MerLean-example

31 Lem 3: Relative Vertex-to-Edge Expansion

Definition 356 Edge Boundary at Vertex

Let \(G = (V, E)\) be a simple graph and \(S \subseteq V\). For a vertex \(v \in V\), the edge boundary at \(v\) is the set of edges in the edge boundary \(\delta S\) that are incident to \(v\):

\[ (\delta S)_v \; :=\; \{ e \in \delta S \mid v \in e \} . \]
Definition 357 High Expansion Vertices

Let \(G = (V, E)\) be a simple graph, \(S \subseteq V\), \(s \in \mathbb {N}\), and \(b \in \mathbb {R}\). The set of high expansion vertices is

\[ A \; :=\; \{ v \in S \mid s - b \; \leq \; |(\delta S)_v| \} , \]

i.e., the vertices in \(S\) whose local boundary degree is at least \(s - b\).

Lemma 358 \(A \subseteq S\)

For any graph \(G\), subset \(S \subseteq V\), and parameters \(s \in \mathbb {N}\), \(b \in \mathbb {R}\), the set of high expansion vertices satisfies \(A \subseteq S\).

Proof

Since \(A\) is defined as a filtered subset of \(S\) (i.e., \(A = S.\operatorname {filter}(\cdots )\)), the inclusion \(A \subseteq S\) follows immediately from the fact that any filtered subset of a finset is contained in that finset.

Lemma 359 Nonemptiness of High Expansion Vertices

If there exists a vertex \(v \in S\) with \(|(\delta S)_v| \geq s - b\), then \(A\) is nonempty.

Proof

Assume there exist \(v \in S\) and a proof \(h_{vb}\) that \(s - b \leq |(\delta S)_v|\). Decomposing the hypothesis, we obtain the vertex \(v\), its membership \(v \in S\), and the bound \(h_{vb}\). Then \(v \in A\) follows directly from the membership criterion \(A = S.\operatorname {filter}(s - b \leq |(\delta S)_\cdot |)\), and hence \(A\) is nonempty.

Lemma 360 Partition Sum of Edge Boundary

For any graph \(G\) and subset \(S \subseteq V\), the edge boundary decomposes as a disjoint union over vertices in \(S\):

\[ |\delta S| \; =\; \sum _{v \in S} |(\delta S)_v|. \]
Proof

We first show that \(\delta S = \bigcup _{v \in S} (\delta S)_v\). For the forward direction: given \(e \in \delta S\), by the characterization of the edge boundary (using mem_edgeBoundary), \(e = \{ a, b\} \) with \(a \in S\) and \(b \notin S\); thus \(e \in (\delta S)_a\) and \(a \in S\), so \(e\) lies in the big union. For the reverse direction: if \(e \in (\delta S)_v\) for some \(v \in S\), then \(e \in \delta S\) by the filter condition.

Next, we show the family \(\{ (\delta S)_v\} _{v \in S}\) is pairwise disjoint. Suppose \(e \in (\delta S)_v \cap (\delta S)_w\) with \(v \neq w\). Then \(e \in \delta S\), and by mem_edgeBoundary, \(e = \{ a,b\} \) with \(a \in S\), \(b \notin S\). Both \(v \in e\) and \(w \in e\) hold. By the membership condition for \(\operatorname {Sym2}\), each of \(v, w\) equals either \(a\) or \(b\). Since \(b \notin S\) but \(v, w \in S\), both \(v = a\) and \(w = a\), contradicting \(v \neq w\).

Having established the disjoint partition, the cardinality identity follows from the finset big-union cardinality formula for pairwise disjoint families.

Let \(G\) be an \(s\)-regular graph. For any \(S \subseteq V\) and \(v \in V\),

\[ |(\delta S)_v| \; \leq \; s. \]
Proof

We compute:

\[ |(\delta S)_v| \; \leq \; |G.\operatorname {incidenceFinset}(v)| \; =\; \deg _G(v) \; =\; s. \]

The first inequality holds because \((\delta S)_v \subseteq G.\operatorname {incidenceFinset}(v)\): any edge \(e \in (\delta S)_v\) lies in the edge boundary \(\delta S \subseteq E(G)\) (by edgeBoundary_subset_edgeFinset) and satisfies \(v \in e\), hence \(e \in G.\operatorname {incidenceFinset}(v)\). The equality of the incidence finset cardinality with \(\deg _G(v)\) follows from the standard identity, and \(\deg _G(v) = s\) by \(s\)-regularity.

Lemma 362 Relative Cheeger with Non-strict Inequality

Let \(G\) be a finite, connected, \(s\)-regular graph on \(|V| \geq 2\) vertices with second-largest eigenvalue \(\lambda _2\). For \(\alpha \in (0,1)\) and \(S \subseteq V\) with \(0 {\lt} |S|\) and \(|S| \leq \alpha |V|\),

\[ (1 - \alpha )(s - \lambda _2)|S| \; \leq \; |\delta S|. \]
Proof

We have \(|S| {\gt} 0\) (given) and \(|V| {\gt} 0\) (since \(|V| \geq 2\)). Since \(|S| \leq \alpha |V|\) and \(\alpha {\lt} 1\), we deduce \(|S| {\lt} |V|\), so \(S\) is a proper nonempty subset.

We split into two cases based on the sign of \(s - \lambda _2\).

Case 1: \(s - \lambda _2 \geq 0\). Apply the spectral Laplacian bound (spectralLaplacianBound):

\[ (s - \lambda _2)|S|\! \left(1 - \frac{|S|}{|V|}\right) \; \leq \; |\delta S|. \]

Since \(|S|/|V| \leq \alpha \) (as \(|S| \leq \alpha |V|\)), we have \(1 - \alpha \leq 1 - |S|/|V|\). Multiplying both sides of the latter by the nonneg quantity \((s-\lambda _2)|S| \geq 0\) preserves the inequality, yielding

\[ (1-\alpha )(s-\lambda _2)|S| \; \leq \; (s-\lambda _2)|S|\! \left(1-\frac{|S|}{|V|}\right) \; \leq \; |\delta S|. \]

Case 2: \(s - \lambda _2 {\lt} 0\). Then \((1-\alpha )(s-\lambda _2) {\lt} 0\) (since \(1-\alpha {\gt} 0\)), so \((1-\alpha )(s-\lambda _2)|S| {\lt} 0 \leq |\delta S|\) by linear arithmetic.

Let \(G\) be an \(s\)-regular graph, \(S \subseteq V\), \(b \in \mathbb {R}\), and \(A = \{ v \in S \mid |(\delta S)_v| \geq s - b\} \). Then

\[ |\delta S| \; \leq \; s|A| + (s - b)(|S| - |A|). \]
Proof

Let \(B := S \setminus A\). Since \(A \subseteq S\), we have the disjoint decomposition \(S = A \sqcup B\).

By the partition sum identity (Lemma 360):

\[ |\delta S| \; =\; \sum _{v \in S} |(\delta S)_v| \; =\; \sum _{v \in A} |(\delta S)_v| + \sum _{v \in B} |(\delta S)_v|. \]

For the sum over \(A\): since each \(|(\delta S)_v| \leq s\) by \(s\)-regularity (Lemma 361), the bounded sum inequality gives \(\sum _{v \in A} |(\delta S)_v| \leq s|A|\).

For the sum over \(B\): for any \(v \in B = S \setminus A\), we have \(v \notin A\), meaning \(|(\delta S)_v| {\lt} s - b\) (i.e., the defining condition for \(A\) fails). Hence \(|(\delta S)_v| \leq s - b\) (strict inequality implies \(\leq \)), and again by the bounded sum inequality, \(\sum _{v \in B} |(\delta S)_v| \leq (s-b)|B|\).

Finally, \(|B| = |S \setminus A| = |S| - |A|\) (since \(A \subseteq S\)), obtained by the cardinality identity for set differences. Combining all estimates gives \(|\delta S| \leq s|A| + (s-b)(|S| - |A|)\).

Let \(G\) be a finite, connected, \(s\)-regular graph on vertex set \(V\) with \(|V| \geq 2\) and second-largest eigenvalue \(\lambda _2\). Let \(\alpha \in (0,1)\), \(b \in \mathbb {R}\) with \(0 {\lt} b \leq s\), and \(S \subseteq V\) with \(|S| \leq \alpha |V|\). Define

\[ A \; :=\; \bigl\{ v \in S \; \big|\; |(\delta S)_v| \geq s - b \bigr\} , \qquad \beta \; :=\; \frac{(b - \lambda _2) - \alpha (s - \lambda _2)}{b}. \]

Then

\[ |A| \; \geq \; \beta \, |S|. \]
Proof

Step 0: Trivial case. If \(|S| = 0\), then both sides are zero and the inequality holds trivially by simplification. Henceforth assume \(|S| {\gt} 0\).

Step 1: Lower bound on \(|\delta S|\). By Lemma 362 (relative Cheeger with \(\leq \)):

\[ (1-\alpha )(s - \lambda _2)|S| \; \leq \; |\delta S|. \]

Step 2: Upper bound on \(|\delta S|\). By Lemma 363:

\[ |\delta S| \; \leq \; s|A| + (s-b)(|S| - |A|). \]

Step 3: Combine. From Steps 1 and 2 by linear arithmetic:

\[ (1-\alpha )(s-\lambda _2)|S| \; \leq \; s|A| + (s-b)(|S| - |A|). \]

Expanding the right side: \(s|A| + (s-b)|S| - (s-b)|A| = b|A| + (s-b)|S|\). Hence:

\[ (1-\alpha )(s-\lambda _2)|S| - (s-b)|S| \; \leq \; b|A|. \]

Step 4: Simplify the coefficient. We compute by ring arithmetic:

\[ (1-\alpha )(s-\lambda _2) - (s-b) \; =\; (b - \lambda _2) - \alpha (s-\lambda _2). \]

Thus:

\[ \bigl[(b-\lambda _2) - \alpha (s-\lambda _2)\bigr]|S| \; \leq \; b|A|. \]

Step 5: Divide by \(b\). Since \(b {\gt} 0\), dividing both sides by \(b\) gives:

\[ \frac{(b-\lambda _2) - \alpha (s-\lambda _2)}{b} \cdot |S| \; \leq \; |A|, \]

which is the desired conclusion. This final step is performed by rewriting with \(\operatorname {div_mul_eq_mul_div}\), applying \(\operatorname {div_le_iff}\) with \(b {\gt} 0\), and concluding by linear arithmetic.