Motivation
An ordinary differential equation prescribes a rate of change rather than a function. The equation tells us what the slope must be wherever a solution passes. An initial condition selects a point through which the solution must pass. Solving the equation therefore involves several questions: whether a curve exists, whether that curve is uniquely determined, how long it can be continued, and how to represent it usefully.
An explicit formula answers some of these questions only after its domain is examined. The function satisfies and , but its initial value solution has interval . The same algebraic expression on describes a different solution branch. An interval crossing cannot be used, because a differentiable real-valued solution cannot have an infinite value there.
This note develops selected methods through the initial value viewpoint. Separation of variables and integrating factors are followed by linear systems, where the scalar exponential becomes a matrix exponential. The organizing principle is to check every transformation against the original equation and the initial condition.
Definitions
Definition 1.1.1
(Solution and initial value problem)
Let be open and let be continuous. A solution of on an interval is a continuously differentiable map whose graph lies in and which satisfies the equation at every time in . An initial value problem also specifies . A maximal solution cannot be extended to a larger interval while retaining the equation and initial data.
Intervals, rather than arbitrary subsets of the line, matter because the initial condition propagates through connected time. A formula defined on two disjoint intervals does not connect the behavior on one to the behavior on the other. For higher order equations, introduce the state variables . An equation solved for becomes a first order system and requires initial values.
For a scalar separable equation , division by is legitimate only on a region where . Define an antiderivative of there. The chain rule gives , so . Every zero of must also be checked separately: it produces a constant solution that division would discard.
When does separation describe the whole solution?
Suppose is continuous on a time interval , is continuously differentiable on an open state interval , and . Each zero of gives the constant solution . The local uniqueness theorem also says more: a solution starting at a state with cannot meet such a constant solution at any later or earlier time in its interval of existence. Indeed, if two solutions meet, uniqueness makes them equal near the meeting time. Their agreement set is both open and closed relative to their common interval, so it is the whole interval. Thus this solution stays in the same connected component of as .
On the function is continuous and has an antiderivative . Integrating the chain rule with the initial value gives
Conversely, wherever a differentiable function valued in satisfies (S), differentiating recovers the original equation. Since never vanishes and has one sign on the interval , is strictly monotone there; the relation therefore determines at most one value of for each . The initial-value solution is on the connected time interval containing for which the argument lies in the range of , within . This does not by itself say that the maximal time interval is all of : the inverse can cease to exist at a finite time, as in the blow-up example above.
Worked example 1.1.2
(Separation with a time-dependent rate)
Solve , . Here changes sign, but is smooth. Uniqueness keeps the solution in . Integrating with the initial value gives
The formula is defined for every real and direct differentiation verifies the equation. Its derivative is negative for and positive for : uniqueness prevents crossing an equilibrium, but it does not force a nonautonomous solution to be monotone in time.
Common mistake
(Recovering the equilibrium is not recovering every solution)
Checking the zeros of after dividing restores the constant solutions, but it need not restore every solution. For with , both and solve the problem on . The latter crosses the equilibrium at zero. For any , another solution is for and for ; its derivative agrees at the join. Each is verified in the original equation. Division by describes the nonzero pieces but misses the joining. The state function is continuous, yet it is not locally Lipschitz at zero, so the uniqueness argument above cannot be used there.
Quick check
(For , , find the separated solution and its maximal interval containing zero. Is it defined for every real time?)
The initial value lies above the equilibrium , so work on the component .
Solution · Answer
Separation gives . Its denominator is at least , so the solution is actually defined on all ; its values stay above . This illustrates why the domain must be checked from the resulting formula, rather than guessed from an initial value above equilibrium.
Another explicit method recognizes a total derivative. An equation is exact on a domain if a differentiable potential satisfies and there. Along a differentiable solution graph, the chain rule then gives . Its curves lie on level sets of . For continuously differentiable coefficients, equality is necessary. On a rectangle it is sufficient, as one can construct the potential by integrating along horizontal and vertical segments. A domain with holes requires additional care; equality of the cross derivatives alone is not a global potential theorem on every open set.
For example, has potential . The implicit family supplies solution graphs locally wherever , by the implicit function theorem. At a point where that derivative vanishes, the level set still exists as a set, but the equation cannot simply be divided by to obtain a regular equation for . This separates the geometry of level sets from the existence of a graph with the chosen independent variable.
Comparing a solution with barriers
Sometimes a differential equation cannot be solved in elementary functions, yet an inequality is enough to answer a useful question. A lower solution has a rate no greater than the rate prescribed by the equation at its own state; an upper solution has a rate no smaller. The following result makes this comparison precise, including when the initial values coincide.
Theorem 1.1.3
(Scalar comparison on a common interval)
Let be continuous, where is an interval, and suppose that for all and . If take values in and satisfy
then throughout . The statement assumes that both curves already exist on this common interval; it does not assert global existence.
Proof
(The positive part of the difference)
Put and . The function is absolutely continuous. At almost every time at which it is differentiable, its derivative is at most : when , the differential inequalities give ; when the positive part is locally zero; at almost every zero its derivative is zero. Thus almost everywhere. Integrating and using gives . Since , it vanishes, proving the claim.
Proof X-ray
Why the first-touch picture needs an estimate
The naive argument looks at a first time where overtakes and compares slopes there, but at a contact the slopes may be equal, so slopes alone decide nothing. The proof instead tracks the whole excess . The two differential inequalities are used only where , and the Lipschitz hypothesis is used exactly once, to turn into . That makes nonincreasing, and the argument closes because it starts from . Without the Lipschitz bound this step has no constant, and the counterexample below shows that the conclusion then fails.
Worked example 1.1.4
(Bounding a harvested population without solving it)
Let , , where is continuous and nonnegative. On any common forward interval of existence, compare with the unharvested solution , . On the compact ranges of the two curves, has a finite Lipschitz constant. Since and , comparison gives
The formula for follows by separation or direct substitution. This upper bound does not guarantee that stays nonnegative: the model must be reconsidered if the computed population reaches zero while harvesting continues.
Counterexample mode
Continuity without a Lipschitz bound can lose comparison
On , take , and . Both curves solve and begin at zero, yet for . Thus even two exact solutions need not remain ordered from equal initial data when the state dependence is merely continuous. Here is not Lipschitz near zero.
Quick check
(A solution of starts at . Show on any interval on which it exists that .)
Use as a lower solution.
Solution · Answer
Let and . Then , while and . The derivative of is bounded on the compact ranges of and , so the comparison theorem gives . The inequality is a lower bound; it is not an exact solution formula.
Definition 1.1.5
(Linear system and fundamental matrix)
A linear system has the form , where is a continuous matrix function and a continuous vector function on an interval. A fundamental matrix is an invertible matrix-valued solution of . Its columns form a basis of homogeneous solutions.
Invertibility at one time implies invertibility throughout the coefficient interval, as the uniqueness argument below shows. Coefficients are assumed continuous on that interval; no constant-matrix assumption is hidden in this definition.
Theorem/Proposition
Theorem 1.1.6
(Local existence and uniqueness)
Suppose is continuous on a closed cylinder , contained in , and is Lipschitz in its second variable there with constant . Then the initial value problem has a unique local solution through . In particular, a continuously differentiable on an open domain has a unique local solution through each point of that domain.
The Lipschitz hypothesis compares two states at the same time: . Continuity alone ensures local existence under the usual finite-dimensional existence theorem, but it does not ensure uniqueness. The last sentence follows by taking a smaller convex cylinder and bounding the derivative with respect to the state on its compact closure. It is a sufficient local criterion, not a necessary one.
Theorem 1.1.7
(Linear solution formula)
Let and be continuous on an interval containing . There is a unique solution of , , throughout . If is any fundamental matrix, then
For constant , one may use , where .
The formula distinguishes propagation of the initial state from accumulated forcing. The factor propagates an input applied at time to time . For time dependent matrices it generally cannot be replaced by : matrices at different times need not commute. For a constant matrix, the power series converges absolutely in every matrix norm on bounded time intervals, and termwise differentiation gives .
Superposition applies to homogeneous solutions: any linear combination still solves the homogeneous equation. For a fixed nonzero forcing, arbitrary linear combinations of particular solutions need not preserve the equation. Instead, the difference of any two particular solutions is homogeneous. Thus the full inhomogeneous solution set is one particular solution plus the homogeneous solution space, an affine set of dimension whose elements correspond one-to-one to the initial states .
Concept lensStructural
An integrating factor is a change of coordinates
The substitution measures a solution relative to the moving homogeneous basis. With no forcing, is constant; with forcing, only the coordinate vector evolves according to . The scalar integrating factor performs the same operation in one dimension. This viewpoint explains why the homogeneous evolution cancels before integration, rather than leaving a formula to memorize.
Proof sketch or proof idea
Proof
(A contraction on a set of curves)
For local existence, rewrite the differential equation as the integral equation
Let bound on the cylinder, and choose with , , and . On the closed set of continuous curves satisfying for , the right hand side defines a map . The bound keeps inside this set. The integral estimate makes it a contraction. The contraction theorem supplies a unique fixed point. Continuity of the integrand then makes this fixed point continuously differentiable, so it solves the original equation.
This also explains successive approximation: begin with the constant curve and repeatedly apply . Each new curve uses the preceding curve's rate of change. Contraction controls the whole curve in the supremum norm, rather than merely comparing successive values at a chosen time. Local uniqueness on overlapping intervals allows the local solutions to be joined consistently. The union of all their compatible intervals is the maximal interval.
Proof X-ray
Two estimates do different jobs
The inequality keeps each iterate inside the cylinder where the hypotheses are known. The separate inequality makes differences shrink in the complete function space. Neither estimate replaces the other: a map may contract distances but leave the admissible set, or preserve that set without being a contraction. The fixed point gives a curve first; the integral equation then supplies its differentiability.
Worked example: Picard iteration for
Consider the initial value problem , . Integrating from to turns it into the fixed-point equation , where
Start from and set . The first three iterates are
The pattern is for . The case is the formula for . If the formula holds for some , then
which completes the induction. These are partial sums of the exponential series with the terms and removed, so for every real , uniformly on each bounded interval. Direct substitution shows that satisfies and on all of . This global conclusion comes from the explicit computation, not from the contraction argument.
A constructive local guarantee
Let and , a closed subset of the complete space . On the rectangle , we have , and , so the right-hand side is Lipschitz in the state with constant . For of either sign, with . Hence for , the function is continuous and
So maps into itself with contraction factor . The iterates from converge uniformly on to the unique fixed point of in . That fixed point is continuous, so by the fundamental theorem of calculus it satisfies in the interior of and ; endpoint derivatives are one-sided. The argument says nothing about larger intervals.
For the error estimate, and . For , telescoping gives
Letting yields for every . This is a guaranteed upper bound, not the actual error.
Numerical check. For the bound is . Since , the triangle inequality bounds its absolute value by the same tail at , where equality holds. Thus the maximum on occurs at , where it equals . The bound is valid but about a hundred times larger than the true error.
Quick check
(Using only the geometric estimate, find the smallest that guarantees on .)
Solve for an integer .
Solution · Answer
We need , that is, . Since and , the estimate certifies and no smaller . The true error of is much smaller. The exact solution can be used to calculate the true error; the geometric estimate does not use it.
Gronwall's inequality in integral form
The continuation and continuous-dependence arguments in this chapter rest on one principle: a nonnegative quantity bounded by a constant plus an integral of its own past values cannot grow faster than an exponential. This section proves that principle.
Lemma 1.1.8
(Grönwall's inequality)
Let , and let be continuous. If
then
Proof
(An integrating factor for the bound)
Define . Because is continuous, the fundamental theorem of calculus shows that is continuously differentiable, with and . The hypothesis says exactly that . Since , multiplying by preserves the inequality, so .
Put , so . Then
Thus is nonincreasing, and . Hence . No step divides by , so the case is included and gives .
Comparing two solutions
Let and solve on a common interval with and . Suppose both graphs remain in a set on which for all , where is continuous and nonnegative on all of . Subtracting the integral equations and estimating the integrand gives
and the lemma with yields
This finite-time estimate is conditional: it holds only on an interval where both solutions exist and stay in . If , it forces on , which is forward uniqueness. For times , apply the same argument to and , which solve ; this is justified only if the Lipschitz bound holds there with a continuous nonnegative bound on the backward interval.
Example
Consider for with initial values and , where are arbitrary real numbers. The solutions are and . With the estimate gives , which equals exactly, so the bound is sharp here. On each the solution depends continuously on the initial value, with constant . That constant is unbounded as , and indeed the zero solution is unstable: continuous dependence on finite intervals does not imply global stability.
Quick check
(Two solutions of satisfy . Bound for .)
First check that both solutions exist on .
Solution · Answer
By the mean value theorem, for all real , so we may take and . The smooth field is globally Lipschitz. Also , so a solution cannot escape in finite time; local existence and uniqueness then extend both solutions throughout . The estimate gives .
Proof
(Global existence and invertibility for linear systems)
This proves the linear solution formula. For linear equations, the Lipschitz constant on each compact time interval is bounded by . Integral estimates prevent finite blow-up inside that compact interval: if and , then, for ,
One obtains this bound from the integral equation and Grönwall's inequality. Continuing across compact subintervals gives existence on all of . To see why boundedness permits continuation, suppose a finite endpoint lies inside . The displayed bound keeps bounded there, and the equation then bounds . The solution has a finite limit at the endpoint, from which local existence extends it. This contradicts maximality.
Construct by solving the homogeneous system with initial columns equal to the standard basis at . If , the homogeneous solution vanishes at . Uniqueness makes it identically zero, so . Thus is invertible. Put in the inhomogeneous equation; cancellation of leaves . Integration gives the formula. This proves both its existence and its dependence on the initial data.
Worked examples
Worked example 1.1.9
(Separation, equilibria, and a finite maximal endpoint)
Consider with . The constant solutions are and . Away from them,
Solving and imposing the initial condition gives
The formula also evaluates correctly at the two equilibrium initial values, but their existence was established before division. If , the denominator is , so the maximal interval containing is . The solution tends to at its right endpoint. If , the denominator remains positive for all real time, and the solution increases to as . The algebraic formula therefore describes several qualitatively different domain and limit behaviors.
Worked example 1.1.10
(An integrating factor is a product rule in reverse)
Solve with . Multiplication by gives . Integration from to , followed by integration by parts, yields
For a general equation with continuous coefficients, use . Then . Since never vanishes, multiplication and division preserve every solution on the coefficient interval. Different solutions here differ by a multiple of ; the forcing fixes the common nonhomogeneous part.
The general initial-value formula deserves a closer look. Suppose are continuous on an interval containing , and . Integrating the product rule from to gives
The upper limit may lie on either side of : the definite integral is oriented. The first term propagates the initial value; the integral accumulates the forcing after weighting it by the same integrating factor. Conversely, differentiating (1) gives the equation and substituting gives . Thus the formula constructs a solution throughout and determines it uniquely: any two solutions have a difference whose product with is constant and initially zero.
Worked example 1.1.11
(A variable coefficient and its initial value)
Solve , . Here , so . Integrating from to gives , hence on all of . The same formula solves the equation for negative ; no separate sign choice is needed. Direct differentiation gives and verifies the equation.
A balance law: a tank with changing volume
An integrating factor becomes useful only after the unknown and its units have been chosen. Suppose a tank initially contains liters of pure water. Brine containing grams of salt per liter enters at liters per minute, and the well-mixed liquid leaves at liters per minute. Assume there is enough capacity and that the rates and perfect mixing persist. Let be the mass of salt in grams after minutes, for . The volume is liters, so the concentration inside is grams per liter. The balance is rate in minus rate out, not concentration in minus concentration out:
Both terms on the right have units of grams per minute. The coefficient varies because the volume changes. On the integrating factor is positive. Multiplication and integration give
where the constant follows from . Thus
The concentration rises toward grams per liter, even though the salt mass grows without bound as the tank volume grows. A limiting concentration is not a limiting mass. The initial value and the differential equation can be checked directly; continuous coefficients give uniqueness on the physical time interval.
Exercise 1.1.12
(Keep the same initial volume and incoming brine, but let liters per minute leave and suppose initially the tank contains grams of salt. Find the amount up to the time the tank empties. Why can the formula not by itself describe the mixture after that time?)
Core
Solution(3)
- 1
Write the balance law with the new volume
Now , and for the balance is with .
- 2
Find a solution and invoke uniqueness
Key step
The function satisfies and , so it is the solution there: the initial concentration already equals the incoming concentration.
Because: The coefficient is continuous while .
- 3
Identify where the model stops
At the volume vanishes, so the expression and hence the well-mixed outflow model cease to be defined. The algebraic continuation past is not a physical solution of this model.
Counterexample mode
Dividing by the leading coefficient may remove the initial point
Consider with initial time . Dividing by gives only on intervals that do not contain zero, so the continuous-coefficient theorem and formula (1) cannot be applied across zero. In the original equation every function is a solution on with ; that datum therefore does not give uniqueness. If instead , the original equation at zero requires , so no differentiable solution exists. A zero leading coefficient requires a fresh check of the original equation: failure of a theorem's hypotheses alone proves neither nonexistence nor nonuniqueness.
Worked example 1.1.13
(A Jordan block and resonant forcing)
Let
Write , where . Commutativity of and gives . Further,
Variation of constants therefore gives
Substitution checks the second row first and then the first. Repeated eigenvalues do not by themselves tell us the number of independent eigenvectors; the nilpotent part produces a polynomial factor. Despite those factors, both components tend to zero because exponential decay dominates every fixed polynomial.
Common mistakes
Dividing by the state can erase an equilibrium; recover it by checking every excluded value in the original equation, as in the separation section above. Conversely, an implicit equation obtained after integration is not automatically a graph near every point; recovering as a function requires an appropriate nonzero derivative or a direct local analysis.
Common mistake
(Reading a maximal interval off a formula's singularities)
A finite singularity of a formula must be examined before declaring a maximal interval. A removable singularity may merely reflect an inconvenient representation, whereas divergence of the state prevents continuation. For a linear system with continuous coefficients on all of , finite state blow-up is excluded by the preceding estimate, even when the solution grows very rapidly at infinite time.
Common mistake
(Using scalar exponential rules for matrices)
The product equals when and commute. Without that hypothesis the scalar exponential rule is unreliable. A diagonalization formula is valid only after an invertible eigenvector matrix has been found. A defective matrix should be handled by a Jordan chain, a direct differential equation, or its power series. For example, with and one has , , while .
Summary
The initial value problem includes its solution interval. Explicit methods are reversible only on the domains where their divisions and substitutions make sense. Local Lipschitz control supplies uniqueness, while linear equations with continuous coefficients extend across every compact part of their coefficient interval. Fundamental matrices describe homogeneous propagation, and variation of constants adds the accumulated effect of forcing.
Exercises
Quick check
(Solve with and give its maximal interval containing zero.)
Identify the first singularity in each time direction.
Solution · Answer
Separation gives . Its denominator vanishes at , and its maximal interval through zero is . Differentiation gives . The divergence at prevents a finite-valued extension.
Quick check
(Solve , , and explain the polynomial factor.)
Use the integrating factor .
Solution · Answer
The product equation is , so on all of . The forcing has the same exponential rate as the homogeneous solution. In the transformed equation it becomes a constant, whose integral supplies the factor .
Quick check
(For , compute and solve with .)
The matrix square is zero.
Solution · Answer
The power series terminates: . Hence . The second coordinate is constant and the first integrates that constant, directly confirming both equations.
Quick check
(Let two solutions of remain in one region with Lipschitz constant on . What initial error guarantees their distance at time is at most ?)
Use the estimate from comparing two solutions, with .
Solution · Answer
The estimate is . Thus an initial distance at most suffices. This is an upper bound supplied by the Lipschitz information, not a claim that every pair of trajectories achieves the bound.