A modulus finer than a Lipschitz constant
The usual local Lipschitz condition makes uniqueness of an initial-value problem accessible through Grönwall's inequality. It is sufficient, but it is not the only way to prevent two solutions from separating. The decisive issue is whether the permitted rate of separation is strong enough to move away from zero in finite time. To measure that rate we replace the linear bound by a general function of .
Definition 1.3.1
(Osgood modulus of continuity)
Let . A function is an Osgood modulus if it is continuous and nondecreasing, , for , and
The notation in (1) means that as for some, hence every, . A function on a rectangle has modulus in if for all and all , with the same for every .
The Lipschitz case is ; the definition keeps only the two features of that the uniqueness argument actually uses, monotonicity and the divergence (1).
Theorem 1.3.2
(Osgood's uniqueness criterion)
Let be continuous on and have an Osgood modulus on in . If and are classical solutions of on an interval , both with values in , and , then on .
This is a uniqueness statement; existence still requires a separate theorem.
Why divergence prevents separation
The proof compares the distance between the two solutions with a scalar quantity that obeys a differential inequality, and then measures how long that quantity would need to leave zero.
Proof
(Comparison with a regularized travel time)
Set
Subtracting the integral equations and the same for , and using the modulus, gives . The function is and nonnegative because is continuous, and
where we extend as the constant above if necessary. This extension is only for the comparison quantity , which may exceed ; it does not change the hypothesis on .
Suppose for some . For any , apply the chain rule to
Since , integration from to gives
As , monotone convergence makes the left side tend to , a contradiction. Therefore on , so there.
Proof X-ray
Where each hypothesis enters
The proof has three moves. First, the modulus converts the vector question into the scalar inequality ; this uses that the same works for every , since the integral runs over . Second, monotonicity of turns into ; without it, a larger argument would not give a larger bound. Third, the non-obvious move: one cannot divide by and integrate from , because and, under (1), is infinite for every . Adding makes a finite function, so the chain rule is legitimate.
Divergence is used only at the last line, where it makes the finite bound impossible. If (1) failed, the same inequality would read , which allows ; that is exactly the room exploited by the waiting-time solutions below.
Examples on either side of the threshold
The criterion is tested by computing the integral in (1) for concrete moduli and by checking that a given really satisfies the bound.
Worked example 1.3.3
(A Lipschitz bound is the basic case)
Let on the rectangle with . Take . It is continuous, increasing, zero only at , and
So Osgood's criterion contains the Lipschitz uniqueness theorem. Edge case: if does not depend on , the best Lipschitz constant is , but is not allowed because is required for . Any positive multiple of still bounds the zero difference, so works and uniqueness follows; indeed both solutions equal .
Worked example 1.3.4
(A non-Lipschitz modulus that still diverges)
Assume and set for , . Since there and as , it is continuous and increasing. The substitution , , gives
Now let on and . It is increasing, and , so is concave: increments over intervals of length decrease as the interval moves right, giving . Hence has modulus . It is not Lipschitz at zero, because . The theorem therefore shows that is the only solution of , , with values in .
Worked example 1.3.5
(Power moduli and the exponent one)
For with and ,
If , then and the limit is the finite number : (1) fails. If , then and the integral diverges. At the integral is , which diverges. So among powers, (1) holds exactly when . The case is degenerate: forces , so does not depend on . The interesting non-Lipschitz moduli therefore lie strictly between () and , as does.
For the failure of (1) does not by itself prove nonuniqueness for every equation with that modulus, but nonuniqueness can occur.
Counterexample mode
Continuity alone does not give uniqueness
Claim: any continuous modulus with gives uniqueness. Take and , the waiting-time example of the preceding initial-value note. For , , so has the modulus on any interval ; its reciprocal is integrable near zero. Both and solve the problem: and for . Two different solutions leave the same point, so the claim is false. The repair is the divergence (1); the example shows it cannot simply be dropped, while failure of (1) alone does not imply nonuniqueness.
Scope and direction of time
The proof works on a common interval on which both solutions remain in the rectangle. Local applications can shrink the rectangle and time interval around the initial point. For uniqueness backward in time, require to satisfy the same bound on , replace by , and apply the argument to . The hypotheses are uniform in on the chosen rectangle; pointwise continuity in without a common modulus is not the same assertion.
Common mistake
(Reading the criterion as a characterization)
A non-Osgood modulus does not prove that solutions branch. For any modulus must satisfy , so and has no Osgood modulus. Yet for , , every solution satisfies , so is nonincreasing from and forward in time. The criterion is sufficient, not necessary.
Common mistake
(Integrating from zero)
Under (1), for every , so the function that a direct separation of variables would use is not finite, and has no meaning. The proof must either integrate from a positive level or, as above, regularize with and let only at the end.
Practice and solutions
Quick check
(Does satisfy the divergence condition near zero?)
Solution · Answer
Yes: . As in the power example, a function with this modulus does not depend on , so the conclusion is valid but uninformative.
Quick check
(Does satisfy it?)
Solution · Answer
No: .
Quick check
(If with , identify a modulus and conclude uniqueness.)
Solution · Answer
, so use . Its reciprocal integral diverges, so two solutions with the same initial value coincide.
Quick check
(In the proof, always. Why may the comparison quantity still exceed ?)
Solution · Answer
accumulates over time; the only a priori bound is , which can be larger than on a long interval. That is why is extended by the constant before writing .