Short answer. The epsilon delta definition says the limit of f at a point a is L when every positive output tolerance epsilon has a positive input window delta so that f(x) is less than epsilon away from L whenever x is less than delta away from a, and x is not a. Search for one numerical window on the epsilon delta finder. A sample is not a proof.

The epsilon delta definition looks like a wall of symbols because it packs a negotiation into one sentence. Once you hear the negotiation, the symbols become a checklist rather than a curse. The claim $\lim_{x\to a}f(x)=L$ means this. No matter how tightly someone demands that $f(x)$ stay near $L$, you can answer with a restriction on how close $x$ must be to $a$, excluding $x=a$ itself if needed.

That last clause is the whole point of a limit. Nearby values matter. The value at the point is a separate question, which is where continuity enters later. This page is a first reading in the explainer library for students who have used limit notation and now need the precise sentence. The wider map is calculus foundations, and the proof habit it belongs to is real analysis and proof habits. The study route that treats this definition as a skill, not a slogan, is the intro to proofs pathway.

The claim as a conversation

  • Your challenger picks a tolerance $\varepsilon>0$ for the outputs: “Keep $|f(x)-L|$ smaller than this.”
  • You reply with a window $\delta>0$ for the inputs: “If $0<|x-a|<\delta$, I guarantee your output demand.”

The order matters. Epsilon comes first because the definition must work for every positive tolerance, including absurdly small ones. Delta is allowed to depend on epsilon, and also on $a$ and $f$. A single lucky $\delta$ for $\varepsilon=0.1$ does not finish the proof. You need a rule that produces a valid $\delta$ from an arbitrary $\varepsilon>0$.

In symbols, the same conversation is: for every $\varepsilon>0$ there exists $\delta>0$ such that if $0<|x-a|<\delta$, then $|f(x)-L|<\varepsilon$. The “for every” is the challenger. The “there exists” is you. Swap the order and you have promised a window that works for all tolerances at once, which is a different and usually false claim.

A linear example you can finish in one line

Example Let $f(x)=3x+1$ and $a=2$, so the candidate limit is $L=7$. Then

$$|f(x)-7|=|3x+1-7|=3|x-2|.$$

If the challenger gives $\varepsilon$, choose $\delta=\varepsilon/3$. Whenever $0<|x-2|<\delta$, you get $|f(x)-7|<\varepsilon$. The factor $3$ is the slope. Steeper functions need tighter input windows for the same output tolerance. Flatter functions give you more room.

Notice that $x=2$ never entered the algebra. You used the nearby distance $|x-2|$. If someone redefined $f(2)$ to be $100$, the limit would still be $7$. The graph would have a hole or a spike at a single point. Limits do not see that spike.

A quadratic example that needs a bound

Example Claim $\lim_{x\to 3}x^2=9$. Then $|x^2-9|=|x-3|\,|x+3|$. The extra factor $|x+3|$ is not constant. If $x$ is allowed to wander across the whole line, that factor can be huge, and no single $\delta$ would work. So you first restrict the neighborhood. Require $|x-3|<1$, which forces $x$ between $2$ and $4$, hence $|x+3|<7$. Then

$$|x^2-9|<7|x-3|.$$

Given $\varepsilon>0$, choose $\delta=\min\bigl(1,\varepsilon/7\bigr)$. The $1$ keeps the convenience bound. The $\varepsilon/7$ meets the output demand. Suppose $0<|x-3|<\delta$. Then both controls apply, and $|x^2-9|<\varepsilon$.

That two part choice of $\delta$ is the standard scratch move for polynomials and rational functions. Bound the messy factor on a small interval around $a$. Then shrink $\delta$ so that the remaining $|x-a|$ factor finishes the $\varepsilon$ demand.

Pitfall. Finding one $\delta$ that works for a specific $\varepsilon$ does not finish the proof. A table of nearby values does not finish it either. You need a rule that produces a valid $\delta$ for an arbitrary $\varepsilon>0$. Also, $\delta$ is not unique. If $\delta=0.01$ works, then $\delta=0.001$ works too. You are not hunting a “correct” window. You are exhibiting one window that keeps the promise.

A hole the algebra is allowed to ignore

Example Let $f(x)=(x^2-4)/(x-2)$ for $x\neq 2$. The formula is undefined at $x=2$. For every other $x$, the difference of squares factors, and $f(x)=x+2$. The candidate limit as $x$ approaches $2$ is $4$. That is the value of the simplified line at $2$, even though $f$ itself never arrives there.

$$|f(x)-4|=|x+2-4|=|x-2|,$$

provided $x\neq 2$. Given $\varepsilon>0$, choose $\delta=\varepsilon$. If $0<|x-2|<\delta$, then $|f(x)-4|<\varepsilon$. The proof never plugged in $x=2$. The definition forbade that plug in. Someone who says the function is undefined, so the limit does not exist, has confused the value at the point with the nearby behavior. Someone who cancels and then evaluates the simplified formula at $2$ has computed the limit and has not computed $f(2)$. Those are different sentences.

Continuity would ask for both, and this $f$ is not continuous at $2$ because $f(2)$ does not exist. If you redefine the function so that $f(2)=4$, the same $\delta$ works and the function becomes continuous at $2$. If you redefine $f(2)=100$, the limit stays $4$ and continuity fails. The epsilon delta work did not change. Only the value at the single point changed. Limits do not see that point. The glossary line for continuity is the extra demand, not a second definition of the limit.

A denominator you must keep away from zero

Example Claim $\lim_{x\to 3}\dfrac{1}{x-1}=\dfrac12$. Then

$$\left|\frac{1}{x-1}-\frac12\right|=\frac{|x-3|}{|2(x-1)|}.$$

The factor $|x-1|$ in the denominator is dangerous near $x=1$, but the point under study is $x=3$. A small window around $3$ can stay away from $1$. Require $|x-3|\le \tfrac12$, so $x$ sits between $2.5$ and $3.5$, and $|x-1|\ge 1.5$. Then $|2(x-1)|\ge 3$, and the fraction is at most $|x-3|/3$. Given $\varepsilon>0$, choose $\delta=\min\bigl(\tfrac12,\, 3\varepsilon\bigr)$. If $0<|x-3|<\delta$, both controls apply, and $|f(x)-1/2|<\varepsilon$.

The convenience radius is part of the proof, not decoration. A radius of $3$ around $x=3$ would reach $x=0$, and the denominator $x-1$ would be allowed to get small. The bound would explode, and the same algebra would no longer produce a legal $\delta$. You choose the first restriction so the messy factor stays controlled. Then you spend the rest of $\delta$ on the $\varepsilon$ demand. That is the same two part move as the quadratic, now used to protect a denominator.

When the two sides refuse to share a limit

Example Let $g(x)=0$ when $x<1$ and $g(x)=2$ when $x>1$. The value $g(1)$ can be anything, or undefined. You do not need it. Suppose someone claims the two sided limit is a number $L$. Outputs on the left sit at $0$. Outputs on the right sit at $2$. Those numbers are $2$ apart. Choose $\varepsilon=1$.

Every positive $\delta$ opens a window that contains points on both sides of $1$. At a left point, $|g(x)-L|=|L|$. At a right point, $|g(x)-L|=|2-L|$. Both distances would need to be smaller than $1$. That asks $L$ to sit within $1$ of $0$ and within $1$ of $2$ at the same time, strictly. No real number does that. So no candidate $L$ survives, and no $\delta$ can rescue a candidate the two sides refuse to share. The two sided limit does not exist.

The limit from the left is $0$. The limit from the right is $2$. Each one sided claim has its own easy proof, because on that side the function is constant. The two sided claim is the demand that those one sided limits exist and are equal. They exist. They are not equal. A table that samples only the right of $1$ will look calm and will lie about the two sided question. That is why a numerical check has to approach from both sides before you invest in a $\delta$ rule.

Pitfall. Canceling a factor of $x-a$ is legal only for the nearby points the definition already uses. After you cancel, do not pretend the original formula was defined at $a$. Also do not swap the quantifiers. A single $\delta$ that works for every $\varepsilon$ at once would force $f(x)=L$ on a whole punctured window. Ordinary limits do not promise that. Epsilon is chosen first. Delta may depend on it.

Continuity is the limit plus the value

The limit definition is silent about $f(a)$. Continuity at $a$ is three demands together. The limit exists. The number $f(a)$ exists. Those two numbers are equal. In epsilon delta language you may drop the exclusion $0<|x-a|$, because the point $x=a$ now has to obey the same output tolerance, and it does exactly when $f(a)=L$.

Example Let $h(x)=|x|$. Claim continuity at $0$ with value $0$. Then $|h(x)-0|=|x|$. Given $\varepsilon>0$, choose $\delta=\varepsilon$. If $|x|<\delta$, then $|h(x)|<\varepsilon$. This includes $x=0$. So $h$ is continuous at $0$. The same function is not differentiable at $0$. Slopes from the left and from the right are $-1$ and $1$. A derivative is a limit of a difference quotient, and that quotient limit fails here even though the function limit for $h$ itself succeeds. Finish that distinction on what a derivative actually measures. Epsilon delta is the shared grammar under both claims.

A quieter continuity failure is a single lifted point. Let $p(x)=0$ for $x\neq 0$ and $p(0)=1$. The limit at $0$ is $0$, by the same $\delta=\varepsilon$ argument you would use for the constant function $0$ on the punctured line. But $p(0)=1$, so $p$ is not continuous at $0$. The graph is a flat line with one point lifted. Limits ignore that point. Continuity does not.

A decision framework for a written proof

  1. Name $a$ and a candidate $L$. If you do not have a candidate, simplify the formula for $x\neq a$, or read a two sided table, before you promise a number.
  2. Compare the story from the left with the story from the right. If those candidates disagree, stop. Write that the two sided limit does not exist. Do not hunt for a $\delta$.
  3. Form $|f(x)-L|$ and factor out $|x-a|$ when you can.
  4. If the other factor is a positive constant $M$, choose $\delta=\varepsilon/M$. A steeper $M$ needs a tighter window. The linear example used $M=3$.
  5. If the other factor still depends on $x$, restrict to a convenience neighborhood, often $|x-a|<1$, and bound that factor by some $M$ there. Then choose $\delta=\min(1,\varepsilon/M)$.
  6. If a denominator appears, choose the convenience neighborhood so the denominator stays away from zero. Shrink until a lower bound appears. Then spend the rest of $\delta$ on $\varepsilon$, as in the example $1/(x-1)$ at $x=3$.
  7. Write the proof forward. Let $\varepsilon>0$. Choose $\delta=\ldots$. Suppose $0<|x-a|<\delta$. Then the bound applies, hence $|f(x)-L|<\varepsilon$.
  8. Add a continuity sentence only when the problem asked for continuity. That sentence is $f(a)=L$, together with the same window after you drop the exclusion of $x=a$.

Scratch work may run backwards, from the output demand toward a window. The written proof runs forwards. A reader is checking the forward version. If your scratch $\delta$ used a bound you never earned, the forward writeup will have a gap, and that gap is the thing to repair.

The same negotiation for sequences

A sequence limit replaces the input window with an index. The challenger still names $\varepsilon>0$. You answer with an integer $N$ such that $n>N$ forces $|a_n-L|<\varepsilon$. There is no $\delta$, because the input is already a discrete list. The logic is the same negotiation. Every tolerance gets a reply, and the reply may depend on the tolerance.

Partial sums are a sequence. An infinite series converges when that sequence of partial sums has a limit in this sense. The page sequences, series, and infinite sums is about that meaning. It will send you onward to tests when you cannot name the partial sums. Do not treat a series total as a different kind of addition. It is this limit, wearing a summation sign.

If you cannot name those partial sums, convergence tests for series say which test is allowed to decide, and when the test stays silent. The sequence and series calculator adds a finite arithmetic or geometric list. That number is one partial sum. It does not decide the infinite series.

How to know you understood

  1. Say the conversation out loud. The challenger names $\varepsilon$. You answer with $\delta$. Then say what fails if you swap those two quantifiers.
  2. For $f(x)=3x+1$ at $a=2$, produce $\delta=\varepsilon/3$ and name the factor $3$ as the slope.
  3. For $x^2$ at $3$, explain why $\delta=\min(1,\varepsilon/7)$, and what job the $1$ is doing.
  4. For $(x^2-4)/(x-2)$, explain why the limit at $2$ is $4$ even though the original formula is undefined there.
  5. For the jump that equals $0$ on the left of $1$ and $2$ on the right, show that $\varepsilon=1$ defeats every candidate $L$.
  6. State continuity as the limit statement plus $f(a)=L$, and give $|x|$ at $0$ as a function that is continuous there.

How to practice without panic

  1. Start with linear functions, then absolute values, then quadratics, then rational functions where you bound the denominator away from zero near $a$.
  2. Scratch work may assume $|x-a|$ is already smaller than $1$, so that extra factors stay controlled. Then tighten $\delta$ to satisfy both the convenience bound and the $\varepsilon$ demand.
  3. Write the polished proof forward: “Let $\varepsilon>0$. Choose $\delta=\ldots$. Suppose $0<|x-a|<\delta$. Then … hence $|f(x)-L|<\varepsilon$.”
  4. If a numerical table looks wild as $x$ approaches $a$ from two sides, stop proving and ask whether the two sided limit exists at all.

This habit transfers to continuity arguments, which add the demand that $f(a)=L$, and to derivatives, which are limits of difference quotients. The numerical derivative calculator estimates that quotient with a small step, which can suggest a slope and cannot prove it. The same language sits under what a derivative actually measures. After you can write a $\delta$ rule, an implicit curve is a later place the chain rule hides a limit. That story is implicit differentiation intuition. For the wider map, return to calculus foundations and real analysis and proof habits.

Try it with the matching tool

A numerical table is not a proof. It is still a good way to see the negotiation before you write $\delta$. The matching tool is the epsilon delta finder. You type $f(x)$, the point $a$, a proposed limit $L$, and a tolerance $\varepsilon$. It searches for a large input window where the sampled outputs stay inside that tolerance, and it draws the horizontal epsilon band and the vertical delta band.

Try it. Open the epsilon delta finder on the linear preset, $f(x)=3x+1$ at $a=2$ with $L=7$. For $\varepsilon=0.1$ the search should return a $\delta$ near $0.03333$, which is $\varepsilon/3$. Then load the quadratic preset. A written proof often picks a smaller convenient $\delta$, namely $\min(1,\varepsilon/7)$. The finder is hunting a large window that the samples support. Load the jump preset last. No positive $\delta$ survives, because the two sides do not share a limit.

The numerical limit estimator is the companion picture. It only asks whether nearby outputs seem to settle. After the picture looks calm, come back and write a $\delta$ rule by hand for the linear or quadratic case. A sampled window is not a proof.

Find a numerical delta for one epsilon, then write the rule a proof would use.

Open the epsilon delta finder

Citations and further reading

  • Abbott, S. Understanding Analysis. Springer. Clear motivation for $\varepsilon$ arguments in early chapters. Springer link
  • MIT OpenCourseWare, 18.100A Real Analysis, lecture notes and assignments on limits. ocw.mit.edu
  • Stewart, J. Calculus, sections on the precise definition of a limit, for the standard calculus framing.
  • Paul’s Online Math Notes, the definition of the limit. tutorial.math.lamar.edu

Last reviewed: 21 September 2026.