In [metric spaces](/page/Metric%20Space), [sequences](/page/Sequence) characterise the [topology](/page/Topology) completely: a set is [closed](/page/Closed%20Set) if and only if it contains the limits of all its convergent sequences, a function is [continuous](/page/Continuous) if and only if it preserves sequential limits, and [compactness](/page/Compact%20Space) is equivalent to sequential compactness. These sequential characterisations are the workhorse of classical analysis, underpinning everything from the Bolzano--Weierstrass theorem to the Arzel\`a--Ascoli theorem. In general topological spaces, however, sequences fail to detect the topology. The fundamental obstruction is that a topology may have "too many" open sets relative to the countable index set $\mathbb{N}$: a sequence, being a function from $\mathbb{N}$, can only "probe" countably many neighbourhoods, and in spaces that are not [first-countable](/page/First-Countable%20Space), this is not enough to determine convergence. This failure is not a curiosity confined to pathological spaces --- it arises in mainstream mathematics whenever one works with uncountable products, [weak topologies](/page/Weak%20Topology) on non-separable Banach spaces, or spaces of distributions. [example: Sequential Closure Failure in an Uncountable Product] Consider the product space $X = \{0, 1\}^{\mathbb{R}}$, the set of all functions $f: \mathbb{R} \to \{0, 1\}$ equipped with the [product topology](/page/Product%20Topology). Each factor $\{0, 1\}$ carries the discrete topology. Define the subset \begin{align*} A := \{f \in X : f(t) = 0 \text{ for all but finitely many } t \in \mathbb{R}\}, \end{align*} consisting of functions with finite support. We claim that the constant function $g \equiv 1$ belongs to the [closure](/page/Closure) $\overline{A}$ but is not the limit of any sequence in $A$. **Closure membership.** A basic open set in the product topology has the form \begin{align*} U = \prod_{t \in \mathbb{R}} U_t, \quad \text{where } U_t = \{0, 1\} \text{ for all but finitely many } t. \end{align*} Suppose $g \in U$. Then $U_t = \{1\}$ for only finitely many values of $t$, say $t_1, \ldots, t_k$. The function $f$ defined by $f(t_i) = 1$ for $i = 1, \ldots, k$ and $f(t) = 0$ for all other $t$ belongs to both $A$ and $U$. Hence every basic open neighbourhood of $g$ meets $A$, so $g \in \overline{A}$. **No sequence in $A$ converges to $g$.** Suppose for contradiction that $(f_n)_{n \in \mathbb{N}}$ is a sequence in $A$ with $f_n \to g$ in the product topology. Each $f_n$ has finite support, so the set $S := \bigcup_{n=1}^\infty \operatorname{supp}(f_n)$ is a countable union of finite sets, hence countable. Since $\mathbb{R}$ is uncountable, there exists $t_0 \in \mathbb{R} \setminus S$. For this $t_0$, we have $f_n(t_0) = 0$ for every $n$, so $\pi_{t_0}(f_n) = 0$ for all $n$. But $\pi_{t_0}(g) = 1$, and since the projection $\pi_{t_0}: X \to \{0,1\}$ is continuous, convergence $f_n \to g$ would require $\pi_{t_0}(f_n) \to \pi_{t_0}(g) = 1$, a contradiction. [/example] This example reveals a genuine gap in the theory: the [closure](/page/Closure) page on Androma notes that in metric spaces, $x \in \overline{A}$ if and only if some sequence in $A$ converges to $x$, but in the product space above, the sequential closure of $A$ (the set of all sequential limits from $A$) is strictly smaller than $\overline{A}$. Sequences cannot reach every point of the closure. Two solutions to this problem emerged in the 1930s and 1940s. **Nets**, introduced by Moore and Smith (1922) and systematised by Kelley (1955), generalise sequences by replacing the index set $\mathbb{N}$ with an arbitrary directed set --- a partially ordered set in which every finite subset has an upper bound. **Filters**, introduced by Cartan (1937), take a dual approach: instead of tracking a generalised sequence of points, one tracks a family of "large" subsets of the space that are eventually entered. Both frameworks recover the full power of sequential characterisations in arbitrary topological spaces, and they are equivalent in a precise sense: every net determines a filter (its "eventuality filter"), and every filter determines a class of nets (any net whose tails generate the filter). This page develops both theories in parallel, emphasising the problems each solves and the contexts in which each is most natural. ## Definition The key idea behind nets is to replace the index set $\mathbb{N}$, ordered by $\leq$, with a more general directed set. The directed set must be rich enough to "probe" all neighbourhoods of a point, which requires that any two elements have a common upper bound --- this is the property that $\mathbb{N}$ uses implicitly every time we write "for all $n \geq N$." [definition: Directed Set] A **directed set** is a pair $(D, \preceq)$ where $D$ is a non-empty set and $\preceq$ is a binary relation on $D$ satisfying: 1. **Reflexivity:** $\alpha \preceq \alpha$ for all $\alpha \in D$. 2. **Transitivity:** If $\alpha \preceq \beta$ and $\beta \preceq \gamma$, then $\alpha \preceq \gamma$. 3. **Directedness:** For every $\alpha, \beta \in D$, there exists $\gamma \in D$ with $\alpha \preceq \gamma$ and $\beta \preceq \gamma$. A directed set is a preordered set (reflexive and transitive, but not necessarily antisymmetric) with the additional property that every pair of elements has an upper bound. [/definition] The natural numbers $(\mathbb{N}, \leq)$ are the prototypical directed set, but the definition is far more permissive. The neighbourhood filter of a point, ordered by reverse inclusion, is a directed set --- and this is precisely the directed set that makes nets powerful in general topology. [example: The Neighbourhood Directed Set] Let $(X, \tau)$ be a topological space and let $x \in X$. Define $\mathcal{N}(x) := \{U \in \tau : x \in U\}$, the collection of all open neighbourhoods of $x$. The relation \begin{align*} U \preceq V \quad :\Longleftrightarrow \quad V \subset U \end{align*} (reverse inclusion) makes $(\mathcal{N}(x), \preceq)$ into a directed set: reflexivity and transitivity are immediate from the corresponding properties of $\subset$, and directedness holds because $U \cap V$ is an open neighbourhood of $x$ with $U \preceq U \cap V$ and $V \preceq U \cap V$. In a metric space, the countable subfamily $\{B(x, 1/n) : n \in \mathbb{N}\}$ is cofinal in $\mathcal{N}(x)$ (every open neighbourhood of $x$ contains some $B(x, 1/n)$). This is why sequences, indexed by $\mathbb{N}$, suffice in metric spaces. In a non-first-countable space, $\mathcal{N}(x)$ has no countable cofinal subset, and $\mathbb{N}$ is too small to serve as an index set. [/example] With directed sets in hand, the definition of a net is immediate. [definition: Net] Let $(X, \tau)$ be a topological space and $(D, \preceq)$ a directed set. A **net** in $X$ is a function \begin{align*} s: D &\to X \\ \alpha &\mapsto s_\alpha. \end{align*} We write $(s_\alpha)_{\alpha \in D}$ for the net, in analogy with the sequence notation $(a_n)_{n \in \mathbb{N}}$. The net $(s_\alpha)_{\alpha \in D}$ **converges** to a point $x \in X$, written $s_\alpha \to x$, if for every open set $U$ containing $x$, there exists $\alpha_0 \in D$ such that $s_\alpha \in U$ for all $\alpha \succeq \alpha_0$. In this case, $x$ is called a **limit** of the net. [/definition] When $D = \mathbb{N}$ with the usual ordering, a net is a sequence and this definition reduces to the standard definition of sequential convergence. In a [Hausdorff space](/page/Hausdorff%20Space), limits of nets are unique (the proof is identical to the sequential case: if $x \neq y$, choose disjoint open sets $U \ni x$ and $V \ni y$, then the net is eventually in both $U$ and $V$, contradicting directedness). [remark: Nets vs Sequences --- What Changes] The passage from sequences to nets is not merely a matter of replacing $\mathbb{N}$ with a larger index set. There are several subtle differences: 1. **Uncountable index sets.** A net may be indexed by an uncountable directed set. The net "probing all neighbourhoods" of a point $x$ in a non-first-countable space requires an uncountable index set. 2. **Non-linear ordering.** A directed set need not be totally ordered. Two indices $\alpha$ and $\beta$ may be incomparable (neither $\alpha \preceq \beta$ nor $\beta \preceq \alpha$), so a net does not have a well-defined "order of evaluation." 3. **Non-antisymmetric ordering.** Different indices $\alpha \neq \beta$ may satisfy both $\alpha \preceq \beta$ and $\beta \preceq \alpha$, so the preorder on $D$ is not necessarily a partial order. Despite these differences, the "eventually" quantifier --- "there exists $\alpha_0$ such that for all $\alpha \succeq \alpha_0$..." --- works exactly as it does for sequences. [/remark] ## Characterising Topology via Nets The central payoff of nets is that they recover the full sequential characterisations of closure, continuity, and compactness, without any first-countability hypothesis. In metric spaces, these characterisations are routine consequences of the countable neighbourhood base at each point. In general topological spaces, they require the richer indexing that nets provide. ### Closure The [closure](/page/Closure) of a set $A$ in a metric space consists of all sequential limits from $A$. In a general topological space, the sequential closure may be strictly smaller than the topological closure, as we saw in the opening example. Nets repair this gap. [quotetheorem:1045] The forward direction uses the neighbourhood directed set from the example above: if $x \in \overline{A}$, then every open neighbourhood $U$ of $x$ meets $A$, so we can choose $s_U \in U \cap A$ for each $U \in \mathcal{N}(x)$. The resulting net $(s_U)_{U \in \mathcal{N}(x)}$, indexed by $(\mathcal{N}(x), \preceq)$ with reverse inclusion, converges to $x$ by construction: given any open set $V \ni x$, the net is eventually in $V$ because for all $U \succeq V$ (i.e., $U \subset V$), we have $s_U \in U \subset V$. The converse is immediate: if $s_\alpha \to x$ and each $s_\alpha \in A$, then every open set containing $x$ eventually contains $s_\alpha \in A$, hence meets $A$, so $x \in \overline{A}$. This characterisation has a direct corollary for closed sets: a set $A$ is closed if and only if it contains the limits of all convergent nets with values in $A$. The analogous sequential statement fails in general --- a set can be sequentially closed (closed under sequential limits) without being closed --- but the net statement holds universally. ### Continuity In metric spaces, continuity is equivalent to preserving sequential limits. This equivalence extends to nets without restriction. [quotetheorem:1046] The forward direction is immediate: if $f$ is continuous and $V$ is an open neighbourhood of $f(x)$, then $f^{-1}(V)$ is an open neighbourhood of $x$, so the net is eventually in $f^{-1}(V)$, hence the image net is eventually in $V$. The converse proceeds by contradiction: if $f$ is not continuous, there exists an open $V \subset Y$ with $f^{-1}(V)$ not open in $X$. Then there exists $x \in f^{-1}(V)$ such that no open neighbourhood of $x$ is contained in $f^{-1}(V)$. Constructing a net indexed by the neighbourhood directed set of $x$, choosing $s_U \in U \setminus f^{-1}(V)$ for each $U$, yields a net converging to $x$ whose image is never in $V$, contradicting the hypothesis. This theorem is particularly useful in functional analysis when verifying continuity of maps between spaces carrying weak or weak-* topologies, where the topology is defined by families of seminorms or functionals rather than by a single metric. ### Hausdorffness In Hausdorff spaces, limits of sequences are unique. The same holds for nets, and the converse is also true --- a fact that has no sequential analogue. [quotetheorem:1047] The forward direction is identical to the sequential case. For the converse, suppose $X$ is not Hausdorff: there exist distinct $x, y \in X$ such that every pair of open neighbourhoods $U \ni x$, $V \ni y$ satisfies $U \cap V \neq \varnothing$. Consider the directed set $D = \mathcal{N}(x) \times \mathcal{N}(y)$, ordered by $(U_1, V_1) \preceq (U_2, V_2)$ iff $U_2 \subset U_1$ and $V_2 \subset V_1$. For each $(U, V) \in D$, choose $s_{(U,V)} \in U \cap V$. This net converges to both $x$ and $y$, so limits are not unique. The sequential analogue fails: there exist non-Hausdorff spaces in which every convergent sequence has a unique limit (for example, the cocountable topology on an uncountable set). ## Subnets and Cluster Points In the theory of sequences, subsequences play a central role: the Bolzano--Weierstrass theorem guarantees that every bounded sequence in $\mathbb{R}^n$ has a convergent subsequence, and this subsequential compactness is the foundation of many existence arguments. Extending this idea to nets requires care, because the naive generalisation --- restricting a net to a cofinal subset of the index set --- is too restrictive. <!-- NOTATION PROPOSAL: We use the notation (s_\alpha)_{\alpha \in D} consistently for nets, with Greek letters for directed-set indices and Roman letters for sequence indices. This is standard in Kelley and Willard but not yet recorded on the Androma Notation page. --> [definition: Subnet] Let $(s_\alpha)_{\alpha \in D}$ be a net in a set $X$, and let $(E, \preceq_E)$ be another directed set. A **subnet** of $(s_\alpha)_{\alpha \in D}$ is a net $(s_{\varphi(\beta)})_{\beta \in E}$ where $\varphi: E \to D$ is a function satisfying: 1. **Monotonicity:** If $\beta_1 \preceq_E \beta_2$, then $\varphi(\beta_1) \preceq \varphi(\beta_2)$. 2. **Cofinality:** For every $\alpha \in D$, there exists $\beta \in E$ with $\varphi(\beta) \succeq \alpha$. Equivalently, $\varphi$ is an order-preserving map whose image is cofinal in $D$. [/definition] The cofinality condition ensures that the subnet "eventually reaches" every part of the original directed set, while monotonicity ensures that the subnet respects the directional structure. A subsequence of a sequence $(a_n)_{n \in \mathbb{N}}$ is a subnet in this sense: the map $\varphi: \mathbb{N} \to \mathbb{N}$ defined by $\varphi(k) = n_k$ is order-preserving (since $n_1 < n_2 < \cdots$) and cofinal (since $n_k \geq k$). [remark: Willard's Broader Definition] Some authors, notably Willard, use a weaker definition of subnet that relaxes monotonicity to a weaker "eventual" condition: $\varphi: E \to D$ is required only to satisfy the property that for every $\alpha \in D$, there exists $\beta_0 \in E$ such that $\varphi(\beta) \succeq \alpha$ for all $\beta \succeq_E \beta_0$. Under this definition, the composition $s \circ \varphi$ is still eventually in every set that the original net is eventually in, which suffices for the convergence theory. The Kelley (order-preserving) definition is more commonly used in modern treatments and is the one adopted here. [/remark] The concept of a cluster point captures what it means for a net to "visit a neighbourhood infinitely often" --- the net-theoretic analogue of a subsequential limit point. [definition: Cluster Point of a Net] Let $(s_\alpha)_{\alpha \in D}$ be a net in a topological space $(X, \tau)$. A point $x \in X$ is a **cluster point** of the net if for every open set $U$ containing $x$ and every $\alpha_0 \in D$, there exists $\alpha \succeq \alpha_0$ with $s_\alpha \in U$. Equivalently, $x$ is a cluster point if and only if the net is **frequently** in every neighbourhood of $x$: the set $\{\alpha \in D : s_\alpha \in U\}$ is cofinal in $D$ for every open $U \ni x$. [/definition] The relationship between cluster points and subnets mirrors the relationship between subsequential limits and subsequences: a point $x$ is a cluster point of a net if and only if some subnet converges to $x$. This equivalence is fundamental to the net-theoretic proof of Tychonoff's theorem. [quotetheorem:1048] The construction of the subnet from a cluster point mirrors the "diagonal" argument for sequences but uses a more elaborate directed set. Define $E = \{(\alpha, U) \in D \times \mathcal{N}(x) : s_\alpha \in U\}$, ordered by $(\alpha_1, U_1) \preceq_E (\alpha_2, U_2)$ iff $\alpha_1 \preceq \alpha_2$ and $U_2 \subset U_1$. The directedness of $E$ uses the cluster point hypothesis: given $(\alpha_1, U_1)$ and $(\alpha_2, U_2)$, choose $\gamma \succeq \alpha_1, \alpha_2$ in $D$ (by directedness of $D$), then use the cluster point property to find $\alpha \succeq \gamma$ with $s_\alpha \in U_1 \cap U_2$; the pair $(\alpha, U_1 \cap U_2)$ is an upper bound. The map $\varphi: E \to D$, $(\alpha, U) \mapsto \alpha$ is order-preserving and cofinal, and the subnet $s_{\varphi(\beta)}$ converges to $x$ by construction. ## Filters and Filter Bases Nets generalise sequences by enriching the index set. Filters take a fundamentally different approach: instead of tracking individual points visited by a generalised sequence, a filter tracks the *collections of sets* that a convergent object must eventually enter. This shift from "which points are visited" to "which sets are large" turns out to be algebraically and set-theoretically cleaner in many contexts, particularly in compactness arguments. The motivation comes from examining what convergence of a net actually requires. When a net $(s_\alpha)_{\alpha \in D}$ converges to $x$, the essential information is not the values $s_\alpha$ themselves, but the collection of sets that the net is eventually in. This collection --- the "eventuality filter" --- is a family of subsets of $X$ that is closed under finite intersections and supersets, and that does not contain the empty set. Abstracting these properties gives the definition of a filter. [definition: Filter] Let $X$ be a non-empty set. A **filter** on $X$ is a non-empty collection $\mathcal{F} \subset \mathcal{P}(X)$ of subsets of $X$ satisfying: 1. **Non-degeneracy:** $\varnothing \notin \mathcal{F}$. 2. **Upward closure:** If $A \in \mathcal{F}$ and $A \subset B \subset X$, then $B \in \mathcal{F}$. 3. **Finite intersection:** If $A, B \in \mathcal{F}$, then $A \cap B \in \mathcal{F}$. [/definition] The axioms encode the three properties that the collection of "eventual" sets of a convergent net must possess. Non-degeneracy says that the net actually visits something (the empty set is never entered). Upward closure says that if the net is eventually in $A$, it is eventually in any superset of $A$. Finite intersection says that if the net is eventually in both $A$ and $B$, it is eventually in $A \cap B$ (choose an index beyond the thresholds for both). [example: The Neighbourhood Filter] Let $(X, \tau)$ be a topological space and $x \in X$. The **neighbourhood filter** of $x$ is \begin{align*} \mathcal{N}(x) := \{A \subset X : \text{there exists an open } U \text{ with } x \in U \subset A\}. \end{align*} This is a filter on $X$: the empty set contains no neighbourhood of $x$; any superset of a neighbourhood is a neighbourhood; and the intersection of two neighbourhoods is a neighbourhood (since $U_1 \cap U_2$ is open and contains $x$). [/example] The neighbourhood filter is a filter on $X$ defined using the ambient topology. A quite different filter arises from any net, by recording which sets the net eventually enters. [example: The Eventuality Filter of a Net] Let $(s_\alpha)_{\alpha \in D}$ be a net in $X$. For each $\alpha_0 \in D$, define the **tail** (or **residual set**) \begin{align*} T_{\alpha_0} := \{s_\alpha : \alpha \succeq \alpha_0\}. \end{align*} The **eventuality filter** (or **filter generated by the tails**) is \begin{align*} \mathcal{F}_s := \{A \subset X : T_{\alpha_0} \subset A \text{ for some } \alpha_0 \in D\}. \end{align*} This is a filter: $\varnothing$ contains no tail (since every tail is non-empty by directedness); supersets of eventual sets are eventual; and if $T_{\alpha_0} \subset A$ and $T_{\beta_0} \subset B$, then choosing $\gamma \succeq \alpha_0, \beta_0$ gives $T_\gamma \subset T_{\alpha_0} \cap T_{\beta_0} \subset A \cap B$. The net $(s_\alpha)$ converges to $x$ if and only if $\mathcal{N}(x) \subset \mathcal{F}_s$ --- every neighbourhood of $x$ is an eventual set of the net. This reformulation connects nets to filters. [/example] In practice, specifying a filter by listing all its members is unwieldy. Just as a topology is often given via a basis, a filter is often given via a *filter base*. [definition: Filter Base] Let $X$ be a non-empty set. A **filter base** on $X$ is a non-empty collection $\mathcal{B} \subset \mathcal{P}(X)$ satisfying: 1. **Non-degeneracy:** $\varnothing \notin \mathcal{B}$. 2. **Finite intersection property:** For every $B_1, B_2 \in \mathcal{B}$, there exists $B_3 \in \mathcal{B}$ with $B_3 \subset B_1 \cap B_2$. The **filter generated** by $\mathcal{B}$ is \begin{align*} \mathcal{F}(\mathcal{B}) := \{A \subset X : B \subset A \text{ for some } B \in \mathcal{B}\}. \end{align*} [/definition] With filters and filter bases in hand, we can define convergence and cluster points purely in terms of set membership, without any reference to an index set or an ordering. [definition: Convergence and Cluster Points of Filters] Let $(X, \tau)$ be a topological space and $\mathcal{F}$ a filter on $X$. 1. The filter $\mathcal{F}$ **converges** to $x \in X$, written $\mathcal{F} \to x$, if $\mathcal{N}(x) \subset \mathcal{F}$ --- every neighbourhood of $x$ belongs to $\mathcal{F}$. 2. A point $x \in X$ is a **cluster point** of $\mathcal{F}$ if $U \cap F \neq \varnothing$ for every open $U \ni x$ and every $F \in \mathcal{F}$. [/definition] Convergence of a filter to $x$ means that the filter is "at least as fine as" the neighbourhood filter of $x$: it contains all the sets that a convergent object should eventually enter. A cluster point is a weaker notion --- the filter "frequently" visits every neighbourhood, but need not be eventually in each one. ## Ultrafilters and Compactness The most striking application of filters is to compactness, where they provide a proof of [Tychonoff's theorem](/page/Product%20Topology) that is both cleaner and more direct than the net-based argument. The key tool is the ultrafilter --- a filter that is "maximally fine" in the sense that it cannot be enlarged without becoming degenerate. [definition: Ultrafilter] A filter $\mathcal{U}$ on a set $X$ is an **ultrafilter** if it is not properly contained in any other filter on $X$. Equivalently, $\mathcal{U}$ is an ultrafilter if and only if for every $A \subset X$, either $A \in \mathcal{U}$ or $X \setminus A \in \mathcal{U}$. [/definition] The equivalence between maximality and the "dichotomy" property is the crucial characterisation. If a filter $\mathcal{F}$ is not an ultrafilter, there exists a set $A$ such that neither $A$ nor $X \setminus A$ belongs to $\mathcal{F}$. In this case, $\mathcal{F}$ can be enlarged: since $F \cap A \neq \varnothing$ for all $F \in \mathcal{F}$ (otherwise $F \subset X \setminus A$ would force $X \setminus A \in \mathcal{F}$ by upward closure), the collection $\mathcal{F} \cup \{A\}$ has the finite intersection property and generates a strictly larger filter. [example: Principal and Non-Principal Ultrafilters] For any $x \in X$, the collection $\mathcal{U}_x := \{A \subset X : x \in A\}$ is an ultrafilter, called the **principal ultrafilter** at $x$. The dichotomy is immediate: for any $A \subset X$, either $x \in A$ or $x \in X \setminus A$. On a finite set, every ultrafilter is principal: if $\mathcal{U}$ is an ultrafilter on a finite set $X = \{x_1, \ldots, x_n\}$, then $X = \{x_1\} \cup \cdots \cup \{x_n\}$ belongs to $\mathcal{U}$, and the dichotomy property applied iteratively forces some singleton $\{x_k\} \in \mathcal{U}$, making $\mathcal{U} = \mathcal{U}_{x_k}$. On an infinite set, **non-principal** (or **free**) ultrafilters exist but cannot be constructed explicitly --- their existence requires Zorn's Lemma (or equivalently, the Boolean Prime Ideal Theorem, which is strictly weaker than the full Axiom of Choice). A non-principal ultrafilter $\mathcal{U}$ on $\mathbb{N}$ contains the cofinite filter $\{A \subset \mathbb{N} : \mathbb{N} \setminus A \text{ is finite}\}$ but no finite set. Such ultrafilters are the foundation of ultraproduct constructions in model theory and nonstandard analysis. [/example] The connection between ultrafilters and compactness is remarkably clean. [quotetheorem:1049] The proof reveals why the open-cover definition and the filter-convergence definition are two faces of the same coin. Suppose $X$ is compact and $\mathcal{U}$ is an ultrafilter that does not converge. Then for each $x \in X$, there is an open $U_x \ni x$ with $U_x \notin \mathcal{U}$, so $X \setminus U_x \in \mathcal{U}$ by the dichotomy. The cover $\{U_x\}_{x \in X}$ has a finite subcover $\{U_{x_1}, \ldots, U_{x_k}\}$, giving $X \setminus U_{x_1} \cap \cdots \cap X \setminus U_{x_k} = X \setminus (U_{x_1} \cup \cdots \cup U_{x_k}) = \varnothing \in \mathcal{U}$, contradicting non-degeneracy. Conversely, if every ultrafilter converges and $\{V_i\}_{i \in I}$ is an open cover with no finite subcover, then $\{X \setminus V_i\}_{i \in I}$ has the finite intersection property, generates a filter, which extends to an ultrafilter $\mathcal{U}$ by Zorn's Lemma. This ultrafilter converges to some $x$, and since $x \in V_{i_0}$ for some $i_0$, we have $V_{i_0} \in \mathcal{U}$ (as a neighbourhood of $x$) and $X \setminus V_{i_0} \in \mathcal{U}$ (by construction), so $\varnothing = V_{i_0} \cap (X \setminus V_{i_0}) \in \mathcal{U}$, a contradiction. This characterisation makes the proof of Tychonoff's theorem almost immediate: an ultrafilter $\mathcal{U}$ on $\prod_{i \in I} X_i$ projects to an ultrafilter $\pi_i(\mathcal{U})$ on each factor $X_i$ (where $\pi_i(\mathcal{U}) := \{\pi_i(U) : U \in \mathcal{U}\}$, which is an ultrafilter because $\pi_i$ is surjective and preserves the dichotomy). If each $X_i$ is compact, $\pi_i(\mathcal{U})$ converges to some $x_i \in X_i$. The point $x = (x_i)_{i \in I}$ is then the limit of $\mathcal{U}$ in the product topology, because the subbasis sets $\pi_i^{-1}(V)$ (with $V$ open in $X_i$ containing $x_i$) belong to $\mathcal{U}$ (since $V \in \pi_i(\mathcal{U})$ means $\pi_i^{-1}(V) \in \mathcal{U}$), and these subbasis sets generate the product topology. The filter proof is arguably the most natural proof of Tychonoff's theorem because the Axiom of Choice enters in exactly one place --- the existence of ultrafilter extensions --- rather than being distributed throughout a transfinite induction or Zorn's Lemma argument on open covers. ## The Equivalence Between Nets and Filters The theories of nets and filters are not competing frameworks but two languages for the same underlying concept. Every net determines a filter (its eventuality filter), and every filter determines a canonical net. These correspondences preserve convergence and cluster points. [quotetheorem:1050] The verification is a direct unwinding of the definitions. For part (1), the eventuality filter converges to $x$ iff every neighbourhood of $x$ is an eventual set of the net, which is exactly the definition of net convergence. For part (2), the canonical net $(s_{(y,F)})$ converges to $x$ iff for every open $U \ni x$, there exists $(y_0, F_0) \in D_\mathcal{F}$ such that $s_{(y,F)} = y \in U$ whenever $(y,F) \succeq (y_0, F_0)$, i.e., whenever $F \subset F_0$. Taking $F_0 = U$ (which belongs to $\mathcal{F}$ since $\mathcal{F}$ converges to $x$, so $U \in \mathcal{F}$), every $(y, F)$ with $F \subset U$ satisfies $y \in F \subset U$, confirming convergence. [explanation: When to Use Nets vs Filters] Both nets and filters characterise the topology of any space, so the choice between them is one of convenience, not of power. Some guidelines: **Nets are more natural when:** the convergence argument involves a concrete limiting process --- taking finer partitions, smaller parameters, larger truncations --- where the directed set has a natural interpretation. Most arguments in analysis use nets (or sequences), because the objects under study (functions, operators, measures) are naturally indexed by parameters. **Filters are more natural when:** the argument is about set-theoretic or lattice-theoretic structure --- compactness, ultrafilter extensions, products. Filters treat convergence as a property of a family of sets, which meshes well with arguments involving the Axiom of Choice, Zorn's Lemma, and Boolean algebra. The ultrafilter proof of Tychonoff's theorem is the paradigmatic example. **In functional analysis:** nets are standard. Weak convergence $x_\alpha \rightharpoonup x$ in a Banach space $X$ means that $f(x_\alpha) \to f(x)$ for every $f \in X^*$, and this is most naturally phrased in terms of nets when the weak topology is not metrizable. However, the Eberlein--Smulian theorem shows that for bounded subsets of Banach spaces, weak sequential convergence and weak net convergence coincide, so sequences often suffice in practice. **In general topology:** filters are standard for compactness arguments (Tychonoff, Stone--Cech compactification, ultrafilter convergence), while nets are standard for separation axioms and continuity arguments. [/explanation] ## Net-Theoretic Compactness The net-theoretic characterisation of compactness is the direct generalisation of the Bolzano--Weierstrass theorem: a space is compact if and only if every net has a convergent subnet. This provides an alternative to the open-cover definition that is closer in spirit to the sequential compactness used in metric-space analysis. [quotetheorem:1051] The equivalence with the ultrafilter characterisation is immediate: if every ultrafilter converges, then every net has a cluster point (the eventuality filter of the net extends to an ultrafilter, which converges to some $x$, and $x$ is a cluster point of the net). Conversely, an ultrafilter *is* the eventuality filter of its canonical net, so if every net has a cluster point, every ultrafilter has a cluster point, and a cluster point of an ultrafilter is a limit (since for any open $U \ni x$, either $U \in \mathcal{U}$ or $X \setminus U \in \mathcal{U}$; if $X \setminus U \in \mathcal{U}$, then the cluster point condition requires $U \cap (X \setminus U) \neq \varnothing$, a contradiction). This characterisation recovers the Bolzano--Weierstrass theorem as a special case: in $\mathbb{R}^n$ with the standard topology, a subset $K$ is compact if and only if every sequence in $K$ has a convergent subsequence (since $\mathbb{R}^n$ is metrizable, sequential compactness and compactness coincide, and subsequences are subnets). In non-metrizable spaces like the weak-* topology on the dual of a non-separable Banach space, the net characterisation provides what sequential compactness cannot: the [Banach--Alaoglu theorem](/page/Compact%20Space) guarantees that the closed unit ball of $X^*$ is weak-* compact, meaning every net in the ball has a convergent subnet, even though convergent subsequences may not exist. [example: Compactness of $\{0,1\}^I$ via Nets] We return to the space $X = \{0,1\}^I$ for an arbitrary index set $I$, equipped with the product topology. By Tychonoff's theorem, $X$ is compact (each factor $\{0,1\}$ is compact). We illustrate the net characterisation directly. Let $(f_\alpha)_{\alpha \in D}$ be a net in $X$. For each coordinate $i \in I$, the projected net $(\pi_i(f_\alpha))_{\alpha \in D}$ is a net in the finite discrete space $\{0,1\}$. We construct a convergent subnet by a transfinite selection argument. Well-order $I = \{i_\beta : \beta < \kappa\}$ for some ordinal $\kappa$. Define a decreasing family of cofinal subsets of $D$ by transfinite induction: set $D_0 = D$. Given $D_\beta$ cofinal in $D$, the net restricted to $D_\beta$ takes values in $\{0, 1\}$ at coordinate $i_\beta$. Since $\{0,1\}$ is finite, at least one value --- say $c_\beta \in \{0,1\}$ --- is assumed cofinally, meaning $D_{\beta+1} := \{\alpha \in D_\beta : f_\alpha(i_\beta) = c_\beta\}$ is cofinal in $D_\beta$ (hence in $D$). At limit ordinals $\lambda$, define $D_\lambda$ by choosing a cofinal subset common to all $D_\beta$ with $\beta < \lambda$ (this requires the Axiom of Choice, applied to the directed set structure). The net restricted to $D_\kappa$ converges to the point $g \in X$ defined by $g(i_\beta) = c_\beta$: for each basic open set $\pi_{i_{\beta_1}}^{-1}(c_{\beta_1}) \cap \cdots \cap \pi_{i_{\beta_k}}^{-1}(c_{\beta_k})$ containing $g$, the net is eventually in this set once the index exceeds the threshold from step $\max(\beta_1, \ldots, \beta_k)$. [/example] ## Standard Techniques with Nets and Filters Working with nets and filters in practice requires a small toolkit of recurring argument patterns. These techniques appear throughout general topology, functional analysis, and abstract measure theory whenever sequential methods are insufficient. ### The Subnet Extraction Argument This is the net-theoretic analogue of the "extract a convergent subsequence" argument that pervades metric-space analysis. The pattern is: 1. **Start with a net.** Typically arising from a limiting process (finer and finer approximations, or elements of a directed family). 2. **Apply compactness** to extract a convergent subnet. 3. **Identify the limit** using continuity or a uniqueness argument. 4. **Promote to full convergence.** If the original net has a unique cluster point (e.g., because the limit is determined by a weak formulation or a uniqueness theorem), then the full net converges, not just the subnet. The last step deserves emphasis: in a compact space, if every convergent subnet of a net converges to the same limit $x$, then the original net converges to $x$. (If not, there is a neighbourhood $U$ of $x$ such that the net is frequently outside $U$; extracting a subnet from the complement and using compactness gives a subnet converging to some $y \neq x$, contradicting uniqueness.) ### The Filter Refinement Argument In filter-based arguments, the analogue of "extract a subsequence" is "extend to a finer filter." The pattern is: 1. **Start with a filter** $\mathcal{F}$ (often the eventuality filter of a net, or the neighbourhood filter of a point). 2. **Extend to an ultrafilter** $\mathcal{U} \supset \mathcal{F}$ using Zorn's Lemma. 3. **Apply the ultrafilter convergence characterisation** of compactness to find a limit. 4. **Transfer properties** from $\mathcal{U}$ back to $\mathcal{F}$ (the limit of $\mathcal{U}$ is a cluster point of $\mathcal{F}$, since $\mathcal{F} \subset \mathcal{U}$). ### The Weak and Weak-* Convergence Pattern In functional analysis, the [weak topology](/page/Weak%20Topology) on a Banach space $X$ and the weak-* topology on $X^*$ are typically not metrizable on the whole space. Net convergence is the correct language: a net $(x_\alpha)_{\alpha \in D}$ in $X$ converges weakly to $x$ if and only if $f(x_\alpha) \to f(x)$ for every $f \in X^*$. When the space is separable and reflexive, the Eberlein--Smulian theorem guarantees that bounded weakly convergent nets can be replaced by sequences, but the general theory requires nets. The [Banach--Alaoglu theorem](/page/Compact%20Space) states that the closed unit ball $B_{X^*}$ is compact in the weak-* topology. For non-separable $X$, this compactness is genuinely net-theoretic: every net in $B_{X^*}$ has a weak-* convergent subnet, but there may exist sequences in $B_{X^*}$ with no weak-* convergent subsequence. This distinction is critical in the theory of $L^\infty$ and spaces of measures, where the predual is typically not separable. ### Verifying Continuity via Nets To verify that a map $f: X \to Y$ between topological spaces is continuous, one can verify the net criterion: show that $s_\alpha \to x$ implies $f(s_\alpha) \to f(x)$ for every net. In practice, this is most useful when $X$ carries a topology defined by a family of seminorms or by duality (e.g., weak, weak-*, or initial topologies), where the open sets are difficult to describe explicitly but convergence has a clean characterisation. [example: Continuity of the Evaluation Map] Let $X$ be a Banach space and consider the evaluation map at a fixed functional $f \in X^*$: \begin{align*} \hat{f}: (X, \text{weak}) &\to \mathbb{R} \\ x &\mapsto f(x). \end{align*} To verify continuity of $\hat{f}$ with respect to the weak topology on $X$, let $(x_\alpha)_{\alpha \in D}$ be a net in $X$ converging weakly to $x$. By definition of the weak topology, $x_\alpha \rightharpoonup x$ means $g(x_\alpha) \to g(x)$ for every $g \in X^*$. In particular, taking $g = f$, we obtain $\hat{f}(x_\alpha) = f(x_\alpha) \to f(x) = \hat{f}(x)$. So $\hat{f}$ preserves net convergence, hence is continuous. This argument works for any net, regardless of whether the weak topology on $X$ is metrizable. [/example] ## References Kelley, J. L., *General Topology* (1955). Willard, S., *General Topology* (1970). Schechter, E., *Handbook of Analysis and Its Foundations* (1997). Munkres, J. R., *Topology* (2nd ed., 2000). Folland, G. B., *Real Analysis: Modern Techniques and Their Applications* (2nd ed., 1999). Megginson, R. E., *An Introduction to Banach Space Theory* (1998).