An improper integral converges absolutely when the integral of converges, and converges conditionally when it converges only because positive and negative pieces cancel. Absolute convergence is the stronger of the two: it always implies convergence, and it is the version you can test with the comparison tests, which require a nonnegative integrand.
| Item | Statement |
|---|---|
| Converges absolutely | converges |
| Converges conditionally | converges but diverges |
| Main theorem | absolute convergence convergence, never the reverse |
| Engine of the proof | |
| Standard absolute test | find with and convergent |
| Model absolutely convergent | |
| Model conditionally convergent | (value ) |
| The whole family | : absolute for , conditional for , divergent for |
| Block fact used constantly | |
| Tool for conditional convergence | integration by parts, which trades for an absolutely convergent tail |
| Comparison tests need | , so apply them to first |
| Pitfall | diverges at , not at |
| Pitfall | convergence of does not force |
| Useful corollary | absolutely convergent and bounded absolutely convergent |
The Two Kinds of Convergence
The comparison tests of the previous section require the integrand to be nonnegative. When changes sign, we distinguish two kinds of convergence: one where the integral of also converges, and one where convergence relies on cancellation between the positive and the negative parts.
Let be integrable on every bounded subinterval of .
- The integral converges absolutely if converges.
- The integral converges conditionally if converges but diverges.
Read the two clauses as a pair of independent questions about the same integrand: does converge, and does converge? Because , its integral either converges to a finite number or increases to ; there is no oscillation to worry about. The two answers together produce exactly three possibilities, and the next theorem rules out a fourth.
The picture below is the example to keep in mind for the rest of the section. The graph of crosses the axis at every multiple of , so the area it encloses splits into lobes that alternate in sign. Write for the (unsigned) area of the th lobe.
The signed areas of the lobes of on alternate, and they shrink, but only about as fast as . Numerically , , , , , . Adding the unsigned areas gives ; adding the signed areas gives .
The same two definitions apply verbatim to Type II improper integrals (those with an infinite discontinuity at an endpoint) and to integrals over : in every case, replace by and ask the same question.
Absolute Convergence Implies Convergence
The key fact is that absolute convergence is the stronger condition. It is stronger in the strict sense: it implies convergence, and there are convergent integrals that are not absolutely convergent.
Absolute Convergence Implies Convergence. If converges, then converges.
In words: if the total unsigned area is finite, then the signed area is finite too. Cancellation can only help.
Proof
Since , adding throughout gives
The point of this step is that is a nonnegative function, so the Direct Comparison Test applies to it even though it does not apply to .
Because converges, the Direct Comparison Test gives convergence of . Writing and using linearity:
\begin{aligned} \int_a^\infty f(x)\,dx &= \int_a^\infty \bigl(f(x)+|f(x)|\bigr)\,dx \\ &\quad - \int_a^\infty |f(x)|\,dx, \end{aligned}a difference of two convergent integrals. Therefore converges.
Two consequences are worth stating separately.
- The contrapositive. If diverges, then diverges. In other words, a divergent integral can never be absolutely convergent.
- The converse is false. Divergence of tells you nothing about ; the integral may still converge conditionally, as does. This is the single most common misreading of the theorem.
A convenient bound comes free with the theorem: when converges,
which is the improper-integral version of the triangle inequality. It often gives a numerical estimate at no extra cost.
Worked Examples
The first example is the standard pattern for absolute convergence: bound by a -integral. The second and third are the standard pattern for conditional convergence: get convergence from integration by parts, then get divergence of from a block-by-block comparison with the harmonic series.
Show that converges absolutely.
Solution
We have
because for every . Since converges (the -test with ), the Direct Comparison Test gives convergence of . By definition, the integral converges absolutely.
By the theorem above, the original integral therefore converges as well, and the triangle inequality bounds its value: .
Show that converges conditionally.
Solution
That the integral converges was established with the comparison tests: the singularity at is removable because as , and integration by parts on turns the tail into an absolutely convergent integral. See the worked example on the Dirichlet integral. It remains to show that diverges.
On each interval , for , the function traces one full arch, so
independently of . Also on this interval, since never exceeds there. Replacing by that smallest value can only decrease the integral:
\begin{aligned} \int_{n\pi}^{(n+1)\pi}\frac{|\sin x|}{x}\,dx &\ge \frac{1}{(n+1)\pi}\int_{n\pi}^{(n+1)\pi}|\sin x|\,dx \\ &= \frac{2}{(n+1)\pi}. \end{aligned}The picture below is exactly this estimate: each shaded arch is the graph of flattened down to the constant height , and it sits under the true curve.
Summing the block estimate over gives a lower bound for the integral over :
\bbox[8px, #E6F0FA, border: 3px solid #0066CC]{\int_\pi^{(N+1)\pi}\frac{|\sin x|}{x}\,dx \ \ge\ \frac{2}{\pi}\sum_{n=1}^{N}\frac{1}{n+1}}The sum on the right is a tail of the harmonic series, which diverges, so the right side tends to as . Therefore diverges, and the convergence of is conditional.
The next example runs the same two-step argument from scratch, with the integration by parts written out in full. It also shows that nothing about the method depends on the decay being a power of .
Show that converges conditionally.
Solution
(i) Convergence. Integrate by parts with and , so that
Then for every ,
\begin{aligned} \int_2^b \frac{\sin x}{\ln x}\,dx &= \left[-\frac{\cos x}{\ln x}\right]_2^b \\ &\quad - \int_2^b \frac{\cos x}{x(\ln x)^{2}}\,dx. \end{aligned}The boundary term is , and as , so it tends to .
The remaining integral converges absolutely, because
Both pieces have limits, so converges. Integration by parts did the essential work: it converted an integrand decaying like into one decaying like , which is small enough to be handled by comparison.
(ii) Non-absolute convergence. Since , every interval with lies inside . On such an interval , so , and
\begin{aligned} \int_{n\pi}^{(n+1)\pi}\frac{|\sin x|}{\ln x}\,dx &\ge \frac{1}{\ln\bigl((n+1)\pi\bigr)}\int_{n\pi}^{(n+1)\pi}|\sin x|\,dx \\ &= \frac{2}{\ln\bigl((n+1)\pi\bigr)} \\ &\ge \frac{2}{(n+1)\pi}, \end{aligned}where the last step uses . Summing over again produces a multiple of a partial sum of the harmonic series, which grows without bound. Hence diverges and the convergence is conditional.
Why Conditional Convergence Depends on Cancellation
It is worth seeing why the comparison tests cannot possibly settle . The natural bound is
and the graph of really does live inside the envelope formed by and , touching it near each peak and trough.
The curve trapped between on . The envelope controls the size of the oscillations but is itself not integrable.
The trouble is that the envelope is , and diverges (the -test with ). A divergent upper bound gives no information, and part (b) of the Direct Comparison Test does not apply either, because is not bounded below by any fixed multiple of (it vanishes at every multiple of ). So no comparison against the envelope can decide the question, in either direction. Something finer is needed, and that finer thing is cancellation.
Here is the accounting, in terms of the lobe areas from the first figure. The unsigned total is
because the block estimate shows and the harmonic series diverges. The signed total is the alternating sum
which converges because the decrease steadily to . This is precisely the situation of the alternating harmonic series , which converges while does not. The integral is called the Dirichlet integral; its exact value requires techniques beyond this text, but the convergence is elementary.
Why does the distinction matter in practice? Because an absolutely convergent integral behaves like a proper integral: its value does not depend on the order in which you accumulate area, and estimates such as are available. A conditionally convergent integral has none of that stability. Its value is the limit of a delicately balanced cancellation, and small changes can destroy it. For instance, multiplying by the bounded function gives
whose integral over diverges, since diverges while converges. Multiplying an absolutely convergent integrand by a bounded function never causes such a collapse, as you will prove in the exercises.
A Strategy for Integrands That Change Sign
The following procedure, the Absolute-First Strategy, is what to do when you meet an integrand that is not of one sign. It is ordered so that the cheapest test comes first.
- Locate every improper point. Check each endpoint and every interior point where the integrand blows up, and split the integral so that each piece is improper at one point only. A single divergent piece makes the whole integral divergent.
- Test first. On each piece, apply the Direct Comparison Test or the Limit Comparison Test to , which is nonnegative and therefore legal input for those tests. The usual comparison functions are the -integrals and the exponentials .
- If converges, stop. The integral converges absolutely, hence converges, and gives you a bound for free.
- If diverges, conclude nothing yet. Return to itself and try integration by parts, taking to be the oscillating factor ( or ) and the slowly decaying factor. If the boundary term has a limit and the new integral converges absolutely, the original integral converges conditionally.
- If integration by parts fails, look for divergence. The standard argument: find intervals marching off to infinity on which keeps one sign and for some fixed . If converged to , these increments would have to tend to , so no such could exist.
Two traps.
- Forgetting the origin. The integral has a perfectly well behaved tail, since . It still diverges, because near the integrand behaves like . Step 1 exists precisely to catch this.
- Assuming the integrand must vanish. For series, convergent forces . For integrals there is no such rule: converges although keeps swinging between and forever. Convergence of an improper integral is a statement about accumulated area, not about the size of the integrand.
Five Integrals Classified
The table applies the strategy to five standard integrands. Read each row left to right: the two middle columns are the two independent questions, and the verdict is determined by the pair.
| Integral | Does converge? | Does converge? | Verdict |
|---|---|---|---|
| yes | no | converges conditionally | |
| yes | yes | converges absolutely | |
| yes | no | converges conditionally | |
| yes | yes | converges absolutely | |
| yes | yes | converges absolutely |
Notice that the answer "no" in the third column never by itself settles anything: rows one and three are conditional, but an integral such as also has divergent and is itself divergent. The fourth logical combination, convergent while diverges, never appears: it is exactly what the theorem of this section forbids.
Summary. An absolutely convergent improper integral behaves like a proper integral: the value is not affected by the order of integration or by cancellation between positive and negative regions. A conditionally convergent integral converges only because of cancellation, and it diverges if the integrand is replaced by its absolute value. When applying comparison tests, check whether ; if not, consider first whether converges (absolute convergence) before working with directly.
Exercises
Determine whether converges absolutely, conditionally, or diverges.
Answer
Converges absolutely.
Solution
We have and converges (the -test with ). By the Direct Comparison Test, converges. Since absolute convergence implies convergence, the integral converges, and it converges absolutely.
Determine whether converges absolutely, conditionally, or diverges.
Answer
Converges conditionally.
Solution
(i) Convergence. We use integration by parts with and , giving and . Then:
\begin{aligned} \int_1^b \frac{\sin x}{\sqrt{x}}\,dx &= \Bigl[-\frac{\cos x}{\sqrt{x}}\Bigr]_1^b \\ &\quad - \int_1^b \frac{\cos x}{2x^{3/2}}\,dx. \end{aligned}As : (since ), and the remaining integral converges absolutely since and converges. Therefore the integral converges.
(ii) Non-absolute convergence. On we have , so:
\begin{aligned} \int_{n\pi}^{(n+1)\pi}\frac{|\sin x|}{\sqrt{x}}\,dx &\ge \frac{1}{\sqrt{(n+1)\pi}}\int_{n\pi}^{(n+1)\pi}|\sin x|\,dx \\ &= \frac{2}{\sqrt{(n+1)\pi}}. \end{aligned}Since diverges (comparison with the harmonic series via , or more directly since diverges), we conclude that diverges. Therefore the convergence is conditional.
Show that converges absolutely, and evaluate it.
Answer
.
Solution
Since and converges, the Direct Comparison Test gives convergence of . Therefore the integral converges absolutely.
To evaluate the integral, we integrate by parts twice. Let .
First integration by parts. Let and , so and :
\begin{aligned} \int_0^b e^{-x}\sin x\,dx &= \bigl[-e^{-x}\sin x\bigr]_0^b + \int_0^b e^{-x}\cos x\,dx \\ &= -e^{-b}\sin b + \int_0^b e^{-x}\cos x\,dx. \end{aligned}Second integration by parts. Let and , so and :
\begin{aligned} \int_0^b e^{-x}\cos x\,dx &= \bigl[-e^{-x}\cos x\bigr]_0^b - \int_0^b e^{-x}\sin x\,dx \\ &= (-e^{-b}\cos b + 1) - \int_0^b e^{-x}\sin x\,dx. \end{aligned}Solving for . Substituting back:
Adding to both sides:
Since as , taking the limit gives , so .
Determine whether converges absolutely, conditionally, or diverges.
Answer
Converges conditionally.
Solution
(i) Convergence. We use integration by parts with and , so and . Then:
\begin{aligned} \int_1^b \frac{\cos(\pi x)}{x}\,dx &= \Bigl[\frac{\sin(\pi x)}{\pi x}\Bigr]_1^b \\ &\quad + \int_1^b \frac{\sin(\pi x)}{\pi x^{2}}\,dx. \end{aligned}As : , and since , so the boundary term tends to . For the remaining integral, and converges, so by the Direct Comparison Test the remaining integral converges absolutely. Therefore converges.
(ii) Non-absolute convergence. The function has period , so on each interval with it completes exactly half a period and traces one full arch. Hence , the same value for every . Also on . Therefore:
Summing over :
Since the harmonic series diverges, diverges. The convergence of is therefore conditional.
Determine whether converges absolutely, conditionally, or diverges. If it converges, give a numerical bound for its value.
Answer
Converges absolutely, and its value is at most in absolute value.
Solution
The integrand changes sign, so we test first. Since and for ,
Since converges (the -test with ), the Direct Comparison Test gives convergence of . The integral therefore converges absolutely, and by the theorem of this section it converges.
For the bound, use the triangle inequality together with the sharper comparison function :
\begin{aligned} \left|\int_1^\infty \frac{\cos x}{x^{2}+1}\,dx\right| &\le \int_1^\infty \frac{dx}{x^{2}+1} \\ &= \lim_{b\to\infty}\bigl[\arctan x\bigr]_1^b \\ &= \frac{\pi}{2}-\frac{\pi}{4} \\ &= \frac{\pi}{4}. \end{aligned}This is a typical dividend of absolute convergence: the same comparison that proves convergence also produces an error bound.
Show that converges absolutely, and evaluate it.
Answer
.
Solution
Absolute convergence. Since we have for , and
converges. By the Direct Comparison Test, converges, so the integral converges absolutely.
Evaluation. Let . Integrate by parts with , , so and :
\begin{aligned} \int_0^b e^{-x}\cos x\,dx &= \bigl[-e^{-x}\cos x\bigr]_0^b - \int_0^b e^{-x}\sin x\,dx \\ &= \bigl(1-e^{-b}\cos b\bigr) - \int_0^b e^{-x}\sin x\,dx. \end{aligned}Now integrate the remaining integral by parts with , , so and :
\begin{aligned} \int_0^b e^{-x}\sin x\,dx &= \bigl[-e^{-x}\sin x\bigr]_0^b + \int_0^b e^{-x}\cos x\,dx \\ &= -e^{-b}\sin b + \int_0^b e^{-x}\cos x\,dx. \end{aligned}Substituting the second computation into the first:
\begin{aligned} \int_0^b e^{-x}\cos x\,dx &= 1-e^{-b}\cos b + e^{-b}\sin b \\ &\quad - \int_0^b e^{-x}\cos x\,dx. \end{aligned}Moving the last integral to the left side gives
Since as , we get , so .
Note that this is the same value as from an earlier exercise; the two integrals are equal, although the functions are not.
Show that converges, and decide whether the convergence is absolute or conditional. Comment on why the result is surprising.
Answer
Converges conditionally. It is surprising because does not tend to as .
Solution
Substitution. On any finite interval , put . Then , and since we have . The limits become and , so
As we also have , so the left side has a limit exactly when the right side does:
Convergence. The integral was shown to converge in the exercise on sine over the square root, by integration by parts. Hence converges.
Not absolutely. The same substitution applied to gives
and part (ii) of that same exercise shows the right side is . So the convergence is conditional.
Why it is surprising. For an infinite series, convergence of forces . Here returns to whenever and to whenever , so it never settles down. What does happen is that the successive zeros of , namely , crowd together as grows: consecutive zeros are about apart. The lobes become narrow fast enough that their areas shrink, and the alternating signs finish the job. Numerically the integral equals about .
Let denote the greatest integer less than or equal to . Determine whether converges absolutely, conditionally, or diverges.
Answer
Converges conditionally.
Solution
Reduce to a series. For we have , so the integrand equals on that interval. The integrand has a jump at each integer but is bounded, so it is integrable on every bounded subinterval, and
Adding these for gives
Convergence. The terms are positive, strictly decreasing (because decreases), and tend to . By the Alternating Series Test the series converges, so the integrals over integer upper limits converge to some number . For a general upper limit with ,
so has the same limit as through all real values. The integral converges.
Not absolutely. Since for every ,
which diverges by the -test with . Therefore the convergence is conditional.
This integral is the exact analogue of the alternating harmonic series: it converges for precisely the same reason, and it fails to converge absolutely for precisely the same reason. Its value can be evaluated with Wallis's product and equals .
Determine whether converges or diverges. Be careful to check both ends of the interval.
Answer
Diverges, because of the singularity at . The tail by itself converges absolutely.
Solution
Step 1: split at the two improper points. The integral is improper at , where the integrand is unbounded, and at . Split at :
Step 2: the tail is harmless. For we have , and converges. By the Direct Comparison Test the tail converges absolutely.
Step 3: the origin is not. Near we have , so the integrand behaves like . Confirm this with the Limit Comparison Test using , which is legitimate because on , so the integrand is positive there:
Since lies in and diverges (the -test for Type II integrals with ), the piece diverges.
Conclusion. One divergent piece makes the whole integral divergent, so diverges.
Compare this with , where the extra factor of in the denominator is absent and the singularity at the origin is removable, since . Raising the exponent from to improves the behaviour at infinity and ruins it at the origin.
True or false: if is continuous on and converges, then . Justify your answer, and state a condition under which the conclusion does hold.
Answer
False. The conclusion does hold if, in addition, is monotone on , and also if is uniformly continuous on .
Solution
A first counterexample. By the exercise on , that integral converges, yet takes the value at and the value at for every . So has no limit at all as , let alone the limit .
A second counterexample, with spikes. Let be the function that equals on each interval for and equals elsewhere. (These intervals are disjoint, since .) On any bounded interval has finitely many discontinuities, so it is integrable there, and for every ,
The function is nondecreasing and bounded above, so it has a limit; the integral converges (to ). But for every integer , so . Replacing the rectangles by narrow continuous triangles of height and base makes the same example continuous.
When the conclusion is true. If is monotone on and converges, then : a monotone function has a limit at infinity, and would force the integral of a function eventually bounded away from to diverge. The same conclusion holds if is uniformly continuous on : uniform continuity stops the spikes from becoming arbitrarily narrow, since if for some sequence , then on an interval of one fixed length around each , and those intervals contribute the same positive amount of area infinitely often.
The moral is that convergence of an improper integral is a statement about accumulated area, not about the size of the integrand. This is the main way integrals differ from series.
Suppose converges absolutely and is integrable on every bounded subinterval of with for all . Prove that converges absolutely, and show by example that the statement fails if is only conditionally convergent.
Answer
The product integral converges absolutely, with . The statement fails for conditionally convergent : take and on .
Solution
Proof. The product is integrable on every bounded subinterval, being a product of two such functions. For every ,
and is nonnegative with convergent by hypothesis. Since , the Direct Comparison Test applies and gives convergence of . By the theorem of this section, converges, and it converges absolutely. Taking absolute values through the limit gives the bound
Failure for conditional convergence. Let , and , so is bounded and converges conditionally. Their product is
using the identity . Now converges (integrate by parts exactly as for ), while diverges. A divergent integral minus a convergent one is divergent, so diverges.
Multiplying by a bounded function therefore destroyed the convergence. This is the precise sense in which conditional convergence is fragile: it depends on cancellation, and a bounded factor can remove the cancellation.
Classify each integral as absolutely convergent, conditionally convergent, or divergent.
Answer
(a) converges conditionally. (b) converges absolutely. (c) diverges.
Solution
(a) First test . On the graph of runs from down through a zero at the midpoint to , so traces two half-arches whose areas together give , and there. Hence
and diverges (its terms exceed those of the harmonic series), so diverges. Absolute convergence fails.
Now test itself by parts, with and , so and :
\begin{aligned} \int_1^b \frac{\cos x}{x^{1/3}}\,dx &= \left[\frac{\sin x}{x^{1/3}}\right]_1^b \\ &\quad + \frac{1}{3}\int_1^b \frac{\sin x}{x^{4/3}}\,dx. \end{aligned}The boundary term tends to , since . The remaining integral converges absolutely because and . So the integral converges, and since diverges the convergence is conditional.
(b) Since for and ,
and converges. By the Direct Comparison Test the integral converges absolutely. (This is also the previous exercise applied with , which is absolutely convergent, and the bounded factor .)
(c) Here the oscillation grows instead of decaying. An antiderivative of is , as differentiating confirms:
Therefore
Taking gives for the first bracket, while gives . The partial integrals oscillate with unbounded amplitude, so no limit exists and the integral diverges.
Let be a real number. Classify
as absolutely convergent, conditionally convergent, or divergent, according to the value of .
Answer
Absolutely convergent for ; conditionally convergent for ; divergent for .
Solution
Case 1: , absolute convergence. Since ,
and converges by the -test because . The Direct Comparison Test gives convergence of , so the integral converges absolutely.
Case 2: , conditional convergence. Two things must be shown.
Convergence. Integrate by parts with and , so and :
\begin{aligned} \int_1^b \frac{\sin x}{x^{p}}\,dx &= \left[-\frac{\cos x}{x^{p}}\right]_1^b \\ &\quad - p\int_1^b \frac{\cos x}{x^{p+1}}\,dx. \end{aligned}Because , the boundary term satisfies , so it tends to . For the remaining integral, with , so converges absolutely. Both parts have limits, so the integral converges. This is the payoff of integration by parts: it raises the exponent from to , moving it across the threshold .
Non-absolute convergence. On we have , so
\begin{aligned} \int_{n\pi}^{(n+1)\pi}\frac{|\sin x|}{x^{p}}\,dx &\ge \frac{1}{\bigl((n+1)\pi\bigr)^{p}}\int_{n\pi}^{(n+1)\pi}|\sin x|\,dx \\ &= \frac{2}{\pi^{p}(n+1)^{p}}. \end{aligned}Summing over bounds below by , and the series diverges for . Hence diverges and the convergence is conditional.
Case 3: , divergence. Write , so the integrand is . On we have and (since and ), so
for every . Now suppose the integral converged, say as . Then the difference of the values at and would tend to . But that difference is exactly the integral above, which is at least for every . This contradiction shows the integral diverges.
Summary of the trichotomy. Reading the three cases together: the exponent separates absolute from conditional convergence, and the exponent separates conditional convergence from divergence. The case is the tail of the Dirichlet integral of this section, and is the exercise on sine over the square root. Every classification question about reduces to locating relative to these two thresholds.