Every solver you have written assumes that the initial value problem has one answer, so that “the” solution is something Euler, RK4 and RK45 can all approximate. Nothing in the code checks it. You give the integrator and a starting point, and it returns numbers.
Two loose ends from earlier lessons are paid off here. In 3.3 Separable Equations the solution of reached infinity at a finite time, although is as smooth as a function can be. In 3.6 Exact Equations & Substitutions the equation had two solutions from , and a solver started at exactly “sees slope 0” and never finds the second. One idea explains both: a speed limit on how fast the slope field may tilt.
Transcript
- A solver returns a number whatever you give it. Hand it y prime equals f of t and y, plus a start, and it never checks whether there is one answer, many, or one that stops early. When can we trust it?
- Fix t. The slope is a function of y. Lipschitz is a speed limit on how fast it tilts: through a point of the graph, lines of slope plus and minus L form a cone, and the graph must stay inside. The line y does. The cube root, vertical at zero, escapes.
- Take secants from the origin on grids of spacing 0.2, 0.02 and 0.002. Their slopes are 2.92, 13.57 and 62.996, and keep growing. Any L is eventually beaten: the cube root is not Lipschitz at zero.
- Now y prime equals cube root of y, from zero. Zero is a solution. So is a branch that waits until time c, then leaves: two thirds of t minus c, to the power three halves, either sign. At t equals one, c zero gives 0.5443, c one half gives 0.1925, and everything between plus and minus 0.5443 is reached.
- Existence can be local. y prime equals one plus y squared, from zero, has the solution tangent t, reaching infinity at pi over two, 1.5708, and backward at minus pi over two. That open interval is the maximal interval of existence.
- A smooth right side does this too. For y prime equals y squared the solution ends at one over y0: start higher, end sooner. The theorem never promised more than local.
- What does the theorem promise? On a box of half-height b, the slope is at most y0 plus b, squared, so the guaranteed length is b over that. For y0 one it peaks at b one, giving one quarter; in general one over four y0.
- Chain those guarantees along y equals one over one minus t. Each step lands on the curve, and the totals run 0.25, 0.4375, 0.5781, 0.6836, 0.7627. The full sum is exactly one over y0, the blow-up time. Local guarantees, chained, land on the wall.
- Solutions cannot cross. For y prime equals y times one minus y, zero and one are solutions, so a solution starting between them is trapped in the strip. Bounded, it cannot blow up: it exists for all time, and we never solved anything.
- Where Lipschitz fails, this changes. Run any cube-root branch backward: it reaches zero with slope zero and merges into the zero solution. Curves can merge exactly where f stops being Lipschitz.
- Picard iteration: feed a guess into one plus the integral of f. For y prime equals y, start at one, and get one plus t, then plus t squared over two, then the cubic term. These are the Taylor partial sums, converging to e to the t.
- Without Lipschitz, the machine picks a side. From a guess of zero it returns zero forever. From plus one thousandth, 200 passes give 0.5442 at t equals one, the c zero branch. One answer of infinitely many, silently.
- Nearby starts for y prime equals L y separate like e to the L t. From one and one plus ten to the minus eight, the gap at t equals two is ten to the minus eight times e to the two L. Euler's gap over that is 0.999, 0.975, 0.905 for L one, five, ten; for L ten it reached 4.39. Large L is a warning, and previews stiffness.
- Navier-Stokes has this shape. In three dimensions smooth solutions are known for a short time. For all time is open, a Millennium Prize worth one million dollars. In two dimensions, yes. As analogy only, y prime equals y squared is a local solution that ends.
What the question is
For the IVP , , three things can go wrong: no solution, several solutions, or exactly one that exists only for a short time. The direction field of 3.2 Direction Fields & Euler's Method Revisited suggests “one curve through every point”, the picture Euler’s method relies on. The theorem below says when it is trustworthy.
Lipschitz: a speed limit on tilting
Fix . The slope is a function of alone, and moving from height to tilts the arrow from to . We say is Lipschitz in with constant on a region if
The left side divided by is the slope of a secant line of the graph of . So no secant is steeper than : the arrows cannot tilt faster than linearly as you move in .
When exists, the mean value theorem gives for some between them. So if on the region, is Lipschitz with constant , and a continuous is bounded on any small closed box. Two cautions. A continuous is sufficient, not necessary: has no derivative at , yet every secant slope lies in , so . And Lipschitz is not the same as differentiable. Failure is always a secant slope that grows without limit.
The theorem, and what it does not say
Existence and uniqueness (Picard and Lindelof). Let be continuous on the box and Lipschitz in on (for instance, continuous on ). Let . Then the IVP has exactly one solution on , where .
The shape of has a reason: the slope never exceeds , so in time the graph moves at most and stays in the box where the hypotheses hold. The theorem is local. It guarantees some interval around and says nothing about the solution on the rest of the box, or for all . Continuity alone gives existence (Peano) but not uniqueness, as the next section shows.
- and . One thick curve from any start; the readout gives for every box, even for at , where does not exist. Tick also draw solutions from other y and the curves never meet.
- and . grows with the box, since . The dashed lines mark where the thick solution ends; move the start and the end moves.
- and . Click near the -axis (it snaps onto ): is unbounded for every box and several solutions leave one point. Click slightly above the axis: a small box has a finite and one solution.
Non-uniqueness:
Take , , with the real cube root. Dividing by loses , which solves the equation. For ,
Check: for , , and so is . Now set for and the branch for . Both pieces have value and slope at , so the glued function solves the equation. For every and each sign this is a solution of , plus : infinitely many through one point. At the branch gives , gives , and gives ; as runs over and the sign over both choices, takes every value in .
Where Lipschitz fails. is infinite at . The secant between and has slope , which is at and unbounded as .
Why that lets a solution leave zero. Suppose and for near . For a solution with , while it stays near ,
since . A nonnegative quantity that starts at and never increases stays at : a Lipschitz equation cannot lift a solution off an equilibrium. At distance the speed is at most , too small to start. For the speed is , which beats for small .
The loose end from lesson 3.6
The same thing explains , with right side . The part has derivative , harmless. The part has derivative , infinite at , so the theorem is silent at , and there are two solutions, and . The plainest case is : both and for solve it.
A solver sees slope 0. Euler at computes and does not move, at any step size. Start at for and Euler with reaches at , on its way to the branch . Starting values of and of a trillionth give and , and neither run warns you.
Existence is local, even for a smooth
Now the other loose end. For , is continuous everywhere, so the theorem gives a unique solution near every point. Yet the solution through is , which blows up at . And , has the solution , which exists on and ends at both sides. Through the solution is , so the interval has length whatever is.
The maximal interval of existence is the largest open interval around on which the solution is defined. It belongs to the solution, not the equation (lesson 3.3). A solution can end only by growing without bound or by reaching the edge of where is defined and locally Lipschitz; we use that standard fact without proof.
Why the theorem allows this. For , , the box has , so the guaranteed length is largest at : , after which . Repeat from there. With current value the best box has and , so the guaranteed step is , after which the solution is . Values grow by and steps shrink by , so the guaranteed time adds up to
exactly the blow-up time (60 steps from sum to ). The mechanism: grows with height, so each application guarantees less, and the guarantees never sum past .
The linear contrast. For the secant slope of in is at every height, so the box can be as tall as we like, and the formula of lesson 3.4 exists on every interval where and are continuous. Linear equations have global solutions; nonlinear ones in general do not.
Solutions cannot cross
Suppose two solutions of an equation satisfying the theorem meet at . Both solve the IVP with that data, so they agree near . The set where they agree is open, and closed by continuity, so they agree on their whole common interval. Two solution curves either never touch or coincide.
This is usable. For , the constants and are solutions and is continuous. A solution with starts between them and cannot cross either, so it is trapped in the strip for as long as it exists. Being bounded, it cannot blow up, so it exists for all . We learned this without solving anything. It is the seed of 3.8 Autonomous Equations, Equilibria & Stability, which turns equilibria and the strips between them into a theory of long-time behaviour.
Without uniqueness the picture changes. For , every nonzero solution run backward reaches at its own , meeting the zero solution with the same slope , and merges into it (tick the other-starts box in the sim). Curves can touch and merge exactly where stops being Lipschitz.
Picard iteration: building the solution
The proof of the theorem is a construction, and the construction is an algorithm. Integrating the ODE gives the equivalent integral equation . The unknown is on both sides, so treat the right side as a machine: feed in a guess, get a better one.
Worked example: , . Start from :
By induction , since integrating gives . The iterates are the Taylor partial sums of : each pass buys one more term.
Why it converges. Let be the largest gap on . The Lipschitz bound inside the integral gives , and repeating, , the general term of the series for . Terms of a convergent series go to zero, so the iterates converge. For each pass shrinks the gap: the machine is a contraction, with exactly one fixed point. That is existence and uniqueness at once.
Without Lipschitz. For , , the guess gives : the zero function is a fixed point, and the machine reports forever. Start from a positive constant such as and it converges to a different fixed point: 200 passes on a 2000-point grid give , matching the branch value up to grid error. A negative start gives . Without a contraction, each fixed point attracts some guesses, and the output depends on where you begin.
What this means for a numerical solver
(a) A solver returns a number, always. Euler on from gives at for and . From it gives and , and from the negatives. The true values at fill the interval , and the solver reported one point with no flag. This is lesson 3.3’s RK4 printing finite values past a blow-up: a finite output is not evidence that the IVP has an answer.
(b) sets how fast errors can grow. For two exact solutions in a region where has constant , , so grows at most like . For the gap is exactly , so the bound is attained. Run Euler on from and with : at the gap divided by is for , for and for , and for the gap of has become . A large means nearby solutions fly apart: the first warning of the stiffness met as in 3.2 Direction Fields & Euler's Method Revisited and treated in 4.5 Stability Regions & Stiffness. This bound is claimed only for exact solutions (proved above) and for Euler on the linear equation (computed below); error bounds for general methods are not derived here.
(c) Near a non-Lipschitz point, convergence theorems stop applying. Their proofs contain . With there is no bound, and (a) is the picture: which branch a solver selects depends on rounding noise.
Exercise: measuring L and running Picard
Implement a secant-slope estimate of the Lipschitz constant and the Picard machine on a grid, and use them on the lesson’s equations.
Problems
For , , use the box , . (a) Find and . (b) What interval does the theorem guarantee? (c) RK4 gives and reaches near . Does this contradict (b)?
Show worked solution
(a) , so works; , so .
(b) : one solution on .
(c) No. The theorem gives a lower bound, not a forecast; other boxes, or restarting as in the lesson, certify more.
For , : (a) name constant solutions that bracket and show the solution exists for all ; (b) find ; (c) compute from the closed form .
Show worked solution
(a) and are solutions, and gives uniqueness, so the solution stays in . Bounded, it cannot blow up: it exists for all .
(b) On , , so increases to a limit . If , then for , where , so is bounded below by a positive constant and , a contradiction. Hence .
(c) , matching RK4 with . It is below , matching .
Let (real cube root squared), . (a) Show and both solve it. (b) They cross at the origin: does this contradict the lesson? (c) Show that for and for solves it for each , and find all possible values of .
Show worked solution
(a) For , , so for either sign of . And .
(b) No. is unbounded at , so the hypotheses fail there and crossing is allowed there only.
(c) For , ; at both pieces have value and slope . Since , every solution is nondecreasing, so . While , , so . The family gives for , filling out , which is therefore the set of possible values.
For , , start from and compute . Compare with : through which power does each iterate agree?
Show worked solution
. . .
agrees with through , through , through : its coefficient is , not . In general is correct through (exact fractions confirm this up to ). The series has radius , the blow-up time, so these polynomials approximate the solution only inside its interval of existence.
Key Takeaways
- The IVP has exactly one solution near if is continuous and Lipschitz in (for instance continuous), on an interval of guaranteed length . It is local.
- Lipschitz means no secant of is steeper than . A bound on suffices, but is Lipschitz without being differentiable, and a continuous can fail it.
- , has infinitely many solutions. A Lipschitz equation cannot lift a solution off an equilibrium; and can, which is why had a second solution.
- Existence is local even for smooth : ends at and gives . Linear equations are global.
- Solution curves cannot cross where uniqueness holds, so a solution between two others stays between them (lesson 3.8). Where Lipschitz fails, curves can merge.
- Picard iteration is a contraction on a short interval; for it gives the Taylor series of .
- A solver always returns numbers, and at a non-unique IVP it silently picks a branch. governs how fast nearby solutions separate, like .
- 3D Navier-Stokes: short-time smooth solutions are known, global ones are open. The same local-versus-global gap as , as an analogy only.