幂律分布：all you need to know about power law distribution

For power law distribution
$$p_k=\frac{\nu -1}{k_{min}}(\frac{k}{k_{min}}{)}^{-\nu }$$
To estimate $\nu$ by moment estimation method
$$\begin{gathered} Formethodofmoments \\ E[x]=\frac{\nu -1}{\nu -2} \\ E[x^2]=\frac{\nu -1}{\nu -3}{x}_{min}^2 \\ {\hat{x}}_{min}=\frac{\sum x_i^2(\nu -3)}{\sum x_i(\nu -2)} \\ {\hat{v}}_{MME}=\frac{-2(\sum x_i{)}^2+2n\sum x_i^2\pm \sqrt{-n(\sum x_i{)}^2\sum x_i^2+n^2(\sum x_i^2{)}^2}}{-(\sum x_i{)}^2+n\sum x_i^2} \\ Formaximizinglog-likelihood \\ p(x)=\prod _{i=1}^n\frac{\nu -1}{x_{min}}(\frac{x}{x_{min}}{)}^{-\nu } \\ logp(x)=nlog(\nu -1)-nlog{x}_{min}-\nu \sum _{i=1}^{\infty }log\frac{x_i}{x_{min}} \\ \frac{\partial logp(x)}{\partial x_{min}}=\frac{-n}{x_{min}}+\frac{\nu }{x_{min}}\sum _{i=1}^{\infty }\frac{1}{x_i}=0 \\ \frac{\partial logp(x)}{\partial \nu }=\frac{n}{\nu -1}-\sum _{i=1}^{\infty }log\frac{x_i}{x_{min}} \\ {\hat{\nu }}_{MLE}=1+n[\sum _{i=1}^{n}ln\frac{x_i}{x_{min}}{]}^{-1} \end{gathered}$$
To find the statistical properties of these estimator including unbiasedness and consistency
$$\begin{gathered} E[{\hat{\nu }}_{MME}]=\frac{-2(\frac{n}{\nu -1}{)}^2+2n\frac{n^2}{(\nu -1)(\nu -2)}\pm \sqrt{-n(\frac{n}{\nu -1}{)}^2\frac{n^2}{(\nu -1)(\nu -2)}+n^2(\frac{n^2}{(\nu -1)(\nu -2)}{)}^2}}{-(\frac{n}{\nu -1}{)}^2+n\frac{n^2}{(\nu -1)(\nu -2)}} \\ whenn\to \infty ,E[{\hat{\nu }}_{MME}]\to 3or1 \\ E[{\hat{\nu }}_{MLE}]=1+nE[\sum _{i=1}^{n}ln\frac{x_i}{x_{min}}{]}^{-1}] \\ =1+E[\frac{n}{z}] \\ =1+\frac{n(\nu -1)}{n-1} \\ =\frac{n\nu }{n-1}-\frac{1}{n-1} \\ whenn\to \infty ,E[{\hat{\nu }}_{MLE}]\to \nu \end{gathered}$$
So MME estimator and MLE estimator are both biased estimator and MLE estimator is a consisent estimator while MME estimator is not.

For joint PDF
$$\begin{gathered} f(x;\nu )=(\nu -1{)}^n{x}_{min}^{-n(\nu -1)}\prod _{i=1}^n{x}_i^{-\nu }{1}_{x_i\ge x_{min}} \\ =c(\nu )h(x)exp[\sum _{j=1}^k{q}_j(\nu )t_j(x)] \\ =1_{x_i\ge x_{min}}({\nu }^n{x}_{min}^{n(\nu -1)})exp[-\nu \sum _{i=1}^{n}ln{x}_i] \end{gathered}$$
So ${\sum }_{i=1}^{n}ln{x}_i$ is complete and sufficient.

Let $ Y=ln\frac{x}{x_{min}}$,so $x=x_{min}{e}^Y$
$$\begin{gathered} f_X(x)=(\nu -1)x_{min}^{\nu -1}{x}^{-\nu }{1}_{x\ge x_{min}} \\ {f}_Y(y)=f_X(x_{min}{e}^Y)|\frac{dx}{dy}| \\ =(\nu -1)x_{min}^{\nu -1}(x_{min}{e}^Y{)}^{-\nu }{1}_{x_{min}{e}^Y\ge x_{min}}{x}_{min}{e}^Y \\ =(\nu -1)e^{-y(\nu -1)}{1}_{Y\ge 0} \end{gathered}$$
So y follows the exponential distribution with rate $\nu -1$
$$\begin{gathered} z=\sum _{i=1}^n{Y}_i=\sum _{i=1}^{n}ln{X}_i-ln{X}_{min} \\ z\sim \Gamma (n.\nu -1) \\ E(Y_i)=\frac{1}{\nu -1} \\ E(z)=\frac{n}{\nu -1} \end{gathered}$$
Therefore $\frac{1}{n}{\sum }_{i=1}^n{Y}_i=\frac{1}{\nu -1}$ which means $\frac{1}{n}{\sum }_{i=1}^n{Y}_i$ is an unbiased estimator for $\frac{1}{\nu -1}$, we guess $1+\frac{1}{\frac{1}{n}{\sum }_{i=1}^n{Y}_i}$ is an unbiased estimator for $\nu$.
$$\begin{gathered} E[\frac{n}{z}]=E[\frac{1}{\frac{1}{n}{\sum }_{i=1}^n{Y}_i}] \\ =n{\int }_0^{\infty }\frac{1}{z}\frac{1}{\Gamma (n)}(\nu -1{)}^n{z}^{n-1}{e}^{-(\nu -1)z}dz \\ =n\frac{(\nu -1)\Gamma (n-1)}{\Gamma (n)}{\int }_0^{\infty }\frac{1}{\Gamma (n-1)}(\nu -1{)}^{n-1}{z}^{n-2}{e}^{-(\nu -1)z}dz \\ =\frac{n(\nu -1)}{n-1} \\ AccordingtoLehmann-ScheffeTheorem \\ \sum _{i=1}^{n}ln{x}_{i}iscompleteandsufficient \\ \because T=\frac{n-1}{n}\frac{1}{\frac{1}{n}{\sum }_{i=1}^n{Y}_i}+1isunbiasedfor\tau (\nu )=\nu \end{gathered}$$
So an unbiased for $\nu$ is ${\hat{\nu }}_{good}=\frac{n-1}{n}\frac{1}{\frac{1}{n}{\sum }_{i=1}^n{Y}_i}+1=\frac{n-1}{{\sum }_{i=1}^{n}ln{X}_i-ln{X}_{min}}+1$.

To find CRLB for $\nu$
$$\begin{gathered} CRLB(\nu )=\frac{{\tau }^{\prime }(\nu {)}^2}{nE[\frac{\partial }{\partial \nu }lnf(x;\nu ){]}^2} \\ \tau (\nu )=\nu \\ {p}_k=\frac{\nu -1}{k_{min}}(\frac{k}{k_{min}}{)}^{-\nu } \\ lnf(x;\nu )=ln(\nu -1)-(\nu -1)ln{x}_{min}-\nu lnx \\ \frac{\partial }{\partial \nu }lnf(x;\nu )=\frac{1}{\nu -1}-ln{x}_{min}-lnx \\ CRLB(\nu )=\frac{1}{nE[\frac{1}{\nu -1}-Y{]}^2} \\ E[\frac{1}{\nu -1}-Y{]}^2=\frac{1}{(\nu -1{)}^2}-\frac{2}{\nu -1}\frac{1}{\nu -1}+Var[y^2]+E[y{]}^2 \\ =\frac{1}{(\nu -1{)}^2} \\ \therefore CRLB(\nu )=\frac{(\nu -1{)}^2}{n} \end{gathered}$$

$$\begin{gathered} E[{\hat{\nu }}_{MLE}]=E[1+\frac{1}{\frac{1}{n}{\sum }_{i=1}^n{Y}_i}]=E[\frac{n}{z}]+1 \\ =\frac{n(\nu -1)}{n-1}+1 \\ E[{\hat{\nu }}_{MLE}^2]=E[1+\frac{2}{\frac{1}{n}{\sum }_{i=1}^n{Y}_i}+(\frac{1}{\frac{1}{n}{\sum }_{i=1}^n{Y}_i}{)}^2] \\ =1+\frac{2n(\nu -1)}{n-1}+n^2{\int }_0^{\infty }\frac{1}{z^2}\frac{1}{\Gamma (n)}(\nu -1{)}^n{z}^{n-1}{e}^{-(\nu -1)z}dz \\ =1+\frac{2n(\nu -1)}{n-1}+\frac{n^2(\nu -1{)}^2}{(n-1)(n-2)} \\ Var({\hat{\nu }}_{MLE}^2)=E[{\hat{\nu }}_{MLE}^2]-E[{\hat{\nu }}_{MLE}{]}^2 \\ =1+\frac{2n(\nu -1)}{n-1}+\frac{n^2(\nu -1{)}^2}{(n-1)(n-2)}-(\frac{n(\nu -1)}{n-1}+1{)}^2 \\ =\frac{n^2(\nu -1{)}^2}{(n-1{)}^2(n-2)} \\ \therefore CRLB(\nu )=\frac{(\nu -1{)}^2}{n}\le \frac{n^2(\nu -1{)}^2}{(n-1{)}^2(n-2)}=Var({\hat{\nu }}_{MLE}^2) \end{gathered}$$

$$\begin{gathered} Var({\hat{\nu }}_{UMVUE})=Var(\frac{n-1}{n}\frac{1}{\frac{1}{n}{\sum }_{i=1}^n{Y}_i}+1) \\ =\frac{(n-1{)}^2}{n^2}Var({\hat{\nu }}_{MLE}) \\ =\frac{(n-1{)}^2}{n^2}\ast \frac{n^2(\nu -1{)}^2}{(n-1{)}^2(n-2)} \\ =\frac{(\nu -1{)}^2}{n-2} \\ \therefore CRLB(\nu )=\frac{(\nu -1{)}^2}{n}\le \frac{(\nu -1{)}^2}{n-2}=Var({\hat{\nu }}_{UMVUE}^2) \end{gathered}$$

To estimate $\nu$ by confidence interval
$$\begin{gathered} Y=ln\frac{x}{x_{min}}whichfollowsexponentialdistrbution \\ 100(1-\alpha )\%=P({\Gamma }_{\frac{\alpha }{2}}(n,\nu -1)\le Y\le {\Gamma }_{1-\frac{\alpha }{2}}(n,\nu -1)) \\ =P[\frac{n}{{\Gamma }_{1-\frac{\alpha }{2}}(n,\nu -1)}+1\le {\hat{\nu }}_{MLE}\le \frac{n}{{\Gamma }_{\frac{\alpha }{2}}(n,\nu -1)}+1] \\ CI=(\frac{n}{{\Gamma }_{1-\frac{\alpha }{2}}(n,\nu -1)}+1,\frac{n}{{\Gamma }_{\frac{\alpha }{2}}(n,\nu -1)}+1) \end{gathered}$$
According to CLT
$$\begin{gathered} {\hat{\nu }}_{MLE}=1+n[\sum _{i=1}^{n}ln\frac{x_i}{x_{min}}{]}^{-1} \\ \frac{{\sum }_{i=1}^n{X}_i}{n}=e^{\frac{n}{\hat{\nu }-1}}\frac{{\sum }_{i=1}{X}_{min}}{n} \\ \frac{\overline{X}-e^{\frac{n}{\hat{\nu }-1}}\frac{{\sum }_{i=1}{X}_{min}}{n}}{S/\sqrt{n}}\sim t(n-1) \\ 100(1-\alpha )\%=P(-t_{1-\frac{\alpha }{2}}(n-1)\le \frac{\overline{X}-e^{\frac{n}{\hat{\nu }-1}}\frac{{\sum }_{i=1}{X}_{min}}{n}}{S/\sqrt{n}}\le t_{1-\frac{\alpha }{2}}(n-1)) \\ =P(1+\frac{n}{ln\frac{\overline{X}+t_{1-\frac{\alpha }{2}}(n-1)\frac{S}{\sqrt{n}}}{\frac{{\sum }_{i=1}^n{X}_{min}}{n}}}\le \hat{\nu }\le 1+\frac{n}{ln\frac{\overline{X}-t_{1-\frac{\alpha }{2}}(n-1)\frac{S}{\sqrt{n}}}{\frac{{\sum }_{i=1}^n{X}_{min}}{n}}}) \\ CI=(1+\frac{n}{ln\frac{\overline{X}+t_{1-\frac{\alpha }{2}}(n-1)\frac{S}{\sqrt{n}}}{\frac{{\sum }_{i=1}^n{X}_{min}}{n}}},1+\frac{n}{ln\frac{\overline{X}-t_{1-\frac{\alpha }{2}}(n-1)\frac{S}{\sqrt{n}}}{\frac{{\sum }_{i=1}^n{X}_{min}}{n}}}) \end{gathered}$$
For hypothesis test, test $H_0:\nu >{\nu }_0,H_1:\nu <{\nu }_0$ at significant level $\alpha$.
$$\begin{gathered} \alpha =P[1+n[\sum _{i=1}^{n}ln\frac{x_i}{x_{min}}{]}^{-1}\ge {\nu }_0] \\ =P[\sum _{i=1}^{n}ln\frac{x}{x_{min}}\le \frac{n}{{\nu }_0-1}] \\ =P[{\Gamma }_{1-\alpha }(n,\nu -1)\le \frac{n}{{\nu }_0-1}] \\ If{\Gamma }_{1-\alpha }(n,\hat{\nu }-1)\le \frac{n}{{\nu }_0-1},reject{H}_0 \\ \beta =P(TypeII)=P(accept{H}_0|H_1) \\ =P[{\Gamma }_{1-\alpha }(n,\hat{\nu }-1)>\frac{n}{{\nu }_0-1}|\nu <{\nu }_0] \end{gathered}$$
Similarly, according to CLT
$$\begin{gathered} \alpha =P(\frac{\overline{X}-e^{\frac{n}{\hat{\nu }-1}}\frac{{\sum }_{i=1}{X}_{min}}{n}}{S/\sqrt{n}}\ge t_{1-\alpha }(n-1)) \\ =P[\nu \ge 1+\frac{n}{ln\frac{\overline{X}-t_{1-\alpha }(n-1)\frac{S}{\sqrt{n}}}{\frac{{\sum }_{i=1}^n{X}_{min}}{n}}})] \\ If\hat{\nu }\ge 1+\frac{n}{ln\frac{\overline{X}-t_{1-\alpha }(n-1)\frac{S}{\sqrt{n}}}{\frac{{\sum }_{i=1}^n{X}_{min}}{n}}},reject{H}_0 \end{gathered}$$

estimation method	function	unbiasedness	consistency	effectiveness
measurement	calculate by function F	calculate the E(F)	calculate E(F) given $n\to \infty$	compared with $CRLB(\nu )=\frac{(\nu -1{)}^2}{n}$
MME	$\frac{-2(\sum x_i{)}^2+2n\sum x_i^2\pm \sqrt{-n(\sum x_i{)}^2\sum x_i^2+n^2(\sum x_i^2{)}^2}}{-(\sum x_i{)}^2+n\sum x_i^2}$	biased	$E[{\hat{\nu }}_{MME}]\to 3or1$	bigger than $\frac{(\nu -1{)}^2}{n}$
MLE	$1+n[{\sum }_{i=1}^{n}ln\frac{x_i}{x_{min}}{]}^{-1}$	$E[{\hat{\nu }}_{MLE}]=\frac{n\nu }{n-1}-\frac{1}{n-1}$	$E[{\hat{\nu }}_{MLE}]\to \nu$	$\frac{n^2(\nu -1{)}^2}{(n-1{)}^2(n-2)}\ge \frac{(\nu -1{)}^2}{n}$
unbised estimator	$1+\frac{n-1}{{\sum }_{i=1}^{n}ln{X}_i-ln{X}_{min}}$	$E[{\hat{\nu }}_{UMVUE}]=\nu$	$E[{\hat{\nu }}_{UMVUE}]\to \nu$	$\frac{(\nu -1{)}^2}{n-2}\ge \frac{(\nu -1{)}^2}{n}$

其他相关：
Covid19：传染病模型SIR/SEIR的估计与拟合
Covid19：基于网络的传染病模型Virus spread on network
Covid19：传染病模型SIR/SEIR on Network
Covid19：传染病模型模拟Agent based simulation for SIR/SEIR models.

欢迎关注二幺子的知识输出通道：