The series everything else is measured against

The square with a corner left uncut settles one series exactly. Its terms halve, the pieces fit inside one square, and the total is one.

That is a complete answer to a question nobody often asks. The questions that arise are about series whose terms do something less tidy — $n/2^n$, or $1/n^2$, or $1/n!$ — and for those there is no dissection and no closed form. What there is instead is the observation that none of that matters if the terms eventually fit under a geometric series’ terms, because a total sitting under a total that exists is a total that exists.

That is the comparison test, and the geometric series is the thing compared against because it is the only family whose total can be written down without doing any work. Everything else in elementary convergence testing is a way of arranging that comparison.

Why a bound is enough #

The step that looks like sleight of hand is the one where a bound becomes an answer, and it is worth being precise about it.

A series of non-negative terms has partial sums that never decrease. A sequence that never decreases does one of exactly two things: it climbs past every bound, or it settles. Nothing else is available — there is no oscillation, no wandering, and no third case. So for a series of non-negative terms, staying under some bound is the same statement as converging.

That equivalence is the whole reason a comparison is worth anything, and it is where the hypothesis about signs earns its place. The alternating series $1 - 1 + 1 - 1 + \cdots$ has partial sums that never exceed one and does not converge, because its partial sums are not increasing; the argument breaks at the first sentence rather than at the last.

It is also why the dissection worked. A square is a bound, and the pieces being disjoint parts of it is the statement that the partial sums cannot exceed its area — so that figure proves the same thing this comparison does, in the one case where the bound can be drawn.

What the comparison actually requires #

Three details in the figure are doing work and each corresponds to a place the test is often misstated.

The domination need only hold eventually. The bars in the hero figure are above the curve at the first term, and nothing is wrong. A finite number of terms is a finite number, and adding one to a total that exists gives a total that exists; only the tail decides. The figure finds the index from which the domination holds rather than assuming one, and verifies it at every term afterwards.

The comparison series’ total has to be computable, not merely finite. The bound in the caption is the head plus $c\,r^{k}/(1-r)$, and the second piece is the closed form the dissected square derived in one subtraction. Comparing against a convergent series nobody can sum gives a proof of convergence and no number; comparing against a geometric series gives both.

And the terms must be non-negative. Everything above is about size, and with signs allowed there are convergent series whose terms are not dominated by any convergent series of positive terms — the alternating harmonic series is one, and rearranging it reaches any total at all. That failure is not a technicality: it is the difference between a series that converges and one that converges absolutely, and the comparison test is a test for the second.

That second figure is the first sign of trouble. The comparison succeeds and the bound it gives is nothing like the true total, because $1/n^2$ and a geometric series are not really the same kind of object — one shrinks by a factor each step and the other by a proportion that tends to nothing. The looseness is a symptom of a mismatch that the next section turns into a failure.

The ratio test is one question #

The comparison test asks for a geometric series to be produced. The ratio test is the observation that there is a mechanical way to look for one: ask whether the series is approximately geometric, by measuring the ratio of each term to the one before.

If the ratios settle on a number $L$ below one, then from some point on every ratio is below some $r$ with $L < r < 1$, so from that point the terms are dominated by a geometric series of ratio $r$, and the comparison applies. If $L$ is above one the terms eventually grow and the total cannot exist, since a series whose terms do not go to nothing has nowhere to settle.

And if $L$ is one, the test has no information. That is not a limitation to be worked around later; it is the whole of the difficulty, and the figure exhibits it: $1/n^2$ and $1/n$ have ratios climbing to one, they lie on top of each other on the right of the plot, and one of them converges while the other does not.

The reason is structural rather than technical. A ratio tending to one says the terms are not approximately geometric, and the yardstick has been asked a question about an object outside its family. Comparing $1/n^2$ against a geometric series was already loose in the figure above; comparing $1/n$ against one fails outright, because no geometric series has terms that shrink as slowly.

Which is most of the interesting cases #

It is worth stating how large the silent region is, because the ratio test is usually taught as the general method and it is not.

Every series of the form $1/n^p$ has ratio $(n/(n+1))^p \to 1$, for every $p$. So the ratio test decides none of them — not $1/n^2$, not $1/n^{1.001}$, not $1/\sqrt{n}$ — and the family it cannot decide contains the boundary the whole subject is organised around. The root test, which looks at the $n$-th root of the $n$-th term instead, is strictly stronger and is silent on exactly the same family, for the same reason.

Where the ratio test is decisive is where a factorial or an exponential is present, because those produce ratios that genuinely settle away from one. $1/n!$ has ratio $1/(n+1) \to 0$; $n^{10}/2^n$ has ratio tending to a half; and a power series’ radius of convergence is found by exactly this computation, which is the test’s real home. A Taylor series’ interval is a ratio-test calculation and nothing else — and the endpoints of that interval, where the ratio is one, are precisely the cases the test hands back undecided.

What the yardstick is really measuring #

There is a way of reading all of this that explains why the geometric series occupies the position it does, and it is about rates rather than about totals.

A geometric series’ tail after $k$ terms is exactly $r^k/(1-r)$ — a fixed proportion of the last term, for ever. So knowing the ratio is knowing the error at every stage, which is what makes the family the natural standard: it is not that geometric series are common, it is that they are the ones whose remainder is a formula.

A series with a ratio tending to $L<1$ inherits that. Its tail eventually behaves like a geometric one, so its error after $k$ terms shrinks by a roughly constant factor each step, and computing it to a given accuracy costs a number of terms proportional to the number of digits wanted. That is what converges means in practice, and it is what the ratio test is really certifying.

A series with ratio tending to one inherits none of it. $\sum 1/n^2$ converges, and its tail after $k$ terms is about $1/k$ — so a thousand terms buy three digits and a million buy six. The comparison test says the total exists; nothing about the total’s existence says it can be computed, and the rate is the quantity that matters once it can.

So the ratio test’s silence at one is not only a logical gap. It is the boundary between series that can be added up and series that cannot, and on the far side of it both answers — converges, diverges — come with the same unhelpful arithmetic.

The comparison that is actually made #

In practice almost nobody produces a dominating geometric series by hand, and the reason is that a slightly weaker statement does the same work with none of the effort.

The limit comparison test: if two series of positive terms have $a_n/b_n$ tending to a positive number, then either both converge or neither does. No domination has to be exhibited, no constant has to be found, and no index from which it holds has to be located — the limit does all three at once, because a ratio tending to $c$ means the terms are eventually between $c/2$ and $2c$ times each other, which is a domination in both directions.

That is how every $1/n^p$ question is settled once one of them is known. To decide $\sum 1/(n^2 + 3n + 7)$, divide it by $1/n^2$; the quotient tends to one; so the two behave alike, and the polynomial’s lower terms have been disposed of without a single inequality being written.

It also explains why the boundary cases resist. The test needs a known series to compare against, and inside the family $1/n^p$ every member’s ratio against every other tends to nought or to infinity rather than to a positive number — the test compares a series to one of the same rate, and the family’s members have no two the same. So the limit comparison test settles everything given one member of the family, and decides nothing about the family. Which member is the boundary has to come from elsewhere, and that is the gap condensation fills.

The order of business is therefore the opposite of how the tests are usually listed. The ratio test is fast and narrow, the comparison tests are general and need an input, and the input — the threshold inside $1/n^p$ — is the one thing none of them produces.

What the yardstick says about the terms going to nothing #

There is a standing confusion the comparison makes short work of, and it is worth clearing because it is the single most common error about series.

Terms going to nothing is necessary and is not sufficient. If the terms do not go to nothing the partial sums cannot settle, so the total does not exist; that direction is immediate. The reverse fails, and the harmonic series is the standard counterexample — its terms go to nothing and its total passes every bound.

What the yardstick adds is why the reverse fails, in a form worth remembering. A geometric series’ terms go to nothing at a fixed proportion per step, and that rate is what pays for the total. The harmonic series’ terms go to nothing at a rate that itself slows down: the step from $1/n$ to $1/(n+1)$ is a proportion $1/(n+1)$ of the term, and that proportion vanishes. So the terms shrink and the shrinking stops being worth anything, which is exactly what the ratio climbing to one records.

Read that way the ratio test’s verdict is not about convergence at all in the first instance. It measures whether the shrinking is sustained, and convergence is a consequence when it is. At the boundary the shrinking is not sustained, the consequence is unavailable, and whether the total exists becomes a question about how fast the sustaining fails — which is a second-order question and needs a second-order instrument.

What is left when it is silent #

Three things, and each is a separate instrument.

Condensation. If a geometric series is the wrong comparison, compare the series with itself at doubled spacing. That is Oresme’s grouping — used to show the harmonic series has no total — turned into a general instrument, and it decides the whole family $1/n^p$ in one line by turning it back into a geometric series. It is the subject of the repair at the boundary.

Integration. A series of decreasing terms is bracketed by two integrals of the same function, and $\int_1^\infty x^{-p}\,dx$ is elementary. That gives the same threshold with a different instrument, and it gives the tail estimate as well.

The test’s own history, which explains its shape. The ratio criterion is d’Alembert’s, from 1768, and it was written for power series rather than for numerical ones — the question it was built to answer is for which $x$ does this expansion mean anything, and there the ratio genuinely settles, because the $x$ carries it. Cauchy’s root test came fifty years later and is strictly stronger, in the precise sense that whenever the ratio test decides, the root test decides the same way, and there are series it decides that the ratio test does not. Both are silent on $1/n^p$. A test built for power series being useless on the boundary family is not a defect in it; it is the family being outside what the test was ever asked about.

And an exact answer, occasionally. $\sum 1/n^2$ is $\pi^2/6$, which no comparison would ever have produced — a comparison bounds, and a bound is not a value. That the constant is $\pi^2/6$ rather than something unremarkable is Euler’s, and it is one of the places where a question about a total turns out to have an answer from an entirely different subject.

Twenty terms against a claim about all of them #

Every figure here is drawn at twenty terms and every claim is about infinitely many. The hero’s domination is verified at each drawn term and argued in prose for the rest, which is where the argument actually lives: the claim is that $n/2^n \le c\,(0.75)^n$ for all $n$ past a point, and that is an inequality about a function rather than an observation about fourteen bars.

A ratio tending to one is a statement about a limit, and a plot shows a finite stretch. The two curves that converge on one in the hero’s companion are at $0.95$ and $0.96$ at the right-hand edge, which is compatible with settling at one and equally compatible with settling at $0.96$ — and $0.96$ would be a verdict. Nothing in the drawing distinguishes the two, and the distinction is the whole content.

And the bound the comparison produces is drawn and the true total is not. For $1/n^2$ the comparison gives a bound near two and the answer is $1.6449$; the figure shows the first number and has no way of showing the second, because computing it needs a different subject entirely.

Where the yardstick runs out #

The ratio test is silent on every $1/n^p$, which is the family the boundary between convergence and divergence runs through. The repair is a comparison a geometric series cannot make — the series against itself at doubled spacing — and gets the threshold exactly, at $p = 1$.

What that repair then reveals is stranger than the repair. Applying it a second time produces series between $1/n$ and every $1/n^{1+\varepsilon}$; applying it again produces series between those; and the process does not terminate, so there is no slowest divergent series and no fastest convergent one to compare anything against. The yardstick is not merely the wrong shape at the boundary — there is no single right one, and that is what comparing a series with itself at doubled spacing is about.

A standard, and the shape it imposes #

The habit worth carrying is to notice what a standard of comparison assumes about what it is comparing.

A geometric series is the yardstick because its remainder has a closed form. Everything it measures well is something that behaves like it — a fixed proportion lost each step — and everything it measures badly shares the property that the proportion is not fixed. The test’s silence is therefore not an accident of how it was formulated; it is the standard reporting that the object is not of its kind, which is the most useful thing a measurement can say when it cannot say a number.

The general version of that is worth having outside this subject. When an instrument returns no verdict on a whole family of cases, the first question is not how to strengthen it. It is what the instrument was built to recognise, and whether the family it cannot see is a family it was never about.