Derivatives of Inverse Functions

Suppose you know that y=x3+2xy = x^3 + 2x is a smooth, strictly increasing function of xx. Its inverse x=f−1(y)x = f^{-1}(y) exists. But you were never handed a formula for it. Solving a cubic explicitly is unpleasant. For more general functions it may be outright impossible. Can you still find the derivative of f−1f^{-1} at a point? You would use only what you know about ff itself. This is one of the oldest and most useful questions in differential calculus. The answer is a strikingly clean formula. It lets you differentiate an inverse without ever writing it down explicitly. The idea matters because inverse functions are everywhere. log⁡\log is the inverse of exp⁡\exp. arcsin⁡\arcsin is the (restricted) inverse of sin⁡\sin. nnth roots are inverses of nnth powers. In probability the quantile function is the inverse of a cumulative distribution function. In every one of these cases, the classical definitions of the derivatives of log⁡\log, arcsin⁡\arcsin, xn\sqrt[n]{x} are obtained by inverting a known derivative. They are not re-derived from a limit from scratch. The big picture is geometric. The graph of f−1f^{-1} is the graph of ff reflected across the line y=xy = x. So a tangent line of slope mm at a point (a,f(a))(a, f(a)) becomes something new after reflection. It becomes a tangent line of slope 1/m1/m at (f(a),a)(f(a), a) — provided m≠0m \neq 0. Making this reflection argument rigorous requires three ingredients already in your toolkit. First, the link between Monotonicity and the Sign of the Derivative and injectivity. Second, the algebra of limits packaged in The Chain Rule. Third, the topological fact — Darboux's Theorem: The Intermediate Value Property of Derivatives. It states that derivatives cannot jump over zero without vanishing somewhere. This article proves the one-variable Inverse Function Theorem. Under a hypothesis of continuous strict monotonicity and a nonvanishing derivative, f−1f^{-1} is differentiable. It satisfies (f−1)′(f(a))=1/f′(a)(f^{-1})'(f(a)) = 1/f'(a). We build the theorem from first principles. We examine why each hypothesis is load-bearing with an explicit counterexample. We also work two nontrivial numerical examples by hand.

Key takeaways
  • If ff is continuous and strictly monotonic on an interval, f−1f^{-1} exists and is continuous on f(I)f(I).
  • If in addition ff is differentiable at aa with f′(a)≠0f'(a) \neq 0, then f−1f^{-1} is differentiable at b=f(a)b = f(a).
  • The formula is (f−1)′(b)=1f′(a)=1f′(f−1(b))(f^{-1})'(b) = \dfrac{1}{f'(a)} = \dfrac{1}{f'(f^{-1}(b))}.
  • The hypothesis f′(a)≠0f'(a) \neq 0 cannot be dropped: f(x)=x3f(x) = x^3 at a=0a=0 shows f−1f^{-1} can fail to be differentiable.
  • The proof is a direct ε\varepsilon–δ\delta/sequential limit computation, not a symbolic "solve for yy" trick.
  • Key idea: reciprocal difference quotients converge to reciprocal derivatives, because f−1f^{-1} is continuous and injective.

Where this lives today

The formula (f−1)′(b)=1/f′(a)(f^{-1})'(b) = 1/f'(a) is the mechanism behind essentially every "inverse special function" derivative learned in a first calculus course: ln⁡′(x)=1/x\ln'(x) = 1/x from exp⁡′(t)=exp⁡(t)\exp'(t) = \exp(t), arctan⁡′(x)=1/(1+x2)\arctan'(x) = 1/(1+x^2) from tan⁡′(t)=sec⁡2t\tan'(t) = \sec^2 t, and the derivative of xn\sqrt[n]{x} from the derivative of t↦tnt \mapsto t^n. In numerical analysis, Newton's method for solving f(x)=0f(x) = 0 implicitly relies on local invertibility and the reciprocal-derivative estimate to control convergence rates. In statistics and probability, the derivative of a quantile function (inverse CDF) used in the change-of-variables formula for densities is exactly an instance of this theorem, since a strictly increasing CDF with positive density is precisely a differentiable strictly monotonic function with nonvanishing derivative. In economics, inverse demand and inverse supply curves are differentiated via this same reciprocal rule to compute marginal quantities. In cryptography and coding theory, monotone bijections and their smooth inverses appear in the analysis of certain hash and mixing constructions where local invertibility (a discrete cousin of this theorem) controls distortion. Finally, this one-variable result is the base case from which the multivariable Inverse Function Theorem (with the Jacobian matrix inverse replacing 1/f′1/f') is generalized in advanced analysis and differential geometry, built on The Chain Rule in higher dimensions together with a fixed-point argument.

Ask the tutor: Can you show me how this same reciprocal-slope idea generalizes to the multivariable Inverse Function Theorem with Jacobian matrices?

DefinitionInvertibility and the derivative of the inverse

Let I⊆RI \subseteq \mathbb{R} be an interval and let f:I→Rf : I \to \mathbb{R} be a function. 1. ff is strictly increasing (resp. strictly decreasing) on II if for all x1,x2∈Ix_1, x_2 \in I with x1<x2x_1 < x_2 we have f(x1)<f(x2)f(x_1) < f(x_2) (resp. f(x1)>f(x2)f(x_1) > f(x_2)). ff is strictly monotonic if it is strictly increasing or strictly decreasing on II. 2. If ff is strictly monotonic on II, then ff is injective, so f:I→f(I)f : I \to f(I) has a well-defined inverse function f−1:f(I)→If^{-1} : f(I) \to I satisfying

f−1(f(x))=xfor all x∈I,f(f−1(y))=yfor all y∈f(I).f^{-1}(f(x)) = x \quad \text{for all } x \in I, \qquad f(f^{-1}(y)) = y \quad \text{for all } y \in f(I).

3. Given aa interior to II (or an endpoint, with one-sided derivative), ff is differentiable at aa with derivative f′(a)f'(a) if

f′(a)=lim⁡x→af(x)−f(a)x−af'(a) = \lim_{x \to a} \frac{f(x) - f(a)}{x - a}

exists as a finite real number. 4. Given J=f(I)J = f(I) and b=f(a)∈Jb = f(a) \in J with bb a limit point of JJ, we say f−1f^{-1} is differentiable at bb with derivative (f−1)′(b)(f^{-1})'(b) if

(f−1)′(b)=lim⁡y→bf−1(y)−f−1(b)y−b(f^{-1})'(b) = \lim_{y \to b} \frac{f^{-1}(y) - f^{-1}(b)}{y - b}

exists as a finite real number.

Intuition

Picture the graph of a strictly increasing differentiable function ff passing through (a,b)(a, b) with b=f(a)b = f(a), and suppose the tangent line there has slope f′(a)=m≠0f'(a) = m \neq 0. Reflecting the entire plane across the line y=xy = x swaps the roles of horizontal and vertical, turning the graph of ff into the graph of f−1f^{-1}, and turning the point (a,b)(a,b) into (b,a)(b,a). A line of slope mm reflects to a line of slope 1/m1/m — this is because reflection across y=xy=x swaps Δx\Delta x and Δy\Delta y in any secant line, so the slope Δy/Δx\Delta y/\Delta x becomes Δx/Δy=1/(Δy/Δx)\Delta x/\Delta y = 1/(\Delta y/\Delta x). Passing to the limit, the tangent line to f−1f^{-1} at (b,a)(b,a) should have slope 1/m1/m. This reflection picture is the entire content of the theorem; the proof below simply makes "passing to the limit" honest. The reciprocal relationship also explains geometrically why f′(a)=0f'(a) = 0 must be excluded: a horizontal tangent line reflects to a vertical line, which is not the graph of a function — f−1f^{-1} genuinely fails to be differentiable (though it may still have infinite one-sided "derivative," a vertical tangent) at such a point.

Figure
Reflecting the graph of ff across y=xy=x turns slope mm into slope 1/m1/m
x y y = x y = f(x) (a, b), slope m y = f⁻¹(x) (b, a), slope 1/m

Theorem and proof

TheoremOne-variable Inverse Function Theorem

Let I⊆RI \subseteq \mathbb{R} be an open interval and let f:I→Rf : I \to \mathbb{R} satisfy:

(H1) ff is continuous on II;

(H2) ff is strictly monotonic on II;

(H3) ff is differentiable at some point a∈Ia \in I;

(H4) f′(a)≠0f'(a) \neq 0. Let J=f(I)J = f(I) and b=f(a)b = f(a). Then: (i) JJ is an interval, f−1:J→If^{-1} : J \to I exists, is strictly monotonic (of the same type as ff), and is continuous on JJ; (ii) bb lies in the interior of JJ, and f−1f^{-1} is differentiable at bb, with

(f−1)′(b)=1f′(a).(f^{-1})'(b) = \frac{1}{f'(a)}.

Consequently, if ff is differentiable on all of II with f′(x)≠0f'(x) \neq 0 for every x∈Ix \in I, then f−1f^{-1} is differentiable at every point of JJ and

(f−1)′(y)=1f′(f−1(y))for all y∈J.(f^{-1})'(y) = \frac{1}{f'(f^{-1}(y))} \qquad \text{for all } y \in J.
Proof

We give a complete proof organized into the four claims of the statement, using direct limit arguments (no unproved "solve for the inverse" heuristics).

Step 1: J=f(I)J = f(I) is an interval, and f−1f^{-1} exists. Since ff is strictly monotonic on II (H2), it is injective: if x1≠x2x_1 \neq x_2 then, say, x1<x2x_1 < x_2, so strict monotonicity gives f(x1)≠f(x2)f(x_1) \neq f(x_2) (either f(x1)<f(x2)f(x_1) < f(x_2) or f(x1)>f(x2)f(x_1) > f(x_2) according to whether ff increases or decreases). Injectivity means f:I→Jf : I \to J is a bijection, so f−1:J→If^{-1} : J \to I is a well-defined function. To see JJ is an interval, let y1,y2∈Jy_1, y_2 \in J with y1<y2y_1 < y_2, and let y1<y0<y2y_1 < y_0 < y_2; we must show y0∈Jy_0 \in J. Write y1=f(x1)y_1 = f(x_1), y2=f(x2)y_2 = f(x_2) for x1,x2∈Ix_1, x_2 \in I. Since ff is continuous on II (H1), the closed interval [x1,x2][x_1,x_2] (or [x2,x1][x_2,x_1]) lies in II because II is an interval, and the Intermediate Value Theorem applied to ff on this closed interval produces x0x_0 strictly between x1x_1 and x2x_2 with f(x0)=y0f(x_0) = y_0. Hence y0∈Jy_0 \in J. This proves JJ is an interval.

Step 2: f−1f^{-1} is strictly monotonic of the same type as ff. Assume ff is strictly increasing (the decreasing case is symmetric, replacing ff by −f-f). Let y1<y2y_1 < y_2 in JJ, and set x1=f−1(y1)x_1 = f^{-1}(y_1), x2=f−1(y2)x_2 = f^{-1}(y_2). If we had x1≥x2x_1 \geq x_2, then since ff is strictly increasing, f(x1)≥f(x2)f(x_1) \geq f(x_2), i.e. y1≥y2y_1 \geq y_2, contradicting y1<y2y_1 < y_2. Hence x1<x2x_1 < x_2, i.e. f−1(y1)<f−1(y2)f^{-1}(y_1) < f^{-1}(y_2). So f−1f^{-1} is strictly increasing on JJ.

Step 3: f−1f^{-1} is continuous on JJ. Fix y0∈Jy_0 \in J and let x0=f−1(y0)x_0 = f^{-1}(y_0). We show f−1f^{-1} is continuous at y0y_0; treat the case y0y_0 interior to JJ (endpoints are handled identically with one-sided limits, using that JJ is an interval by Step 1). Let ε>0\varepsilon > 0. Shrinking ε\varepsilon if necessary we may assume [x0−ε,x0+ε]⊆I[x_0 - \varepsilon, x_0+\varepsilon] \subseteq I (possible since II is open). Since ff is strictly increasing (WLOG), we have

f(x0−ε)<f(x0)=y0<f(x0+ε).f(x_0 - \varepsilon) < f(x_0) = y_0 < f(x_0+\varepsilon).

Let δ=min⁡{y0−f(x0−ε), f(x0+ε)−y0}>0\delta = \min\{ y_0 - f(x_0-\varepsilon),\, f(x_0+\varepsilon) - y_0 \} > 0. If y∈Jy \in J with ∣y−y0∣<δ|y - y_0| < \delta, then

f(x0−ε)<y<f(x0+ε).f(x_0 - \varepsilon) < y < f(x_0 + \varepsilon).

Applying f−1f^{-1}, which is strictly increasing by Step 2, and using f−1(f(x0±ε))=x0±εf^{-1}(f(x_0\pm\varepsilon)) = x_0 \pm \varepsilon, we get

x0−ε<f−1(y)<x0+ε,x_0 - \varepsilon < f^{-1}(y) < x_0 + \varepsilon,

i.e. ∣f−1(y)−x0∣<ε|f^{-1}(y) - x_0| < \varepsilon. This proves f−1f^{-1} is continuous at y0y_0. Since y0∈Jy_0 \in J was arbitrary, f−1f^{-1} is continuous on JJ. This establishes part (i) of the theorem.

Step 4: b=f(a)b = f(a) is interior to JJ. Since II is open and a∈Ia \in I, there is η>0\eta > 0 with (a−η,a+η)⊆I(a-\eta, a+\eta) \subseteq I. By strict monotonicity and Step 1's IVT argument applied on [a−η,a+η][a-\eta,a+\eta], the image f((a−η,a+η))f((a-\eta,a+\eta)) is an open interval around bb contained in JJ — open because ff restricted to [a−η,a+η][a-\eta, a+\eta] is a continuous strictly monotonic bijection onto [f(a−η),f(a+η)][f(a-\eta),f(a+\eta)] (increasing case; endpoints swap in the decreasing case) by the same IVT argument, and f(a)=bf(a)=b is strictly between the images of the endpoints. Hence bb is interior to JJ, so it makes sense to ask for the two-sided derivative of f−1f^{-1} at bb.

Step 5: differentiability of f−1f^{-1} at bb, and the formula. We compute the limit defining (f−1)′(b)(f^{-1})'(b) directly. Let (yn)(y_n) be any sequence in J∖{b}J \setminus \{b\} with yn→by_n \to b; by the sequential characterization of limits it suffices to show

f−1(yn)−f−1(b)yn−b⟶1f′(a).\frac{f^{-1}(y_n) - f^{-1}(b)}{y_n - b} \longrightarrow \frac{1}{f'(a)}.

Set xn=f−1(yn)∈Ix_n = f^{-1}(y_n) \in I, so f(xn)=ynf(x_n) = y_n and f−1(b)=af^{-1}(b) = a. Claim: xn≠ax_n \neq a for all nn, and xn→ax_n \to a. Since f−1f^{-1} is injective (it is a function inverse) and yn≠by_n \neq b, we get xn=f−1(yn)≠f−1(b)=ax_n = f^{-1}(y_n) \neq f^{-1}(b) = a. Since f−1f^{-1} is continuous at bb (Step 3) and yn→by_n \to b, we get xn=f−1(yn)→f−1(b)=ax_n = f^{-1}(y_n) \to f^{-1}(b) = a. Now rewrite, using xn≠ax_n \neq a so the denominator below is nonzero:

f−1(yn)−f−1(b)yn−b=xn−af(xn)−f(a)=1f(xn)−f(a)xn−a.\frac{f^{-1}(y_n) - f^{-1}(b)}{y_n - b} = \frac{x_n - a}{f(x_n) - f(a)} = \frac{1}{\dfrac{f(x_n) - f(a)}{x_n - a}}.

Since xn→ax_n \to a, xn≠ax_n \neq a, and ff is differentiable at aa (H3) with f′(a)≠0f'(a) \neq 0 (H4), the difference quotient in the denominator satisfies

f(xn)−f(a)xn−a⟶f′(a)≠0.\frac{f(x_n) - f(a)}{x_n - a} \longrightarrow f'(a) \neq 0.

By the algebraic limit theorem for quotients (valid since the limit f′(a)f'(a) of the denominator sequence is nonzero, so the denominator sequence is eventually nonzero and reciprocals converge to the reciprocal — this is exactly the continuity of t↦1/tt \mapsto 1/t at t=f′(a)≠0t = f'(a) \neq 0, an instance of the composition-of-limits fact underlying The Chain Rule), we conclude

f−1(yn)−f−1(b)yn−b=1f(xn)−f(a)xn−a⟶1f′(a).\frac{f^{-1}(y_n) - f^{-1}(b)}{y_n - b} = \frac{1}{\dfrac{f(x_n)-f(a)}{x_n - a}} \longrightarrow \frac{1}{f'(a)}.

Since (yn)(y_n) was an arbitrary sequence in J∖{b}J\setminus\{b\} converging to bb, the sequential criterion for limits gives

lim⁡y→bf−1(y)−f−1(b)y−b=1f′(a),\lim_{y \to b} \frac{f^{-1}(y) - f^{-1}(b)}{y - b} = \frac{1}{f'(a)},

which is precisely the statement that f−1f^{-1} is differentiable at bb with (f−1)′(b)=1/f′(a)(f^{-1})'(b) = 1/f'(a). This proves part (ii).

Step 6: the global formula. If ff is differentiable on all of II with f′(x)≠0f'(x) \neq 0 for every x∈Ix \in I, apply Steps 1–5 with aa replaced by an arbitrary point of II. Every y∈Jy \in J is f(a)f(a) for a unique a=f−1(y)∈Ia = f^{-1}(y) \in I (by bijectivity from Step 1), and (H3)–(H4) hold at that aa by assumption. Hence f−1f^{-1} is differentiable at every y∈Jy \in J, with

(f−1)′(y)=1f′(a)=1f′(f−1(y)).(f^{-1})'(y) = \frac{1}{f'(a)} = \frac{1}{f'(f^{-1}(y))}.

■\blacksquare

∎
CorollaryNonvanishing derivative from monotonicity plus Darboux

Let ff be differentiable on an open interval II with f′(x)≠0f'(x) \neq 0 for all x∈Ix \in I. Then ff is automatically strictly monotonic on II (so hypothesis (H2) of the theorem is implied by (H3)+(H4) holding at every point, given continuity of ff, which differentiability provides).

Proof

By Darboux's Theorem: The Intermediate Value Property of Derivatives, f′f' has the intermediate value property on II: it cannot take both a positive and a negative value on II without taking the value 00 somewhere in between. Since f′(x)≠0f'(x) \neq 0 everywhere on II by hypothesis, f′f' cannot change sign on II; that is, either f′>0f' > 0 throughout II or f′<0f' < 0 throughout II. By the standard consequence of the Mean Value Theorem relating sign of derivative to monotonicity — Monotonicity and the Sign of the Derivative — ff is strictly increasing on II in the first case and strictly decreasing in the second. ■\blacksquare

∎

This corollary is why, in practice, one often verifies the Inverse Function Theorem's hypotheses by checking only that ff is differentiable with f′≠0f' \neq 0 everywhere on an interval: monotonicity, and hence invertibility, is then free.

Worked examples

Example 1 — a genuinely non-symmetric cubic. Let f(x)=x3+2x+1f(x) = x^3 + 2x + 1 on I=RI = \mathbb{R}. We find (f−1)′(4)(f^{-1})'(4). First, f′(x)=3x2+2>0f'(x) = 3x^2 + 2 > 0 for all xx, so by the Corollary, ff is strictly increasing on all of R\mathbb{R}; it is a polynomial, hence continuous and differentiable everywhere, so hypotheses (H1)–(H4) of the theorem hold at every point of II. We must locate aa with f(a)=4f(a) = 4:

a3+2a+1=4⟹a3+2a−3=0.a^3 + 2a + 1 = 4 \quad\Longrightarrow\quad a^3 + 2a - 3 = 0.

Testing a=1a = 1: 1+2−3=01 + 2 - 3 = 0. So a=1a = 1 is a root. (Factoring, a3+2a−3=(a−1)(a2+a+3)a^3+2a-3 = (a-1)(a^2+a+3), and the quadratic factor a2+a+3a^2+a+3 has discriminant 1−12=−11<01 - 12 = -11 < 0, so a=1a=1 is the only real root — consistent with injectivity.) Now compute f′(1)=3(1)2+2=5f'(1) = 3(1)^2 + 2 = 5. By the theorem,

(f−1)′(4)=1f′(1)=15.(f^{-1})'(4) = \frac{1}{f'(1)} = \frac{1}{5}.

Note we never had to solve for f−1(y)f^{-1}(y) as an explicit algebraic formula in yy (which would require the general cubic formula) — only for the specific preimage a=1a=1 of b=4b=4, and then a derivative evaluation at that single point.

Example 2 — an inverse trigonometric derivative from scratch, on a restricted domain. Let f(x)=sin⁡xf(x) = \sin x restricted to I=(−π/2,π/2)I = (-\pi/2, \pi/2), so f−1=arcsin⁡f^{-1} = \arcsin on J=(−1,1)J = (-1,1). We verify differentiability of arcsin⁡\arcsin at b=32b = \tfrac{\sqrt3}{2} and compute the derivative there directly from the theorem, without assuming the usual arcsine derivative formula in advance. Check hypotheses: f′(x)=cos⁡xf'(x) = \cos x, and cos⁡x>0\cos x > 0 for all x∈(−π/2,π/2)x \in (-\pi/2,\pi/2), so by the Corollary ff is strictly increasing on II; ff is differentiable everywhere (standard fact about sin⁡\sin), continuous, and f′(x)≠0f'(x) \neq 0 on II. All hypotheses (H1)–(H4) hold at every a∈Ia \in I. We need a∈(−π/2,π/2)a \in (-\pi/2,\pi/2) with sin⁡a=32\sin a = \tfrac{\sqrt3}{2}: this is a=π/3a = \pi/3. Then

f′(a)=cos⁡(π/3)=12.f'(a) = \cos(\pi/3) = \frac{1}{2}.

By the theorem,

(f−1)′ ⁣(32)=arcsin⁡′ ⁣(32)=1f′(π/3)=11/2=2.(f^{-1})'\!\left(\tfrac{\sqrt3}{2}\right) = \arcsin'\!\left(\tfrac{\sqrt3}{2}\right) = \frac{1}{f'(\pi/3)} = \frac{1}{1/2} = 2.

As a sanity check against the closed-form identity arcsin⁡′(y)=1/1−y2\arcsin'(y) = 1/\sqrt{1-y^2} (itself a corollary of this same theorem applied at a general point, using cos⁡(arcsin⁡y)=1−y2\cos(\arcsin y) = \sqrt{1-y^2} for y∈(−1,1)y \in (-1,1)): with y=3/2y = \sqrt3/2, 1−y2=1−3/4=1/41 - y^2 = 1 - 3/4 = 1/4, so 1/1/4=21/\sqrt{1/4} = 2. The two computations agree.

CounterexampleWhy f′(a)≠0f'(a) \neq 0 cannot be dropped

Consider f(x)=x3f(x) = x^3 on I=RI = \mathbb{R}, and take a=0a = 0, b=f(0)=0b = f(0) = 0. Hypotheses (H1)–(H3) all hold: ff is continuous on R\mathbb{R}; ff is strictly increasing on R\mathbb{R} (if x1<x2x_1 < x_2 then x23−x13=(x2−x1)(x22+x1x2+x12)>0x_2^3 - x_1^3 = (x_2-x_1)(x_2^2+x_1x_2+x_1^2) > 0 since the second factor is nonnegative and vanishes only when x1=x2=0x_1=x_2=0, so strictly positive whenever x1≠x2x_1 \ne x_2 except possibly needing separate care — direct check: cubing is strictly increasing on R\mathbb{R} by elementary inequalities); and ff is differentiable everywhere with f′(x)=3x2f'(x) = 3x^2. But hypothesis (H4) fails at a=0a=0: f′(0)=0f'(0) = 0. The inverse function is f−1(y)=y1/3f^{-1}(y) = y^{1/3}, defined and continuous on all of R=J\mathbb{R} = J. We test differentiability of f−1f^{-1} at b=0b=0 directly from the definition:

lim⁡y→0f−1(y)−f−1(0)y−0=lim⁡y→0y1/3y=lim⁡y→01y2/3=+∞.\lim_{y \to 0} \frac{f^{-1}(y) - f^{-1}(0)}{y - 0} = \lim_{y\to 0} \frac{y^{1/3}}{y} = \lim_{y \to 0} \frac{1}{y^{2/3}} = +\infty.

The limit does not exist as a finite real number — it diverges to +∞+\infty. So f−1(y)=y1/3f^{-1}(y) = y^{1/3} is not differentiable at y=0y=0, even though ff was continuous, strictly monotonic, and differentiable at the corresponding point a=0a=0. Geometrically this is exactly the "horizontal tangent reflects to vertical tangent" picture from the Intuition section: ff has a horizontal tangent at the origin, and f−1f^{-1} has a vertical tangent there, which is incompatible with (finite) differentiability. This shows hypothesis (H4) is not a technical convenience — it is essential, and no weaker replacement using only (H1)–(H3) can repair the theorem's conclusion (ii).

Scope, edge cases & common traps

SituationWhat happens
f′(a)≠0f'(a) \neq 0, all other hypotheses holdf−1f^{-1} differentiable at b=f(a)b=f(a), formula holds exactly as stated
f′(a)=0f'(a) = 0f−1f^{-1} may fail to be differentiable at bb (vertical tangent); see counterexample
ff merely injective, not monotonic, on a general domainf−1f^{-1} may not even be continuous; the theorem requires strict monotonicity on an interval
II not open, aa an endpointUse one-sided derivatives throughout; conclusion (ii) becomes a one-sided differentiability statement
ff differentiable at aa only, not nearbyConclusion (ii) still holds at the single point b=f(a)b=f(a); the theorem does not require differentiability of ff near aa, only continuity + monotonicity near aa and differentiability at aa
f′(x)≠0f' (x)\ne0 for all xx but ff not assumed monotonic in advanceAutomatic by the Corollary (Darboux + sign-of-derivative), so this case reduces to the main theorem

The mathematician behind it

Origins · 1 minAugustin-Louis Cauchy and how this was discovered
AC
Augustin-Louis Cauchy
1789–1857 · Early 19th-century Paris, foundational rigor in analysis

Cauchy was among the first to insist that calculus be built on a precise theory of limits rather than intuitive appeals to infinitesimals or geometric plausibility. His lecture notes at the École Polytechnique — the Cours d'Analyse (1821) and later works on differential calculus — gave the first widely circulated rigorous treatments of continuity, the derivative as a limit of difference quotients, and consequences such as the behavior of inverse functions. The style of proof used above — reducing a derivative claim to a sequential ε\varepsilon-δ\delta limit computation, with every algebraic step justified by a previously proved limit law — is squarely in the tradition Cauchy established and that Weierstrass later sharpened into the fully modern ε\varepsilon-δ\delta formalism.

History
- 1821 — Cauchy's Cours d'Analyse gives a rigorous limit-based definition of continuity and derivative, the substrate on which the inverse function computation depends.
- 1837 — Dirichlet clarifies the modern function concept, sharpening what it means for f−1f^{-1} to be "a function" rather than a multivalued relation.
- 1861–1872 — Weierstrass's Berlin lectures formalize the ε\varepsilon-δ\delta definition of limit used to make the sequential argument in Step 5 fully airtight.
- 1872 — Darboux proves that derivatives satisfy the intermediate value property, supplying the tool used in the Corollary to derive monotonicity from f′≠0f'\neq 0.
- 1900s–present — The one-variable theorem becomes the template for the multivariable Inverse Function Theorem (Fréchet derivatives, Banach spaces), taught in every advanced real and functional analysis course.
Why isn't it enough to just differentiate f(f−1(y))=yf(f^{-1}(y)) = y using the Chain Rule?

That computation presupposes f−1f^{-1} is differentiable at yy — you cannot legally apply the Chain Rule to a function you have not yet shown is differentiable. The genuine content of the Inverse Function Theorem is proving differentiability of f−1f^{-1} from the definition of the derivative as a limit of difference quotients, which is what Step 5 of the proof accomplishes. Once differentiability of f−1f^{-1} is established independently, the Chain-Rule computation becomes a valid shortcut for re-deriving the formula, but it cannot serve as the proof of existence.

What happens at points where f'(a) = 0?

The inverse function can fail to be differentiable there, typically because the graph of f−1f^{-1} has a vertical tangent line at the corresponding point. The standard example is f(x)=x3f(x)=x^3 at a=0a=0: f−1(y)=y1/3f^{-1}(y)=y^{1/3} has an infinite one-sided (and two-sided) derivative at 00, meaning the limit defining (f−1)′(0)(f^{-1})'(0) diverges to +∞+\infty rather than existing as a finite number. So the conclusion of the theorem genuinely requires f′(a)≠0f'(a)\ne 0; it is not merely a simplifying convenience.

Do we need f differentiable on a whole neighborhood of a, or just at the single point a?

Only at the single point aa, together with continuity and strict monotonicity of ff on an open interval around aa. The proof only ever uses differentiability of ff at the one point aa to evaluate the limit of the difference quotient there; the surrounding hypotheses (continuity, monotonicity) are used only to guarantee that f−1f^{-1} is well-defined and continuous near bb, which in turn is needed to show the auxiliary sequence xn→ax_n \to a.

How does this relate to the multivariable Inverse Function Theorem?

The multivariable version replaces the scalar f′(a)f'(a) with the Jacobian matrix Df(a)Df(a), and the hypothesis f′(a)≠0f'(a) \ne 0 becomes invertibility of that matrix. The conclusion becomes D(f−1)(b)=[Df(a)]−1D(f^{-1})(b) = [Df(a)]^{-1}, a direct matrix analogue of 1/f′(a)1/f'(a). The one-variable proof by explicit sequential limits does not generalize directly — the higher-dimensional proof instead uses a contraction mapping / fixed-point argument — but the one-variable case is the essential intuition-building base case and the formula's shape is identical.

Is strict monotonicity really necessary, or would plain injectivity on an interval suffice?

On an interval, continuity plus injectivity actually forces strict monotonicity — this is a classical fact (a continuous injective function on an interval cannot fail to be monotonic, else the Intermediate Value Theorem would produce a contradiction to injectivity). So assuming continuity, requiring strict monotonicity is not an extra restriction beyond injectivity; it is automatic. The explicit strict monotonicity hypothesis is stated mainly to make the proof of continuity of f−1f^{-1} transparent and self-contained.

Can (f−1)′(b)(f^{-1})'(b) ever equal f′(a)f'(a) instead of its reciprocal — is there a common sign error to watch for?

The formula is always the reciprocal, 1/f′(a)1/f'(a), never f′(a)f'(a) itself, and this is exactly what the reflection-across-y=xy=x picture predicts: slope mm becomes slope 1/m1/m. A common computational slip is to forget to invert and to write (f−1)′(b)=f′(a)(f^{-1})'(b) = f'(a) directly; always sanity-check by confirming the sign and rough magnitude match the reciprocal relationship, e.g. a steep tangent for ff should correspond to a shallow tangent for f−1f^{-1} at the reflected point.

Does the theorem require ff to be differentiable everywhere on I, or can hypotheses be checked pointwise?

The full "global" formula (f−1)′(y)=1/f′(f−1(y))(f^{-1})'(y) = 1/f'(f^{-1}(y)) for all y∈Jy \in J does require ff differentiable with f′≠0f'\ne 0 at every point of II. But the "local" version of the theorem, as stated, only needs continuity and strict monotonicity of ff on an interval around aa, plus differentiability with f′(a)≠0f'(a)\ne0 at the single point aa; this yields differentiability of f−1f^{-1} at the single corresponding point b=f(a)b=f(a). The global formula is simply this local statement applied at every point simultaneously.

Before this