Existence and uniqueness

Most equations y’ = f(x, y) have no formula for their solutions. When does an initial value problem have a solution, when is it the only one, and how can it be built without a formula?

Where you are 6 in First-order equations

Why this exists

Separation in Separable equations in general and integrating factors in Linear equations and integrating factors end in formulas, and a formula shows by itself that a solution exists and is the only one. Most equations, such as y' = x^2 + y^2 in Direction fields and solution curves, have no formula. Without another way to settle both questions, a sketch, a numerical solution or a model rests on nothing.

Picard’s theorem settles both. Its proof builds the solution as a limit of explicit approximations, and it proves the existence and no-crossing facts that Direction fields and solution curves stated ahead of time. Throughout, y is a function of x, or of t when the variable is time.

Definition

In plain words: f is Lipschitz in y when changing y changes f by at most a fixed multiple of the change.

An initial value problem y' = f(x, y), y(x_0) = y_0, and its solutions are as in Direction fields and solution curves; on a closed interval, the derivative at an end is one-sided. Fix a, b > 0 and the closed rectangle

R = \{(x, y) : |x - x_0| \le a,\ |y - y_0| \le b\}.

A function f is Lipschitz in y on a set S in the plane if there is a constant L \ge 0, a Lipschitz constant, with

|f(x, y_1) - f(x, y_2)| \le L\,|y_1 - y_2|

whenever (x, y_1) and (x, y_2) lie in S. Continuity of f means continuity as a function of two variables (Limits and continuity in several variables).

For f continuous on R, the Picard iterates, named for Émile Picard, are

\begin{aligned} \varphi_0(x) &= y_0, \\ \varphi_{n+1}(x) &= y_0 + \int_{x_0}^{x} f\big(t, \varphi_n(t)\big)\,dt, \end{aligned}

defined as long as the graph of \varphi_n stays in R.

Intuition

The fundamental theorem of calculus turns the equation into y(x) = y_0 + \int_{x_0}^{x} f(t, y(t))\,dt (Result 1). The right side is a machine that takes a function and returns another, and a solution is a function the machine returns unchanged. Picard iteration feeds the machine its own output, starting from the constant y_0.

The Lipschitz condition makes the machine pull functions together. If two inputs differ by at most \delta, the integrands differ by at most L\delta, so the outputs differ by at most L\delta\,|x - x_0|. Fed back n times, the gap is multiplied by \dfrac{(L|x - x_0|)^n}{n!}, which tends to 0: the iterates settle down, and two solutions, both returned unchanged, must coincide.

The interval comes from the slopes. If |f| \le M on R, a solution from (x_0, y_0) stays between the lines of slope \pm M through that point (Figure 2), and these reach the top and bottom of R at \displaystyle |x - x_0| = \frac{b}{M}.

Results

1. The integral equation. Let f be continuous on R, J an interval containing x_0, and \varphi continuous on J with graph in R. Then \varphi solves y' = f(x, y), y(x_0) = y_0 on J if and only if, for all x \in J,

\varphi(x) = y_0 + \int_{x_0}^{x} f\big(t, \varphi(t)\big)\,dt.

2. A Lipschitz constant from the partial derivative. If \dfrac{\partial f}{\partial y} (Partial derivatives) exists and is continuous on R, then f is Lipschitz in y on R with L = \max_R \left|\dfrac{\partial f}{\partial y}\right|.

3. Picard’s theorem. Let f be continuous on R and Lipschitz in y on R with constant L, let M > 0 with |f| \le M on R, and put

\begin{aligned} h &= \min\left(a, \frac{b}{M}\right), \\ I_h &= [x_0 - h, x_0 + h]. \end{aligned}

(a) The iterates are defined on I_h with graphs in R and converge to a solution \varphi of the initial value problem on I_h, with

|\varphi(x) - \varphi_n(x)| \le \frac{M L^n |x - x_0|^{n+1}}{(n+1)!}.

(b) Every solution on an interval J \subseteq I_h containing x_0, with graph in R, equals \varphi on J.

4. The local theorem. Let D be an open set in the plane, f and \dfrac{\partial f}{\partial y} continuous on D, and (x_0, y_0) \in D. Then the initial value problem has a solution on an open interval containing x_0, and any two of its solutions agree on some open interval containing x_0.

Result 4 is Result 3 of Direction fields and solution curves, so the no-crossing and order results proved there from it now rest on a proof.

Why it is true

One fact from multivariable calculus is used without proof: a continuous function on a closed rectangle is bounded and attains its maximum. Also, if \psi is continuous with graph in R, then t \mapsto f(t, \psi(t)) is continuous, as a composition of continuous functions.

Result 1. If \varphi solves, then \varphi'(x) = f(x, \varphi(x)) is continuous on J, and the fundamental theorem gives \varphi(x) - \varphi(x_0) = \int_{x_0}^{x} \varphi'(t)\,dt, which is the integral equation. Conversely, if the integral equation holds, its integrand is continuous, so the fundamental theorem makes \varphi differentiable with \varphi'(x) = f(x, \varphi(x)), and \varphi(x_0) = y_0. \square

Result 2. Fix x and y_1, y_2 \in [y_0 - b, y_0 + b]. The mean value theorem for y \mapsto f(x, y) gives f(x, y_1) - f(x, y_2) = \dfrac{\partial f}{\partial y}(x, \eta)\,(y_1 - y_2) for some \eta between y_1 and y_2, and (x, \eta) \in R. \square

Result 3. The iterates stay in R. Suppose \varphi_n is continuous on I_h with graph in R, as \varphi_0 is. Then f(t, \varphi_n(t)) is continuous with absolute value at most M, so \varphi_{n+1} is defined and differentiable on I_h, and for x, x' \in I_h

|\varphi_{n+1}(x) - \varphi_{n+1}(x')| \le M|x - x'|.

In particular |\varphi_{n+1}(x) - y_0| \le M|x - x_0| \le Mh \le b.

The key bound (*). Let \psi be continuous on an interval J \subseteq I_h containing x_0, with graph in R and \psi(x) = y_0 + \int_{x_0}^{x} f(t, \psi(t))\,dt on J. We prove by induction that for every n \ge 0 and x \in J

|\psi(x) - \varphi_n(x)| \le \frac{M L^n |x - x_0|^{n+1}}{(n+1)!}.

For n = 0, |\psi(x) - y_0| = \left|\int_{x_0}^{x} f(t, \psi(t))\,dt\right| \le M|x - x_0|. If (*) holds for n, the Lipschitz condition gives

\begin{aligned} |\psi(x) - \varphi_{n+1}(x)| &\le \left|\int_{x_0}^{x} L\,|\psi(t) - \varphi_n(t)|\,dt\right| \\ &\le \left|\int_{x_0}^{x} \frac{M L^{n+1} |t - x_0|^{n+1}}{(n+1)!}\,dt\right|, \end{aligned}

which is (*) for n + 1. The same steps prove |\varphi_{n+1}(x) - \varphi_n(x)| \le c_n = \dfrac{M L^n h^{n+1}}{(n+1)!} on I_h: for n = 0 it is the bound on \varphi_1 - y_0, and \varphi_{n+2} - \varphi_{n+1} is the integral of f(t, \varphi_{n+1}) - f(t, \varphi_n).

Existence. If L > 0, then \dfrac{c_{n+1}}{c_n} = \dfrac{Lh}{n+2} \to 0, so \sum c_n converges by the The ratio and root tests; if L = 0, every term past c_0 is 0. For each x \in I_h the series y_0 + \sum_{k \ge 0} \big(\varphi_{k+1}(x) - \varphi_k(x)\big), whose partial sums are the \varphi_n(x), converges absolutely by comparison. Call its sum \varphi(x). Then for every x \in I_h

|\varphi(x) - \varphi_n(x)| \le \varepsilon_n = \sum_{k \ge n} c_k,

and \varepsilon_n \to 0 as the tail of a convergent series. One bound serves every x: this is uniform convergence (Pointwise and uniform convergence). Letting n \to \infty in the two inequalities for \varphi_{n+1} above gives |\varphi(x) - \varphi(x')| \le M|x - x'| and |\varphi(x) - y_0| \le b, so \varphi is continuous with graph in R. Since |f(t, \varphi_n(t)) - f(t, \varphi(t))| \le L\varepsilon_n, the integrals of these two functions from x_0 to x differ by at most L\varepsilon_n h \to 0. Letting n \to \infty in the definition of \varphi_{n+1} gives the integral equation for \varphi, and by Result 1, \varphi solves on I_h. The bound in (a) is (*) with \psi = \varphi.

Uniqueness. A solution \psi as in (b) satisfies the integral equation by Result 1, so (*) holds. Its right side is at most c_n, which tends to 0, so \psi(x) = \lim \varphi_n(x) = \varphi(x). \square

Result 4. Since D is open, it contains a closed rectangle R centered at (x_0, y_0). On R, |f| \le M for some M > 0, and f is Lipschitz in y by Result 2. Result 3 gives a solution on I_h, so on (x_0 - h, x_0 + h). Two solutions are continuous with value y_0 at x_0, so for some \delta \in (0, h] both have graphs in R over (x_0 - \delta, x_0 + \delta), and by Result 3(b) both equal \varphi there. \square

Worked examples

Example 1

Iterates that rebuild the tangent

Compute the Picard iterates \varphi_1, \varphi_2, \varphi_3 for y' = 1 + y^2, y(0) = 0, and compare them with the solution \tan x.

Solution

Here \varphi_0 = 0, and each iterate integrates 1 + \varphi_n(t)^2 from 0:

\begin{aligned} \varphi_1(x) &= \int_0^x 1\,dt = x, \\ \varphi_2(x) &= \int_0^x (1 + t^2)\,dt = x + \frac{x^3}{3}, \\ \varphi_3(x) &= \int_0^x \left(1 + t^2 + \frac{2t^4}{3} + \frac{t^6}{9}\right)dt \\ &= x + \frac{x^3}{3} + \frac{2x^5}{15} + \frac{x^7}{63}. \end{aligned}

The function \tan x solves the problem on \displaystyle \left(-\frac{\pi}{2}, \frac{\pi}{2}\right), since (\tan x)' = \sec^2 x = 1 + \tan^2 x, and its Maclaurin series begins

\begin{aligned} \tan x = x &+ \frac{x^3}{3} + \frac{2x^5}{15} \\ &+ \frac{17x^7}{315} + \cdots. \end{aligned}

So \varphi_3 matches it through x^5 and misses at x^7, where \displaystyle \frac{1}{63} = \frac{5}{315}. Near 0 the iterates close in on \tan x (Figure 1).

−1 1 −4 −2 2 4 x y
\varphi_1
\varphi_2
\varphi_3
\tan x
Figure 1The Picard iterates \varphi_1 = x, \displaystyle \varphi_2 = x + \frac{x^3}{3} and \varphi_3 of y' = 1 + y^2, y(0) = 0 (dashed), against the solution \tan x (thick), as in Example 1.
Example 2

The guaranteed interval

For the same problem, find the largest h that Result 3 gives from rectangles centered at (0, 0), and bound |\tan x - \varphi_3(x)| on I_h.

Solution

On R the right side 1 + y^2 is at most M = 1 + b^2, and \displaystyle \frac{\partial f}{\partial y} = 2y gives L = 2b by Result 2. So h = \min\left(a, \dfrac{b}{1 + b^2}\right). Since 1 + b^2 - 2b = (b - 1)^2 \ge 0, the quotient is at most \displaystyle \frac{1}{2}, with equality only at b = 1. With a = b = 1, so that M = 2 and L = 2, Result 3 gives \displaystyle h = \frac{1}{2}, and no rectangle does better (Figure 2).

−1 −0.5 0.5 1 −1 1 x y R
Figure 2The rectangle R with a = b = 1 for y' = 1 + y^2 (Example 2). Slopes on R are at most M = 2, so a solution through the origin stays between the lines y = \pm 2x, which reach the top and bottom of R at \displaystyle x = \pm\frac{1}{2} = \pm h. The solution \tan x leaves R at \displaystyle x = \pm\frac{\pi}{4}.

Result 3(a) with n = 3 gives, for \displaystyle |x| \le \frac{1}{2},

\begin{aligned} |\tan x - \varphi_3(x)| &\le \frac{2 \cdot 2^3\,x^4}{4!} \\ &= \frac{2}{3}x^4 \le \frac{1}{24}. \end{aligned}

The actual error at \displaystyle x = \frac{1}{2} is \displaystyle \tan\frac{1}{2} - \varphi_3\left(\frac{1}{2}\right) \approx 0.00035: the bound is safe but crude. So is the interval: the solution lives on \displaystyle \left(-\frac{\pi}{2}, \frac{\pi}{2}\right), more than three times as long as I_h, but Result 3 sees only the rectangle, which \tan x leaves at \displaystyle x = \frac{\pi}{4}.

Example 3

Lipschitz without a derivative

Show that \displaystyle y' = \sqrt{x^2 + y^2}, y(0) = 0, has exactly one solution on I_h for the rectangle with a = b = 1, although \displaystyle \frac{\partial f}{\partial y} does not exist at (0, 0), and prove that this solution is odd.

Solution

On the line x = 0, f(0, y) = |y|, which has no derivative at y = 0, so Results 2 and 4 do not apply. The Lipschitz condition holds anyway, with L = 1 on the whole plane. If y_1 and y_2 are both 0 there is nothing to prove. Otherwise put \displaystyle S = \sqrt{x^2 + y_1^2} + \sqrt{x^2 + y_2^2}, which is at least |y_1| + |y_2| > 0. Multiplying and dividing by S,

\begin{aligned} |f(x, y_1) - f(x, y_2)| &= \frac{|y_1^2 - y_2^2|}{S} \\ &\le \frac{|y_1 - y_2|\,(|y_1| + |y_2|)}{|y_1| + |y_2|}, \end{aligned}

which is |y_1 - y_2|. On R, \displaystyle M = \sqrt{2}, so Result 3 gives \displaystyle h = \frac{1}{\sqrt{2}} and exactly one solution \varphi on I_h with graph in R. Its first iterate is \displaystyle \varphi_1(x) = \int_0^x |t|\,dt = \frac{x|x|}{2}.

Proof that \varphi is odd. Put \psi(x) = -\varphi(-x) on I_h, which is symmetric about 0. Then \psi(0) = 0, the graph of \psi lies in R, and by the chain rule

\psi'(x) = \varphi'(-x) = \sqrt{x^2 + \psi(x)^2}.

So \psi solves the same problem, and Result 3(b) gives \psi = \varphi, that is \varphi(-x) = -\varphi(x). \square

Example 4

Priya and Samir’s tank

At the firehouse, Priya drains a training tank through a valve at its base. By Torricelli’s law the water depth y, in meters, t minutes after she opens the valve, obeys

y' = -\frac{1}{10}\sqrt{y}, \qquad y \ge 0.

(a) The tank starts 1.44 meters deep. When is it empty? (b) The next morning Samir finds the tank empty at 9:00 (take t = 0 then) and wonders when it was full. Show that the equation cannot tell him.

Solution

Here y is a function of t. For each constant c let

y_c(t) = \begin{cases} \dfrac{(c - t)^2}{400}, & t \le c, \\[4pt] 0, & t > c. \end{cases}

For t < c, \displaystyle \sqrt{y_c} = \frac{c - t}{20} and \displaystyle y_c' = -\frac{c - t}{200} = -\frac{1}{10}\sqrt{y_c}. For t > c both sides are 0, and at t = c both one-sided derivatives are 0, as is \displaystyle -\frac{1}{10}\sqrt{0}. So every y_c solves on \mathbb{R}.

(a) The solution with y(0) = 1.44 is y_c with \displaystyle \frac{c^2}{400} = 1.44 and c > 0, so c = 24: the tank is empty after 24 minutes. While y > 0, the right side has the continuous derivative \displaystyle -\frac{1}{20\sqrt{y}}, so Result 4 and no crossing make this the only solution until then; after that, y' \le 0 and y \ge 0 keep y = 0.

(b) Every y_c with c \le 0, and the zero function, has y(0) = 0. The solution y_c was 1.44 meters deep at t = c - 24, so the tank may have been full at any time 24 or more minutes before 9:00, or never (Figure 3). Result 3 fails at (0, 0) because

\frac{|f(y) - f(0)|}{|y - 0|} = \frac{1}{10\sqrt{y}}

is unbounded as y \to 0^+: no Lipschitz constant exists near y = 0. Forward in time the tank stays empty; Exercise 11 explains why uniqueness survives in that direction.

−40 −30 −20 −10 0 10 0.5 1 1.5 t y
Figure 3Solutions y_c of the tank equation \displaystyle y' = -\frac{1}{10}\sqrt{y} for c = -20, -10, 0 (Example 4). All of them, and the zero solution, have y(0) = 0; the dashed line marks a full tank at 1.44 meters.
Example 5

A short proof: solutions on a whole strip

Let f be continuous on the strip S = [\alpha, \beta] \times \mathbb{R} and Lipschitz in y on S with constant L, and let x_0 \in [\alpha, \beta]. Prove that the initial value problem has a solution on all of [\alpha, \beta], and that every solution on an interval J \subseteq [\alpha, \beta] containing x_0 agrees with it on J.

Solution

Since f is defined for every y, the iterates need no rectangle: by induction each \varphi_n is continuous on [\alpha, \beta], as in the proof of Result 3. Let \ell = \beta - \alpha and let K be the maximum of the continuous function |f(x, y_0)| on [\alpha, \beta]. Then |\varphi_1(x) - \varphi_0(x)| \le K|x - x_0|, and the induction of Result 3 gives

\begin{aligned} |\varphi_{n+1}(x) - \varphi_n(x)| &\le \frac{K L^n |x - x_0|^{n+1}}{(n+1)!} \\ &\le \frac{K L^n \ell^{n+1}}{(n+1)!}. \end{aligned}

These bounds form a convergent series, so as in Result 3 the iterates converge to a function \varphi with |\varphi - \varphi_n| \le \varepsilon_n on [\alpha, \beta], where \varepsilon_n \to 0. Continuity of \varphi takes one more step: given \eta > 0, choose n with \varepsilon_n < \eta. Then \varphi(x') and \varphi(x) are each within \eta of \varphi_n(x') and \varphi_n(x), which are within \eta of each other for x' near x. Passing to the limit in the integral as in Result 3 gives the integral equation, so \varphi solves on [\alpha, \beta].

For uniqueness, let \psi solve on J, let [c, d] \subseteq J be a closed interval containing x_0, and let B be the maximum of |\psi - y_0| on [c, d]. The case n = 0 being the definition of B, the induction of (*) gives, on [c, d],

|\psi(x) - \varphi_n(x)| \le \frac{B L^n |x - x_0|^n}{n!}.

The right side tends to 0, so \psi = \varphi on [c, d], and every point of J lies in such an interval. \square

Boundaries

  • Continuity alone gives existence, not uniqueness. Peano’s theorem says that f continuous on R already gives a solution on I_h; its proof, by the Arzelà-Ascoli theorem, is beyond this lesson. Example 4 and Exercise 7 show uniqueness failing, and without continuity a solution need not exist (Exercise 12).
  • The Lipschitz condition is sufficient, not necessary. Exercise 13 has a unique solution with no Lipschitz constant, and Exercise 11 gets uniqueness in one direction from monotonicity.
  • The theorem is local. h can be far shorter than the solution’s interval (Example 2), and a solution can become unbounded, as \tan x does. A Lipschitz condition on a whole strip (Example 5) gives a solution on the whole interval at once.

Where this goes next

Autonomous equations and stability uses uniqueness at every turn: a solution that starts between two equilibria never reaches either. The proof of Result 3 works word for word for a system \mathbf{x}' = \mathbf{F}(t, \mathbf{x}), with |\cdot| read as the length of a vector, and a second-order equation is a system of two first-order ones; Constant-coefficient equations and Linear systems of differential equations rely on this. Pointwise and uniform convergence treats uniform limits in general, and Euler's method approximates the solution this lesson guarantees.

Exercises

Picard iterates

  1. For y' = y + 1, y(0) = 0, find \varphi_1, \varphi_2, \varphi_3, prove by induction that \displaystyle \varphi_n(x) = \sum_{k=1}^{n} \frac{x^k}{k!}, and identify the limit.
  2. Find \varphi_1, \varphi_2, \varphi_3 for y' = x - y^2, y(0) = 0.

Intervals and error bounds

  1. For y' = x + y^2, y(0) = 0, and the rectangle with a = b = 1, find M, L and h, and bound |\varphi(x) - \varphi_2(x)| on I_h.
  2. For y' = y^3, y(0) = 1, with a = 1, find the b that makes Result 3’s h largest, and that h. Then solve the problem and give its maximal interval.
  3. For Exercise 1 with \displaystyle a = \frac{1}{2} and b = 1, find h, bound |\varphi - \varphi_4| on I_h, and compare with the actual error at \displaystyle x = \frac{1}{2}.

Hypotheses

  1. Decide whether f is Lipschitz in y on the given set, and find the smallest Lipschitz constant when it is: (a) f = x\sin y on [-2, 2] \times \mathbb{R}; (b) f = \dfrac{1}{1 + y^2} on the plane; (c) \displaystyle f = \sqrt[3]{y} on [-1, 1] \times [-1, 1]; (d) f = y^2 on the plane.
  2. With the real cube root, \displaystyle (y - 2)^{2/3} = \left(\sqrt[3]{y - 2}\right)^2. (a) Show that Result 3 applies to y' = 3(y - 2)^{2/3}, y(0) = 2, on no rectangle, and find infinitely many solutions on \mathbb{R}. (b) Find two different solutions on \mathbb{R} with y(0) = 10.
  3. Show that Result 3 applies to y' = 1 + |y|, y(0) = 0, although \displaystyle \frac{\partial f}{\partial y} does not exist on the line y = 0, and solve the problem on \mathbb{R}.
  4. Use Example 5 to show that y' = \cos(xy), y(0) = 0, has exactly one solution on \mathbb{R}, and prove that it is odd.

Proofs

  1. Let f be continuous on S = [x_0, \beta] \times \mathbb{R} and Lipschitz in y on S with constant L, and let \psi_1 and \psi_2 solve y' = f(x, y) on [x_0, \beta]. With \delta = |\psi_1(x_0) - \psi_2(x_0)|, prove that |\psi_1(x) - \psi_2(x)| \le \delta\,e^{L(x - x_0)} there.
  2. Let f be continuous on [x_0, \beta] \times \mathbb{R} and nonincreasing in y for each x. Prove that two solutions on [x_0, \beta] with the same value at x_0 are equal. Explain how this gives the uniqueness for t \ge 0 in Example 4, and why it says nothing for t \le 0.
  3. Let f(x, y) = -1 for y \ge 0 and f(x, y) = 1 for y < 0. Prove that y' = f(x, y), y(0) = 0, has no solution on any open interval containing 0.
  4. Refute: “If an initial value problem has exactly one solution, then f is Lipschitz in y near the initial point.” Use f(x, y) = y\ln|y| for y \ne 0, f(x, 0) = 0, at (0, 0), and Result 3 of Separable equations in general.
  5. Let p and q be continuous on an open interval I containing x_0. Use Example 5 to prove that y' + p(x)\,y = q(x), y(x_0) = y_0, has a solution on all of I and that every solution on an interval J \subseteq I containing x_0 is its restriction to J.
Answers
  1. \varphi_1 = x, \displaystyle \varphi_2 = x + \frac{x^2}{2}, \displaystyle \varphi_3 = x + \frac{x^2}{2} + \frac{x^3}{6}. If the formula holds for \varphi_n, then \varphi_{n+1} is the integral from 0 to x of \displaystyle 1 + \sum_{k=1}^{n} \frac{t^k}{k!}, which is \displaystyle x + \sum_{k=1}^{n} \frac{x^{k+1}}{(k+1)!}, the formula for n + 1. The limit is e^x - 1 (The common Maclaurin series), which solves: (e^x - 1)' = e^x = (e^x - 1) + 1.
  2. \displaystyle \varphi_1 = \frac{x^2}{2}, \displaystyle \varphi_2 = \frac{x^2}{2} - \frac{x^5}{20}, and since \displaystyle \varphi_2(t)^2 = \frac{t^4}{4} - \frac{t^7}{20} + \frac{t^{10}}{400}, \displaystyle \varphi_3 = \varphi_2 + \frac{x^8}{160} - \frac{x^{11}}{4400}.
  3. M = 2, attained at |x| = |y| = 1; L = 2 from \displaystyle \frac{\partial f}{\partial y} = 2y; \displaystyle h = \min\left(1, \frac{1}{2}\right) = \frac{1}{2}. With n = 2, \displaystyle |\varphi - \varphi_2| \le \frac{2 \cdot 2^2 |x|^3}{3!} = \frac{4}{3}|x|^3 \le \frac{1}{6}.
  4. On R, M = (1 + b)^3, so \displaystyle h = \min\left(1, \frac{b}{(1 + b)^3}\right). The derivative of b(1 + b)^{-3} is (1 - 2b)(1 + b)^{-4}, zero at \displaystyle b = \frac{1}{2}, so the largest h is \displaystyle \frac{1}{2}\left(\frac{2}{3}\right)^3 = \frac{4}{27} \approx 0.148. Separating, \displaystyle -\frac{1}{2y^2} = x - \frac{1}{2}, so \displaystyle y = \frac{1}{\sqrt{1 - 2x}} on \displaystyle \left(-\infty, \frac{1}{2}\right), reaching more than three times as far right as h.
  5. M = 2, L = 1, \displaystyle h = \min\left(\frac{1}{2}, \frac{1}{2}\right) = \frac{1}{2}, and \displaystyle |\varphi - \varphi_4| \le \frac{2|x|^5}{5!} \le \frac{1}{1920} \approx 0.00052. At \displaystyle \frac{1}{2}, e^{1/2} - 1 \approx 0.64872 and \varphi_4 \approx 0.64844, so the actual error is about 0.00028.
  6. (a) Yes, L = 2: |x\sin y_1 - x\sin y_2| \le |x|\,|y_1 - y_2| by the mean value theorem for \sin, and no smaller constant works, since at x = 2 the ratio \displaystyle \frac{2\sin y}{y} tends to 2 as y \to 0. (b) Yes, \displaystyle L = \frac{3\sqrt{3}}{8} \approx 0.6495, the maximum of \displaystyle \left|\frac{\partial f}{\partial y}\right| = \frac{2|y|}{(1 + y^2)^2}, reached at \displaystyle y = \pm\frac{1}{\sqrt{3}}, where difference quotients tend to it, so none smaller works. (c) No: \displaystyle \frac{|f(x, y) - f(x, 0)|}{|y|} = |y|^{-2/3} is unbounded as y \to 0. (d) No: \displaystyle \frac{|y_1^2 - y_2^2|}{|y_1 - y_2|} = |y_1 + y_2| is unbounded, though on |y| \le b the constant 2b works.
  7. (a) Near y = 2 the ratio \displaystyle \frac{|f(y) - f(2)|}{|y - 2|} = 3|y - 2|^{-1/3} is unbounded, so f is Lipschitz on no rectangle around (0, 2). For each c \ge 0, the function equal to 2 for x \le c and to 2 + (x - c)^3 for x > c solves: on the cubic, (y - 2)^{2/3} = (x - c)^2, so y' = 3(x - c)^2 = 3(y - 2)^{2/3}, and both one-sided derivatives at c are 0. (b) y = 2 + (x + 2)^3 on \mathbb{R}, and the function equal to 2 for x \le -2 and to 2 + (x + 2)^3 for x > -2.
  8. \big||y_1| - |y_2|\big| \le |y_1 - y_2|, so L = 1 on every rectangle, and |f| \le 1 + b there; but 1 + |y| has no derivative in y at y = 0. The solution is y = e^x - 1 for x \ge 0 and y = 1 - e^{-x} for x < 0. Each piece solves on its side, where |y| = y and |y| = -y respectively, and both one-sided derivatives at 0 are 1 = f(0, 0). By Example 5 on each strip [-a, a] \times \mathbb{R} it is the only solution.
  9. On [-a, a] \times \mathbb{R}, |\cos(xy_1) - \cos(xy_2)| \le a|y_1 - y_2| by the mean value theorem for \cos. Example 5 gives a unique solution \varphi_a on each [-a, a]; for a < a', \varphi_{a'} solves on [-a, a], so it equals \varphi_a there. Together they define one solution \varphi on \mathbb{R}, and any solution on \mathbb{R} equals \varphi_a on each [-a, a]. For oddness, \psi(x) = -\varphi(-x) has \psi(0) = 0 and \psi'(x) = \varphi'(-x) = \cos\big(x\psi(x)\big), since (-x)\,\varphi(-x) = x\,\psi(x). So \psi = \varphi by uniqueness. \square
  10. The proof of Result 1 works on S, so D = |\psi_1 - \psi_2| satisfies D(x) \le \delta + L\int_{x_0}^{x} D(t)\,dt. Call the right side U(x). Then U' = LD \le LU, so e^{-L(x - x_0)}U has derivative e^{-L(x - x_0)}(U' - LU) \le 0, and e^{-L(x - x_0)}U(x) \le U(x_0) = \delta. Hence D \le U \le \delta e^{L(x - x_0)}. \square (This is Grönwall’s inequality.)
  11. Let d = \psi_1 - \psi_2. Then (d^2)' = 2d\,\big(f(x, \psi_1) - f(x, \psi_2)\big) \le 0, because f nonincreasing in y makes the two factors of opposite signs or zero. So d^2 is nonincreasing, with d^2(x_0) = 0 and d^2 \ge 0, hence d = 0. \square In Example 4, extend the right side by 0 for y < 0; it stays continuous and nonincreasing, so the solution with y(0) = 0 is unique for t \ge 0. Backward in time, z(s) = y(-s) solves \displaystyle z' = \frac{1}{10}\sqrt{z}, whose right side is nondecreasing, so the argument does not apply, and uniqueness indeed fails.
  12. Suppose \varphi solves on an open interval containing 0. Then \varphi'(0) = f(0, 0) = -1 < 0, so \varphi(x) < \varphi(0) = 0 for all x in some (0, \delta). There \varphi'(x) = f(x, \varphi(x)) = 1, so for 0 < x < \delta the mean value theorem on [0, x] gives \varphi(x) = \varphi(0) + x\,\varphi'(\xi) = x > 0, a contradiction. \square
  13. f is continuous, since y\ln|y| \to 0 as y \to 0, but not Lipschitz near (0, 0): \displaystyle \frac{|f(x, y) - f(x, 0)|}{|y|} = \big|\ln|y|\big| \to \infty. Yet the solution is unique. With g = 1 and h(y) = y\ln|y|, h \ne 0 on \displaystyle \left(0, \frac{1}{2}\right], and the integral of \displaystyle \frac{1}{s|\ln s|} from \varepsilon to \displaystyle \frac{1}{2} is \ln|\ln\varepsilon| - \ln\ln 2, which tends to \infty. By Result 3(a) of Separable equations in general, no solution takes both the value 0 and a value in \displaystyle \left(0, \frac{1}{2}\right); h is odd, so the same holds for \displaystyle \left(-\frac{1}{2}, 0\right). A solution with y(0) = 0 that is not identically 0 would take values in one of these intervals by the intermediate value theorem. So y = 0 is the only solution, and the claim is false.
  14. Fix a closed [\alpha, \beta] \subseteq I containing x_0. Then f(x, y) = q(x) - p(x)\,y is continuous on [\alpha, \beta] \times \mathbb{R}, and |f(x, y_1) - f(x, y_2)| = |p(x)|\,|y_1 - y_2|, so L = \max_{[\alpha, \beta]} |p| works. Example 5 gives one solution on [\alpha, \beta], and two such solutions agree on the overlap of their intervals by its uniqueness part. Closed intervals [\alpha_k, \beta_k] containing x_0, increasing with union I, thus give one solution \varphi on I. A solution \psi on J equals \varphi on every closed [c, d] \subseteq J containing x_0, since [c, d] lies in some [\alpha_k, \beta_k]. \square This reproves Result 1 of Linear equations and integrating factors without its formula.
math.wiki · CC-BY-SA-4.0