Must epsilon be extremely small?

No. It is a chosen positive radius, although mathematical arguments often consider arbitrarily small radii.

Is every epsilon-ball circular?

No. Its geometry depends on the metric and dimension.

Does a point at exactly epsilon count?

It counts in a closed ball and does not count in an open ball. The convention must be stated.

AI Glossary / Graphs and symbolic AI / Intermediate

Epsilon-ball

Definition

An epsilon-ball is a neighborhood of points within a specified distance epsilon of a chosen center. For a metric d and positive radius epsilon, the open ball contains points x with d(x, center) < epsilon. A closed ball instead uses ≤ and includes its boundary. In AI, balls help describe nearby examples, local search regions, and bounded input changes.

What it means

Epsilon is the radius, not a fixed universal constant. The meaning of “nearby” also depends on the distance measure. In ordinary two-dimensional Euclidean geometry, a ball is a disk. Other metrics produce different shapes even when the center and numerical radius stay the same.
The mathematical ball includes all qualifying points in the underlying space. A dataset's neighborhood contains only the observed records that fall inside that ball. Confusing those two sets can lead to claims about coverage or robustness that a finite collection of examples does not support.

How it works

Choose the space, center, distance function, radius, and boundary convention. Compute the distance from each candidate point to the center and compare it with the radius. A point on the boundary is excluded under the open convention and included under the closed convention.
For numerical features, preprocessing is part of the definition in practice. Measuring one feature in a larger numerical unit can change the neighborhood unless the metric or scaling accounts for it. When a ball is used as an allowed perturbation region, membership describes permitted mathematical changes; additional constraints may be needed for the changed input to remain meaningful.

Three practical examples

A local prediction method looks for training records inside a radius around a new feature vector, then uses those records according to a stated prediction rule.
A robustness experiment restricts test perturbations to a closed norm ball and reports the norm, radius, and valid-input constraints.
An optimization exercise examines whether a candidate point has a better objective value anywhere in a sufficiently small neighborhood.

Think of it this way

Think of drawing a service radius around a location, then checking which destinations lie inside it. The analogy explains a center and a distance limit. It is incomplete because road distance, straight-line distance, and travel time produce different neighborhoods, and AI feature spaces need not correspond to physical maps at all.

Worked example

Use center (0, 0) and radius 5. Under Euclidean distance, point A = (3, 4) is exactly 5 units away because the square root of 3 squared plus 4 squared is 5. A is outside the open ball but inside the closed ball. Point B = (1, 1) is about 1.414 units away and belongs to both.
Now use Manhattan distance, which adds absolute coordinate differences. A is 7 units away and belongs to neither radius-5 ball. Under maximum-coordinate distance, A is 4 units away and belongs even to the open ball.
The point did not move; the metric changed. These calculations show why a statement such as “all examples within epsilon were tested” is incomplete without its distance definition. They also show why a round drawing should not be treated as the shape of every ball.

Illustrative case study

In a fictional tutorial, the Birch team builds a nearest-neighborhood demonstration with two features: delivery distance and item weight. It initially uses raw values without documenting units. Changing distance from kilometers to meters causes the same radius to include a very different set of records.
The team keeps the original records, states a scaling rule, and regenerates the neighborhoods using the intended feature representation. It compares the selected records before interpreting predictions. The lesson concerns the definition of similarity; the example makes no claim that one scaling choice is universally correct or that the resulting predictions improve.

Common misconceptions and mistakes

A ball is not just its outer shell; the shell is a sphere under the corresponding distance. A numerical radius does not identify a neighborhood without a metric and center. Open versus closed boundaries matter in exact definitions, even when finite-precision calculations make comparisons delicate.
Do not infer that every nearby feature vector is a valid real-world example. A continuous change to a categorical encoding may not represent an allowed category. Likewise, successful tests on sampled points inside a ball do not prove that a model behaves correctly at every point in that ball.

Business use

The term helps teams specify local similarity and the scope of robustness tests. A useful description names the representation, normalization, distance, radius, and boundary rule. For a radius-based service, it should also say what happens when no records fall inside the neighborhood.
When reviewing a model evaluation, ask whether the neighborhood represents meaningful variations for the application. A mathematically convenient ball may be appropriate for one research question and inadequate for another. The concept supports precise questions; it does not independently validate the selected distance or business interpretation.

Check your understanding

With center 10 on the real line, ordinary absolute-distance metric, and radius 2, does 12 belong to the open epsilon-ball?

Answer: No. Its distance is exactly 2, whereas the open ball requires a distance strictly below 2. It belongs to the corresponding closed ball.

Related terms

Learn next

Learn next about distance metrics. Calculate the radius-2 neighborhoods of a small two-feature dataset under Euclidean and maximum-coordinate distance. Mark boundary points separately, then explain how a change of units affects the result when no normalization is applied.

Sources

Paige Bright, MIT original Introduction to Metric Spaces lecture transcript: https://ocw.mit.edu/courses/18-s190-introduction-to-metric-spaces-january-iap-2023/1LSNMR_R8hNIrja0hmgoOOqLgELeW1GMI_transcript.pdf University of Oklahoma, original metric-space exercise solutions defining open balls: https://math.ou.edu/~dmccullough/teaching/s10-4853/hw_soln2.pdf

Previous: Epoch Next: Epsilon greedy policy