A bound for inverse accuracy
An invertible Jacobian at one point supplies a local inverse under the inverse function theorem's regularity assumptions. A uniform comparison with one fixed invertible matrix can say more on a convex region: distinct inputs stay a definite distance apart after applying the map. The same inequality converts an output residual into a bound on reconstruction error.
This lesson builds on the mean-value and Taylor section, the differentiability section, and implicit functions and local inversion.
Vector outputs need an inequality, not a shared intermediate point
Definition 4.5.1
(Operator-norm bound)
For a real matrix , the statement means for all vectors ; the operator norm is the least such .
Lemma 4.5.2
(Vector mean-value inequality)
Let be convex, and let be on an open neighborhood of . If for all and , then for all .
Proof
(Vector mean-value inequality by a scalar projection)
For , the segment stays in . If the desired estimate is immediate. Otherwise set
The scalar mean value theorem gives a point such that
The final step uses Cauchy–Schwarz and . There is no assertion that every output component has the same intermediate point in its own mean-value equality. The projection is chosen after the endpoint difference is known.
Proof X-ray
Why the projection is chosen last
The proof converts a vector inequality into a scalar one. Convexity of is used only to keep the whole segment inside the region where the derivative bound holds. The non-obvious move is to choose the unit vector after the endpoint difference is known, so that equals the norm itself rather than one component of it. Cauchy–Schwarz then turns into a bound by . Without the segment the derivative bound has no force at intermediate points, and without the projection the components would need different intermediate points .
A convenient sufficient assumption above is a uniform operator-norm bound on the Jacobian.
An upper bound alone does not imply injectivity. A constant map has derivative zero and identifies every input. Nor may one compare points in different components merely from a derivative bound: the segment condition is essential.
Counterexample mode
A derivative bound without a segment
Take , which is not convex, and on , on . is on the open set with , so holds with . For and , however, while . The segment from to leaves , so the derivative bound says nothing about it; the inequality needs convexity.
Subtract one fixed linear map
Theorem 4.5.3
(Fixed-matrix comparison)
Let be near a convex set , and let be a fixed matrix with for all , where . If for all , then for all , . In particular, if then is injective on .
Proof
(Applying the mean-value inequality to F minus a linear map)
Let be near a convex set , and choose a fixed square matrix . Suppose
Apply the preceding inequality to . Its derivative is , so
Assume constants and satisfy
Triangle and reverse-triangle inequalities now give
When , the lower constant is positive. Thus forces , proving injectivity on all of . This is a quantitative conclusion from a uniform estimate, rather than an inference from a nonsingular Jacobian at one point.
Invertibility of a real square matrix supplies such an : one may take . In concrete symmetric examples, decomposing a vector into perpendicular eigendirections often gives a sharper transparent calculation.
Worked example 4.5.4
(A small sine coupling of the identity)
On let . Show that is injective and bound its inverse.
Take , so . Then , and for , , so on the convex set . Hence for all . The lower constant is positive, so is injective, and its inverse on is Lipschitz with constant .
What the inverse statement actually guarantees
Corollary 4.5.5
(Lipschitz inverse and invertible Jacobian)
Under the hypotheses of the fixed-matrix comparison with , the inverse of on satisfies , and for every , so is invertible.
Proof
(Reading the lower bound as a bound on the inverse)
Once is injective on , its inverse is defined on its image . For any , the lower bound gives
This proves a Lipschitz bound on the inverse's stated domain. It does not say that every proposed output belongs to the image. Injectivity and existence of a preimage are separate tasks.
At every , the same comparison, applied to , gives
Thus the square Jacobian has zero kernel and is invertible. The local inverse function theorem then supplies an open neighborhood of the output with a inverse. This local conclusion does not by itself identify the whole image or show that an arbitrary distant target is attained.
Common mistake
(Inferring injectivity from an invertible Jacobian at every point)
For the Jacobian determinant is everywhere, yet . Pointwise invertibility gives local inverses only; a uniform estimate such as on a convex set is what gives global injectivity there.
For a closed domain, a boundary point has no open input neighborhood inside that domain. A theorem about a neighborhood for the ambient map and a theorem about its restriction to the closed set are different statements.
Residual versus reconstruction error
Suppose an exact but unknown input satisfies . A proposed reconstruction has residual . The two-point lower bound yields
The residual is measured in output space; the conclusion is measured in input space. Keeping the denominator is necessary: a small output error can correspond to a larger input error when the map nearly compresses a direction.
The certificate assumes that an exact preimage exists in the controlled region. If none exists, there is no to compare with, even if a residual is small. A numerical procedure producing a candidate and an error certificate about an existing solution answer related but different questions.
Worked example 4.5.6
(A residual certificate with a nonidentity matrix)
Let on , with for an unknown . A candidate has residual . Bound .
Take , so and . Here has operator norm at most . The lower constant is , so .
Common mistake
(Dividing by the largest stretch)
In the example above the certificate divides by , which involves the smallest stretch of , not . Using would give the too-small value , which is not justified.
The sufficient bound can fail while injectivity survives
If , the displayed lower bound becomes zero and proves no separation. This does not establish noninjectivity. A map can remain one-to-one while compressing some small input differences faster than linearly.
Conversely, a bound with is strong enough to exclude this behavior uniformly on the controlled region.
Worked example 4.5.7
(The borderline case )
Let on and take .
Here , so and the lower constant is : the comparison proves no separation. Nevertheless is strictly increasing, hence injective. It is not stably invertible, though: for , while , so no constant gives for all small , and the inverse is not Lipschitz at .
Checklist
- State the convex comparison region and the open neighborhood on which the map is .
- Bound using one fixed matrix , not a separately chosen matrix for each pair.
- Prove the matrix lower bound, and verify before claiming a positive separation constant.
- Use a scalar projection to prove a vector norm estimate; do not postulate a common componentwise mean-value point.
- State the inverse's domain as the actual image. Check existence separately.
- For a residual estimate, place both the exact and proposed input in the region where the derivative bound holds.
Quick checks
Quick check
(On a convex set, with . What lower constant and inverse Lipschitz constant follow?)
Find the smallest stretch of .
Solution · Answer
and , so , and the inverse is Lipschitz with constant .
Quick check
(Show that is injective on using the fixed-matrix comparison.)
Take .
Solution · Answer
, so , and . Then , which forces whenever .
Quick check
(A candidate has residual and . What does the certificate give, and what must be true for it to apply?)
Which two points must lie in the controlled region?
Solution · Answer
. The bound applies only if an exact solution with exists, and both and lie in the convex region where .