1. Definition and axioms
1.1 Formal definition of a pseudometric
Let \(X\) be a set. A function \(d : X \times X \to [0,\infty)\) is called a pseudometric on \(X\) if for all \(x,y,z \in X\) it satisfies:
- Nonnegativity: \(d(x,y)\ge 0\).
- Symmetry: \(d(x,y)=d(y,x)\).
- Triangle inequality: \(d(x,z)\le d(x,y)+d(y,z)\).
- Weakened identity of indiscernibles: it is not required that \(d(x,y)=0\) implies \(x=y\). Instead, it is only allowed that different points may have zero distance.
The key modification from a metric is the fourth axiom: distinct points can be treated as “distance-indistinguishable.”
1.2 Comparison with metrics
A metric is precisely a pseudometric in which the condition \(d(x,y)=0\) forces equality \(x=y\). Thus, every metric is a pseudometric, but not every pseudometric is a metric. The pseudometric can be viewed as collapsing some pairs of points into indistinguishable classes while preserving all other metric-like inequalities.
1.3 Relation to semimetrics and generalized distances
Terminology varies across fields. A semimetric typically drops or relaxes one of the metric axioms (commonly either symmetry or the triangle inequality). A pseudometric retains nonnegativity, symmetry, and the triangle inequality while relaxing only the identity requirement. More broadly, “generalized distance” functions arise in settings where constraints are imposed only partially—for example, allowing zero distances for distinct points, allowing asymmetry, or allowing infinite values—depending on the modeling needs.
2. Basic properties
2.1 Nonnegativity and symmetry
Nonnegativity ensures that the “distance-like” quantity is never negative, enabling interpretations in terms of magnitudes or costs. Symmetry means that exchanging the roles of the two points does not alter the value; this matches many geometric and measurement scenarios.
2.2 Triangle inequality consequences
The triangle inequality provides a form of consistency: traveling directly between two points cannot be more expensive than taking an intermediate route. As a consequence, pseudometric spaces inherit many structural features familiar from metrics, such as the stability of bounds and the behavior of balls under inclusion.
2.3 Zero-distance pairs and equivalence classes
Because \(d(x,y)=0\) may occur for distinct \(x\neq y\), the relation \[ x \sim y \quad \Longleftrightarrow \quad d(x,y)=0 \] is central. In a pseudometric space, \(\sim\) is an equivalence relation: it is reflexive (from \(d(x,x)\ge 0\) and triangle inequality considerations), symmetric (from symmetry), and transitive (from the triangle inequality: if \(d(x,y)=0\) and \(d(y,z)=0\), then \(d(x,z)\le d(x,y)+d(y,z)=0\)).
2.4 Induced preorders or quotient behavior (intuition)
A pseudometric does not necessarily distinguish points, so many statements are naturally formulated “up to” the equivalence relation \(\sim\). Intuitively, points inside the same zero-distance class behave like the same location in the induced geometry. This motivates viewing the pseudometric space as a metric space with some points “identified” or “glued together.”
3. Induced quotient metric
3.1 Equivalence relation from zero pseudodistance
Using \(x \sim y\) defined by \(d(x,y)=0\), one forms the set of equivalence classes \(X/{\sim}\). Each class consists of points that are indistinguishable from the perspective of the pseudometric.
3.2 Construction of the quotient space
Let \([x]\) denote the equivalence class of \(x\). Define a function \(\bar d : (X/{\sim}) \times (X/{\sim}) \to [0,\infty)\) by \[ \bar d([x],[y]) = d(x,y). \] This is well-defined: if \(x'\sim x\) and \(y'\sim y\), then triangle inequality and zero-distance properties show \(d(x',y')=d(x,y)\).
3.3 Proof sketch that the quotient distance is a metric
The constructed \(\bar d\) inherits nonnegativity, symmetry, and the triangle inequality directly from \(d\). For the identity axiom, if \(\bar d([x],[y])=0\), then \(d(x,y)=0\), hence \(x\sim y\) and \([x]=[y]\). Therefore \(\bar d\) is a metric on the quotient set.
3.4 Universality of the quotient (mapping viewpoint)
The quotient metric can be characterized by a universal property: any map from \(X\) to a metric space that “respects” zero-distance identification factors uniquely through the quotient. Concretely, if a function \(f:X\to Y\) satisfies \(d_X(x,x')=0 \Rightarrow f(x)=f(x')\), then there is an induced map \(\tilde f: X/{\sim}\to Y\). This reflects that the only ambiguity in \(X\) not visible to \(d\) is precisely the zero-distance equivalence.
4. Topological structure
4.1 Open balls and basic open sets
For a pseudometric space \((X,d)\), define open balls by \[ B_d(x,r)=\{y\in X : d(x,y)<r\}. \] These balls generate a natural collection of open sets. Because different points can have zero distance, balls centered at different but equivalent points coincide: if \(d(x,x')=0\), then \(B_d(x,r)=B_d(x',r)\) for all \(r>0\).
4.2 Topology induced by a pseudometric
The family of unions of open balls defines a topology on \(X\). This topology depends on the pseudometric, not merely on the quotient metric. However, the topology on \(X\) typically fails to be Hausdorff when there are nontrivial zero-distance identifications: distinct points may not be separable by disjoint neighborhoods because they share identical neighborhood structures whenever their distance is zero.
4.3 Sequences, limits, and continuity
Convergence can be defined using the induced topology: a sequence \(x_n\) converges to \(x\) if \(d(x_n,x)\to 0\). If \(x'\) lies in the same zero-distance class as \(x\), then \(x_n\to x\) implies \(x_n\to x'\) as well, since \(d(x_n,x')\le d(x_n,x)+d(x,x')=d(x_n,x)\) when \(d(x,x')=0\). Continuity of a map \(f:(X,d_X)\to (Y,d_Y)\) can be characterized via \(\varepsilon\)-\(\delta\) conditions expressed in terms of pseudometric distances.
4.4 Separation properties (T0/T1/T2) connections
Because points with zero distance cannot be separated by positive-distance neighborhoods, the separation axioms reflect the structure of the pseudometric:
- The topology is typically at least \(T_0\): if \(d(x,y)=0\) and \(d(y,x)=0\) (symmetry is automatic), then they cannot be distinguished by open sets in a stronger sense, but the exact relationship depends on whether the pseudometric distinguishes points via open balls.
- If the pseudometric is a metric, the induced topology becomes Hausdorff (\(T_2\)).
- Nontrivial zero-distance classes often prevent \(T_2\), while still allowing weaker separation depending on how equivalence classes overlap in terms of neighborhoods.
5. Cauchy sequences and completeness
5.1 Cauchy criteria under pseudometrics
A sequence \((x_n)\) in \((X,d)\) is Cauchy if for every \( \varepsilon>0 \) there exists \(N\) such that for all \(m,n\ge N\), \[ d(x_m,x_n)<\varepsilon. \] This definition matches the metric case word-for-word. The difference is that “Cauchy” may not force convergence to a unique point, because multiple points can represent the same limiting position in terms of distance.
5.2 Completeness notions and their behavior under quotienting
A pseudometric space is complete if every Cauchy sequence converges (in the pseudometric sense) to some point in the space. Because convergence is compatible with the quotient by zero-distance identification, completeness is closely tied to the completeness of the induced quotient metric space. In fact, passing to \(X/{\sim}\) often makes the completeness behavior easier to state: any Cauchy sequence in \(X\) induces a Cauchy sequence in the quotient, and convergence in the quotient corresponds to convergence up to zero distance in the original space.
5.3 Completeness vs. total boundedness (overview)
In metric geometry, total boundedness and completeness combine to yield compactness via standard theorems. For pseudometric spaces, an analogous relationship holds after quotienting, since compactness and coverings are sensitive to the separation properties lost under pseudometrics. Thus, it is common to formulate compactness-type statements for the quotient metric, where the topology becomes Hausdorff.
5.4 Examples illustrating different outcomes
- If the pseudometric collapses many points into equivalence classes, sequences may converge even when they do not converge to a single distinguishable point in \(X\).
- A pseudometric space can be non-complete because its quotient metric space is non-complete, even if some Cauchy behavior seems “stable” among representatives.
- Conversely, incompleteness in \(X\) may disappear after identifying zero-distance classes, producing a complete quotient.
6. Convergence and Cauchy filters
6.1 Filter-based reformulation
Filters generalize sequences and can capture convergence without relying on a countable index set. In a pseudometric space, a point \(x\) is a limit of a filter \(\mathcal{F}\) if for every \(r>0\), the set \(B_d(x,r)\) is in \(\mathcal{F}\). A filter is Cauchy if for every \(\varepsilon>0\) there exists some set \(A\in\mathcal{F}\) such that \(d(a,a')<\varepsilon\) for all \(a,a'\in A\).
6.2 Relationship to uniform structures
Pseudometrics naturally induce a uniform structure. This uniformity encodes all statements of the form “points get close” (or “images under a map get close”) without picking a specific notion of neighborhoods. Consequently, convergence, Cauchy filters, and uniform continuity can be treated uniformly across different spaces using the same underlying uniform framework.
6.3 Stability under operations (restrictions, products)
Uniform structures also behave well under standard constructions:
- Restricting a pseudometric to a subset yields an induced pseudometric, and Cauchy/convergence properties restrict accordingly.
- Products of pseudometric spaces can be endowed with natural product pseudometrics (e.g., using sums or maxima of component distances), under which product convergence and product Cauchy behavior mirror the component-wise behavior.
7. Examples and standard constructions
7.1 Trivial and near-trivial pseudometrics
Two extreme examples illustrate the concept:
- The zero pseudometric defined by \(d(x,y)=0\) for all \(x,y\) satisfies all pseudometric axioms. Its quotient has a single point.
- A pseudometric can be “almost metric” if it differs from a metric only by collapsing a small collection of pairs into zero-distance equivalence classes.
7.2 Quotient-derived pseudometrics
Starting from a metric space and identifying points via an equivalence relation that merges only points at zero metric distance, one obtains a pseudometric on the original set or a metric on the quotient set. This viewpoint is common in applications where observations are considered identical up to a chosen rule.
7.3 Pullbacks along functions
Given a function \(f:X\to Y\) and a pseudometric \(d_Y\) on \(Y\), one can define a pseudometric on \(X\) by \[ d_X(x,x') = d_Y(f(x),f(x')). \] Such a construction is called a pullback. It measures distance in \(X\) by how far images are in \(Y\). If \(f\) is not injective, distinct points with the same image have zero distance in \(X\).
7.4 Product and sum constructions
For pseudometric spaces \((X,d_X)\) and \((Y,d_Y)\), a pseudometric on the product \(X\times Y\) can be defined by combining component distances, such as \[ d((x,y),(x',y')) = d_X(x,x') + d_Y(y,y') \] or \[ d((x,y),(x',y')) = \max\{d_X(x,x'),\, d_Y(y,y')\}. \] Both preserve the pseudometric axioms and lead to convergence/Cauchy notions consistent with coordinate-wise behavior.
7.5 Pseudometrics from norms/seminorms
In linear spaces, a seminorm \(p\) may vanish on nonzero vectors. Defining \(d(x,y)=p(x-y)\) yields a pseudometric. The zero set of the seminorm determines which vectors are treated as equivalent. This is a standard route in functional analysis, where kernels of linear maps or constraints often naturally produce seminorms rather than norms.
8. Uniform continuity and Lipschitz maps
8.1 Lipschitz maps between pseudometric spaces
A map \(f:(X,d_X)\to (Y,d_Y)\) is Lipschitz if there exists \(L\ge 0\) such that \[ d_Y(f(x),f(x')) \le L\, d_X(x,x') \] for all \(x,x'\in X\). If \(L=0\), the map is constant on each zero-distance class of \(X\). Lipschitz maps preserve Cauchy sequences up to appropriate notions and are automatically uniformly continuous.
8.2 Uniform continuity characterization
A function \(f\) is uniformly continuous if for every \(\varepsilon>0\) there exists \(\delta>0\) such that whenever \(d_X(x,x')<\delta\), one has \(d_Y(f(x),f(x'))<\varepsilon\). The uniform version does not depend on the particular point in \(X\), making it well-suited to the uniform structure induced by pseudometrics.
8.3 Isometries and “pseudoisometries”
An isometry between pseudometric spaces usually means distance-preserving: \[ d_Y(f(x),f(x')) = d_X(x,x'). \] Such a map may still fail to be injective, because points in the same zero-distance class are indistinguishable under the pseudometric. In metric language this would collapse to an isometric embedding with no ambiguity; in the pseudometric context it is normal for the map to identify entire equivalence classes.
8.4 Behavior under quotient metrics
If one has a Lipschitz (or uniformly continuous) map \(f\) that identifies zero-distance points appropriately, it descends to a map on the quotient space. Distance estimates on \(X\) translate into corresponding estimates on \(X/{\sim}\), often simplifying proofs because the quotient is a genuine metric space.
9. Common applications in analysis and geometry
9.1 Seminormed spaces and induced pseudometrics
Many analytic constructions begin with seminorms. In a seminormed vector space, distance between two elements can be defined via the seminorm of their difference. This converts algebraic structure into geometric-like notions (balls, convergence, continuity), while respecting that some directions carry no “distance weight” because the seminorm vanishes there.
9.2 Function spaces with pseudometrics
On spaces of functions, pseudometrics commonly appear by measuring discrepancy modulo an equivalence relation, such as agreement almost everywhere or agreement up to a negligible set. For instance, two functions may differ on a set deemed irrelevant by the modeling framework, yet still be treated as having zero distance under the chosen pseudometric.
9.3 Identifying indistinguishable states in models
In modeling, one often records only what can be observed through measurements. If two states produce identical observations under the measurement procedure, they are effectively indistinguishable. A pseudometric captures this by assigning zero distance to states that cannot be told apart, while retaining meaningful distances between genuinely different observable behaviors.
9.4 Approximation and error measurement frameworks
Approximation schemes measure error using distance-like quantities. When errors are measured only up to an equivalence (e.g., gauge choices, symmetries, or representation redundancies), the induced notion of discrepancy is frequently only a pseudometric. This allows numerical or theoretical methods to remain consistent even when multiple representations correspond to the same underlying object.
10. Computation and practical considerations
10.1 Estimating pseudodistances from data
In data analysis, a pseudometric can represent a dissimilarity score where certain distinctions are not resolved by the available information. Estimation from samples may naturally yield zero scores for different objects if the features used cannot separate them, producing behavior consistent with pseudometric structure.
10.2 Regularization effects and identifiability
Regularization in inverse problems and machine learning sometimes induces equivalence classes: multiple parameter settings may lead to nearly identical predictions under the model. Distances measured between outputs can then behave pseudometrically on parameter space, reflecting limited identifiability—zero or near-zero distances indicate that parameters are not distinguishable given the imposed structure and data noise.
10.3 Numerical examples (conceptual)
Conceptually, one may consider an algorithm that compares objects only through a compressed representation. If the representation map identifies objects that share the same compressed features, then the distance computed in feature space pulls back to a pseudometric on the original set of objects. Computations then reflect indistinguishability as exact zero (or extremely small values) rather than enforcing strict separation.
11. Further topics
11.1 Pseudometric spaces and uniform spaces
Pseudometrics are a concrete way to realize uniform structures. Many theorems about uniform continuity and Cauchy behavior can be stated in the language of uniform spaces, with pseudometrics serving as a generating example class.
11.2 Gromov–Hausdorff-type ideas (high-level)
Metrics enable comparisons between spaces up to notion of closeness, such as Gromov–Hausdorff distance. Pseudometric generalizations appear in related ideas when one compares spaces modulo identifications or when correspondences induce only distance-like discrepancies rather than strict metric separation.
11.3 Extensions to extended pseudometrics
One can extend the codomain from \([0,\infty)\) to \([0,\infty]\), allowing infinite “distances.” The resulting structure is useful when some points are considered too far to relate via the given framework, while still preserving a form of triangle inequality. Such extensions require care in defining topology and convergence.
11.4 Related notions: ultrapseudometrics and embeddings (overview)
In model-theoretic and geometric analysis contexts, ultrapseudometrics arise when ultralimits are taken, producing distance structures suited to limiting processes. Embedding and approximation questions also relate: one studies when a pseudometric space can be faithfully represented inside a metric space, potentially after quotienting by zero-distance equivalence.