Curvature-Distortion Numbers of Graphs: Prescribed Lin–Lu–Yau Curvature Lower Bounds

The sharp feasible interval is Proposition 4.1; the case K_2 is included in Theorem 4.5. For |E(T)|\ge2, the fixed-point formula and coordinatewise minimality are Theorem 4.5, uniqueness of the normalized k-optimum is Theorem 4.7, complementarity is Corollary 4.9, and monotonicity, continuity, and the endpoint constant-curvature statement are Theorem 4.10.

Let x=x_0,\ldots,x_n=y be a shortest path from x to y, and fix 0<\varepsilon<k. Since, by (?),

for every i=1,\ldots,n, for \alpha<1 sufficiently close to 1 we have, simultaneously for all i,

Combining with (?), we have

Writing \delta_z for the Dirac measure at z, we have W_1(\delta_z,\mu_z^{\alpha,w})=1-\alpha. Hence, by the triangle inequality,

Since d(x,y)=n, the preceding estimates give

Thus n(k-\varepsilon)\le2. Letting \varepsilon\downarrow0 gives d(x,y)\le2/k.

For k=0, the assertion is exactly the 2-core reduction of [Xia2026InvariantI], so assume k>0.

Let w be k-admissible on G, and aggregate the first-layer weights at each x\in A_G by setting

By Lemma 2.4, core curvatures are unchanged. If e_C is the boundary edge of a component C\in\mathcal C_x, Theorem 3.2 gives

Summing over C\in\mathcal C_x gives b_x/d_x^w\ge kh_G(x)/2. Set

Since aggregation preserves d_x^w, the resulting proxy weight \widehat w is k-loaded admissible.

Conversely, let \widehat w be k-loaded admissible and fix x\in A_G. Write \mathcal C_x=\{C_1,\ldots,C_q\}, t_i=|V(C_i)|, and b_x=\widehat w_{xx^\partial}. By (?), choose positive b_i with

Assign b_i to the boundary edge of C_i and apply the converse part of Theorem 3.2 independently to each component, then we can reconstruct a positive weight w_G on G such that every edge of G\setminus K has curvature exactly k and the boundary edge of C_i has curvature at least k. The total first-layer load at every core vertex remains b_x, so Lemma 2.4 preserves all core curvatures. The reconstructed weight w_G on G is therefore k-admissible.

By Corollary 3.1, G is finite. Note that D=\mathrm{DN}_{\mathrm{LLY}}^k(G). Proposition 2.2 gives,

The forest obstruction from [Xia2026InvariantI, Theorem 1.1] says that either \mathcal S(G) is a forest or G=C_n for some n\ge6. The cycle case is impossible for k>0 by (?), whose left-hand side would be at least nk>0 but equals 0. Thus \mathcal S(G) is a forest.

Let T be a nontrivial component. By Theorem 5.1, \mathrm{DN}_{\mathrm{LLY}}^{k}(T)\le D. Theorem 1.1 gives

Finally, by Proposition 2.2,

so the branch-vertex bound from [Xia2026InvariantI, Theorem 1.5] yields |B(T)|\le\lceil D\rceil.

If w is k-admissible, (?) gives

Since k>0, the right-hand side is positive, so |E(G)|<|V(G)|. Connectedness gives |E(G)|\ge|V(G)|-1, hence |E(G)|=|V(G)|-1 and G is a tree. Then Theorem 1.1 gives k\le2/|E(G)|. Conversely, the assertion follows directly from Theorem 1.1.

The case G=K_1 is immediate. Assume E(G)\ne\varnothing. Choose a k-admissible weight w_0 of finite distortion and normalize it so that \min_e w_{0,e}=1. Put M=\operatorname{dist}(w_0). It suffices to minimize over the nonempty set

By continuity of w\mapsto(\kappa_{\mathrm{LLY}}^w(e))_{e\in E(G)}, \mathcal A_M is closed and hence compact. On \mathcal A_M, \operatorname{dist}(w)=\max_{e\in E} w_e, a continuous function, so the infimum is attained. The monotonicity follows because every k_2-admissible weight is k_1-admissible.

A finite-distortion k-admissible weight would give finite diameter by Proposition 1.2; a locally finite connected graph of finite diameter is finite.

Put c_k=1+k/2. Since the boundary edge e and every edge of C are tree-like, (?) applies. Summing it over the t-1 internal edges of C gives

Applying (?) to e=xr yields

proving necessity; (3.2) is algebraically equivalent.

For sufficiency, (3.2) implies kt\le2. Root C at r. For v\in V(C) let t_v be the number of vertices in the descendant subtree rooted at v and set

Since t_v\le t and kt\le2, one has \sigma_v>0. For each vertex u\in V(C), let e_u be its parent edge, with e_r=e. Set [u]:=\{v\in V(C): v\ \text{is a child of}\ u\}, and define recursively

Since the child subtree sizes sum to t_u-1,

Hence for every internal parent–child edge uv,

so its curvature is k by (?). At the root, p_r^w(e)=1-k(t-1)/2; together with p_x^w(e)\ge kt/2, this gives curvature at least k on e by (?), with equality when the inequality in (?) is sharp.

Let x\in A_G. By (?),

It is strictly less than 1 because x also carries positive core-edge weight.

If every edge has curvature at least k, then (?) gives mk\le2. Conversely, by the constant-curvature existence result of Lin–Liu [LinLiu2026, Theorem 3.1], there is a positive weight w on T such that

Hence this weight is k-admissible for every 0\le k\le2/m.

The first two assertions are immediate. For k=0,

which is nondecreasing in each variable.

Assume k>0. For (a,b)\ne(0,0), direct differentiation gives

If a\ge b, this is clearly positive. If a<b, then

and hence

Thus

By symmetry,

At the origin, strict increase along the coordinate axes follows from

Therefore \Psi_k is strictly increasing in each variable for k>0.

Substituting (2.1) and clearing denominators gives

Its unique nonnegative root is (?).

By Proposition 4.1, choose a normalized k-admissible vector x^{(0)}. Iterating x^{(n+1)}=F_{T,k}(x^{(n)}) gives a coordinatewise decreasing sequence bounded below by \mathbf1, because F_{T,k} is order preserving. Its limit is a fixed point by continuity. Since x^{(0)} is normalized, some coordinate initially equals 1; that coordinate remains 1 throughout the iteration. Hence the limit is normalized.

For uniqueness, the case k=0 is exactly the finite-tree fixed-point result of [Xia2026InvariantI], by (?). Let k>0 and let x,y be normalized fixed points. Set

Without loss of generality, let R\ge1. If R>1, choose e=u\sim v attaining it. Then x_e=Ry_e>1, and homogeneity and monotonicity of \Psi_k give

Equality holds throughout. Strict monotonicity therefore forces

The equality of the corresponding sums, together with x_f\le R y_f termwise, forces x_f=R y_f on every adjacent edge f. Repeating this along the connected edge set of the tree gives x=Ry, contradicting normalization. Exchanging x and y gives y\le x, and hence x=y.

Starting the iteration from any normalized k-admissible vector z produces the unique fixed point below z, proving coordinatewise minimality. Since the fixed point is itself k-admissible, (4.5) follows by Proposition 2.2. For T=K_2, the assertion is immediate by Proposition 4.1.

At k=0 this is the uniqueness assertion in the finite-tree fixed-point theorem of [Xia2026InvariantI]. The case T=K_2 is immediate, so assume k>0 and |E(T)|\ge2. Let w be normalized and k-optimal. By coordinatewise minimality,

Put D=\mathrm{DN}_{\mathrm{LLY}}^{k}(T)=\max_e w_{T,e}^{\star,k}. If D=1, both weights are the unit weight.

Assume D>1 and suppose that

for some f\in E(T). Let e=uv be adjacent to f, with f\ni u, and assume w_{T,e}^{\star,k}>1. Since w\ge w_T^{\star,k},

By k-admissibility of w and strict monotonicity of \Psi_k,

Thus strict inequality propagates from f to every adjacent edge whose canonical weight is greater than 1.

For every non-pendant edge e=uv,

and hence

Choose e_0 such that

Let

be a shortest sequence of pairwise adjacent edges. Every f_j, 1\le j<r, is non-pendant, and w_{T,e_0}^{\star,k}=D>1. Applying the preceding comparison successively gives

and therefore

contrary to \max_{e\in E(T)}w_e=D. Hence

The unnormalized statement follows from scale invariance.

If w_{T,e}^{\star,k}>1, the second term in (4.4) is active, so Lemma 4.3 gives curvature exactly k. If the curvature is strictly larger than k, the contrapositive gives w_{T,e}^{\star,k}=1.

The case K_2 is immediate. Let m\ge2 and 0\le k_1\le k_2\le2/m. The k_2-canonical weight is k_1-admissible, so minimality gives

coordinatewise.

If k_j\to k, then by (Proof),

Hence the sequence \{w_T^{\star,k_j}\} lies in the compact set

Let a convergent subsequence satisfy

Since

for every \ell, and since (k,x)\mapsto F_{T,k}(x) is continuous, passing to the limit gives

Thus w is a fixed point of F_{T,k}. By uniqueness of the fixed point,

Thus every convergent subsequence of \{w_T^{\star,k_j}\} has the same limit w_T^{\star,k}. Since the sequence lies in a compact set, it follows that

Hence the map

is continuous on [0,2/m]. In particular, the map

is continuous on [0,2/m] for every e\in E(T). Since E(T) is finite and

continuity of each k\mapsto w_{T,e}^{\star,k} implies continuity of \mathcal D_T. Since each map k\longmapsto w_{T,e}^{\star,k} is nondecreasing, the function \mathcal D_T is nondecreasing as well by (?).

At k=2/m, all curvatures are at least 2/m and their sum is 2 by (?), so all are equal. If k<2/m and all curvatures equaled k, the same identity would give 2=mk<2, a contradiction.

If E(H)=\varnothing, then H=K_1 and the claim follows from \mathrm{DN}_{\mathrm{LLY}}^k(K_1)=1\le\mathrm{DN}_{\mathrm{LLY}}^k(G). If \mathrm{DN}_{\mathrm{LLY}}^k(G)=\infty, there is nothing to prove. Otherwise, restrict any finite-distortion k-admissible weight w on G to H. For any edge xy\in E(H),

so (2.1) shows that its curvature cannot decrease. Restriction also cannot increase distortion. Indeed, \operatorname{dist}(w|_H)\le\operatorname{dist}(w). Taking infima gives the result.

For a normalized k-optimal weight w, 1\le w_e\le D. Hence on a tree-like edge xy, we have

By Lemma 4.3 and the monotonicity of \Psi_k,

This proves (5.2).