥″. The state is 𝑋 =(𝑥,𝑣,𝑤), and an initial condition prescribes 𝑥(𝑡0), 𝑥′(𝑡0), and 𝑥″(𝑡0).

Systems of higher-order equations work the same way: we record derivatives for every unknown, as many as each one needs. Consider

𝑥′=𝑦+𝑧,𝑦″=−𝑥+𝑧′,𝑧‴=𝑥𝑦,

where 𝑥 appears to first order, 𝑦 to second order, and 𝑧 to third order. For 𝑥 we record only its value. For 𝑦 we also record 𝑢 =𝑦′. For 𝑧 we record 𝑝 =𝑧′ and 𝑞 =𝑧″. The system becomes

𝑥′=𝑦+𝑧,𝑦′=𝑢,𝑢′=−𝑥+𝑝,𝑧′=𝑝,𝑝′=𝑞,𝑞′=𝑥𝑦.

The state 𝑋 =(𝑥,𝑦,𝑢,𝑧,𝑝,𝑞) has 1 +2 +3 =6 coordinates, and an initial condition prescribes 𝑥, 𝑦, 𝑦′, 𝑧, 𝑧′, and 𝑧″ at 𝑡0.

Every equation we have met fits this pattern. Suppose the unknowns are 𝑦1,…,𝑦𝑚, where 𝑦𝑖 appears to order 𝑛𝑖, and the equations give each highest derivative in terms of time and lower derivatives:

𝑦(𝑛𝑖)𝑖=𝐺𝑖(𝑡, 𝑦1,…,𝑦(𝑛1−1)1, …, 𝑦𝑚,…,𝑦(𝑛𝑚−1)𝑚),𝑖=1,…,𝑚.

Record every derivative below the highest one, for every unknown, in a single vector

𝑋=(𝑦1,…,𝑦(𝑛1−1)1, …, 𝑦𝑚,…,𝑦(𝑛𝑚−1)𝑚)∈ℝ𝑁,𝑁=𝑛1+⋯+𝑛𝑚.

Differentiating each coordinate of 𝑋 gives either the next coordinate or one of the 𝐺𝑖, so 𝑋 satisfies a first-order system.

Definition 5.6 (First-order system). A first-order system for an unknown vector-valued function 𝑋 =(𝑥1,…,𝑥𝑁) is an equation

𝑋′=𝐹(𝑡,𝑋),

where 𝐹 is a given function of time and 𝑋, with values in ℝ𝑁. In coordinates, it says 𝑥′𝑘 =𝐹𝑘(𝑡,𝑥1,…,𝑥𝑁) for 𝑘 =1,…,𝑁.

The first-order system obtained by recording derivatives this way is the first-order form of the original equation. Converting an equation to this system is called writing it in first-order form.

Theorem 5.7 (Every equation has a first-order form). Functions 𝑦1,…,𝑦𝑚 solve the equations above on an interval 𝐼 exactly when the vector 𝑋 built from them solves their first-order form on 𝐼. Every solution of the first-order form arises this way. So the two solution sets correspond one-to-one, and prescribing 𝑋(𝑡0) is the same as prescribing 𝑦𝑖(𝑡0),…,𝑦(𝑛𝑖−1)𝑖(𝑡0) for every 𝑖.

Proof. If the 𝑦𝑖 solve the original equations, then each coordinate of 𝑋 differentiates to the next one by definition, and the remaining equations are the original equations. Conversely, suppose 𝑋 solves the first-order form. For each unknown, the equations say that each recorded coordinate is the derivative of the one before it. So those coordinates really are 𝑦𝑖,𝑦′𝑖,…,𝑦(𝑛𝑖−1)𝑖, and the remaining equation says 𝑦(𝑛𝑖)𝑖 =𝐺𝑖. ∎

Writing equations in first-order form pays off numerically too. In section 4.3.1 we used Euler's method on the coffee and the room, and later on the spring: compute every rate from the current values, then advance every quantity by one short step. In first-order form, that recipe becomes a single update,

𝑋𝑘+1=𝑋𝑘+ℎ𝐹(𝑡𝑘,𝑋𝑘).

The whole state moves at once, using rates computed at the current time. Any equation we can write in first-order form, we can approximate this way. For the spring, 𝑋 =(𝑥,𝑣) and 𝐹(𝑡,𝑋) =(𝑣, −𝑥). Starting from 𝑋0 =(1,0) with step ℎ =0.1,

𝑋1=(1,0)+0.1(0,−1)=(1,−0.1),𝑋2=(1,−0.1)+0.1(−0.1,−1)=(0.99,−0.2).

The position begins to fall and the velocity becomes negative, just as a spring released from rest at 𝑥 =1 should.

Figure 5.3 Euler's method for the spring in first-order form, starting from (𝑥,𝑣) =(1,0). On the left, the gold points are the Euler approximations to 𝑥(𝑡) and 𝑣(𝑡), drawn over the true solutions. On the right, the same points trace a polygon in the state plane, drawn over the true circle. Move the step-size slider: every panel updates together, and the approximation stays closer to the truth as ℎ shrinks.

Writing an equation in first-order form also tells us what its initial conditions are. An initial condition prescribes the value 𝑋(𝑡0), which is exactly the state at 𝑡0 in the sense of definition 1.3: the complete collection of current values we must prescribe.

Definition 5.8 (Initial-condition space). For an equation with first-order form 𝑋′ =𝐹(𝑡,𝑋), 𝑋 ∈ℝ𝑁, the initial-condition space I is the set of possible initial conditions 𝑋(𝑡0). It is all of ℝ𝑁, or the part of ℝ𝑁 on which 𝐹 is defined.

For a single equation of order 𝑛, a point of I is the list (𝑦(𝑡0),𝑦′(𝑡0),…,𝑦(𝑛−1)(𝑡0)). So I is ℝ for 𝑦′ =𝑦, ℝ2 for the spring, ℝ3 for 𝑥‴ =𝑥″, and ℝ6 for the system 𝑥′ =𝑦 +𝑧, 𝑦″ = −𝑥 +𝑧′, 𝑧‴ =𝑥𝑦.

So far this is only a definition: it says what we prescribe. Whether prescribing it actually picks out a solution is a separate question.

5.3Existence, Uniqueness, and Initial Data

Every equation now comes with two spaces: the solution set S, whose points are whole functions, and the initial-condition space I, whose points are lists of numbers. There are two natural ways to move between them.

Going from S to I is easy. Given a solution, evaluate it at 𝑡0 and record its initial condition.

Definition 5.9 (Evaluation map). The evaluation map 𝐸𝑡0 :S →I sends each solution to its initial condition, 𝐸𝑡0(𝑋) =𝑋(𝑡0). For a scalar equation of order 𝑛, it sends 𝑦 to (𝑦(𝑡0),𝑦′(𝑡0),…,𝑦(𝑛−1)(𝑡0)).

Going the other way is the interesting direction. Given an initial condition, we would like to return the solution that starts there: a solution map from I back to S. But for this to make sense, two things must be true. Some solution must start at that initial condition, and only one may. These are exactly the questions of existence and uniqueness.

We saw at the end of the first section that chapter 3 answers both questions for a single first-order equation 𝑦′ =𝐹(𝑡,𝑦). The same theorem holds for systems 𝑋′ =𝐹(𝑡,𝑋). Its hypotheses now ask how each component of 𝐹 depends on each component of the state.

Write 𝐹 =(𝐹1,…,𝐹𝑁) and 𝑋 =(𝑥1,…,𝑥𝑁). The partial derivative 𝜕𝐹𝑖/𝜕𝑥𝑗 measures how the 𝑖th rate changes as we vary the 𝑗th state coordinate, holding time and the other coordinates fixed. There are 𝑁2 such derivatives.

Theorem 5.10 (Existence and uniqueness for systems). Suppose 𝐹(𝑡,𝑋) and all partial derivatives 𝜕𝐹𝑖/𝜕𝑥𝑗 are continuous on an open set 𝐷 ⊆ℝ ×ℝ𝑁, and let (𝑡0,𝑋0) ∈𝐷.

  • Existence: the initial value problem 𝑋′ =𝐹(𝑡,𝑋), 𝑋(𝑡0) =𝑋0 has a solution on some open interval containing 𝑡0.

  • Uniqueness: any two solutions of this initial value problem agree on every interval where both are defined.

So every initial condition has exactly one solution, at least near 𝑡0. We will use this theorem without proving it. The hypotheses are sufficient conditions: an equation that fails them might still have unique solutions, but the theorem says nothing about it.

Corollary 5.11 (Evaluation is one-to-one). Under the hypotheses of theorem 5.10, the evaluation map 𝐸𝑡0 :S →I is one-to-one.

Indeed, if two solutions in S have the same initial condition, they solve the same initial value problem on 𝐼, so by uniqueness they agree on all of 𝐼.

For example, the equation 𝑥″ =1 −(𝑥′)2 has first-order form 𝑥′ =𝑣, 𝑣′ =1 −𝑣2. Here 𝐹1(𝑥,𝑣) =𝑣 and 𝐹2(𝑥,𝑣) =1 −𝑣2, with partial derivatives

𝜕𝐹1𝜕𝑥=0,𝜕𝐹1𝜕𝑣=1,𝜕𝐹2𝜕𝑥=0,𝜕𝐹2𝜕𝑣=−2𝑣.

Everything is continuous everywhere. So every choice of 𝑥(𝑡0) and 𝑥′(𝑡0) gives exactly one local solution, by theorem 5.7. We have settled this for every initial position and velocity without solving the equation for any of them.

The same check works for any equation in first-order form. The equations that record derivatives, like 𝑥′ =𝑣, have a single coordinate on the right-hand side, so they always pass. Only the equations carrying the original 𝐺𝑖 need checking.

The theorem does not promise that solutions last. Existence is only local: even when the hypotheses hold everywhere, a solution need not exist for all time. We already know this from 𝑦′ =𝑦2. The solution with 𝑦(0) =𝑎 ≠0 is 𝑦 =𝑎/(1 −𝑎𝑡), which blows up at 𝑡 =1/𝑎.

This affects the evaluation map. On 𝐼 =( −1,1), the solution with 𝑦(0) =5 blows up at 𝑡 =1/5, so no solution in S has initial condition 5. Evaluation is one-to-one, but it need not be onto.

Definition 5.12 (Solution map). The solution map sends each initial condition whose solution lasts across 𝐼 to that solution in S. On these initial conditions, it is the inverse of the evaluation map.

Every initial condition does have a solution on its own largest interval containing 𝑡0. But to compare solutions as points of one space, we fix a single interval, and an initial condition whose solution doesn't last that long has no point of S to go to.

When the hypotheses fail, uniqueness can fail too. Both 𝑦 =0 and 𝑦 =𝑡3 solve 𝑦′ =3𝑦2/3 with 𝑦(0) =0. An initial condition can't tell these two solutions apart, so for this equation evaluation is not one-to-one.

The corollary has a satisfying consequence for our examples. We checked that every function 𝐴cos⁡𝑡 +𝐵sin⁡𝑡 solves the spring equation. Are there any other solutions? Take any solution 𝑥 on ℝ, and let 𝐴 =𝑥(0) and 𝐵 =𝑥′(0). Then 𝑥 and 𝐴cos⁡𝑡 +𝐵sin⁡𝑡 have the same initial condition, so by the corollary they are the same function. There are no missing spring solutions! The plane of functions in example 5.3 is the entire solution set.

The same argument works for 𝑥″ =1 −(𝑥′)2. Any solution on ℝ has some initial condition (𝑎,𝑏), and by uniqueness it agrees with 𝑥𝑎,𝑏 wherever both are defined. So every solution on ℝ is one of the functions 𝑥𝑎,𝑏. This also forces |𝑏| ≤1, since for |𝑏| >1 the function 𝑥𝑎,𝑏 blows up, while a solution on all of ℝ cannot.

We now have two ways to describe a solution: as a whole function in S, or by its initial condition in I. Under the existence and uniqueness hypotheses, that initial condition pins down the function: one number for 𝑦′ =𝑦, two for the spring, six for our system with 𝑥, 𝑦, and 𝑧. We know how many values to prescribe before finding a single formula. On a fixed interval 𝐼, some initial conditions may have no solution that lasts long enough, but each has at most one.

What this picture does not give us is a way to find the solutions. To compute the solution map at one initial condition, we still have to solve that initial value problem. And in general, solving it once doesn't help with the next one. For 𝑦′ =𝑦2, the solution from 𝑦(0) =1 is 1/(1 −𝑡), but doubling it does not give the solution from 𝑦(0) =2, which is 2/(1 −2𝑡).

For some equations, though, known solutions do help. For the spring, adding and scaling solutions produces new ones, and cos⁡𝑡 and sin⁡𝑡 alone produce every solution. This suggests that linear algebra might help us solve differential equations. In the next part of the book, we develop that idea: the structure of linear equations lets us organize their solutions, find formulas, and understand the motions they describe.

In the final part, linear systems themselves become a tool. We use linear approximations to understand nonlinear systems near equilibria, then combine that local information with geometry to investigate their behavior more broadly. The progression is from linear algebra to linear differential equations, and from linear differential equations to nonlinear dynamics.

Problems

The objects in this chapter live in several different spaces. A state is a list of numbers, a solution is a function, and a solution set is a collection of functions. Keep track of which object you are working with; many of the calculations below become more useful once we know what they say about the whole collection.

Check Your Understanding

1A point, a graph, and a function

Consider the family 𝑥𝐴,𝐵(𝑡) =𝐴cos⁡𝑡 +𝐵sin⁡𝑡 of solutions to 𝑥″ = −𝑥, and the corresponding system state 𝑋 =(𝑥,𝑥′).

(a)

Choose (𝐴,𝐵) =(1,0). Write the resulting scalar function 𝑥(𝑡) and vector-valued function 𝑋(𝑡). Which of these objects is a point in 𝐶2(ℝ)? Which is a point in 𝐶1(ℝ,ℝ2)?

(b)

At 𝑡 =𝜋/2, record three points: the parameter point (𝐴,𝐵), the graph point (𝑡,𝑥(𝑡)), and the state point (𝑥(𝑡),𝑥′(𝑡)). Explain what information each point records.

(c)

Hold (𝐴,𝐵) fixed and vary 𝑡. Which of those points moves? Then hold 𝑡 =𝜋/2 fixed and vary 𝐴. Are you moving along one solution or selecting different solutions? Use figure 5.1 to check your interpretation.

2How many independent parameters?

(a)

The formulas 𝑦𝐶(𝑡) =𝐶𝑒𝑡 and 𝑧𝐴,𝐵(𝑡) =(𝐴 +𝐵)𝑒𝑡 describe families in 𝐶1(ℝ). Do they describe the same set of functions? Give two different pairs (𝐴,𝐵) that describe the same function, and explain how many independent parameters the set has.

(b)

Show that if 𝐴cos⁡𝑡 +𝐵sin⁡𝑡 =𝐶cos⁡𝑡 +𝐷sin⁡𝑡 for every 𝑡, then 𝐴 =𝐶 and 𝐵 =𝐷. Which two measurements of a function recover its parameters?

(c)

For |𝑎| <1, consider 𝑦𝑎(𝑡) =𝑎/(1 −𝑎𝑡) on ( −1,1). Recover 𝑎 from the function and show that different parameters give different functions. Does having one independent parameter imply that this family is a vector subspace? Test a specific pair or a scalar multiple.

(d)

In 𝑦(𝑡) =𝐶𝑒𝑟𝑡, explain the different roles of 𝐶 and 𝑟 when we are describing solutions of the fixed equation 𝑦′ =2𝑦.

3Recognizing linear structure

For each equation in parts (a)–(d), decide whether it is linear as written. If it is, identify its linear differential operator and forcing, and decide whether it is homogeneous. State what this tells you about its nonempty solution set on a common interval. If it is nonlinear, identify the term that prevents the linear-operator argument from applying.

(a)

𝑦″ +𝑡2𝑦′ −sin⁡(𝑡)𝑦 =0.

(b)

𝑦′ +3𝑦 =𝑒−𝑡.

(c)

𝑦″ +𝑦2 =0.

(d)

𝑥′ =𝑡𝑥 −2𝑦, 𝑦′ =𝑥 +(1 +𝑡2)𝑦.

(e)

For 𝑦′ =𝑦(1 −𝑦), use the constant solutions zero and one to prove that the solution set on ℝ is not affine. Why would merely finding the zero solution be insufficient to decide its structure?

4Arithmetic in an affine solution set

Let 𝐿 be a linear differential operator, and suppose 𝑢 and 𝑣 solve 𝐿[𝑦] =𝑔 on the same interval. Assume 𝑔 is not the zero function.

(a)

Calculate 𝐿[𝑢 −𝑣] and 𝐿[𝑢 +𝑣]. Which equation does each function solve? Explain why the failure of addition here does not mean 𝐿 is nonlinear.

(b)

For real constants 𝑎,𝑏, show that 𝑎𝑢 +𝑏𝑣 solves 𝐿[𝑦] =𝑔 if and only if 𝑎 +𝑏 =1. What changes when 𝑔 =0?

(c)

Suppose 𝑢𝑝 is one solution of 𝐿[𝑦] =𝑔. Prove that every solution is 𝑢𝑝 +ℎ for some ℎ ∈ker⁡𝐿, and that every such function is a solution.

(d)

Apply the description to 𝑦′ =2𝑡. Identify a particular solution and the homogeneous solution space. Choose two distinct solutions and calculate their difference and their midpoint.

5Write the state, then recover the equation

In each conversion, name every state coordinate and specify how the initial data translate. Do not solve the equations.

(a)

Convert 𝑥″ =1 −(𝑥′)2, with 𝑥(0) =𝑎 and 𝑥′(0) =𝑏, to a first-order system. Starting instead from a solution of your system, show directly that its first component solves the scalar equation.

(b)

Convert

𝑦‴+𝑎2(𝑡)𝑦″+𝑎1(𝑡)𝑦′+𝑎0(𝑡)𝑦=0

to a first-order system. Write its initial state in terms of 𝑦(0), 𝑦′(0), and 𝑦″(0).

(c)

Convert the coupled equations

𝑥″=−𝑥+𝑦,𝑦″=𝑥−𝑦

to a first-order system. How many scalar coordinates does its state have? List the initial data, and explain how a system solution recovers both original equations.

(d)

For the equation 𝑎2(𝑡)𝑦″ +𝑎1(𝑡)𝑦′ +𝑎0(𝑡)𝑦 =𝑔(𝑡), what condition on 𝑎2 lets you carry out the usual conversion throughout an interval? Explain which step can fail at a zero of 𝑎2.

6What does the theorem tell us?

Apply theorem 5.10. Check the right-hand side and its state partial derivatives; do not infer uniqueness merely from how many initial values have been supplied.

(a)

For 𝑥′ =2𝑦, 𝑦′ = −3𝑥, check the four state partial derivatives. Does the theorem apply at every initial state? State precisely its conclusion about the interval of existence.

(b)

Repeat for 𝑥′ =𝑣, 𝑣′ =1 −𝑣2. Translate the conclusion into a statement about initial value problems for the scalar second-order equation.

(c)

Regard 𝑦′ =3𝑦2/3 as a one-component system, where 𝑦2/3 =(3√𝑦)2. Compare the initial values 𝑦(0) =1 and 𝑦(0) =0. Where does the theorem apply? At zero, verify two distinct solutions with the same initial value. Distinguish the failed hypothesis from the separate calculation establishing nonuniqueness.

7Have we found every solution?

For the system 𝑥′ =2𝑦, 𝑦′ = −3𝑥, let 𝜔 =√6 and consider the proposed family

𝑥𝐴,𝐵(𝑡)=𝐴cos⁡(𝜔𝑡)+2𝐵𝜔sin⁡(𝜔𝑡),𝑦𝐴,𝐵(𝑡)=𝐵cos⁡(𝜔𝑡)−3𝐴𝜔sin⁡(𝜔𝑡).
(a)

Verify both differential equations. Recover 𝐴 and 𝐵 by evaluating the solution at zero.

(b)

Let (𝑢,𝑣) be any solution on an interval containing zero. Choose 𝐴,𝐵 from its initial state and use uniqueness to prove that it agrees with the displayed solution near zero. Explain why agreement extends throughout the common interval.

(c)

Write the solution with initial state (𝑥(0),𝑦(0)) =(2, −1). Which part of your argument shows that the formula works, and which shows that there are no other solutions with those data?

8The interval changes the solution set

For 𝑦′ =𝑦2, the solution starting at 𝑦(0) =𝑎 is

𝑦𝑎(𝑡)=𝑎1−𝑎𝑡

on its interval of existence containing zero.

(a)

For which real values of 𝑎 is this solution defined throughout ( −1,1)? Include the endpoint values of your parameter range if they work. How does the answer change if the solution must exist on an open interval containing the closed interval [ −1,1]?

(b)

Answer the first question for ( −2,2), then for ℝ. Explain why a one-parameter local family need not remain a one-parameter family when we require solutions for all time.

9Subspaces and affine sets of functions

Work in the vector space 𝐶1(ℝ). For each subset below, decide whether it is a vector subspace, an affine set that is not a vector subspace, or neither. Justify your answers using the definitions; when closure fails, give explicit functions showing the failure. For an affine set that is not a subspace, write it as 𝑓𝑝 +𝑉 and identify both the particular function 𝑓𝑝 and the vector subspace 𝑉.

(a)

{𝑓 ∈𝐶1(ℝ) :𝑓(0) =0}.

(b)

{𝑓 ∈𝐶1(ℝ) :𝑓(0) =1}.

(c)

{𝑓 ∈𝐶1(ℝ) :𝑓(0)𝑓(1) =0}.

10From higher order to first order

Convert each second-order equation or system below into a first-order system: name the state, and write the system as 𝑋′ =𝐹(𝑡,𝑋). When the system is linear and homogeneous, write it as 𝑋′ =𝐴𝑋 with an explicit matrix.

(a)

Two carts slide without friction on a track. The first is attached to a wall by a spring of stiffness 𝑘1, and the second is attached to the first by a spring of stiffness 𝑘2. If 𝑥1 and 𝑥2 are their displacements from rest, Hooke's law and Newton's second law give

𝑚1𝑥″1=−𝑘1𝑥1+𝑘2(𝑥2−𝑥1),𝑚2𝑥″2=−𝑘2(𝑥2−𝑥1).

Convert this pair to a first-order system and write its matrix. How many standard solutions reconstruct every motion of the carts? Which theorem from the notes guarantees that every motion exists for all time?

(b)

A pendulum of length 𝐿 swinging through the angle 𝜃 satisfies

𝜃″=−𝑔𝐿sin⁡𝜃.

Convert this to a first-order system. Is it linear? Check the hypotheses of the local existence-and-uniqueness theorem in state space, and state what it guarantees about a pendulum released from any angle with any angular velocity. Does the global theorem for linear systems apply?

Explorations

11Can solution graphs cross?

Consider the two spring solutions 𝑢(𝑡) =sin⁡𝑡 and 𝑣(𝑡) = −sin⁡𝑡.

(a)

Verify that both solve 𝑥″ = −𝑥. Find their graph intersections and compare their derivatives at those times.

(b)

Write the corresponding state histories 𝑈 =(𝑢,𝑢′) and 𝑉 =(𝑣,𝑣′). At a time when the scalar graphs cross, do the system solutions have the same state? Explain why uniqueness permits the graph crossings.

(c)

Could two distinct spring solutions have both the same value and the same derivative at the same time? Give an argument using the system theorem rather than the explicit solution formula.

(d)

Show that 𝑈 and 𝑉 trace the same unit circle in the state plane, but never occupy the same point at the same time. Explain the distinction between a geometric curve and a function of time. Why does tracing the same curve not contradict uniqueness?

12How long does the nonlinear family exist?

Return to

𝑥𝑎,𝑏(𝑡)=𝑎+log⁡ℎ𝑏(𝑡),ℎ𝑏(𝑡)=cosh⁡𝑡+𝑏sinh⁡𝑡,

which solves 𝑥″ =1 −(𝑥′)2 wherever ℎ𝑏 >0. We want the largest open interval containing zero on which this formula gives a solution.

(a)

Use the exponential definitions of cosh and sinh to write ℎ𝑏(𝑡) as a linear combination of 𝑒𝑡 and 𝑒−𝑡. Show that ℎ𝑏(𝑡) >0 for every real 𝑡 when |𝑏| ≤1. What are the solutions when 𝑏 =1 and 𝑏 = −1?

(b)

When 𝑏 >1 or 𝑏 < −1, find the time 𝑡∗ at which ℎ𝑏(𝑡∗) =0. Determine on which side of 𝑡∗ the function is positive, and give the requested interval of existence in each case.

(c)

What happens to 𝑥𝑎,𝑏(𝑡) as 𝑡 approaches 𝑡∗ from inside that interval? Explain why the solution cannot extend through 𝑡∗ as a real-valued solution, even though the system's right-hand side and all its state partial derivatives are continuous everywhere.

(d)

Which initial states (𝑎,𝑏) give solutions for every real 𝑡? Which give solutions for every 𝑡 ≥0? Sketch both sets in the initial-data plane. Explain why changing 𝑎 does not change the interval of existence.