Limits & the Epsilon-Delta Definition

The limit of a function $f$ at a point $a$ captures the value that $f(x)$ tends toward as $x$ approaches $a$, regardless of whether — or how — $f$ is defined at $a$ itself. The epsilon-delta definition makes this precise: no matter how small an error tolerance $\epsilon$ a skeptic demands around the proposed limit $L$, we can guarantee that tolerance by restricting $x$ to a small enough punctured interval of radius $\delta$ around $a$. This single definition underpins the entire edifice of calculus, replacing 17th-century appeals to "infinitesimals" with statements quantifiable in terms of inequalities. A surprising fact: the limit can exist at points where the function is undefined, discontinuous, or wildly oscillatory away from $a$ — the limit only "sees" arbitrarily small punctured neighborhoods. Equally surprising, a function can fail to have a limit at every rational point yet behave perfectly at others (the popcorn function).

The Core Insight

A limit is a promise that can always be kept. The challenger names a vertical tolerance band of half-width $\epsilon$ around $L$; the defender must respond with a horizontal punctured interval of half-width $\delta$ around $a$ such that the entire graph over that interval lies inside the band. Crucially, $\delta$ is allowed to depend on $\epsilon$ (smaller $\epsilon$ usually forces smaller $\delta$), and the point $x=a$ is always excluded — what happens at $a$ is irrelevant. The limit is a statement about the neighborhood-deleted-of-its-center, not about the center.

Formal Statement

Limit of a Function — Epsilon-Delta

Let $f$ be a real-valued function defined on a punctured neighborhood of $a$ (i.e., on some set containing $(a-r,a)\cup (a,a+r)$ for some $r>0$). We say the limit of $f(x)$ as $x$ approaches $a$ is $L$, written ${\lim }_{x\to a}f(x)=L$, if:

Here:

• $\epsilon >0$ is the vertical tolerance (how close $f(x)$ must be to $L$),
• $\delta >0$ is the horizontal radius (how close $x$ must be to $a$),
• $0<|x-a|$ enforces $x\ne a$ (the punctured condition),
• $L$ is the proposed limit value.

The condition $0<|x-a|<\delta$ describes the punctured interval $(a-\delta ,a)\cup (a,a+\delta )$, and $|f(x)-L|<\epsilon$ describes the open band $(L-\epsilon ,L+\epsilon )$.

One-Sided Limits

The right-hand limit ${\lim }_{x\to a^{+}}f(x)=L$ requires:

The left-hand limit ${\lim }_{x\to a^{-}}f(x)=L$ uses $a-\delta <x<a$. The two-sided limit exists iff both one-sided limits exist and are equal.

Limits Involving Infinity

Uniqueness of Limits

If ${\lim }_{x\to a}f(x)=L_1$ and ${\lim }_{x\to a}f(x)=L_2$, then $L_1=L_2$.

Algebraic Limit Laws

If ${\lim }_{x\to a}f(x)=L$ and ${\lim }_{x\to a}g(x)=M$, then:

and, provided $M\ne 0$,

Squeeze (Sandwich) Theorem

If $g(x)\le f(x)\le h(x)$ on a punctured neighborhood of $a$ and ${\lim }_{x\to a}g(x)={\lim }_{x\to a}h(x)=L$, then

Why It Holds / Derivation

Uniqueness of Limits

Suppose for contradiction that ${\lim }_{x\to a}f(x)=L_1$ and $=L_2$ with $L_1\ne L_2$.

1. Let $\epsilon =\frac{|L_1-L_2|}{2}>0$.
2. By the limit hypothesis for $L_1$, there is ${\delta }_1>0$ with $0<|x-a|<{\delta }_1⟹|f(x)-L_1|<\epsilon$.
3. By the limit hypothesis for $L_2$, there is ${\delta }_2>0$ with $0<|x-a|<{\delta }_2⟹|f(x)-L_2|<\epsilon$.
4. Let $\delta =\min ({\delta }_1,{\delta }_2)$. Since $f$ is defined on a punctured neighborhood, pick any $x$ with $0<|x-a|<\delta$ (such $x$ exists).
5. Then by the triangle inequality:

1. This gives $|L_1-L_2|<|L_1-L_2|$, a contradiction. Hence $L_1=L_2$. $■$

Sum Law: $\lim(f+g)=L+M$

Let $\epsilon >0$ be given.

1. Since ${\lim }_{x\to a}f=L$, there is ${\delta }_1>0$ with $0<|x-a|<{\delta }_1⟹|f(x)-L|<\epsilon /2$.
2. Since ${\lim }_{x\to a}g=M$, there is ${\delta }_2>0$ with $0<|x-a|<{\delta }_2⟹|g(x)-M|<\epsilon /2$.
3. Let $\delta =\min ({\delta }_1,{\delta }_2)>0$. Then for $0<|x-a|<\delta$, both bounds hold simultaneously.
4. By the triangle inequality:

1. This is exactly the definition with limit $L+M$. $■$

Product Law: a representative argument

Let $\epsilon >0$. We estimate

1. Bound $|f(x)|$ locally. Since $f\to L$, choose ${\delta }_0$ so that $|f(x)-L|<1$, hence $|f(x)|<|L|+1$ for $0<|x-a|<{\delta }_0$.
2. Control the first term. Choose ${\delta }_1$ so that $|g(x)-M|<\frac{\epsilon }{2(|L|+1)}$.
3. Control the second term. Choose ${\delta }_2$ so that $|f(x)-L|<\frac{\epsilon }{2(|M|+1)}$.
4. Let $\delta =\min ({\delta }_0,{\delta }_1,{\delta }_2)$. Then for $0<|x-a|<\delta$:

Karl Weierstrass | 1815–1897

Often called the "father of modern analysis," Weierstrass spent fourteen years as a secondary-school teacher before his groundbreaking lectures at the University of Berlin made rigor the gold standard of analysis. He distrusted the geometric intuition and "infinitely small quantities" pervading earlier calculus. Through his lectures (transcribed and circulated, as he published little) he crystallized the static, inequality-based $\epsilon$–$\delta$ definitions of limit and continuity that bear no reference to motion or infinitesimals. His 1872 construction of a continuous-but-nowhere-differentiable function shattered the intuition that continuity implies smoothness, proving that only airtight definitions — not pictures — could be trusted. His students (Cantor, Schwarz, Frobenius, Kovalevskaya, Mittag-Leffler) carried this rigor across mathematics.

Building Intuition

Think of the limit as a manufacturing tolerance problem. You operate a machine whose dial setting is $x$ and whose output is $f(x)$; you want output near a target $L$. A customer specifies the maximum allowable output error $\epsilon$. Your job: find a dial tolerance $\delta$ such that any setting within $\delta$ of the ideal $a$ (but you needn't hit $a$ exactly — perhaps the dial is broken precisely at $a$) yields acceptable output. The limit "exists and equals $L$" precisely when, for every conceivable customer demand $\epsilon$ (no matter how strict), you can find a working tolerance $\delta$.

The logical order is everything: $\epsilon$ comes first, $\delta$ responds. If you could fix one $\delta$ that works for all $\epsilon$, the function would be locally constant — far too strong. The dependence $\delta =\delta (\epsilon )$ is the whole content. Geometrically: thin the horizontal band ($\delta \to 0$) until the graph — viewed only on the punctured interval — fits inside the vertical band of height $2\epsilon$.

The epsilon-delta band game: the graph over the punctured $\delta$-interval must lie inside the $\varepsilon$-band around $L$.

Worked Examples

Linear limit by epsilon-delta

Quadratic limit requiring a bounding step

Common Misconceptions

A limit that does not exist

Consider $f(x)=\sin (1/x)$ as $x\to 0$. As $x\to 0^{+}$, $1/x\to \infty$ and $\sin (1/x)$ oscillates between $-1$ and $1$ infinitely many times in every punctured neighborhood of $0$.

No candidate $L$ can satisfy the definition: take $\epsilon =1/2$. For any $\delta >0$, the interval $(0,\delta )$ contains points where $f=1$ and points where $f=-1$ (since $1/x$ ranges over an entire interval of length $>2\pi$). At least one of these is more than $1/2$ from $L$. Hence

Contrast: $x\sin (1/x)\to 0$ by the Squeeze Theorem, since $|x\sin (1/x)|\le |x|$. The damping factor tames the oscillation.

Connections

The $\epsilon$–$\delta$ limit is the seed from which the calculus grows:

• Continuity is the special case ${\lim }_{x\to a}f(x)=f(a)$ — the limit equals the function value, removing the "puncture" exclusion's effect.
• The Derivative is a limit of difference quotients, $f^{\prime }(a)={\lim }_{h\to 0}\frac{f(a+h)-f(a)}{h}$.
• The Riemann Integral is a limit of Riemann sums as the mesh tends to zero.
• Sequences and Their Limits use an analogous $\epsilon$–$N$ definition; the Sequential Criterion states ${\lim }_{x\to a}f=L$ iff $f(x_n)\to L$ for every sequence $x_n\to a$ with $x_n\ne a$.
• Topology and Open Sets generalizes $\epsilon$ and $\delta$ neighborhoods to open sets, yielding the topological definition: preimages of open sets are open.
• Metric Spaces replace $|x-a|$ by a distance $d(x,a)$, extending limits to abstract settings.

Real-World Applications

1. Numerical analysis & error bounds. Iterative solvers (Newton's method, fixed-point iteration) converge in the $\epsilon$–$\delta$ sense; stopping criteria are $\epsilon$ tolerances, and convergence theorems guarantee the existence of a corresponding step count or step size ($\delta$).
2. Control engineering. Lyapunov stability is literally an $\epsilon$–$\delta$ statement: for every $\epsilon$-ball around equilibrium there is a $\delta$-ball of initial conditions whose trajectories stay within $\epsilon$.
3. Signal processing. Defining instantaneous frequency and the behavior of filters near cutoff relies on limits of transfer functions; convergence of Fourier partial sums uses limit definitions.
4. Floating-point computation (IEEE 754). Machine epsilon ${\epsilon }_{\text{mach}}$ is the practical analog of the analytic $\epsilon$ — the smallest representable relative gap, bounding round-off propagation.
5. Computer graphics / antialiasing. Continuity and limit arguments justify level-of-detail blending and the convergence of subdivision-surface schemes to smooth limit surfaces.
6. Economics — marginal analysis. Marginal cost and marginal revenue are derivatives, hence limits of average rates; the rigorous existence of these marginals is an $\epsilon$–$\delta$ matter.
7. Physics — instantaneous velocity & fields. Velocity $v={\lim }_{\Delta t\to 0}\Delta x/\Delta t$ and electric field via $E={\lim }_{q\to 0}F/q$ are limits; their well-definedness underpins continuum mechanics and electrodynamics.

Historical Development

• c. 1665–1684 — Newton ("ultimate ratios") and Leibniz (differentials $dx$) develop calculus using infinitesimals, without rigorous foundations.
• 1734 — Bishop George Berkeley's The Analyst derides infinitesimals as "ghosts of departed quantities," exposing the logical gaps.
• 1765–1797 — d'Alembert and Lagrange attempt limit- and algebra-based foundations; Lagrange tries (unsuccessfully) to ground calculus in Taylor series.
• 1817 — Bernard Bolzano gives an essentially modern definition of continuity using inequalities, in Rein analytischer Beweis, but his work goes largely unnoticed.
• 1821 — Augustin-Louis Cauchy's Cours d'analyse defines limits, continuity, and convergence verbally ("indefinitely approaches"), introducing $\epsilon$ and $\delta$ as quantities but not yet fully formalized inequalities.
• 1823 — Cauchy bases the derivative and integral on limits in his lectures, an enormous step toward rigor.
• 1850s–1870s — Karl Weierstrass, in his Berlin lectures, supplies the fully static, quantifier-precise $\epsilon$–$\delta$ definitions still used today, eliminating any appeal to motion or infinitesimals.
• 1872 — Weierstrass exhibits a continuous nowhere-differentiable function, demonstrating that only rigorous definitions can be trusted; Heine publishes related $\epsilon$–$\delta$ formulations crediting Weierstrass.
• 1960 — Abraham Robinson's nonstandard analysis rehabilitates infinitesimals on a rigorous logical basis, showing they were not inherently flawed — but $\epsilon$–$\delta$ remains the standard framework.

Subtleties & Limits

The $\epsilon$–$\delta$ definition, while logically airtight, raises deeper questions and has known limitations:

• Uniform vs. pointwise. Ordinary limits/continuity allow $\delta$ to depend on the point $a$. Uniform continuity demands a single $\delta (\epsilon )$ working for all points — a strictly stronger, often subtle condition (e.g., $1/x$ on $(0,1)$ is continuous but not uniformly so). Distinguishing these correctly trips up many students and matters greatly in integration and approximation theory.
• Order of limits. For functions of two variables, ${\lim }_{x\to 0}{\lim }_{y\to 0}f$ and ${\lim }_{y\to 0}{\lim }_{x\to 0}f$ can differ from each other and from the joint limit. The conditions under which limits and operations (integration, summation, differentiation) may be interchanged form the substance of advanced analysis (dominated/monotone convergence, uniform convergence theorems).
• Constructive content. Classically the definition asserts existence of $\delta$; constructive mathematics demands an explicit algorithm producing $\delta$ from $\epsilon$. Not all classically valid limit statements are constructively provable, raising foundational questions about what "the limit exists" means computationally.
• Generalization frontiers. Extending limits to nets and filters (for general topological spaces) and to nonstandard analysis (with genuine infinitesimals) shows the $\epsilon$–$\delta$ formulation is one of several equivalent — but not always interchangeable — viewpoints, each illuminating different aspects.

See Also