Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.
Updated
Clique path networks
Proved construction exclusion, not a resolution of Problem 617. The hereditary caps exclude a further family containing graphs with claws, not just the line graphs already excluded in Line-graph obstruction.
Fix an integer . Throughout,
A graph satisfies the caps if for all . In fact the proofs below only use sets of order at most .
1. The class and the conclusion
Start with a loopless multigraph on a set of branch vertices, replace every edge by an internally disjoint path of length at least three, and replace every vertex of the resulting graph by a nonempty clique. Two bags are completely joined exactly when their underlying vertices are adjacent. Arbitrary positive, unequal bag orders are allowed.
Theorem. If the resulting graph satisfies the caps, then
There is no edge budget, order bound, or independence-number restriction.
More generally one may add edges between branch vertices provided those direct branch edges form a disjoint union of cliques. The conclusion is unchanged. Parallel subdivided paths are allowed, including paths whose endpoints lie in the same direct branch clique. No other edges between bags are allowed.
Here will be proved by constructing a probability distribution on independent sets in which every vertex has probability of being selected. Multiplying this distribution by gives a fractional coloring of total weight ; taking expected cardinality gives the second conclusion in (1).
In particular no color of an E617 counterexample can itself be one of these graphs. If a color is a disjoint union of components in this class, the same conclusion follows by summing the span inequalities.
2. Local consequences of the caps
There is no , and an outsider has at most one neighbor in a : both assertions follow from . Consequently every bag has order at most , every nonisolated branch bag has order at most , and every internal path bag has order at most .
For the last assertion, an internal bag of order together with one vertex in a neighboring bag is a . A vertex of its other neighboring bag has neighbors in that clique, a contradiction. The same argument shows that a branch bag of order has at most one neighboring bag. If it is incident with a subdivided path, its branch vertex therefore has multidegree one and no additional direct branch neighbor.
Any adjacent pair of internal bags has total order at most . If its total were , a vertex immediately before this pair would force the first bag to have order at most one, and a vertex immediately after it would force the second to have order at most one. This would give .
For three consecutive bags of orders , their only nonedges are the pairs between the first and third bags. Thus
If the right side is negative this is automatic; otherwise it follows from the cap, since . The adjacent-pair clique bounds also show that these triples have order at most .
Write . A useful rearrangement of (2) is
3. Odd paths
Consider a path of length , with endpoint bag orders and internal total . Put .
For length three, write its bag orders as and . The internal-pair bound gives . Applying (2) at both ends and rearranging gives
Dividing by and using yields
The denominator is positive. If , then , so (4) implies . If and , (4) gives ; integrality gives . The other case is symmetric.
The sole remaining case is . The two adjacent internal bags then have orders one, and the branch-degree observation in Section 2 makes this an entire isolated component with bag orders . Its order is , its independence number is two, and it is -colorable: use colors on the first bag, color on the next bag, color on the third bag, and colors on the last bag. Such components can be handled separately.
For every odd path of length at least five, there is a stronger bound:
Write the end triples as and . The middle path between has odd length , where ; write for its internal total. Matching all its bags in consecutive pairs, and then matching only its internal bags, gives respectively
All these pairs consist of internal bags of the original path. Therefore
Using (3) at both ends now gives
Assume . Since the coefficient of is nonnegative, replace by on the right. After subtracting the result is
using . This proves (5).
Thus, except for the explicitly removed isolated components, every odd path satisfies . For length at least five it even satisfies .
4. Even paths
For a path of length , let be the total of all its bag orders and put . Then
For length four, write the orders as . The two end triple inequalities (2) give
If , this is nonpositive. Otherwise is a positive integer and , so integrality gives .
For length six, write the orders as . Applying (2) to the three triples centered at gives
The numerator is bilinear in . Its maximum on this rectangle is
as is seen at the four corners. This proves (6).
For length at least eight, write the end triples as and and retain the even middle path from to . Its corresponding excess satisfies (6) by induction. Equation (3) at the two ends gives
If , the right side is nonpositive. If , it is at most . This proves (6) in every case.
5. Coupling independent sets on a path
The following elementary probability lemma accounts for the endpoint dependencies, rather than treating them as independent without proof. Let and let satisfy . Prescribe an endpoint joint probability satisfying the Fréchet bounds
If is odd, a distribution on independent sets of the path with these marginals and this joint probability exists exactly when
If is even, it exists exactly when
Necessity follows by counting selected vertices in each of the four endpoint states. Here is a self-contained sufficiency proof.
The stable-set polytope of an odd cycle of order is described by nonnegativity, the edge inequalities, and . To check integrality, consider a fractional extreme point. The tight edge equations on fractional coordinates form paths, unless the whole cycle is at , which violates the sum bound. Each path has an alternating perturbation parameter. Two parameters can be perturbed while preserving the sum; one can also be perturbed if the sum bound is slack or its path has even order. In the only remaining case there is one odd-order fractional path and the sum bound is tight. Its sum is an integer plus or , whereas all other coordinates and the total sum are integers, forcing to be integral. This is again a contradiction. Thus every feasible point is a convex combination of independent-set indicators.
For odd , add one vertex adjacent to both endpoints, forming an odd cycle, and give it marginal . Conditions (7)--(8) make the vector feasible in the cycle polytope. The induced path distribution has an endpoint joint , because the added vertex can occur only when both endpoints are absent. On the original path there is also a distribution with joint : choose uniformly in , select even-indexed when , and odd-indexed when . Mixing these two distributions gives joint exactly .
For even , close the path with a new path of length three, whose two internal marginals have sum , the first at most and the second at most . Such choices exist by (7). The resulting odd-cycle vector is feasible by (9). Its endpoint joint , since selecting both endpoints excludes both new vertices. The threshold distribution just described now has joint , because the two endpoints have the same parity. Mixing again gives .
6. The global fractional coloring
Remove the isolated exceptional length-three components from Section 3. In each direct branch clique independently, select a branch bag with probability , selecting none with the remaining probability. These are valid probabilities because the joined branch bags constitute a clique of order at most . For no direct branch edges this simply selects the branch bags independently.
An endpoint pair in different branch cliques has joint probability ; one in the same branch clique has joint zero. Put on every subdivided path. Adjacent marginals sum to at most one. For length three, Section 3 proves (8) for joint , hence also for zero. For longer odd paths, (5) proves (8) for every joint allowed by (7). For even paths, (6) proves (9) for every allowed joint.
Apply the coupling lemma separately on each path. Conditional on the already sampled endpoint states, sample its interior according to the corresponding conditional distribution. Events of zero endpoint probability need no choice. Interior paths have no edges between them, so all the selected bags together form an independent set of the bag graph, and every bag has marginal .
Choose one uniformly random vertex from each selected bag. Every vertex then has marginal . This proves (1), including the direct-branch extension.
There is a useful local form. Start with a distribution on independent sets of the remaining graph in which every vertex has marginal exactly . Any clique path of length at least four, whose internal bags have no neighbors outside their own and adjacent bags, can be inserted without changing this outside distribution. The endpoint joint may be arbitrary. The endpoint bags need only be disjoint cliques; their outside neighborhoods need not be homogeneous, and edges between the endpoint cliques are allowed.
Every graph with admits the required exact-marginal normalization: divide fractional-color weights by , pad with the empty independent set, and independently thin the selected vertices to reduce each marginal to . This normalization can change the initial distribution. A length-three path is not asserted to have the universal extension property.
Scope and the remaining structural gap
The initial odd-subdivision span proof and the enlarged fractional proof were separately audited. The arguments use only the cap, elementary probability, and the odd-cycle polytope proof above; no external classification theorem or finite computation is invoked.
The class contains, for example, positive clique blowups of the graph obtained by replacing each edge of by a path of length three. Those graphs can have claws, so this is not a restatement of the line-graph exclusion. The construction also permits an unbounded root and unbounded residual independence number.
No surviving least-color core has been proved to admit this decomposition. Its private-clique support description does not imply that exception vertices can be organized into the internally disjoint paths used here. Deleting a strip can leave a smaller graph with fractional chromatic number above even when every proper induced set has Hall ratio at most . Thus rank minimality cannot be silently replaced by fractional coloring minimality. This theorem excludes a parameterized construction mechanism; it does not prove any new unrestricted E617 case.
7. Allowing paths of length two
Theorem. Let . Start with a loopless multigraph, replace every edge by an internally disjoint path of length at least two, and replace every vertex by a nonempty clique, with complete joins exactly between consecutive bags. If the resulting graph satisfies the hereditary caps, then
Parallel root edges and unequal bag sizes are permitted. Unlike the length-at-least-three theorem, this extension does not permit direct edges between branch bags. A correlated distribution on the branch bags replaces their independent selection.
Minimum degree and a common correlation parameter
First assume . Put , let be the largest bag order, and set . The clique cap gives . A largest bag has at least common neighbors outside itself. Together they form an -set whose nonedges all lie among those neighbors. Hence
For , this forces , and therefore
Define
Then and
After cross multiplication, the latter inequality is exactly .
Since branch bags are pairwise anticomplete, use the following mixture. With probability , select them independently with probabilities , where is the bag order. With probability , choose a common uniform and select every branch bag with . Each branch bag still has marginal . Endpoint bags of orders have joint selection probability
We check that this single distribution can be extended along every path.
Length two and length three
For a path with bag orders , put . Its only nonedges are the endpoint pairs, so its cap and give . Since
equation (12) gives
This is precisely the even-path coupling requirement (9).
For a length-three path , put . The end-triple caps, as in Section 3, give
The coupling requirement is . To check it for arbitrary unequal endpoints, set , . Then , and the preceding bound implies
For the second inequality, increases with , so replace by , then use . The third follows by monotonicity in . This proves the required upper bound for .
Paths of length at least four extend every feasible endpoint joint, by Sections 3--5. Sample all branch states first, then sample each path interior conditional on its two endpoint states. The interiors have no cross edges, so the samples combine into independent bags with marginal . Choosing one uniform vertex from each selected bag gives each vertex marginal exactly . This proves (10) when the minimum degree is at least .
Removing the minimum-degree assumption
If already has an independent-set distribution with marginals and , then the event that no neighbor of is selected has probability at least . Add on a randomized subevent of probability exactly . This preserves all existing marginals and gives the required marginal at .
Repeatedly delete vertices of degree at most . In the remaining induced graph the bags may have smaller orders or disappear. A nonempty bag with quotient degree zero or one cannot remain: it and its sole neighboring bag, if any, form a clique of order at most , making its vertices' degree at most .
Thus a surviving original internal bag retains both its neighboring bags. Propagating along its original path forces the whole path and both branch bags to survive. The nonempty remainder is still a clique blowup of a root subgraph with every path of length at least two. Apply the minimum-degree construction to it, then restore deleted vertices in reverse order by the preceding extension argument. If the remainder is empty, start with the empty-set distribution.
This proves the fractional conclusion itself, not merely the unweighted span inequality. It does not replace induced-ratio minimality by fractional minimality. No least-color budget or independence restriction is used in this theorem, and no general E617 core is known to have the required subdivision structure.
The correlation bounds, arbitrary-endpoint coupling, preservation of the class under taking the -core, and reverse low-degree extension were checked by the author and are author-recorded.