core29 min studyLesson 171 in Mathematics

Limsup, liminf, and subsequences

Extract persistent upper and lower behavior from an oscillating sequence without pretending it has a single limit.

Before you begin

Limits and continuity: reasoning about approaching

On this page

The question for today

What you’ll understand

Construct decreasing upper-tail bounds and increasing lower-tail bounds.
Characterize real convergence using matching limsup and liminf.
Identify subsequential limits and distinguish them from early outliers.

An approximation can oscillate forever while its long-term range becomes perfectly predictable. A single maximum over all recorded terms gives too much importance to early mistakes, and a running average can conceal persistent oscillation. Limit superior and limit inferior answer a different question: after discarding arbitrarily long finite beginnings, which upper and lower levels remain possible? These notions organize subsequences, convergence tests, and error envelopes without requiring the original sequence to settle down.

Discarding beginnings changes the question

For a bounded real sequence, take the supremum of every tail. Removing the first tail term cannot increase that supremum, so these upper-tail bounds form a decreasing sequence. They are bounded below by any lower bound of the original sequence and therefore converge. The limit is the limsup. Tail infima increase for the same reason, and their limit is the liminf. Neither construction requires a largest or smallest term to be attained.

Consider observations that alternate near two levels, with an unusually large first observation. The early outlier affects the supremum of the whole sequence but disappears from every sufficiently late tail. Oscillations recurring arbitrarily late do not disappear. This distinction explains why limsup describes persistent upper behavior rather than a global maximum or a statistical average. Tail bounds also make finite modifications harmless: changing a thousand initial values leaves both limiting envelopes unchanged.

b_N=\sup_{n\ge N}a_n,\qquad c_N=\inf_{n\ge N}a_n,\qquad\limsup_{n\to\infty}a_n=\inf_N b_N,\quad\liminf_{n\to\infty}a_n=\sup_N c_N

For bounded real sequences, monotonicity of the tail bounds and completeness of the real numbers make both finite limits exist.↔ Scroll wide equations horizontally.

Subsequences reveal the surviving destinations

A subsequence retains terms at strictly increasing indices; it cannot repeatedly reuse one favorable observation. If a bounded sequence has limsup L, choose a sufficiently late term within a shrinking tolerance of its tail supremum. Doing this successively with increasing indices produces a subsequence approaching L. Every other convergent subsequence has its limit between liminf and limsup, because sufficiently late terms lie between the corresponding tail bounds.

The interval between those endpoints need not be filled with subsequential limits. Alternating zero and one has only zero and one as such limits, although its liminf and limsup enclose every number between them. Thus the envelopes bound the cluster set but do not describe its internal geometry. Bolzano–Weierstrass guarantees at least one convergent subsequence for every bounded real sequence; it does not turn oscillating sequences into convergent ones.

a_n\to L\quad\Longleftrightarrow\quad\liminf a_n=\limsup a_n=L

This equivalence concerns a finite real L. Matching upper and lower tail envelopes squeeze all sufficiently late terms into every neighborhood of L.↔ Scroll wide equations horizontally.

Use envelopes without overinterpreting data

For unbounded sequences, extended real values are useful. The sequence n has both limiting envelopes equal to positive infinity, meaning it eventually exceeds every real threshold. The sequence alternating n and minus n has limsup positive infinity and liminf negative infinity. These statements concern eventual inequalities, not arithmetic with an actual infinite term. Subtracting two infinite envelopes is therefore not a legitimate shortcut for analyzing a new sequence.

The optimization lab offers a concrete family of sequences: gradient descent on a quadratic gives successive powers of one minus the step size. Predict whether its iterates approach zero, alternate with fixed magnitude, or grow before pressing play. An oscillating plot can suggest two subsequential limits, but finitely many displayed points cannot establish them. Derive the recurrence first, then use the display to inspect consequences of its parameters and initial condition.

Work it through

A persistent oscillation with a vanishing perturbation

Find the limsup, liminf, and all subsequential limits of $a_n=(-1)^n+2/n$ for n≥1.

Along even indices, $a_{2k}=1+1/k$ , so this subsequence approaches 1 from above. Along odd indices, $a_{2k-1}=-1+2/(2k-1)$ approaches −1.
For every positive tolerance, the perturbation 2/n is eventually smaller than that tolerance. Late even terms therefore approach the upper envelope 1, and late odd terms approach the lower envelope −1.
Any convergent subsequence containing infinitely many even and odd indices would have sub-subsequences approaching two different numbers, which is impossible. Its limit must therefore be 1 or −1; both are attained by the parity subsequences.

The result

The limsup is 1, the liminf is −1, and the cluster set is exactly {−1,1}. The original sequence does not converge.

The living notebook

Explore the model

Interactive lab

Interactive lab · optimization

From the update factor to limiting envelopes

Set step size η=2 and a nonzero initial x₀. Predict the two limiting levels of xₙ=(1−η)ⁿx₀, then step through the iterates. Compare with 0<η<2 and η>2. This quadratic recurrence illustrates oscillation and divergence; it does not compute limsup for arbitrary input sequences.

ObjectiveDiscrete iterate

Position x0: 2
Objective x²/2: 2
Iteration behavior: Stable step

Step size η0.3Initial position x₀2

n = 0 · t = 0 s

Gradient descent on f(x) = x²/2: xₙ = (1 − η)ⁿx₀. Playback advances two discrete iterations per second. Values outside |x| ≤ 4 are explicitly reported. Parameter changes pause playback and recompute the model at the current elapsed time. Reset restores both parameters and returns time to zero. Playback starts paused. Adjust either parameter, then choose Play or Step.

Look forA recurring pair of levels differs from a shrinking oscillation, even when both plots alternate.

Pause & reason

Check your understanding

Interactive checks need JavaScript. You can still read the explanation.

Read the checkpoint explanation

For bounded real sequences the finite envelope endpoints are achieved as subsequential limits. Intermediate limits and restrictions on early terms do not follow.

Your turn

Try it yourself

For $b_n=3+(-1)^n/n$ , determine the limiting envelopes and prove convergence using a direct bound.

Take a moment to make a prediction or work it out on paper. Open a hint when you need one.

Give me a hint

Control the distance from 3 uniformly across all sufficiently late indices.

Compare with the solution

The distance $|b_n-3|$ equals 1/n. Given ε>0, choose an integer N>1/ε; every n≥N then has distance below ε. Both tail envelopes consequently approach 3, and the sequence converges to 3. Its terms oscillate about 3, but their separation shrinks rather than persisting. Replacing a finite initial segment, even with enormous values, changes neither envelope.

Keep the idea

Take it with you

Tail extrema forget finite beginnings but retain recurring late behavior.
For bounded real sequences, envelope endpoints are subsequential limits.
Matching finite limsup and liminf characterize convergence.

Read further

Original sources and deeper explanations to follow at your own pace.

Optional progress, saved in this browser when available. Revisit any lesson at any time.

Lesson not marked complete.