Two Different Kinds of Control
Uniform continuity and absolute continuity both concern increments of a function over short intervals, but they ask for different kinds of control. Uniform continuity makes the increment over each sufficiently short interval small. Absolute continuity controls the sum of increments over any finite collection of intervals whose total length is small. The distinction matters: many short intervals can have a large combined increment even when every individual increment is tiny.
Throughout, functions are real-valued and defined on a compact interval \([a,b]\subset\mathbb{R}\), where \(a<b\). For \(u,v\in[a,b]\), uniform continuity means that for every \(\varepsilon>0\), there is a \(\delta>0\) such that \(|u-v|<\delta\) implies \(|f(u)-f(v)|<\varepsilon\). The same \(\delta\) must work for every pair \(u,v\) in the interval.
The defining distinction is visible in the quantifiers. Uniform continuity tests one interval at a time. Absolute continuity tests all finite families at once, using a single bound on their total length. In particular, absolute continuity implies continuity, but its definition imposes more than continuity at each point.
Absolute Continuity Implies Uniform Continuity
Proof. Let \(\varepsilon>0\). Apply the definition of absolute continuity with this \(\varepsilon\), obtaining \(\delta>0\). If \(u,v\in[a,b]\) and \(|u-v|<\delta\), consider the single interval \([\min(u,v),\max(u,v)]\). Its length is \(|u-v|<\delta\), so the definition gives \(|f(u)-f(v)|<\varepsilon\). This is the definition of uniform continuity, with the same \(\delta\) working for every \(u,v\). \(\square\)
On a compact interval, continuity alone also implies uniform continuity. This fact does not imply absolute continuity. For completeness, the compact-interval argument is as follows. Given a continuous \(f\) and \(\varepsilon>0\), for every \(x\in[a,b]\) choose \(r_x>0\) so that \(|y-x|<2r_x\) implies \(|f(y)-f(x)|<\varepsilon/2\). The relative neighborhoods \((x-r_x,x+r_x)\cap[a,b]\) cover \([a,b]\). Compactness gives a finite subcover, with centers \(x_1,\ldots,x_k\). Set \(\delta=\min_i r_{x_i}>0\). If \(|u-v|<\delta\), choose a covering neighborhood containing \(u\), say the one centered at \(x_i\). Then \(|u-x_i|<r_{x_i}\), and \[ |v-x_i|\le |v-u|+|u-x_i|<\delta+r_{x_i}\le 2r_{x_i}. \] Both \(f(u)\) and \(f(v)\) differ from \(f(x_i)\) by less than \(\varepsilon/2\), so \(|f(u)-f(v)|<\varepsilon\). Thus continuous functions on compact intervals are uniformly continuous. The examples below show why this compactness result cannot be strengthened to absolute continuity.
Worked Example: The Identity Function and Total Increment
Let \(f(x)=x\) on \([0,1]\). For any finite family of intervals \([a_j,b_j]\subseteq[0,1]\) with pairwise disjoint interiors, each increment is \(|f(b_j)-f(a_j)|=|b_j-a_j|=b_j-a_j\). Therefore \[ \sum_{j=1}^{N}|f(b_j)-f(a_j)| =\sum_{j=1}^{N}(b_j-a_j). \] For any \(\varepsilon>0\), choose \(\delta=\varepsilon\). If the total interval length is less than \(\delta\), the total increment is the same sum and is also less than \(\varepsilon\). Hence the identity function is absolutely continuous. It is uniformly continuous as well, since for any \(u,v\), \(|f(u)-f(v)|=|u-v|\).
Check: For the intervals \([0,\tfrac14]\) and \([\tfrac12,\tfrac34]\), the total length is \(\tfrac14+\tfrac14=\tfrac12\), and the total increment is also \(\tfrac14+\tfrac14=\tfrac12\). This agrees with the formula for every finite family, not just this one.
A Uniformly Continuous Function That Is Not Absolutely Continuous
The Cantor function gives a direct comparison of individual increments with their sum. Begin with \([0,1]\), and at each stage remove the open middle third of every interval remaining from the preceding stage. At stage \(n\), there are \(2^n\) remaining closed intervals, each of length \(3^{-n}\). Define the Cantor function \(C\) by the usual continuous, nondecreasing construction: it is constant on each removed open interval, has \(C(0)=0\) and \(C(1)=1\), and on each stage-\(n\) remaining interval its endpoint values differ by \(2^{-n}\).
These properties can be obtained from continuous approximations. At stage \(n\), define \(C_n\) to be linear on each remaining interval, taking endpoint values that increase in order from \(0\) to \(1\) in increments of \(2^{-n}\); on each removed gap, let \(C_n\) be constant at the adjacent endpoint value. The endpoint assignments are consistent from stage to stage. On a stage-\(n\) interval, all later approximations take values between that interval's two assigned endpoint values, whose difference is \(2^{-n}\); on a gap already removed, later approximations remain constant. Consequently \(\sup_{x\in[0,1]}|C_m(x)-C_n(x)|\le 2^{-n}\) whenever \(m\ge n\). Thus the approximations converge uniformly. Their limit \(C\) is continuous: for any \(\varepsilon>0\), choose \(n\) so that \(2^{-n}<\varepsilon/3\), then use continuity of \(C_n\) to make its increment less than \(\varepsilon/3\); the two approximation errors and this increment sum to less than \(\varepsilon\). Since \([0,1]\) is compact, the compact-interval argument above shows that \(C\) is uniformly continuous.
Worked Example: The Cantor Function Fails the Absolute-Continuity Test
At stage \(n\), write the remaining intervals as \(I_{n,1},\ldots,I_{n,2^n}\). Each has length \(3^{-n}\), and the endpoint increment of \(C\) across each is \(2^{-n}\). These intervals have pairwise disjoint interiors, so their total length and total increment are respectively \[ \sum_{k=1}^{2^n}|I_{n,k}|=2^n3^{-n}=\left(\frac23\right)^n, \qquad \sum_{k=1}^{2^n}|\Delta C(I_{n,k})|=2^n2^{-n}=1. \] Here \(|\Delta C([u,v])|\) denotes \(|C(v)-C(u)|\).
Fix \(\varepsilon=\tfrac12\). For every proposed \(\delta>0\), choose \(n\) sufficiently large that \((2/3)^n<\delta\). The stage-\(n\) intervals then have total length less than \(\delta\), but their total increment is exactly \(1\), which is not less than \(\tfrac12\). No \(\delta\) works for this \(\varepsilon\), so \(C\) is not absolutely continuous. Yet each individual increment at stage \(n\) is \(2^{-n}\), which tends to zero. The failure is entirely in the sum: there are \(2^n\) such increments, and their sum remains \(1\).
Check: At stage \(2\), there are \(2^2=4\) intervals, each of length \(3^{-2}=\tfrac19\) and increment \(2^{-2}=\tfrac14\). Their total length is \(4\cdot\tfrac19=\tfrac49\), and their total increment is \(4\cdot\tfrac14=1\), as the general formulas predict.
Oscillations Near an Endpoint Give Another Failure
A second example shows that the failure is not specific to a monotone singular function. Define a real-valued function on \([0,1]\) by \[ f(0)=0,\qquad f(x)=x\sin(1/x)\quad\text{for }0<x\le1. \] For \(x>0\) it is continuous. At zero, \(|f(x)|\le x\), so \(f(x)\to0=f(0)\); hence \(f\) is continuous on \([0,1]\) and therefore uniformly continuous. The following intervals show directly that it is not absolutely continuous.
Worked Example: Short Intervals with Large Total Oscillation
For integers \(n\ge1\), set \(t_n=1/(\pi(n+\tfrac12))\). Then \(t_n\) decreases to zero, and \[ \sin(1/t_n)=\sin\bigl(\pi(n+\tfrac12)\bigr)=(-1)^n, \qquad f(t_n)=(-1)^n t_n. \] The intervals \([t_{n+1},t_n]\) have pairwise disjoint interiors. On each one, the endpoint increment has magnitude \[ |f(t_n)-f(t_{n+1})| =|(-1)^nt_n-(-1)^{n+1}t_{n+1}| =t_n+t_{n+1}. \] For integers \(M\ge N\), their total length is \[ \sum_{n=N}^{M}(t_n-t_{n+1})=t_N-t_{M+1}<t_N. \] Their total increment is \(\sum_{n=N}^{M}(t_n+t_{n+1})\). This sum can be made arbitrarily large by increasing \(M\): indeed, \(t_n=1/(\pi(n+\tfrac12))\ge1/(2\pi n)\) for \(n\ge1\), and the harmonic series \(\sum_{n\ge1}1/n\) diverges.
Now fix \(\varepsilon=1\). Given any \(\delta>0\), choose \(N\) such that \(t_N<\delta\). Since the total increments grow without bound as \(M\) increases, choose \(M\ge N\) such that \(\sum_{n=N}^{M}(t_n+t_{n+1})\ge1\). The selected intervals have total length less than \(\delta\), but their total increment is at least \(1\), not less than \(\varepsilon\). Thus \(f\) is not absolutely continuous. The length calculation and increment calculation are different: the intervals fit inside a short segment near zero, while the function oscillates often enough there for the sum of endpoint changes to remain large.
Check: The first three points are \(t_1=2/(3\pi)\), \(t_2=2/(5\pi)\), and \(t_3=2/(7\pi)\). Their function values are \(-2/(3\pi)\), \(2/(5\pi)\), and \(-2/(7\pi)\). Across the first two intervals, the increment sum is \[ \left(\frac{2}{3\pi}+\frac{2}{5\pi}\right) +\left(\frac{2}{5\pi}+\frac{2}{7\pi}\right) =\frac{4}{3\pi}+\frac{4}{5\pi}+\frac{2}{7\pi}, \] which matches the sum of the two endpoint-change magnitudes computed separately.
The Common Mistake: Adding Individual Bounds
A tempting but invalid argument is: “Uniform continuity makes each increment less than \(\varepsilon\), so their sum is less than \(\varepsilon\).” If there are \(N\) intervals, the bound supplied by that reasoning is only \(N\varepsilon\), not \(\varepsilon\). Moreover, \(N\) is not bounded in the absolute-continuity definition. The Cantor example makes the error concrete: at stage \(n\), every increment is \(2^{-n}\), yet there are \(2^n\) intervals, giving total increment \(2^n\cdot2^{-n}=1\). Smallness of every term does not make the sum small when the number of terms can grow.
Absolute continuity supplies what uniform continuity does not: for a chosen target \(\varepsilon\), it chooses a length threshold \(\delta\) that controls the entire sum over every finite family of intervals. The examples also identify the precise failed estimates. For the Cantor function, total interval length \((2/3)^n\) tends to zero while total increment stays equal to \(1\). For \(x\sin(1/x)\), intervals of total length less than any prescribed \(\delta\) can have total increment at least \(1\). Thus uniform continuity is a necessary consequence of absolute continuity on a compact interval, but it is not a substitute for the increment-sum condition.
Check Your Understanding
For each question, distinguish a bound on one increment from a bound on the sum over a finite family.
- Use the definition of absolute continuity to prove that an absolutely continuous function on \([a,b]\) is uniformly continuous.
- At stage \(n\) of the Cantor construction, how many intervals are used, what is their total length, and what is the sum of their endpoint increments?
- Why does the fact that each Cantor-function increment is \(2^{-n}\) not prove that the function is absolutely continuous?
- For \(t_n=1/(\pi(n+\tfrac12))\), compute the length of the union, up to endpoints, of \([t_{n+1},t_n]\) for \(n=N,\ldots,M\).
- In the absolute-continuity definition, why is a bound on the number of intervals unavailable?