Take for example $\Z_2^d$ with the metric defined in exam. The diameter of $\Z_2^d$ is $d$ and for any $1$-Lipschitz function $f:\Z_2^d\rar\R$ we have $\G(f,f)\leq d/2$ and thus $\G(f/\sqrt{d/2},f/\sqrt{d/2})\leq1$; by corollary it follows that: $$ \mu(|f-\la f\ra| > \e\sqrt d)\leq2e^{-2\vp(\e\sqrt2)}~. $$ Though the space has diameter $d$ most of the points of $\Z_2^d$ can be found in a set where the order of variation of $f$ is at most $\sqrt d$. In particular (remember: $\l_1=2$): $$ \mu(|f-\la f\ra| > d/4)\leq2e^{-\sqrt{d/2}}~. $$ In the subsequent subsection we will improve significantly upon the right hand side and show that it's actually bounded by $e^{-d/16}$.

Curvature - the Bakry-Emery criterion

Following D. Bakry and M. Emery (cf. e.g. M. Ledoux) we say that the curvature of $\D_R$ is bounded from below by $C$ if $$ \forall f\in L_2(G)\,\forall x\in G:\quad \G_2(f,f)(x)\geq C\G(f,f)(x)~. $$ That's the so called (infinite dimensional) Bakry-Emery criterion; it was originally aedined for diffusions on mainfolds. We are going to elaborate what this criterion con produce in a discrete setting! First of all, we notice, that in our case it's enough to check the case $x=e$, i.e.: $$ \forall f\in L_2(G):\quad \G_2(f,f)(e)\geq C\G(f,f)(e)~. $$ By the Spectral Gap Theorem we must have $C\leq\l_1$. Thus if we can find some lower bound $C$ for the curvature, then $\l_1\geq C$. However, we are actually aiming at a stronger form of the concentration of measure result.
Suppose for all $x\in G$: $xRx^{-1}=R$ and for all $x\in R$: $x^2=e$. Then the curvature of $(G,R)$ is bounded from below by $2$, but it might be that this is not the best possible lower bound.
Obviously $\G_2(f,f)\geq\tfrac14\sum_{x\in R}(\nabla_x\nabla_xf)^2$. Since $x^2=e$, we have: $\nabla_x\nabla_xf=2\nabla_xf$.
In particular for the groups $\Z_2$ and $S(n)$ with $R$ the set of transpositions the curvature is bounded from below by $2$.
Find the best possible lower bound for the curvature of $S(3)$.
Find the best possible lower bound for the curvature of $S(4)$.
Suppose for alle $x\in G$: $xRx^{-1}\sbe R$. Verify that if $C$ is a lower bound for the curvature of $(G,R)$, then $C$ is a lower bound for the curvature of $(G\times G,R\times\{e\}\cup(\{e\}\times R)$.
Put $S=R\times\{e\}\cup(\{e\}\times R$, then \begin{eqnarray*} \G_2(f,f)(e,e) &=&\sum_{(x,y)\in S}\sum_{(u,v)\in S}(\nabla_{(x,y)}\nabla_{(u,v)}f(e,e))^2\\ &\geq&\sum_{x\in R}\sum_{u\in R}(\nabla_{(x,e)}\nabla_{(u,e)}f(e,e))^2 +\sum_{y\in R}\sum_{v\in R}(\nabla_{(e,y)}\nabla_{(e,v)}f(e,e))^2\\ &=&\G_2(f(.,e),f(.,e))(e)+\G_2(f(e,.),f(e,.))(e)\\ &\geq&C\sum_x(\nabla_{(x,e)}f(e,e))^2+C\sum_x(\nabla_{(e,x)}f(e,e))^2 \end{eqnarray*}
Prove that a lower bound for the curvature of $(G,R)$ for $R=G\sm\{e\}$ is given by $C=|G|/2$.
A lower bound for the curvature for $\Z_3$ with $R=\{\pm1\}$ is thus given by $3/2$.
Compute a lower bound for the curvature of $(\Z_4,R)$ with $R=\{\pm1\}$.
Compute a lower bound for the curvature of $(\Z_5,R)$ with $R=\{\pm1\}$.
Verify that for $p\geq6$ $(\Z_p,R)$, $R=\{\pm1\}$ cannot have a strictly positive lower bound of curvature.
Now we are going to apply the Bakry-Emery criterion:
1. If the curvature of $\D_R$ is bounded from below by $C > 0$, then $$ \G(P_tf,P_tf)\leq e^{-2Ct}P_t\G(f,f)~. $$ 2. For $f\in L_2(G)$ and $\l\geq0$ we have $$ \la\G(f,e^{\l f})\ra\leq\l\la e^{\l f},\G(f,f)\ra~. $$
$\proof$ 2. We will again utilize the Hermite-Hadamard inequality for $\vp(x)=e^x$: \begin{eqnarray*} \G(f,e^{\l f})(y) &=&\tfrac12\sum_{x\in R} (f(xy)-f(y))(e^{\l f(xy)}-e^{\l f(y)})\\ &=&\tfrac12\sum_{x\in R} \l(f(xy)-f(y))^2\frac{e^{\l f(xy)}-e^{\l f(y)}}{f(xy)-f(y)}\\ &\leq&\tfrac14\l\sum_{x\in R}(f(xy)-f(y))^2(e^{\l f(xy)}+e^{\l f(y)}) \end{eqnarray*} Summing over $y\in G$ yields by symmetry of $R$: $$ \la\G(f,e^{\l f})\ra\leq\l\la e^{\l f},\G(f,f)\ra~. $$ 1. By differentiation of the function $F(s)\colon=P_s\G(P_{t-s}f,P_{t-s}f)$ we get for $u=P_{t-s}f$: $$ F^\prime(s) =-P_s\D_R\G(u,u)+P_s\G(u,\D_Ru)+P_s\G(\D_Ru,u) =2P_s\G_2(u,u) \geq2CP_s\G(u,u) =2CF(s) $$ Hence $F(t)\geq F(0)e^{2Ct}$, i.e. for all $f$: $$ \G(P_tf,P_tf)\leq e^{-2Ct}P_t\G(f,f)~. $$ $\eofproof$
Now we will employ Herbst's argument to prove a concentration of measure result:
Let $\mu$ be the normalized counting measure and suppose $\G(f,f)\leq1$. Then for all $\e > 0$: $$ \mu(|f-\la f\ra| > \e)\leq2e^{-\frac12C\e^2}~. $$
$\proof$ W.l.o.g. we may assume that $\la f\ra=0$. Put for $\l\geq0$: $F(t)\colon=\la e^{\l P_tf}\ra$. Then by lemma: \begin{eqnarray*} -F^\prime(t) &=&\l\la\D_RP_tf,e^{\l P_tf}\ra =\l\la\G(P_tf,e^{\l P_tf})\ra\\ &\leq&\l\la\l e^{P_tf},\G(P_tf,P_tf)\ra \leq\l^2e^{-2Ct}\la e^{\l P_tf}\ra =\l^2e^{-2Ct}F(t), \end{eqnarray*} Thus $(\log F)^\prime(t)\leq-\l^2 e^{-2Ct}$ and since $\log F(\infty)=0$ we conclude that $\log F(0)\leq\l^2/2C$, which by Chebyshev's inequality implies: $$ \mu(f > \e) =\mu(\exp(\l f) > \exp(\l\e)) \leq\exp(-\l\e)\la e^{\l f}\ra \leq e^{-\l\e+\l^2/2C}~. $$ Putting $\l=C\e$, we get the desired deviation inequality. $\eofproof$
Let $f:(G,d_R)\rar\R$ be a $1$-Lipschitz function, then $$ \mu(|f-\la f\ra| > t)\leq2e^{-Ct^2/|R|}~. $$
Let $f(x)\colon=d_R(x,e)$ be the distance function. Then for every $(x,y)\in R\times R$ satisfying $d_R(xy,e)=2$ we have $\nabla_x\nabla_yf(e)=0$; it follows that $$ \G_2(f,f)(e) =\tfrac14\sum_{x,y\in R:xy\in R}(\nabla_x\nabla_yf)^2(e) +\tfrac14\sum_{x\in R}(\nabla_x\nabla_{x^{-1}}f)^2(e) =\tfrac14|\{(x,y)\in R\times R:xy\in R\}|+|R| $$ i.e. $$ C\leq2+\frac{|\{(x,y,z)\in R\times R\times R:xyz=e\}|}{2|R|} $$ In order to get a large lower bound for the curvature we need an abundant set of generators! If this set is somehow minimal the best we can haope for is $C=2$.