What "Sufficiently Nice" Really Means in the Determinant-Area Theorem

The textbook's paragraph sits in plain view:

The area of any "sufficiently nice" region R in ℝ² can be approximated as the sum of the areas of rectangles. In fact, as the lengths of the sides of these rectangles approach zero, the sum of the areas of the rectangles approaches the area of R. Hence the area of the image of R under T equals |det A| times the area of R.

The phrase "sufficiently nice" is doing silent work. Behind it lies exactly one hypothesis, with a precise name, and the rest of the proof rests on that name. The goal of this tutorial is to give that hypothesis the rigor the textbook omits.

Two preliminary notes on the paragraph itself. First, the literal wording above comes from a commercial Stewart edition; the freely accessible OpenStax mirror of the same chapter writes the parallelogram argument identically but uses "the side lengths of these rectangles approach zero" in the form lim_{m,n→∞} ΣΣ. Second, every step below uses only tools you already own (linearity, det(AB) = det A · det B, Riemann sums on rectangles) plus one definition you have not met yet: Jordan-measurable.

1. The Base Case: A Linear Map and a Parallelogram

Start with the picture the textbook is leaning on. Let T(x) = Ax be a linear map on ℝ² with standard matrix A. Given two vectors u, v, consider the parallelogram on {u, v}, namely {αu + βv : 0 ≤ α, β ≤ 1}. Linearity says

T(αu + βv) = α Au + β Av,

so T maps the parallelogram on {u, v} to the parallelogram on {Au, Av}. Now compute the area. The signed area of the parallelogram on {u, v} is det[u v]; that of the parallelogram on {Au, Av} is det[Au Av]. Multiplicativity of the determinant gives

$$\det[Au ;; Av] = \det!\big(A,[u ;; v]\big) = (\det A)(\det[u ;; v]).$$

Taking absolute values, the area of the image parallelogram is |det A| times the area of the source parallelogram. Negative determinant means orientation is reversed, but |·| discards the sign so the area scaling is unaffected.

If det A = 0, the matrix collapses ℝ² onto a line, and Au, Av are linearly dependent: the parallelogram on {Au, Av} has area 0, consistent with |det A| · (area of source) = 0.

Let us pin this down with a concrete worked example. Take

$$A = \begin{pmatrix} 2 & 3 \ 0 & 4 \end{pmatrix}, \quad T(x, y) = (2x + 3y, ; 4y).$$

Apply T to the four vertices of the unit square: (0,0) → (0,0); (1,0) → (2,0); (0,1) → (3,4); (1,1) → (5,4). The image is the parallelogram with vertices (0,0), (2,0), (5,4), (3,4), whose base is 2 and height is 4, so area = 8. The determinant is det A = (2)(4) − (3)(0) = 8, which matches the result directly: area scales by |det A| = 8 from the unit square's area of 1.

Caption: Region the unit square is mapped to under T(x, y) = (2x + 3y, 4y). The four vertices (0,0), (2,0), (5,4), (3,4) trace out a parallelogram of base 2 and height 4, giving area 8 = |det A| · 1. Probability Lifesaver

It is worth dwelling for a moment on a second example that looks different: a shear. The matrix $\begin{pmatrix} 1 & 1 \ 0 & 1 \end{pmatrix}$ takes the unit square to a parallelogram with one side tilted, but det = 1, so the new parallelogram has the same area as the unit square. The shape changes visibly; the area does not.

Caption: A shear that visibly tilts one side of the unit square but leaves area unchanged. The determinant of the shear matrix is 1, so the new parallelogram has the same area as the source square. 3Blue1Brown

The lesson is that the determinant measures area scaling, not shape deformation. Two transformations can look nothing alike geometrically and still scale area by the same factor.

That is the entire engine of the base case. From here, the proof extends to any region R by tiling R with tiny rectangles and applying the parallelogram result to each tile. The extension is the topic of the next chapter.

2. Why a Rectangle Tiling Turns One Parallelogram Into Every Region

The textbook's "rectangle approximation" claim has a definite content. Take a bounded region R in ℝ² and tile it with subrectangles of dimensions Δu × Δv, indexed by (i, j). As the mesh (the maximum of Δu and Δv) shrinks to 0, the Riemann sum

$$\sum_i \sum_j f(x_{ij}, y_{ij}), \Delta u, \Delta v$$

converges to the double integral $\iint_R f(x,y), dA$ whenever f is Riemann integrable. Setting f = 1, this gives

$$\text{area}(R) = \lim_{\text{mesh} \to 0} \sum_i \sum_j \Delta u, \Delta v,$$

which is just the textbook's "sum of rectangle areas approaches the area of R" claim written in the mesh → 0 limit language.

The same tiling does the work for the change-of-variables theorem. Tile the parameter region S in the uv-plane with subrectangles S_{ij} of size Δu × Δv. Push each S_{ij} through T. In the limit Δu, Δv → 0, the image R_{ij} of S_{ij} is approximately the parallelogram spanned by Δu r_u and Δv r_v, where r(u, v) = T(u, v) and r_u, r_v are the partial derivatives. The area of that local parallelogram is |J(u_{ij}, v_{ij})| · Δu · Δv, where J is the Jacobian determinant. Sum over all tiles:

$$\iint_R f(x, y), dA = \lim_{\text{mesh} \to 0} \sum_i \sum_j f(g(u_{ij}, v_{ij}), h(u_{ij}, v_{ij})), |J(u_{ij}, v_{ij})|, \Delta u, \Delta v,$$

where T(u, v) = (g(u, v), h(u, v)). OpenStax §5.7 Theorem 5.14

For a linear T, the Jacobian matrix is exactly A and J(u, v) = det A is constant. Each subrectangle's image is then exactly the parallelogram on {Δu · a₁, Δv · a₂}, where a₁, a₂ are the columns of A, and the area of that parallelogram is |det A| · Δu · Δv. Summing, area(T(R)) = |det A| · area(R).

The figure below carries the chain of equalities visually. Subrectangles S_{ij} in the uv-plane (left) push through T to distorted tiles R_{ij} in the xy-plane (right). The inset zoom in the upper-right shows the leading-order parallelogram Δu r_u × Δv r_v that locally replaces R_{ij}.

Caption: The Riemann-sum picture behind the change-of-variables theorem. Region S in the uv-plane is partitioned into subrectangles S_{ij}; T maps each to a tile R_{ij} in the xy-plane. The inset zoom shows the local parallelogram formed by the tangent vectors Δu r_u and Δv r_v. OpenStax Calculus Volume 3 §5.7

A subtle but critical point: this tiling argument is what gives the proof its generality. The base case (Chapter 1) handles a single rectangle; the tiling turns the same parallelogram computation into a sum over a partition of R. But this is also the step that secretly demands a hypothesis on R. The mesh → 0 limit has to converge, and the convergence of rectangle sums is exactly the hypothesis that R is what we will call Jordan-measurable. That is the topic of the next chapter.

3. What "Sufficiently Nice" Actually Means: Bounded, Boundary of Measure Zero

The phrase "sufficiently nice region R" has a precise mathematical translation, and it is one specific condition.

A bounded set S ⊂ ℝⁿ is called Jordan-measurable if its indicator function χ_S is Riemann-integrable. Apostol (Theorem 14.9) shows that this is equivalent to a boundary condition: S is Jordan-measurable iff its topological boundary ∂S has Jordan content (n-dimensional measure) zero. Two equivalent reformulations are useful.

• Inner/outer Jordan measure. The inner Jordan measure is the supremum of areas of simple sets (finite unions of rectangles) contained in S. The outer Jordan measure is the infimum of areas of simple sets containing S. S is Jordan-measurable iff these two agree, and the common value is the Jordan content c(S).
• Boundary criterion. S is Jordan-measurable iff ∂S has n-dimensional content (equivalently, Lebesgue measure) zero.

The Wikipedia article on Peano–Jordan measure gives a clean visual for this construction.

Caption: The inner/outer Jordan-measure construction. The blue curve bounds an arbitrary bounded set B; the green staircase is an inner simple set S₋ ⊂ B; the pink staircase is an outer simple set S₊ ⊃ B. B is Jordan-measurable when the gap between inner and outer approximation vanishes, which happens precisely when the boundary ∂B has measure zero. Wikipedia, Peano–Jordan measure

The picture makes the criterion visual: as the mesh refines, the inner simple set grows toward S from inside, the outer simple set shrinks toward S from outside. They meet at the area c(S) precisely when the boundary ∂S has content zero.

A concrete instance of the same construction, on a familiar shape, is the unit disc approximated by dyadic cubes. At stage k, partition the enclosing square into 2^k × 2^k small squares of side 1/2^k; sum the measures of the cubes contained in the disc to get an inner approximation, and sum those that intersect the disc to get an outer approximation. The two agree as k → ∞ for a Jordan-measurable shape like the disc, whose boundary is the unit circle, a set of content zero.

Caption: The unit disc approximated by a 4 × 4 dyadic cube grid (the Dalessi construction, at stage k = 2). Inner cubes lie entirely inside the disc; outer cubes intersect the disc. As the grid refines, inner and outer sums converge to π — the area of the unit disc. Dalessi ETH thesis

All standard shapes encountered in multivariable calculus are Jordan-measurable. Rectangles, disks, balls, triangles, polygons, Type-I regions bounded above and below by continuous graphs over a closed interval — every one of these has a boundary of content zero. So do finite unions and intersections of such sets.

A common trap. "Jordan-measurable" is not the same as "Lebesgue-measurable." Every Jordan-measurable set is Lebesgue-measurable, and the two measures agree on bounded Jordan-measurable sets. But Jordan-measurable sets do not form a σ-algebra: a countable union of Jordan-measurable sets need not be Jordan-measurable. Lebesgue-measurable sets, by contrast, do form a σ-algebra. The rational numbers Q ∩ [0,1] are the standard witness: each singleton {q} is trivially Jordan-measurable (it has content zero), but the countable union Q ∩ [0,1] is bounded, its boundary is the entire interval [0,1], and it is not Jordan-measurable. We will return to this distinction in Chapter 6.

So "sufficiently nice region R" is not a hand-wave. It is a precise one-liner: bounded, boundary of measure zero. And the rectangle-tiling argument of Chapter 2 is exactly the computation that needs this hypothesis.

4. Where the Rectangle-Sum Argument Collapses: Cantor Sets and Q ∩ [0,1]

The hypothesis is non-negotiable. Here are three sets that are bounded but not Jordan-measurable; in each case the rectangle-sum argument fails, and the textbook's "side lengths → 0" claim breaks.

The standard ε = 1/3 Cantor set

The classic middle-thirds Cantor set C ⊂ [0,1] is built by repeatedly removing the open middle third of each remaining interval. The Lebesgue measure of C is zero, but C is not Jordan-measurable: its complement in [0,1] is dense, so any finite cover of [0,1] \ C covers almost all of [0,1], making the outer Jordan measure of [0,1] \ C equal to 1. The inner Jordan measure of [0,1] \ C, however, is 0 (no finite union of rectangles is entirely inside the dense complement). So c̄(C) = 1 ≠ 0 = c(C).

A more refined way to see this: the boundary of C equals C itself (C is closed and contains all its limit points; every open interval in [0,1] meets C and meets its complement). So ∂C = C has Lebesgue measure zero — good. But ∂C = C has outer Jordan measure 1, since C sits inside [0,1] and cannot be inner-approximated at all. The inner-outer Jordan criterion fails.

The Smith–Volterra–Cantor (fat) Cantor set

The fat Cantor set is built the same way, except at step n one removes only the central portion of width 1/2^{2n} (rather than 1/3). The result is a compact set F ⊂ [0,1] with positive Lebesgue measure (in fact, 1/2), but still not Jordan-measurable: the complement of F is dense, so the inner Jordan measure of F is 0 while the outer Jordan measure equals the positive Lebesgue measure 1/2. Again c̄(F) ≠ c(F), so F is not Jordan-measurable.

The contrast with the standard Cantor set is instructive: F has positive Lebesgue measure, yet still fails Jordan-measurability. "Measure zero" is not enough; "boundary has Jordan content zero" is the right notion.

Caption: The Smith–Volterra–Cantor set under construction. Step 0 (top bar): a single interval. Step 5 (bottom bar): heavily fragmented intervals. At each step, the central 1/2^{2n} of every remaining piece is removed. The complement grows dense, so the inner Jordan measure is 0, but the Lebesgue measure stays positive. Smith–Volterra–Cantor set, Wikimedia Commons

Q ∩ [0,1]

The rationals in [0,1] are the cleanest pathological example. Each singleton {q} is trivially Jordan-measurable: it has content zero (a single point in ℝⁿ has zero measure). But the countable union Q ∩ [0,1] is bounded, its boundary equals the closure minus the interior = [0,1] \ Q = [0,1] (rationals are dense, but [0,1] has empty interior in itself, so closure = [0,1] and interior = ∅... wait: interior of Q ∩ [0,1] is ∅, closure of Q ∩ [0,1] is [0,1], so boundary = [0,1]). The boundary is the entire interval, which has Jordan content 1 ≠ 0. So Q ∩ [0,1] is not Jordan-measurable — despite being a countable union of Jordan-measurable sets. This is exactly the σ-algebra failure flagged in Chapter 3.

The unifying picture: in each case, the boundary has "thick" content (either positive measure or even just not Jordan-content-zero), so the inner and outer rectangle sums do not agree in the limit. The rectangle-tiling approximation gives a lower bound from inside that does not converge to the outer bound, and the change-of-variables formula does not apply.

5. Why an Invertible Linear Map Preserves the Hypothesis

Having named the hypothesis (R Jordan-measurable), we need to verify that it survives the map T(x) = Ax. Without this step, the rectangle-tiling argument of Chapter 2 applies to R but not to T(R), and the conclusion area(T(R)) = |det A| · area(R) does not follow.

The key tool is a property of C¹ maps. If g : A → ℝⁿ is C¹ and E ⊂ A has Lebesgue measure zero, then g(E) has Lebesgue measure zero. This is Lemma 18.1 in Munkres, Analysis on Manifolds. The proof uses the bound |g(x) − g(y)| ≤ nM|x − y| (from the mean-value theorem) on a compact exhaustion of A: a cover of E by small cubes lifts to a cover of g(E) by cubes of total volume smaller by a factor depending on the bound M.

Two consequences follow. First, a C¹ diffeomorphism g satisfies g(Int D) = Int g(D) and g(∂D) = ∂g(D) (Theorem 18.2(a) in Munkres). Second, combining these: if D is rectifiable (Jordan-measurable, in Munkres's terminology; see Theorem 14.1, same source), then ∂D has measure zero, hence g(∂D) has measure zero by Lemma 18.1, hence ∂g(D) = g(∂D) has measure zero, hence g(D) is rectifiable (Theorem 18.2(b)).

A linear map T(x) = Ax with det A ≠ 0 is a C^∞ diffeomorphism of ℝⁿ, so T(R) is rectifiable whenever R is. The rectangle-tiling argument of Chapter 2 applies to T(R) as well, and summing the parallelogram base case (Chapter 1) over a tiling of R gives

$$\text{area}(T(R)) = |\det A| \cdot \text{area}(R).$$

Munkres's Figure 19.3 illustrates a related step in the proof of the change-of-variables theorem — the Fubini reduction that handles the inductive step in Lemma 19.1 — not the "hypothesis survives T" lemma itself. The lemma is Theorem 18.2(b); the proof chain runs through Lemma 18.1 and Theorem 18.2(a).

Caption: The slice diagram for Lemma 19.1 Step 5 (the Fubini reduction in Munkres's proof of the change-of-variables theorem). The compact set S = h⁻¹(Q) is fibred by the last coordinate, so the n-dimensional integral reduces to iterated (n−1)-dimensional integrals. This is one step in the proof machinery that ultimately relies on Theorem 18.2(b) to guarantee rectifiability is preserved under a C¹ diffeomorphism. Munkres, Analysis on Manifolds (1991), Figure 19.3

So the hypothesis is not fragile. Any invertible linear map takes Jordan-measurable regions to Jordan-measurable regions. This is why the determinant-area formula works on every "nice" shape a textbook asks about — rectangles, disks, polygons, Type-I regions — and the formula is consistent across them.

6. Lifting to ℝⁿ: Parallelepipeds, |det A|, and the det A = 0 Case

The whole structure lifts verbatim to ℝⁿ. Replace the parallelogram by a parallelepiped: given n vectors u₁, …, uₙ, the n-parallelepiped they span is

$${\alpha_1 u_1 + \cdots + \alpha_n u_n : 0 \leq \alpha_i \leq 1}.$$

A linear map T(x) = Ax sends this parallelepiped to the parallelepiped spanned by Au₁, …, Auₙ. The n-dimensional volume (Jordan content) of the image is

$$|\det[Au_1 ;\cdots; Au_n]| = |\det(A[u_1 ;\cdots; u_n])| = |\det A| \cdot |\det[u_1 ;\cdots; u_n]|,$$

which is |det A| times the volume of the source parallelepiped. The computation is identical to the 2D case; only the dimension changed.

The lift to 3D is geometrically vivid. The unit cube in ℝ³, spanned by the standard basis vectors i, j, k, maps under T to a slanty "parallelepiped" — a parallelepiped, the three-dimensional analogue of the parallelogram. Its volume is |det A|.

Caption: A linear map takes the unit cube in ℝ³ (teal wireframe) to a slanty parallelepiped (translucent yellow), whose volume is |det A|. The intuition from 2D carries over directly: the determinant is the volume-scaling factor. 3Blue1Brown

Rectangle partitions become box partitions, and the mesh → 0 Riemann-sum argument runs identically in ℝⁿ. The parallelepiped base case is the engine of each tile; summing over a partition of R recovers the full n-dimensional content. The invertibility condition det A ≠ 0 is what makes T a C^∞ diffeomorphism and so preserves Jordan-measurability of R. Apostol Theorem 15.15 gives the linear-case statement directly; Munkres Theorem 20.1 handles both the invertible and degenerate cases.

In the degenerate case det A = 0, the matrix A collapses ℝⁿ onto a proper subspace V of dimension less than n. The image T(R) lies entirely inside V, and any bounded subset of V has n-dimensional content 0. So vol(T(R)) = 0, consistent with |det A| · vol(R) = 0 · vol(R) = 0.

The standard Type-I region in ℝ² — bounded above and below by continuous graphs over a closed interval — is the most common kind of Jordan-measurable region encountered in multivariable calculus. Its boundary is the union of two graphs and two vertical line segments, all of which have 2-dimensional content zero, so the region itself is Jordan-measurable by Apostol Theorem 14.9.

Caption: A Type-I region in ℝ²: S = {(x, y) : a < x < b, φ₁(x) < y < φ₂(x)}, bounded above by the graph y = φ₂(x) and below by y = φ₁(x), with vertical line segments at x = a and x = b. The boundary consists of two graphs plus two segments, all of 2-dimensional content zero, so S is Jordan-measurable. Apostol, Mathematical Analysis (1974), Figure 14.4

Jordan vs. Lebesgue: a short comparison

On bounded Jordan-measurable sets, Jordan and Lebesgue measures agree. But Jordan-measurable sets do not form a σ-algebra, while Lebesgue-measurable sets do.

So the distinction matters precisely when you leave the world of "finite unions of rectangles" and start taking limits. For the change-of-variables theorem, the Jordan framework is enough — the rectangle-tiling construction stays in the finite-covering world throughout, and no countable union is needed.

Summary

The textbook's "sufficiently nice region R" is one precise hypothesis: R ⊂ ℝⁿ is bounded and its boundary ∂R has n-dimensional content zero (which, for a bounded set, is equivalent to Lebesgue measure zero). This is the hypothesis that makes the rectangle-tiling / Riemann-sum argument converge, and the hypothesis that invertible linear maps preserve.

The proof structure that hangs on this hypothesis runs as follows.

1. The parallelogram base case: linearity sends the parallelogram on {u, v} to the parallelogram on {Au, Av}, and |det A| lands via det(AB) = det A · det B.
2. The rectangle tiling: tile R with subrectangles, push each through T, and sum. The local Jacobian factor is constant in the linear case, so each tile's image area is exactly |det A| times the tile area.
3. The hypothesis: this sum converges to area(R) iff R is Jordan-measurable.
4. The preservation: an invertible linear map (det A ≠ 0) is a C^∞ diffeomorphism, so T(R) is Jordan-measurable whenever R is.
5. The n-dimensional lift: the parallelogram becomes a parallelepiped, the rectangle partition becomes a box partition, the mesh → 0 limit runs identically, and det A = 0 gives 0 = 0 by the lower-dimensional-content-vanishes argument.

Everything else in the proof is mechanical: linearity for the base case, multiplicativity for the |det A| factor, Riemann-sum convergence for the tiling step, and the C¹ diffeomorphism property for the preservation step.

参考文献

1. Apostol, Mathematical Analysis (2nd ed., 1974) — Ch. 14 §14.6 (Jordan-measurable sets, Theorem 14.9), Ch. 15 §15.9–15.12 (Theorem 15.15: linear coordinate transformation). PDF mirror
2. OpenStax Calculus Volume 3 §5.7 (Strang/Herman, 2016) — Theorem 5.14 (change of variables for double integrals), Theorem 5.15 (triple-integral analogue), Figures 5.74 and 5.75. OpenStax chapter
3. 3Blue1Brown, "The determinant" (Grant Sanderson, 2016) — Ch. 6 of Essence of Linear Algebra, the canonical visual for det as area/volume scaling. Lesson page
4. Rudin, Principles of Mathematical Analysis (3rd ed., 1976) — Ch. 10 (Definition 10.1, Theorem 10.9 for the |Jacobian| change-of-variables formula) and Ch. 11 (Theorem 11.33, Lebesgue criterion). PDF mirror
5. Wikipedia, "Peano–Jordan measure" — Glossary reference for inner/outer Jordan measure, boundary-measure-zero criterion, counterexamples (standard Cantor set, Smith–Volterra–Cantor, Q ∩ [0,1]), and Jordan-vs-Lebesgue comparison. Wikipedia
6. Munkres, Analysis on Manifolds (1991) — Theorem 14.1 (rectifiable iff boundary measure zero), Lemma 18.1 (C¹ maps preserve measure-zero sets), Theorem 18.2(b) (diffeomorphisms preserve rectifiability), Theorem 20.1 (n-dimensional volume formula, including det A = 0). Internet Archive
7. Dalessi, "An overview of Jordan measure" (ETH Zurich bachelor thesis, April 2023) — Self-contained pedagogical walkthrough; Theorem 4.7 (boundary-measure-zero iff Jordan-measurable); unit-disc dyadic cube approximation. Thesis PDF
8. MIT 18.022 handout (Hesselholt, Fall 2007), "The change of variables theorem" — Compact R² and R³ statement of the change-of-variables theorem under diffeomorphisms, with worked example. Handout PDF
9. Spivak, Calculus on Manifolds (1965) — Theorem 3-13 (rigorous C¹ change-of-variables in ℝⁿ), Theorem 4-6 (determinant as the unique volume element on Λⁿ(ℝⁿ)). PDF mirror
10. Miller, "The Probability Lifesaver: Change of Variable Theorem" (2016/2017) — Intuition-first sketch via Taylor expansion and cross product; §1.3 worked linear 2×2 example (T(x, y) = (2x + 3y, 4y), area 8). Chapter PDF