Talk:Graph homomorphism

Latest comment: 7 years ago by David Eppstein in topic GA Review

Definition issue

edit

The definition is wrong. For example, in the condition

  • δE = δE′ o fE

the range of the left side is sets of E-vertices, while the range of the right side is sets of E'-vertices.

I'm not correcting it right now because

  1. Wikipedia's being slow and flaky
  2. I don't have a reference for this concept.

Anyway, I suspect it could be stated a lot more simply, without so much formalism. Dbenbenn 21:42, 9 Jan 2005 (UTC)

You are right the definition is wrong. I will fix it in the next few days. As for the simplicity: What I really meant to do is draw some commutative diagrams. The obscure equations are just placeholders until I have created the diagrams. MathMartin 13:03, 10 Jan 2005 (UTC)

It appears that the map from a graph to   is not a graph homomorphism. I find that surprising; I think it should be mentioned explicitly. Dbenbenn 15:20, 10 Jan 2005 (UTC)

You are correct again. I have to modify the functions to allow for the removal of edges and arrows by mapping them onto vertices. MathMartin 17:08, 10 Jan 2005 (UTC)

Now the definition should be correct, but I will still try to make it more clear. MathMartin 14:01, 12 Jan 2005 (UTC)

Simpler definition

edit

The new definition looks correct, but scary! What do you think of the following simplification?

A graph homomorphism   from a graph   to a graph   maps vertices of   to vertices of   such that
  1. if   is an edge of  , then   is an edge of  , or  , and
  2. if   is a directed edge of  , then   is a directed edge of   (in the same direction), or  .

Basically, we don't really need to talk about the homomorphism's values on the edges, since that value is determined by the value on vertices. Dbenbenn 14:27, 12 Jan 2005 (UTC)

Your definition is simpler but only valid for homomorphisms between simple graphs. I understand your concerns, I do not like the definition either and I will try to make it clearer. Perhaps it would be best to include both definitions and point out the difference ? MathMartin 16:46, 12 Jan 2005 (UTC)

Oh, right. The problem is that with either loops or multiedges, the definition of "monomorphism" gets broken. How about the following slightly less simple version:
A graph homomorphism   from a graph   to a graph   is a function on the edges and vertices of   such that
  1. vertices of   go to vertices of  ,
  2. if   is an edge of   with endpoints   and   then either   is an edge of   with endpoints   and  , or  , and
  3. if   is a directed edge of   from   to   then either   is a directed edge of   from   to  , or  .
By the way, the definition in the article is still slightly broken. Where does   send a vertex   in the range  ? Dbenbenn 21:41, 12 Jan 2005 (UTC)

I extended your definition slightly to clarify the domain and codomain of f, because later we talk about bijection and it has to be clear where the f is defined. Otherwise I think your definition is easier to understand and the article is now in a useable state.MathMartin 10:19, 25 Jan 2005 (UTC)

WRONG DEFINITION

edit

I do not agree with this definition at all: it might be a good definition for a graph contraction but not for graph homomorphism.

One of the fundamental property of graph homomorphism is that a graph G is k colorable iff G has a homomorphism to the complete graph on k vertices. This is the reason why the property to have a homomorphism to some fixed graph H is called 'H-colorability'. According to your definition, every graph has a homomorphism to  .

Th standard definition of a homomorphism is the following: A homorphism from a graph G to a graph H is a mapping from the vertex set of G to the vertex set of H so that two adjacent vertices of G are mapped to two adjacent vertices of H. This definition can easily be extended to directed graphs and to graphs with loops (if v has a loop attached, v is adjacent to v thus f(v) is adjacent to f(v), that is: f(v) has a loop attached). Extension to multigraphs is not usual at all.

Reference:

P. Hell and J. Nesetril, Graphs and Homorphisms, Oxford Lecture Series in Mathematics and its Applications 28, Oxford University Press, 2004.

pom 11:25, 13 November 2005 (UTC)Reply

I think that terms weak and strong homomorphism need to be introduced. Weak homomorphism does not take the directions of lines into account while strong does. A proper definition should introduce two functions, f1 and f1 where f1 maps vertices and f2 maps lines. If these two functions are bijections and there is a (strong, weak) homomorphism, we have a (strong, weak) graph isomorphism. --212.235.208.165 (talk) 10:41, 24 November 2008 (UTC)Reply

GA Review

edit
GA toolbox
Reviewing
This review is transcluded from Talk:Graph homomorphism/GA1. The edit link for this section can be used to add comments to the review.

Reviewer: David Eppstein (talk · contribs) 06:57, 3 August 2017 (UTC)Reply

Reviewing. But there are multiple whole paragraphs and some even larger sections of the article lacking inline citations: most of the "Definitions" section, most of the "Connection to colorings" section, most of the "Examples" section, etc. Unless these can be fixed, this is far from the verifiability requirement in good article criteria #2 and likely to be a quick fail. —David Eppstein (talk) 06:57, 3 August 2017 (UTC)Reply

Thanks, I'll add some citations later this week (for now: almost all of it is in Hell & Nesetril's Graphs and Homomorphisms, mostly in the publicly available Chapter 1). Tokenzero (talk) 08:29, 3 August 2017 (UTC)Reply
I tried to add citations for everything. I don't have anything for the parts-of-speech tagging example, but in my opinion it is a nice and simple example of how CSPs can express something like G fits the model H, instead of the usual kind of resource allocation examples. And at the same time directed graphs are more natural here. In literature I could only find much more involved applications, nothing in secondary sources, the most similar (but still much more general and too primary source) in here: Padró, Lluís (1996), A Constraint Satisfaction Alternative for POS Tagging (PDF). Tokenzero (talk) 13:26, 6 August 2017 (UTC)Reply

Second reading

edit

Lead:

  • This is significantly less technical than the rest of the article (a good thing) but I think it's mostly because the rest of the article is too technical. More effort should be made in making it readable by non-experts in graph theory. (Let's say, at the level of someone with an undergraduate mathematics degree who hasn't studied any graph theory.) Right now, the lead is at that level but the rest is not. In particular, there is far too much reliance on notation when words would be clearer, and far too complex average sentence length (GAN criteria 1a).
  • Throughout, many technical terms are used with the assumption that the reader either knows what they mean or will follow a link to see what they mean. I think it would be helpful to provide a short gloss here for each such word. A good example: "arc (directed edge)" in definitions. A bad example: "injection" (in the same section).
I tried to split some sentences into two, reduce notation, simplify some wording and add glosses as much as I could. Tokenzero (talk) 23:35, 20 August 2017 (UTC)Reply
  • The lead summarizes the first three sections of the article (definitions, coloring, and applications) but not the rest (structure, incomparability, and complexity) (GAN criteria 1b).
Extended.Tokenzero (talk)

Definitions:

  • Because it only maps vertices to vertices, the definition of homomorphism here seems to be unsuitable for multigraphs, and that is reinforced by the implicit assumption that edges are the same thing as sets of two vertices. Both should be stated more explicitly. This is especially important because "sets of two vertices" is not actually correct: later, we see examples where loops are allowed. So the graphs considered here are not simple graphs, but graphs with loops allowed but with multiple adjacencies disallowed.
Added "In this article, unless stated otherwise, graphs are finite, undirected graphs with loops allowed, but multiple edges (parallel edges) disallowed." Tokenzero (talk)
  • "H-colorable, we shall": comma splice. Also, Wikipedia's manual of style says to avoid first person.
Fixed. Tokenzero (talk)
  • There is a little attempt to distinguish homomorphisms from other types of maps (the description of subgraphs, isomorphisms and covering maps as special types of homomorphisms) but it might be helpful to also discuss homeomorphisms (more than just in the hatnote) and minors at least to the point where someone can distinguish them from homomorphisms.
They're not really related more than also being a graph relation, but ok, I tried to say something. Tokenzero (talk)
  • "If the homomorphism f : G → H is an injection, then G is simply a subgraph of H.": is this a description of all subgraph relations, or does the if-then only go one way? Same for the covering maps in the next sentence.
Fixed. Tokenzero (talk)
  • More examples would go a long way towards making this understandable. In particular, bipartite double cover provides useful and nontrivial examples of covering maps.
Added bipartite double cover. The top figure gives an example, I can't think of anything non-redundant now. (I added a figure of K7 instead of a gloss). Tokenzero (talk)
  • "homomorphically equivalent": this paragraph no footnote.
Added. Tokenzero (talk)
  • Can you please provide an example of a core that is not a complete graph?
Added odd cycles. Tokenzero (talk)
  • "Any finite graph G is..." this is the first time that finiteness or non-finiteness has been mentioned in this article. Some later material seems to assume finiteness (talking about unique cores) without stating it. I think it would be good to be more explicit about this, either stating that graphs are allowed to be infinite unless explicitly stated as finite (and then fixing the later parts that implicitly assume finiteness) or vice versa.
I made 'finite' default, and added "Notably, this is not true in general for infinite graphs." after the statement on unique cores. Tokenzero (talk)

Colorings:

  • "Each k-coloring corresponds to a homomorphism from G to the complete graph": and vice versa?
  • "two colors are different if and only if they are adjacent as vertices of Kk (since Kk has edges between all possible different vertices and no edge from a vertex to itself)": while not wrong mathematically, this is very confusingly worded. This whole paragraph has the feel of someone repeating over and over an obvious mathematical statement, with different but equivalent wording, because the statement is so obvious that they have no idea how to go about actually proving it.
Rephrased this all a bit. Tokenzero (talk)
  • This whole section is seriously incomplete (GAN criteria 3) without any discussion of the Gallai–Hasse–Roy–Vitaver theorem, and especially of the formulation of this theorem that a directed graph G has a homomorphism to a k-vertex transitive tournament iff it has no homomorphism from a k+1-vertex path.
Added. Note this formulation is, as far as I see (and H&N suggest) slightly weaker than GHRV: if you have an non-acyclic orientation with no k+1-vertex path, then you cannot apply the theorem without redoing the 'maximal acyclic' argument. Tokenzero (talk)
  • "General homomorphisms can also be thought of": this paragraph no footnote.
This is not a statement of fact, only a guide for the reader, why would it need a footnote? (And obviously researchers do think about it this way, as the H-coloring name indicates.) Anyway, I made it run into the next paragraph. Tokenzero (talk)

Constraint satisfaction:

  • The first two paragraphs are essentially about distance coloring and its formulation as a homomorphism problem. Maybe this should be at least mentioned?
You mean coloring where vertices at distance at most p get different colors? Or L(h, k)-coloring? Neither is really the same thing. You could define metric spaces and a kind of distance-preserving colorings, but that's far fetched. Anyway this is meant as examples that are as simple as possible, where you can change H to change constraints. Tokenzero (talk)
  • "part-of-speech tagging": unsourced, and I'm skeptical that homomorphism is an effective way of approaching this. The reason is that one needs huge amounts of data to get anywhere with unsupervised learning of language, but huge data is instrinsically noisy, and you have no principled way of throwing out the pairs that are coincidentally next to each other from the ones that you are trying to find. Also the connection between HMMs and homomorphisms is tenous and again needs sourcing.
I removed it. But my intention was to give this obviously simplified example, which would nevertheless show a very different way of thinking about homomorphisms. In practice of course you'd want a kind of fractional weighted homomorphism, and constraints between more than just two words. Then it's exactly what they do in Padró, Lluís (1996), A Constraint Satisfaction Alternative for POS Tagging (PDF). Also I did not connect HMMs and homomorphisms, I meant only that HMMs are the theoretical concept more commonly used for part-of-speech tagging. Tokenzero (talk)
  • "Most algorithmic methods": the methods listed are only naive ones, and for coloring at least other methods including dynamic programming and inclusion-exclusion have also been effective. So is "most" really accurate?
You're right, changed to 'Many', I can make it 'Some' if you want. (Though in a theoretical/worst-case context, it would be hard to find any general method that works for digraphs but not CSPs). Tokenzero (talk)

Structure:

  • This whole section does not carefully distinguish the directed versus the undirected case. Really there are two dense posets and two categories, one for directed graphs and one for undirected graphs. Which of these two are the ones whose properties are described here?
I made 'undirected' default, mentioned 'directed' in Cores, added final paragraphs here and in Incomparability on directed graphs, and added '(undirected)' when stating density to warn the reader this one place is not true for digraphs (they are mooostly dense, but it's complicated). Tokenzero (talk)
  • "equivalence classes, it defines": comma splice.
Fixed. Tokenzero (talk)
  • This appears to be the first time we've seen the slashed-arrow notation. What does it mean? (My guess would be the nonexistence of a homomorphism but this needs to be clearer.)
Replaced with '<' with definition when discussing density, and replaced with words in Incomparability. Tokenzero (talk)

Incomparability:

  • Have we seen a definition for the notation CG used here? Again, are we considering directed or undirected graphs?
Fixed. Tokenzero (talk)

Complexity:

  • "can be solved by brute-force": this time bound needs a source.
Added (but primary, I didn't find an explicit statement in the books). Tokenzero (talk)
  • Why the scare quotes (and why single quotes) on left and right?
Removed, rephrased anyway. Tokenzero (talk)
  • "Hell-Nešetřil", "Feder-Vardi", etc., should use en-dashes, not hyphens.
Fixed. Tokenzero (talk)
  • Why is graph isomorphism mentioned, but not subgraph isomorphism? And this is another paragraph without footnotes.
I removed the subsection altogether, since it probably doesn't make sense to repeat a list of graph relations here. Tokenzero (talk)

David Eppstein (talk) 01:47, 14 August 2017 (UTC)Reply

Many thanks for the suggestions, I've added specific answers above. Tokenzero (talk) 23:35, 20 August 2017 (UTC)Reply

Status query

edit

David Eppstein, where does this review stand now? Are there any issues remaining at this point? Thanks. BlueMoonset (talk) 14:06, 1 October 2017 (UTC)Reply

Third reading

edit

@Tokenzero: Sorry for the delay. @BlueMoonset: still some issues remaining, but it looks like it is converging towards GA.

Here are my comments on the current version. Overall, it looks well-balanced, with good coverage that has been made significantly less technical. At least, I think it should now be readable by someone who already has a little familiarity with graph theory, while before it was a little more advanced than that. I only have some minor issues of writing quality (GA criterion 1a) to address:

Lead: Now at a good balance between avoiding enough jargon to be accessible and, on the other hand, covering the subject. Issues with inadequate summarization of some parts of the article have been addressed.

Definitions: Minor inaccuracies and differentiation from other related notions now addressed. One minor quibble: The article contains 17 instances of the word "any", which often functions as a quantifier, but an ambiguous one: sometimes it means the same thing as "every", and sometimes it means "a single arbitrarily chosen thing". So in "no homomorphism to any proper subgraph", there is no reasonable substitute for "any" (in the second meaning) but in "Any graph G is homomorphically equivalent", I think it would be crisper to replace "Any" by "Every".

Colorings: The first k is not italicized. Re "It is not hard to show that a graph G is k-colorable": I think the clearer statement is that whenever G->H as undirected graphs and H is given an orientation, you can pull the orientation back to G and get a homomorphism of directed graphs. Re "one of the two possible orientation": missing the "s" at the end of "orientations". Re "A folklore theorem states": the scope of the "for all k" tacked at the end of this sentence is unclear. Also the math is badly formatted: the plus signs, equal sign, and minus sign (currently and incorrectly a hyphen) should have spaces around them.

Connection to CSP: "called relational structure" needs either an article or a plural.

Structure: "defined as the disjoint union [G ∪ H]": this is a little confusingly stated, because it skips a step. It's actually the equivalence class of the disjoint union of representative graphs, and it takes an (easy) argument to observe that it doesn't matter which representative graphs G and H you use to construct the disjoint union. The disjoint union of equivalence classes would be [G] ∪ [H] and is not itself an equivalence class unless [G] = [H]. The same issue applies also to the meet, but it's less confusing there because there's no obvious way to define a tensor product of equivalence classes. Re "same definitions apply, in particular": another comma splice. Re "can be though of": should be "can be thought of".

Incomparable graphs: I'm confused about something. Supposedly by considering only the two parameters odd girth and chromatic number (both monotone for homomorphisms) we can generate an infinite antichain. Doesn't that contradict Dickson's lemma, according to which every antichain among pairs of non-negative integers is finite?

Homomorphisms from a fixed family of graphs: "for any class of graphs G, ... [statement that does not refer to G]." What is the class of graphs doing in this sentence?

References: Brown et al 2008 appears not to be used. Should it be moved to a separate "Additional reading" section, or removed, since it isn't actually a reference? Godsil and Royle is only used once; other references that are only used once are detailed in the Notes section rather than given a short reference there and a long reference in the references section. Also, it's a whole book, so it would be helpful if the footnote also included a page number where its claim could be found. Gray 2014 appears not to have been published and while its author seems to have led an interesting life [1] she currently appears to be a student, i.e. not established enough for the "recognized expert" clause of WP:SPS to apply. So I'm skeptical that this is a reliable source, by our standards. And although it doesn't say so, I suspect it may also have been somewhat influenced by our article, raising WP:CIRCULAR issues. Can we find the same claims elsewhere?

David Eppstein (talk) 05:08, 2 October 2017 (UTC)Reply

@David Eppstein: Ok, I've tried to fix it all, see the diff to check if the new phrasings are clear enough now.
Incomparable graphs: G → H implies G has at most the chromatic number of H, but at least the girth of H, so the integer parameters are compared in opposite directions. In terms of comparing integer tuples, the infinite antichain is of the form (k,−k), while Dickson's lemma applies to non-negative integer tuples only.
References:
  • In general, my intention was to separate very focused references (in particular original journal articles) to Notes, while putting surveys, books and such into References. In particular Godsil & Royle is a book with a whole chapter devoted to homomorphisms, I believe the largest beside H&N, with some results (admittedly obscure) not covered by H&N, but of course mostly redundant. Brown et al. is the most comprehensive treatment on the categories of graphs I could find, not just a 'by the way, this is a category' , but seriously going into some details. (It's now also a reference in the technical sense, since I needed to reference one statement from Gray to Brown et al.). I've tried to make that more consistent and ordered now, but I'm open to suggestions (rename Notes to References and References to Further reading?).
  • Gray: It's a useful source in the sense that it is a freely available, concise, elementary exposition with complete proofs for quite a few fundamental statements that appear in the article, but I agree "recognized expert" does not apply. It's apparently a survey done as part of a Vacation Research Scholarship (under 2013/2014, La Trobe), under the supervision of two persons that appear to be recognized experts, which is kind of the best we could hope for. Anyway, I've added other references to the two statements where this was the only one, but kept the references to Gray (the others are more general and require going through their much more technical definitions) and added 'student research report' to Gray. Given the state of the article before 2016 I doubt it's WP:CIRCULAR (Gray's is 2014). (I've seen the Sydney Morning Herald article before, but never associated it with this reference; you have a very good eye for persons behind articles!)
Tokenzero (talk) 16:03, 14 October 2017 (UTC)Reply
All remaining issues addressed; passing. —David Eppstein (talk) 07:13, 16 October 2017 (UTC)Reply