How can we think of nets as a generalization of sequences?
Basically, we can see sequences as a way of enumerating elements of a set: in facty a sequence is a function $x_n:\mathbb{N}\rightarrow X$.
Now, directed set abstract from naturals the following property:
$$\alpha,\beta\in\ D\rightarrow\; \exists \gamma\ st\ \alpha\prec\gamma,\ \beta\prec\gamma$$
Cam we then see nets as a generalization of the ordinary idea of counting, in some sense? What is the intuition behind them and behind directed sets?
Intuition behind nets
general-topologyintuitionnetsorder-theorysoft-question
Related Solutions
Per the OP's request from this comment I am posting some stuff about references I am aware of. Those references are mentioned in my (unfinished) notes on this topic and also in this answer and this conversation in chat. The purpose of reposting them here is that the linked resources have somewhat different scope and focus. Another reason is that if I list there references here, they will become better organized.
I repeat here some stuff mentioned already in Brian M. Scott's answer. But maybe it is not that bad to have one brief and concise answer and also a longer post which digresses a bit to some related stuff.
Since I mention results from various sources, terminology is a bit mixed in this post. But I always use the name filternet in the way you defined it. (This helps to unify terminology a bit; I mention some other places where this notion or some special case is defined, but several different names are used for the same concept in the literature.) If I mention other concepts, I hope it will be clear from the context which type of convergence of filters I mean.
General approach
As already mentioned in Brian M. Scott's answer, Bourbakists used the notion of filternet (in your terminology) in some of their works as a basic notion, which was defined first. (To be more precise, they use filter base more often than filter, but this difference is a very minor one.) They also define other notions related to convergence (such as cluster point or limit superior) in this generality and various types of limits were defined as special cases.
You could probably find this in several of their books. I am aware of this approach being used in:
- N. Bourbaki. Elements of Mathematics. General Topology. Chapters I-IV. Springer-Verlag, Berlin, 1989. See p.70 for the definition of limit of a filternet.
- Jacques Dixmier. General Topology. Springer-Verlag, New York, 1984. Undergraduate Texts in Mathematics.
In Dixmier's book this approach is very prominent. Already in Chapter II the notion of limit along a filter base is defined. (This is precisely your filternet.) Where possible, other notions are defined in a similar manner and results about them are shown in this generality. Cluster points (which the author calls adherence values) are defined in the same chapter. Limit superior and limit inferior is defined in Chapter VII (for $X=\mathbb R$). So if you are looking for a book which really uses the notion of filternet to build all theory related to convergence, this is the book I would certainly recommend.1
Other references which mention this type of convergence are:
- D. H. Fremlin. Measure theory, Volume 2: Broad Foundations. Torres Fremlin, Essex, 2001. In this book there is one paragraph 2A3S in the appendix, where this notion is defined. Some basic results about filternet are shown here. The corresponding notion of limit superior is defined here, too.
There are also some textbooks in mathematical analysis, which introduce the filternet:
- Vladimir A. Zorich. Mathematical Analysis I, Springer, 2004, Undergraduate Text in Mathematics. In this book only the case $X=\mathbb R$ is discussed; see Definition 3.2.12. Again, the idea is to use this as a unifying notion for several other special cases.
- The book M. Gera, V. Ďurikovič: Matematická analýza (Mathematical Analysis, in Slovak) also introduces this notion in the case $I\subseteq\mathbb R$ and $X$ Hausdorff.
Both these books use the notion of filterbase (similarly as Bourbaki). I am aware that the latter will probably not be useful for you (unless you speak Slovak). But I am posting these two references mainly as evidence that this approach in the introductory analysis textbooks is probably not that uncommon and it is likely that some other similar texts exist. (Although it is probably not easy to find them. I knew about one of them since it was the book from which I was taught my second course in analysis. And I found the other one using Google Books.) The reason that this approach appears in some texts is, in my opinion, influence of Bourbaki which lead to striving to greatest possible generality. But I have admit that some proofs might be more economical in this setting.
Gera and Ďurikovič do not have any book by Bourbaki among references. In Zorich I have seen some books by Bourbaki, Cartan and Dieudonné. Maybe it would be possible to check the reference given in these books to see from where the authors took this approach. But this would be rather time-consuming undertaking.
Reduction of the general approach to limit of a filter.
Let me also briefly mention the comparison of the two notions:
- Your definition of filternet.
- Convergence of filter $\mathcal F$ on a topological space $X$ defined as: The filter $\mathcal F$ converges to a points $x$ iff it contains all neighborhoods of the point $x$.
The latter is the notion which, as far as I can say, appears quite frequently in texts on general topology. And it is rather clear that it is a special case of the former (as was explained already in your question.)
But it might be perhaps useful to notice that also the notion of filternet can be obtained relatively easily using the notion of limit of a filter. Namely if we have filternet $(\mathcal F,f\colon I\to X)$, then this filternet converges to the point $x$ if and only the filter generated by $\{f[F]; F\in\mathcal F\}$ converges to $x$.
- This is also discussed in comments to this question.
- This is shown in my notes in the section named "$\mathcal F$-limits and convergence of filters". It is also mentioned in the answer which I have linked to several times.
- This is the way the filternet is defined in Fremlin's book.
The books mentioned above are the only references I am aware of which work in the same generality as suggested in your post. Let me allow to digress a bit from the question to mention two special cases, which are relatively important (and for which there is much more literature).
Ultrafilters and Stone-Čech compactification
In the case that $\mathcal F$ is an ultrafilter on $i$ and $X$ is a compact Hausdorff space, then the convergence of a filternet $\mathcal f: I\to X$ can be also interpreted in this way: If we endow $I$ with the discrete topology, then we can use the Stone-Čech compactification $\beta I$ on this space. The points of $\beta I$ correspond to ultrafilters on $I$. And the limit of an filternet $(\mathcal F, f\colon I\to X)$ is precisely the value of the unique extension of $f$ to $\overline f \colon \beta I \to X$ evaluated at $\mathcal F$. This is explained in more detail in my other answer.
If we know about this connection, it should not be too surprising that the notion of filternet also appears in some texts studying the Stone-Čech compactification. I will mention the two books where I have encountered this:
- N. Hindman, D. Strauss: Algebra in the Stone-Čech compactification; first edition (1998), second edition (2012).
- Stevo Todorčević. Topics in Topology. Springer-Verlag, Berlin{Heidelberg, 1997. Lecture Notes in Mathematics 1652.
In Todorčević's book the notion of filternet is defined in the same way as in your post. (Definition 2 in Section 14.) However, both this notion and the Stone-Čech compactification are used only in one chapter of this book. Although this book only works with $I=\mathbb N$, it is mentioned several times in this chapter that everything works the same for arbitrary set.
The book of Hindman and Strauss is devoted solely to the Stone-Čech compactification. Therefore it devotes a bit more space to filternet than Todorcevic's book. Again, the definition given in this book is exactly the same as yours - see Definition 3.44.
Sequences
If we take $I=\mathbb N$, then we are dealing with sequences. For this case you can find more references.
The case when $\mathcal F$ is ultrafilter and the sequence is bounded is often mentioned in set-theoretical texts. This Google Books search returns Hrbacek-Jech and Komjáth-Totik, to mention just two examples. The first book where I have seen this notion we Balcar-Štěpánek: Teorie Množin (Set Theory, In Czech). I guess that there are some other texts from this area which mention this special case, i.e. limits of bounded sequences along ultrafilters.
Names for this notion
Let me also list here various name which are used for filternet in literature. (To some extent as a curiosity, but perhaps this might be useful when searching for other resources on the topic. Or maybe it might show that searching for something like this on internet might be rather difficult, because every author uses different terminology.)
- Bourbaki calls it limit of $f$ with respect to a filter.
- Dixmier uses the name limit along a filter base.
- Zorich chose limit of a function over a base.
- Hindman and Strauss use the name $p$-limit. Similarly Komjáth and Totik use $D$-limit, Hrbacek and Jech $U$-limit. Of course, in each of these there case $p$, $D$ and $U$ denotes an ultrafilter
- Todorčević does not use any name for this notion, he only introduces the notation $\lim\limits_{n\to\mathcal U} x_n$. Fremlin does not have a name for this notion either and he uses the notation $\lim\limits_{x\to\mathcal F} \phi(x)$.
1 Dixmier does not define the notion of Cauchy function with respect to filter base. A possible approach to Cauchyness is briefly explained in my notes. This is the only notion which appears in these notes for which I do not have a reference in a textbook. A special case of Cauchy condition for $I=X=\mathbb R$ is mentioned in the book by Gera and Ďurikovič.
I dont see the reason to use nets since you are in a normed space and convergence is described nicely by sequences but since you wanna mess around with nets, i will write the first direction by using nets and the other one with sequences.
The direction "$\implies$" doesnt require the assumption of a Hilbert space, it works also for Banach spaces. The argument is the following:
Suppose that $K\in \mathcal{B}(H)$ is a compact operator and that there exists a bounded net, say $(u_i)_{i\in I}$ that converges weakly to $0$ but $\lim_{i}Ku_i \nrightarrow 0$. Then, there is an $\epsilon>0$ and a subnet $(u_{i_\mu})_{\mu\in M}$ of $(u_i)_{i\in I}$ such that $$\tag{1}||Ku_{i_\mu}||\geq \epsilon\ \ \text{for every}\ \mu \in M$$ Now, $(u_i)_{i\in I}$ is bounded, which means that there is a constant $C>0$ such that $||u_i||\leq C$ for every $i\in I$. This means that the net $(u_i)_{i\in I}$ belongs the open ball $B(0,2C)$ centered at $0$ of radius $2C$. Hence, the net $(Ku_{i_\mu})_{\mu \in M}$ belongs to $K\bigl(B(0,2C)\bigr)$. Now, by the compactness of $K$ we know that $\overline{K\bigl(B(0,2C)\bigr)}$ is compact. Hence, the net $(Ku_{i_\mu})_{\mu \in M}$ has a strongly convergent subnet, say $(Ku_{i_{\mu_\sigma}})_{\sigma \in \Sigma}$ to a point $y\in H$. Now, $(u_{i_\mu})_{\mu \in M}$ converges weakly to $0$, which means that $\langle u_{i_{\mu_\sigma}}, h \rangle \to 0\ $ for every $h\in H$. Using the adjoint operator $K^*$ of $K$ we see that $$\lim_{\sigma \in \Sigma}|\langle Ku_{i_{\mu_\sigma}}, h \rangle|=\lim_{\sigma \in \Sigma}|\langle u_{i_{\mu_\sigma}}, K^*h\rangle|=0$$ Which means that $(Ku_{i_{\mu_\sigma}})_{\sigma \in \Sigma}$ converges weakly to $0$. But we already know that $(Ku_{i_{\mu_\sigma}})_{\sigma \in \Sigma}$ converges strongly to $y$. Hence, $y=0$. But this means that $(Ku_{i_{\mu_\sigma}})_{\sigma \in \Sigma}$ can be arbitrary close to $0$ which contradicts $(1)$.
For the other direction "$\impliedby$" suppose that $K$ has the property that $\lim_{i\in I}||Ku_i||=0$ for every bounded net $(u_i)_{i\in I}$ which is weakly convergent to $0$. To show that $K$ is compact we only need to show that $K(B_H)$ is totally bounded (check this) where $B_H$ is the closed unit ball $\{h\in H:\, ||h||\leq 1\}$ of $H$. Supoose that $K(B_H)$ is not totally bounded. Then there exists an $\epsilon>0$ such that $K(B_H)$ cannot be covered by finitely many balls with centers from $K(B_H)$ and radius $\epsilon>0$. This means that for some $x_1 \in B_H$ the open ball $B(Kx_1,\epsilon)$ does not cover $K(B_H)$. Hence, there is some $x_2\in B_H$ such that $||Kx_1-Kx_2||\geq \epsilon$. Now, the open balls $B(Kx_1,\epsilon),\, B(Kx_2,\epsilon)$ cannot cover $K(B_H)$, hence there is some $x_3\in B_H$ such that $||Kx_1-Kx_3||,\, ||Kx_2-Kx_3||\geq \epsilon$. Continuing this way, recursively, we construct a sequence $x_n \in B_H$ such that $$\tag{2} ||Kx_m-Kx_n||\geq \epsilon\ \text{ for every }\ n\neq m$$ Now, $x_n$ belongs to the closed unit ball $B_H$. Since, $H$ is a Hilbert space, it is also reflexive, this means that $B_H$ is weakly compact. Hence, $x_n$ has a weakly convergent subsequence, say $x_{k_n}$ converges weakly to $x$. Then, $z_{n}=x_{k_n}-x$ converges weakly to $0$. From our hypothesis, it follows that $\lim_{n\to \infty}||Kz_n||=0$. But, for every $n\neq m$ $$||Kz_n- Kz_m||=||Kx_{k_n}-Kx_{k_m}||$$ which contradicts $(2)$. Hence, $K(B_H)$ is totally bounded which implies that $\overline{K(B_H)}$ is compact.
Best Answer
To begin a net over the directed set $\mathbb{N}$ is just a sequence. So nets are in a way a generalization of sequences. Let me motivate the analogy further.
Why do we need nets? Let $X$ and $Y$ be topological spaces. When $X$ is first countable the following conditions on a function $f: X \rightarrow Y$ are equivalent:
When the first countability assumption is no longer true, the statement $2. \Rightarrow 1.$
is no longer valid in general. However the following equivalence holds in any case
More generally, almost every property about the topology of metric spaces (where the space is first countable) which has an equivalent formulation using sequences, remains true in general topological spaces but one has to replace sequences by nets.
Let me explain why the property $\forall \alpha, \beta \in D: \exists \gamma \in D: \alpha, \beta \leq \gamma$ is crucial for nets to have any application. In topology it happens a lot that we have a sequence $(x_n)$ such that
If we choose $N = \max \{N_1,N_2\}$ then for all $n \geq N: x_n$ satisfies both property $A$ and $B$. This strategy is applied very often when $A$ and $B$ together with the triangle inequality give something meaningful.
The same strategy can be applied for nets precisely because of the property that I mentioned.