An existence theorem becomes especially useful when its proof tells us how to approximate the object it guarantees. A contraction does exactly that. Iteration produces a Cauchy sequence, completeness supplies a limit, and a strict distance reduction makes that limit unique. Local inversion and ordinary differential equations fit this pattern after we choose the right space and the right self-map.
Motivation
To solve an equation , we might repeatedly correct a guessed input using its current error. To solve , we might repeatedly integrate the velocity predicted by the previous guessed curve. Neither procedure works merely because it can be written down. It must keep its guesses in the allowed domain, reduce their separation by one common factor, and operate in a space where Cauchy sequences have limits.
These requirements explain the local character of the results. Shrinking a ball makes a nonlinear derivative close to its invertible linear part. Shortening a time interval makes integration multiply an error by a small number. “Sufficiently near” and “sufficiently short” therefore refer to estimates that we can inspect, rather than to an unexplained appeal to intuition. Each time this section uses the fixed-point theorem, we should be able to answer three questions: what is the complete space, why does the map stay inside it, and what exactly is the contraction factor?
Definitions
Definition 3.1.1
(Contraction and fixed point)
Let be a metric space. A map is a contraction if there is one number with such that for all . A fixed point is an satisfying .
The factor must be uniform over all pairs. The weaker statement whenever does not automatically give a contraction. The notation is also a substantive assumption: iteration must remain inside the complete space on which the estimate holds.
For a map between finite-dimensional Euclidean spaces, differentiability at means that there is a linear map such that
The notation means that this derivative exists and depends continuously on the point. We use an induced operator norm for matrices. On a convex set, a bound implies by integrating the derivative along the line segment. Convexity guarantees that the whole segment lies in the region where the estimate is available.
Definition 3.1.2
(Local Lipschitz control and ODE solution)
On a time-state rectangle, is uniformly Lipschitz in the state if with one constant for all allowed . A solution of , , is a continuously differentiable curve satisfying both equations at every time in its interval.
Continuity of alone can give existence through other theorems, but it does not give uniqueness. The contraction construction below assumes the stated Lipschitz estimate explicitly. Continuous state derivatives on a neighborhood of a compact convex rectangle are one sufficient way to obtain it, through their boundedness; a continuous function should not be read as automatically Lipschitz.
From partial derivatives to a total derivative
The inverse and implicit theorems below ask for a map. In calculations one usually checks partial derivatives instead. The following argument explains why that check is sufficient, and why derivatives along coordinate axes alone do not suffice.
Theorem 3.1.3
(Continuous partial derivatives give differentiability)
Let be open and . Suppose all component partial derivatives exist on a neighborhood of and are continuous at . Then is differentiable at , with derivative the Jacobian matrix . If these partial derivatives exist and are continuous throughout , then .
Proof
(Proof of differentiability from continuous partials)
Use the Euclidean norm on the domain. It is enough to prove the scalar case for each component. For small , all points formed by changing coordinates of one at a time lie in a ball contained in . Write and . Telescoping and the one-variable mean value theorem on each coordinate segment give
where lies on the segment from to ; if , choose either endpoint. Every as . Subtract and bound the remainder by
There are only finitely many coordinates, and continuity at makes every term tend to zero. Applying this to all finitely many components proves the vector statement. If the partials are continuous everywhere, the entries of are continuous, so is continuous in matrix norm.
Worked example 3.1.4
(Checking differentiability before linearizing)
For , all four partials are continuous on . Thus the theorem justifies the linear approximation
The Jacobian determinant at is , so the inverse theorem below applies locally there. The calculation of a determinant is useful only after checking the differentiability and continuity hypotheses.
Counterexample mode
Partial derivatives at a point do not give differentiability
Claim: if both partial derivatives of a function exist at a point, the function is differentiable there. Define away from the origin and . Both partial derivatives at the origin are zero because the restrictions to the coordinate axes are zero. Yet for , so is not even continuous at the origin. A total derivative would imply continuity by its remainder estimate. The theorem requires partial derivatives on a neighborhood and their continuity at the point; merely finding numbers along the axes does not provide that control.
Quick check
(Let . Find and justify differentiability there.)
Solution · Answer
The partials and are continuous everywhere. Therefore , and the error is .
The chain rule
Let and be open, let be differentiable at , and let be differentiable at . We assume nothing about continuity of partial derivatives. Write and , so that
where , and .
Theorem 3.1.5
(Chain rule)
is differentiable at , and .
Proof
(Proof of the chain rule)
Put . Then
Since , the term is . For the last term, first bound the inner increment. Choose with whenever . Then , so .
The increment may be zero for some , so we must not divide by . Instead, fix and choose such that whenever . This inequality also holds at because . If , then , and therefore
Since is arbitrary, . The whole remainder is therefore .
Order of multiplication. is and is , so is the required composition. The reverse product may be undefined, or may be a different map even when its dimensions permit it. Read the product from right to left. The increment is first transformed by the linearization of , and the result is then transformed by the linearization of .
Worked example 3.1.6
(A chain-rule computation at a point)
Take and at . Here . Then and , so
Quick check
(For and , find .)
Solution · Answer
We have . Since , we get . Also , so the derivative is . As a check, , whose derivative equals at .
The converse fails. Take and . Then is differentiable at , even though is not differentiable at . The hypotheses of the theorem are sufficient, but they are not necessary.
Theorem/proposition
Theorem 3.1.7
(Banach contraction theorem)
A contraction on a nonempty complete metric space has a unique fixed point. For ,
Theorem 3.1.8
(Finite-dimensional inverse and implicit function theorems)
Let be on an open set, with and invertible. There are open neighborhoods of and on which is a bijection with inverse, and
If is near , , and is invertible, then all nearby solutions of have the unique form for a function, with
Theorem 3.1.9
(Local existence and uniqueness by Picard iteration)
Suppose is continuous on , bounded there by , and uniformly Lipschitz in its state with constant . Choose with , and . Then the initial-value problem has a unique solution on among curves staying in that state ball. It is the uniform limit of Picard iterates starting from the constant curve .
The ODE statement assumes positive ; if or , their corresponding inequalities impose no extra restriction. The bounds give a guaranteed interval, which may be smaller than the actual interval of existence.
Proof sketch or proof idea
Proof
(Proof of the contraction theorem and a stability bound)
For the fixed-point theorem, induction gives . Summing the geometric tail yields
Thus is Cauchy. Completeness supplies , and the contraction estimate implies continuity, so . Two fixed points would satisfy , forcing equality. Passing to the limit in the tail estimate gives the first bound. Restarting the same estimate at gives the second. The computable step size therefore certifies an error bound even when the fixed point is unknown.
The same proof gives stability under a controlled change of the map. Suppose and preserve the same complete space, both have contraction factor at most , and their outputs at every input differ by at most . Compare their fixed points by inserting between and . The triangle inequality gives a distance at most times that distance plus , hence an overall bound of . This explains why a factor close to one can make an iteration both slow and sensitive to perturbations. The inequality estimates the fixed points of the stated maps; using rounded computer arithmetic would require a separate bound on the accumulated evaluation errors.
Proof
(Construction of the local inverse)
For inversion, put and define
Its fixed points solve . Since , continuity of the derivative lets us choose a closed ball with there. On this ball is a contraction. Moreover,
If the last term is less than , the map preserves the ball and its fixed point lies in the interior. The ball is complete, so each such has exactly one preimage there. Subtracting two fixed-point equations gives
This proves continuity of the inverse. Restrict the original domain to the interior of the ball and the preimage of a sufficiently small open target ball. The resulting domain is open, and the preceding construction gives the required bijection.
Proof X-ray
What makes the frozen-derivative iteration work
The map freezes the derivative at instead of updating it at every step. Its derivative vanishes at . In this construction, continuity of is used only once: to keep that derivative below on a whole closed ball, which by the convex-set derivative bound makes a contraction there. The radius is fixed first and does not depend on ; only then is restricted, by , so that maps the ball into itself. Completeness enters through the closed ball, and the fixed point lands in the interior, so it is a solution in the open domain. The same factor that gives uniqueness also gives the Lipschitz bound for : subtracting the fixed-point equations leaves . Invertibility of alone would supply no contraction factor away from .
Proof
(Why the local inverse is differentiable)
Throughout the chosen ball, differs from the identity by norm less than one, so it is invertible: the geometric operator series in that difference supplies an inverse. Fix and write . The Lipschitz estimate gives . Differentiability of gives , hence . This is the asserted derivative formula. Continuity follows from continuity of , the inverse curve and matrix inversion on invertible matrices.
For the implicit theorem, apply local inversion to . Its derivative has block form with identity in the upper-left block, zero in the upper-right block and invertible in the lower-right block. The inverse evaluated at gives . Differentiating gives the displayed formula. The conclusion concerns all solutions in a sufficiently small product neighborhood, not all solutions globally.
Proof
(Proof of local existence and uniqueness)
For the ODE theorem, let be the continuous curves on whose uniform distance from is at most . It is a nonempty closed subset of a complete uniform function space. Set
The bound proves that maps into itself. The Lipschitz estimate gives . Banach's theorem supplies a unique fixed curve. Continuity of the integrand and the fundamental theorem of calculus show that it is continuously differentiable and solves the ODE. Conversely, every solution staying in the ball satisfies this integral equation, so fixed-point uniqueness is precisely solution uniqueness on the stated interval.
For practical use, choose the state ball before choosing the time length. Enlarging the state ball may increase both the velocity bound and the Lipschitz bound, so a larger permitted range of values can produce a smaller guaranteed time interval. This is a limitation of the estimate, not a claim that the actual solution ceases to exist earlier. Once a solution is found, a continuation argument can start a fresh local construction from a later point. Such an argument needs the solution to remain inside a region where suitable bounds are still available; it cannot simply repeat the original estimate past a singularity or an escaping trajectory.
Worked examples
Worked example 3.1.10
(Iteration with a numerical error certificate)
The map preserves . Its derivative has magnitude at most , so we may use . The interval is closed and complete. Starting at , iteration converges to the unique fixed point in the interval. If a computed step has , then the true error is at most . A small step alone would not justify that conclusion without the verified contraction factor.
Worked example 3.1.11
(An implicit curve and its chosen branch)
For , the point satisfies and . A unique nearby branch is , with . Near , the same derivative with respect to vanishes; the circle is not a differentiable graph across that point. Switching variables restores a local graph .
Worked example 3.1.12
(Two simultaneous equations and one local branch)
Suppose a parameter and two unknowns satisfy
At both equations hold. Since is polynomial it is on all of . The derivative with respect to the unknown pair, evaluated at this point, is
The implicit theorem therefore supplies an interval about and a neighborhood of in which the solutions are exactly one pair , with . It is the square matrix of derivatives in the unknowns that must be invertible; the derivative with respect to alone is a column and cannot be inverted here.
Differentiating the two equations along this branch at gives
so and . Equivalently, this solves at the base point. The theorem justifies differentiating a branch before we know a formula for it.
For a check, eliminate to obtain . The branch through is
for near zero; its derivatives agree with the calculation. At there is also the solution . It lies outside a sufficiently small neighborhood of and does not contradict local uniqueness. At the two algebraic roots meet and the unknown-derivative determinant vanishes, so the theorem at makes no claim about a branch through that separate point.
Concept lensGeometric
The derivative formula describes a tangent line
Near the base point, the solution set of is the curve , with tangent vector . Differentiating along the curve puts this vector in the kernel of the full derivative
which is the line spanned by . Invertibility of the unknown block says exactly that this kernel is one-dimensional and contains no nonzero vector with -component zero, so it can be normalized to have first component one: that normalization is the implicit derivative formula.
Quick check
(Near , the equations and define as functions of . Justify this claim and find the derivative matrix at , with rows for and columns for .)
Solution · Differentiating a two-equation system
The unknown-derivative matrix at the base point is , whose determinant is . The polynomial equations are , so the implicit theorem applies. Differentiating with respect to gives and , hence . Differentiating with respect to gives and , hence . Thus the requested matrix is . These derivatives are for the branch through ; the equations also have the swapped solution at .
Worked example 3.1.13
(Local existence is not global existence)
For , , use the rectangle , . Here and work. Taking gives and . The theorem guarantees a unique solution there. Direct separation yields , whose maximal interval containing zero is . The theorem's small interval is valid; the blow-up at one shows why local existence must not be called global.
Counterexample mode
Continuity alone does not give uniqueness
The claim that a continuous right-hand side guarantees uniqueness is false. For and on nonnegative time, every waiting time gives a solution that is zero up to and afterwards. The derivatives agree at the join. The square-root function is not Lipschitz near zero because is unbounded. The theorem repairs the false claim by assuming uniform Lipschitz control in the state on the chosen rectangle; continuity alone does not supply that control.
Common mistakes
Common mistake
(Checking the contraction hypotheses)
Before invoking contraction, name the complete space, prove the self-map condition and find one factor below one. A derivative bound on an open interval does not repair a missing boundary fixed point: on contracts but its fixed point zero is excluded.
Common mistake
(Local inversion is neither necessary nor global)
Invertibility of the derivative gives a sufficient local condition. It is not necessary for local one-to-one behavior: is globally bijective, though its derivative vanishes at zero. Its inverse is not differentiable there. Conversely, an invertible derivative at one point gives no global injectivity; a local inverse theorem cannot prevent distant inputs from sharing an output.
Common mistake
(Uniqueness is not length of existence)
For ODEs, distinguish uniqueness of a local trajectory from the length of its existence interval. Also distinguish a Lipschitz condition in the state from regularity in time. The contraction estimate compares two states at the same time and then integrates; continuity in time is enough for that argument. Each hypothesis has a specific job, and a vague appeal to “sufficiently smooth” should not hide it.
Summary
The contraction theorem turns summable successive errors into existence and uniqueness. Inversion applies it to a corrected equation on a closed ball; ODE existence applies it to integration on a closed ball of curves. The same three checks recur: completeness, preservation of the chosen domain and a uniform factor strictly below one. The derivative and time-length hypotheses are methods for obtaining those checks.
Exercises
Quick check
(Give an example of a contraction on a complete metric space whose fixed point is not reached after any finite number of iterates from a suitable starting point.)
Solution · Answer
On , let and . The unique fixed point is zero and . Every finite iterate is positive, but the sequence converges to zero. Existence of a fixed point is not finite termination of the iteration.
Quick check
(For , , compute the first four Picard iterates starting with . On , bound the error of using the contraction theorem.)
Solution · Answer
The first four updates are , , , and . In general . On the complete space of all continuous curves on this interval, the integral map is a self-map and has factor . Since , the first error estimate gives . This bound is valid but less sharp than the factorial tail available from the explicit formula.
Quick check
(For , show that the local implicit theorem applies at every solution. Compute for the branch through .)
Solution · Answer
Here at every real point, so every solution has a unique local branch. The derivative formula gives and hence . Strict monotonicity and the limits of additionally give a global branch, but that requires this extra argument rather than the local theorem alone.
Quick check
(Explain why the map on cannot have a contraction factor less than one, although it preserves a complete space.)
Solution · Answer
For distinct , . Any Lipschitz factor must therefore be at least one. The equation would require , so there is no fixed point. Completeness and the self-map condition do not replace strict contraction.