An invertible derivative makes a map locally equivalent to the identity. A derivative of smaller rank suggests a different picture: some coordinates are genuinely transmitted to the output, while others move inside a level set. The constant-rank theorem makes that picture exact after changing coordinates in both the source and the target.
The word constant matters. A matrix of rank at one point gives first-order information there. A map whose derivative has rank throughout a neighborhood has enough structure to admit a normal form on that neighborhood. We will prove this distinction rather than treating a rank computation at one point as a geometric conclusion.
Motivation
Suppose a map has three input variables but only two independent output directions. A small change invisible to its derivative lies in a one-dimensional kernel. Are those invisible directions tangent to actual curves on which the map remains constant? Under a constant-rank hypothesis, the answer is yes, and every kernel vector comes from such a curve.
The number of displayed equations need not equal the number of independent constraints. Repeating an equation does not reduce a level set's dimension again. Conversely, a singular equation may describe an isolated point even though its derivative vanishes. The theorem will identify exactly when rank counts dimensions and when that inference is unjustified.
All spaces below are finite-dimensional real Euclidean spaces. Domains of differentiable maps are open unless explicitly restricted.
Definitions
Definition 3.5.1
(Rank, immersion and submersion)
For of class , the rank at is the rank of the linear map . The map has constant rank near if this rank equals on some open neighborhood of . Full rank means rank .
It is an immersion at if is injective, so its rank is , and a submersion at if is surjective, so its rank is .
A local coordinate change is a diffeomorphism between open sets: it is bijective, it and its inverse are , and its derivative is invertible. The chain rule shows that changing coordinates in source and target preserves rank.
Definition 3.5.2
(Embedded submanifold and tangent vectors)
A subset is a -dimensional embedded submanifold if, near every , a local ambient coordinate change sends exactly to a coordinate plane of dimension , intersected with the coordinate neighborhood. After permuting coordinates, that plane is .
At , tangent vectors are velocities of curves with . A submanifold chart shows that these velocities form a -dimensional vector space, denoted : they are the inverse-chart derivative applied to the coordinate plane.
Definition 3.5.3
(Regular value)
A value is regular for if is surjective at every . The fiber may be empty, in which case the condition is vacuous. A nonempty regular fiber can occur only when .
Theorem/proposition
Theorem 3.5.4
(Constant-rank normal form)
Suppose is and has constant rank on a neighborhood of . There are local source and target coordinate changes , which may include coordinate permutations, with and , such that
where and are open neighborhoods of zero. The zero output belongs to . Coordinate blocks of dimension zero are permitted.
If is for an integer , the coordinate changes can be chosen .
Theorem 3.5.5
(Fibers, local images and regular values)
Under the constant-rank hypotheses, is locally an embedded submanifold of dimension , and
For a sufficiently small source neighborhood of , the restricted image is an embedded submanifold of the target of dimension : in suitable target coordinates it is a graph for a map . This assertion concerns the restricted image, not all branches of meeting the same target point.
In particular, a nonempty regular fiber is an embedded submanifold of dimension . A map which is a submersion at every point of its domain is an open map.
Theorem 3.5.6
(A necessary multiplier condition)
Suppose has constant rank near , and a real-valued function , of class on an open neighborhood of , has a local maximum or minimum at when restricted to . Then there is such that
The multiplier is unique when is surjective. If the rank is below , the multipliers are nonunique: they form one translate of . This is a necessary condition, not a sufficient extremum test.
Proof sketch or proof idea
Build source coordinates from the independent outputs
Proof
(Source coordinates from an invertible minor)
Work inside the neighborhood where rank is constant, and translate the source and target so that and . Permuting coordinates, choose a nonzero -by- minor in the first rows and columns. For the moment assume . Split inputs as and outputs as
The matrix is invertible. Define
Its derivative has block form
so it is invertible. The inverse function theorem makes a local diffeomorphism. Shrink its image to a product , with open balls about zero, and restrict the source accordingly. In these coordinates,
The first outputs are now exactly the first inputs. Up to this point we used only the rank- minor at the initial point.
Constant rank removes dependence on the remaining inputs
Proof
(Completing the normal form with a target shear)
The chain rule preserves rank, so has rank at every point of the product. But
Subtracting times the top block row from the lower one shows . Hence throughout the product. For fixed , integrate this derivative along the line segment between two points of the convex ball . The result is . Put .
Now use the target coordinate change
It is a diffeomorphism on and sends to . This proves the normal form. Translating and undoing the coordinate permutations gives the original statement.
If , vanishes on a small ball about . Integration along line segments in that ball makes constant there. Source translation and target translation give the asserted zero map. The full-rank cases simply have an empty coordinate block.
For higher regularity, the inverse function theorem retains class . This also follows inductively from , since matrix inversion is smooth on invertible matrices. Compositions and then retain the same class.
Proof X-ray
The pointwise minor starts the proof; constant rank finishes it
An invertible minor lets us use some outputs as coordinates. It does not make the other outputs independent of the remaining inputs. That independence comes from the rank identity forcing everywhere in a neighborhood. Evaluating the rank only at the initial point would establish that equality only there.
Exact level sets and actual tangent curves
Proof
(Proof of the fiber, image and regular-value statements)
In normal coordinates, is equivalent to . Thus the local fiber is
an embedded submanifold of dimension , after swapping the coordinate blocks if necessary. A curve inside the fiber satisfies , so differentiation gives . Every tangent velocity therefore lies in the derivative kernel.
Conversely, let . Differentiating the normal-form identity shows that the first components of are zero. For sufficiently small ,
stays in the coordinate fiber. Its velocity is by the inverse-derivative identity. Thus every kernel vector is realized by an actual fiber curve, proving the tangent equality.
The restricted image in the intermediate target coordinates is the embedded graph . The diffeomorphism carries this exact graph to the coordinate plane , proving the local image statement. When , there is no free source variable , so is one-to-one on the chosen source neighborhood and parametrizes its image regularly. Global injectivity does not follow.
If is surjective, one full-size minor remains nonzero on a neighborhood. Rank cannot exceed , so it is constantly there. Apply the theorem at every point of a regular fiber. Each chart describes the entire fiber exactly within its neighborhood, so these local descriptions establish the embedded-submanifold property at every point of . For a submersion, the normal form is a coordinate projection, whose image of a small product neighborhood is open. Restricting that neighborhood inside any given domain-open set shows that its full image is a union of open target neighborhoods. Hence the map is open.
Extremality annihilates the tangent kernel
Proof
(Proof of the multiplier condition)
If has a constrained local extremum, every fiber curve through makes . The tangent result gives
For a linear map , every functional vanishing on is a linear combination of the rows of . To see this without a dimension guess, define a functional on by . It is well defined because two preimages differ by a kernel vector. Extend to all of by extending a basis. Its coefficients form , giving .
If is surjective, the transpose is injective, so is unique. If its rank is below , every vector in can be added to one multiplier without changing the equation. The multiplier calculation counts independent output directions, not the number of equations written down.
Concept lensStructural
Independent constraints determine codimension
In the normal form, coordinates are fixed and are free along a fiber. Extra output equations may encode functions of those same coordinates. They do not remove more freedom. The rank hypothesis is what turns this local count into an exact geometric statement.
Worked examples
Worked example 3.5.7
(Two outputs, one independent coordinate)
Let . Its first derivative row is and its second is times that row, so the rank is one everywhere. Explicit coordinate changes are
Their inverses are and . Direct substitution gives . The fibers are the lines , and the image is a parabola. Counting two output entries as two independent constraints would give the wrong fiber dimension.
Worked example 3.5.8
(A sphere with a redundant defining equation)
Put and . Near every point of , the row is nonzero, and the rank is constantly one. The fiber is the unit sphere, of dimension , not .
Its tangent space is . Near the north pole, the local graph is , and the tangent plane is horizontal. For the height objective , the north pole is a constrained maximum. The multiplier equation there is
There are infinitely many multipliers because the two displayed constraints are dependent. The value is not a regular value of this map into , but the constant-rank theorem still describes the fiber. Using the single independent equation gives the regular-value formulation.
Worked example 3.5.9
(An immersion can have a crossing in its full image)
Let . Its derivative never vanishes: if , then . The map has rank one everywhere and each sufficiently short source interval maps to a regular curve.
But . The curves and have velocities and there, which are linearly independent. If the entire image were an embedded one-dimensional submanifold near the crossing, both velocities would lie in its one-dimensional tangent space, an impossibility. Local images from two different source neighborhoods can meet with different tangents; the theorem does not combine them into one embedded curve.
Counterexample mode
A singular constraint can hide an isolated feasible point
Take at the origin. Its derivative there has rank zero, but nearby nonzero points have rank one. The zero fiber is the singleton , not a two-dimensional surface predicted by using without a neighborhood hypothesis.
The objective has both a constrained local maximum and minimum on this singleton, but cannot equal . The multiplier theorem's constant-rank hypothesis has failed. A computation at a singular point cannot substitute for the missing local geometry.
Common mistakes
Common mistake
(Intermediate rank at a point need not persist)
Full rank at one point persists nearby, because a nonzero maximal minor stays nonzero. An intermediate rank need not persist: the same minor only prevents rank from becoming smaller nearby; other minors may become nonzero and increase it. Constant rank includes this additional upper control.
Common mistake
(Changing only source coordinates)
The theorem changes both source and target coordinates. Looking only for a source change may leave dependent output functions visible. Likewise, the dimension of a fiber is determined by rank, not by the length of the output vector.
Common mistake
(Local conclusions are not global)
A regular value need not have any preimages. A submersion need not map onto its declared codomain, and a local diffeomorphism need not be globally injective. The assertions are local unless the statement explicitly says otherwise.
The necessary multiplier equation is not a classification of extrema. Once a fiber chart is available, the objective's restriction to the free coordinates still has to be studied. Second-order sufficient conditions and global optimization are separate questions.
Summary
An invertible minor creates source coordinates from independent outputs. Constant rank then removes all remaining free-variable dependence, and a target shear gives the projection normal form. Fibers become coordinate planes, their tangent spaces become derivative kernels, and redundant equations become visible. These local conclusions support multiplier conditions while leaving singular fibers and multiple global image branches outside the theorem's guarantees.
Exercises
Quick check
(For , find a global source coordinate change reducing to projection onto two coordinates. Parametrize the fiber over and compute its tangent line at a general point.)
Solution · Guided solution
Use , with inverse . Both maps are smooth and defined globally, and . The fiber is . At a point of it, the tangent line is spanned by . Directly, the derivative rows are and , whose common kernel is that line. Their rank is two everywhere.
Quick check
(Show that is a local diffeomorphism everywhere. Find its image and all fibers, and explain why there is no global inverse onto that image.)
Solution · Guided solution
The derivative matrix is , with determinant . The inverse function theorem applies at every point. The image is : the factor excludes zero, and every nonzero point has polar form with . For a nonzero target of radius and any chosen argument , the fiber consists of , ; the fiber over zero is empty. Every point fiber is locally zero-dimensional, but the map is periodic in its second input and is not globally one-to-one. Thus it has no global inverse onto its image.
Quick check
(For , compute the rank at every point. Explain why no rank-one normal form exists near the origin, although the zero fiber is a smooth line.)
Solution · Guided solution
The derivative is , so the rank is two when and one when . Every neighborhood of the origin contains both ranks. Coordinate diffeomorphisms preserve rank, so any constant-rank normal form would force rank to be constant on a neighborhood, which this map lacks. In particular, the rank-one form is impossible. Nevertheless is a smooth line. Constant rank is sufficient for the local fiber theorem; it is not necessary for a particular fiber to happen to be smooth.
Quick check
(Prove that is open for every integer . Deduce why full rank at a point supplies the needed local constant rank, whereas intermediate rank at a point does not.)
Solution · Guided solution
For the set is the whole domain; for it is empty. In the remaining cases, rank at least means that some -by- minor has nonzero determinant. Each such determinant is continuous, so its nonzero set is open. Taking the union over the finitely many minors proves the claim. If at a point, the rank cannot rise beyond , and the open lower bound therefore makes it constantly nearby. At an intermediate rank, there is no such automatic upper bound. The previous example has rank one at the origin but rank two arbitrarily close to it.
Quick check
(On the constraint plane , examine at the origin the three objectives , , and . Find their multipliers for and classify their constrained critical points.)
Solution · Guided solution
All three gradients at the origin are , so the unique multiplier is . Restricting to gives , and . The first has a strict local minimum, the second a strict local maximum, and the third a saddle: along the -axis it increases from zero, while along the -axis it decreases. The identical necessary multiplier equation cannot distinguish these outcomes.