Introduction
Pareto models describe heavy-tailed data through a simple distributional form above a lower endpoint. Introduced by Pareto [32 ] to describe wealth, they are also used in economics [35 ] , finance [34 ] , insurance [15 ] , and environmental science [20 ] . For a Pareto type-I random variable X ∼ P ( α ) X \sim\mathcal{P}(\alpha) X ∼ P ( α ) , where α > 0 \alpha> 0 α > 0 is a shape parameter, the distribution function and probability density function are
F ( x ; α ) : = { 1 − 1 x α , x ≥ 1 0 , otherwise , f ( x ; α ) : = { α x α + 1 , x ≥ 1 0 , otherwise , α > 0. (1) F(x;\alpha) :=
\begin{cases}
1-\frac{1}{x^\alpha}, & x \ge1 \\
0, & \text{otherwise},
\end{cases}
\qquad
f(x;\alpha) :=
\begin{cases}
\frac{\alpha}{x^{\alpha+1}}, & x \ge1 \\
0, & \text{otherwise},
\end{cases}
\qquad\alpha> 0. \tag*{(1)} F ( x ; α ) := { 1 − x α 1 , 0 , x ≥ 1 otherwise , f ( x ; α ) := { x α + 1 α , 0 , x ≥ 1 otherwise , α > 0. (1)
Let X 1 , … , X n X_1,\ldots,X_n X 1 , … , X n be independent and identically distributed (iid) random variables with an unknown distribution function F F F supported on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) . We test the composite null hypothesis
H 0 : F ∈ { P ( α ) : α > 0 } H_0:F\in\{\mathcal{P}(\alpha):\alpha>0\} H 0 : F ∈ { P ( α ) : α > 0 }
for some unknown shape parameter α > 0 \alpha>0 α > 0 , against the alternative
H 1 : F ∉ { P ( α ) : α > 0 } , H_1:F\notin\{\mathcal{P}(\alpha):\alpha>0\}, H 1 : F ∈ / { P ( α ) : α > 0 } ,
where the decision is based on the data X 1 , … , X n X_1,\ldots,X_n X 1 , … , X n . For a different positive known lower threshold x m > 0 x_m>0 x m > 0 , the observations should first be divided by x m x_m x m .
Existing Pareto goodness-of-fit procedures use empirical distribution functions [10 ] , probability plots [9 ] , or empirical characteristic functions [27 ] . Characterization-based tests begin with an identity that distinguishes the Pareto family from other distributions; see [29 , 7 ] .
Stein’s method, originally developed for normal approximation [36 ] , characterizes probability distributions through operator identities. Using this approach, Betsch and Ebner [6 ] derived fixed-point characterizations of continuous univariate distributions. For the Pareto type-I distribution, Avhad et al. [3 ] established a Stein-type characterization and proposed goodness-of-fit tests based on integral-type and Cramér–von Mises-type statistics for both complete and right-censored samples. More recently, Bhati and Khandelwal [8 ] combined the Laplace transform with the Stein-based fixed-point characterization of Betsch and Ebner [6 ] . They obtained a characterization of the Pareto type-I distribution and proposed supremum, integral, and L 2 L^2 L 2 -type test statistics based on its empirical counterpart.
In the present paper, we use exponential test functions within the Stein density approach to obtain a Laplace-transform like characterization of the Pareto type-I distribution and construct two test statistics. Integrating the signed empirical discrepancy yields T n T_n T n , a simple moment-type diagnostic, whereas integrating its square yields S n S_n S n . Squaring before integration prevents cancellation between discrepancies at different transform arguments. This distinction has consequences for consistency of the procedures. The test based on T n T_n T n is blind against certain non-Pareto alternatives, whereas the test based on S n S_n S n is consistent against every fixed non-Pareto alternative satisfying the logarithmic moment condition stated below.
A studentized version of T n T_n T n has a standard normal limiting null distribution, permitting direct asymptotic calibration without resampling. In contrast, the limiting null distribution of n S n nS_n n S n is an infinite weighted sum of independent χ 1 2 \chi_1^2 χ 1 2 random variables. The weights are the eigenvalues of the covariance operator of the limiting Gaussian process and depend on the unknown shape parameter α 0 \alpha_0 α 0 . We consider two methods for obtaining critical values. The first uses a parametric bootstrap, with the shape parameter re-estimated in each bootstrap sample. The second replaces α 0 \alpha_0 α 0 by α ^ n \widehat{\alpha}_n α n in the covariance operator, numerically approximates its eigenvalues, and uses the resulting weighted chi-square distribution to obtain approximate critical values. We compare the computational cost and finite-sample performance of these approaches in a simulation study. In the settings considered, the plug-in procedure is computationally faster, particularly for larger samples, whereas the bootstrap generally yields slightly higher empirical power for small samples.
The remainder of the article is organized as follows. Section 2 presents the characterization and constructs the test statistics. Section 3 derives their null limits, identifies the alternatives detected consistently by T n T_n T n , and establishes consistency and calibration results for S n S_n S n . § 4 compares empirical size and power with those of existing procedures. § 5 compares the two calibrations in terms of rejection rates and computation time. § 6 examines threshold sensitivity using Danish fire-insurance losses. § 7 discusses the implications and directions for further research.
New goodness-of-fit tests for the Pareto law
A Stein–Laplace characterization
We specialize the density approach to Stein’s method ([22 , 21 ] ) to the Pareto family ([6 Section 3] ). Exponential test functions yield the following characterization.
Theorem 1. Let α > 0 \alpha> 0 α > 0 and let X X X be a positive random variable. Then X X X has the Pareto type-I distribution P ( α ) \mathcal{P}(\alpha) P ( α ) if and only if, for every t > 0 t > 0 t > 0 ,
α exp ( − t ) − E [ ( t + α + 1 X ) exp ( − t X ) ] = 0. (2) \alpha\exp(-t) - \mathbb{E}\left[\left(t + \frac{\alpha+ 1}{X}\right) \exp(-tX)\right] = 0. \tag*{(2)} α exp ( − t ) − E [ ( t + X α + 1 ) exp ( − tX ) ] = 0. (2)
Proof. Integration by parts proves the forward implication. For X ∼ P ( α ) X \sim\mathcal{P}(\alpha) X ∼ P ( α ) ,
E [ ( t + α + 1 X ) e − t X ] = − ∫ 1 ∞ d d x { α e − t x x − α − 1 } d x = α e − t . \mathbb{E}\left[\left(t + \frac{\alpha+ 1}{X}\right)e^{-tX}\right] = -\int_{1}^{\infty} \frac{\mathrm{d}}{\mathrm{d}x}\left\{\alpha e^{-tx}x^{-\alpha-1}\right\}\mathrm{d}x = \alpha e^{-t}. E [ ( t + X α + 1 ) e − tX ] = − ∫ 1 ∞ d x d { α e − t x x − α − 1 } d x = α e − t .
For the converse, we identify the Laplace transform. Let φ ( t ) : = E [ e − t X ] \varphi(t) := \mathbb{E}[e^{-tX}] φ ( t ) := E [ e − tX ] and suppose (2) holds. The nonnegative terms inside the expectation in (2) are integrable. Tonelli’s theorem therefore gives
t φ ( t ) + ( α + 1 ) ∫ t ∞ φ ( s ) d s = α e − t . t\varphi(t) + (\alpha+ 1)\int_{t}^{\infty}\varphi(s)\,\mathrm{d}s = \alpha e^{-t}. tφ ( t ) + ( α + 1 ) ∫ t ∞ φ ( s ) d s = α e − t .
The Laplace transform φ \varphi φ is differentiable for t > 0 t > 0 t > 0 , since the function x ↦ x e − t x x \mapsto xe^{-tx} x ↦ x e − t x is bounded for each such t t t . Differentiation gives t φ ′ ( t ) − α φ ( t ) = − α e − t t\varphi'(t) - \alpha\varphi(t) = -\alpha e^{-t} t φ ′ ( t ) − α φ ( t ) = − α e − t . Multiplying by t − α − 1 t^{-\alpha-1} t − α − 1 and integrating to infinity, using 0 ≤ φ ( t ) ≤ 1 0 \leq\varphi(t) \leq1 0 ≤ φ ( t ) ≤ 1 , gives
φ ( t ) = α t α ∫ t ∞ u − α − 1 e − u d u = ∫ 1 ∞ e − t x α x − α − 1 d x . \varphi(t) = \alpha t^\alpha\int_{t}^{\infty}u^{-\alpha-1}e^{-u}\,\mathrm{d}u = \int_{1}^{\infty}e^{-tx}\alpha x^{-\alpha-1}\,\mathrm{d}x. φ ( t ) = α t α ∫ t ∞ u − α − 1 e − u d u = ∫ 1 ∞ e − t x α x − α − 1 d x .
The uniqueness theorem for Laplace transforms ([38 Chapter II] ) identifies the law of X X X .
The boundary term in Theorem 1 comes from the lower endpoint of the Pareto support. To see its relation to the density approach, write p α ( x ) : = α x − α − 1 1 { x ≥ 1 } p_\alpha(x) := \alpha x^{-\alpha-1}\mathbb{1}_{\{x\geq1\}} p α ( x ) := α x − α − 1 1 { x ≥ 1 } , where 1 A \mathbb{1}_A 1 A denotes the indicator of a set or event A A A . On the interior of the support, the standard density-Stein operator ([6 Section 3] ) acts on a test function g g g as
( A α g ) ( x ) : = g ′ ( x ) − α + 1 x g ( x ) . (\mathcal{A}_\alpha g)(x) := g'(x) - \frac{\alpha+1}{x}g(x). ( A α g ) ( x ) := g ′ ( x ) − x α + 1 g ( x ) .
For suitable g g g , integration by parts gives
∫ 1 ∞ { g ′ ( x ) p α ( x ) + g ( x ) p α ′ ( x ) } d x = − α g ( 1 ) . \int_{1}^{\infty}\{g'(x)p_\alpha(x) + g(x)p_\alpha'(x)\}\,\mathrm{d}x = -\alpha g(1). ∫ 1 ∞ { g ′ ( x ) p α ( x ) + g ( x ) p α ′ ( x )} d x = − α g ( 1 ) .
Choosing the characterizing function class { g t ( x ) = e − t x : t > 0 } \{g_t(x) = e^{-tx}: t > 0\} { g t ( x ) = e − t x : t > 0 } yields (2).
Two ways to aggregate the discrepancy
Let X 1 , X 2 , … , X n X_1, X_2, \ldots, X_n X 1 , X 2 , … , X n be iid with distribution function F F F on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) . We estimate the shape parameter α \alpha α by the Pareto maximum likelihood estimator (MLE) α ^ n : = n / ∑ j = 1 n log X j \widehat{\alpha}_n := n/\sum_{j=1}^{n}\log X_j α n := n / ∑ j = 1 n log X j , for details see Johnson et al. [19 ] .
Remark 1. For alternatives that may be identical to 1 almost surely, we set α ^ n = 1 \widehat{\alpha}_n=1 α n = 1 on the event ∑ j log X j = 0 \sum_j\log X_j=0 ∑ j log X j = 0 and define the statistics using this convention. If E F [ log X 1 ] > 0 \operatorname{E}_F[\log X_1]>0 E F [ log X 1 ] > 0 , then F ( { 1 } ) < 1 F(\{1\})<1 F ({ 1 }) < 1 and this event occurs only finitely often almost surely. This does not affect any limit below. Under the null hypothesis, it has probability zero.
We replace the expectation in (2) by the sample average and substitute α ^ n \widehat{\alpha}_n α n for α \alpha α , which leads to the empirical discrepancy
D n ( t ) : = α ^ n e − t − 1 n ∑ j = 1 n ( t + α ^ n + 1 X j ) e − t X j , t > 0. (3) D_n(t):=\widehat{\alpha}_n e^{-t}-\frac{1}{n}\sum_{j=1}^{n}\left(t+\frac{\widehat{\alpha}_n+1}{X_j}\right)e^{-tX_j},\qquad t>0. \tag*{(3)} D n ( t ) := α n e − t − n 1 j = 1 ∑ n ( t + X j α n + 1 ) e − t X j , t > 0. (3)
Integrating gives
T n : = ∫ 0 ∞ D n ( t ) d t = α ^ n − ( α ^ n + 2 ) 1 n ∑ j = 1 n X j − 2 . (4) T_n:=\int_{0}^{\infty}D_n(t)\,\mathrm{d}t=\widehat{\alpha}_n-(\widehat{\alpha}_n+2)\frac{1}{n}\sum_{j=1}^{n}X_j^{-2}. \tag*{(4)} T n := ∫ 0 ∞ D n ( t ) d t = α n − ( α n + 2 ) n 1 j = 1 ∑ n X j − 2 . (4)
This statistic reduces the transform identity to a single moment restriction. To retain discrepancies across all values of t t t , we square before integrating and obtain
S n : = ∫ 0 ∞ [ D n ( t ) ] 2 d t = α ^ n 2 2 − 2 α ^ n n ∑ j = 1 n ( 1 ( 1 + X j ) 2 + α ^ n + 1 X j ( 1 + X j ) ) + 1 n 2 ∑ j , k = 1 n ( 2 ( X j + X k ) 3 + ( α ^ n + 1 ) ( α ^ n + 2 ) X j X k ( X j + X k ) ) . (5) \begin{aligned}
S_n&:=\int_{0}^{\infty}[D_n(t)]^2\,\mathrm{d}t\\
&=\frac{\widehat{\alpha}_n^2}{2}-\frac{2\widehat{\alpha}_n}{n}\sum_{j=1}^{n}\left(\frac{1}{(1+X_j)^2}+\frac{\widehat{\alpha}_n+1}{X_j(1+X_j)}\right)\\
&\quad+\frac{1}{n^2}\sum_{j,k=1}^{n}\left(\frac{2}{(X_j+X_k)^3}+\frac{(\widehat{\alpha}_n+1)(\widehat{\alpha}_n+2)}{X_jX_k(X_j+X_k)}\right). \tag*{(5)}
\end{aligned} S n := ∫ 0 ∞ [ D n ( t ) ] 2 d t = 2 α n 2 − n 2 α n j = 1 ∑ n ( ( 1 + X j ) 2 1 + X j ( 1 + X j ) α n + 1 ) + n 2 1 j , k = 1 ∑ n ( ( X j + X k ) 3 2 + X j X k ( X j + X k ) ( α n + 1 ) ( α n + 2 ) ) . (5)
Both expressions can be evaluated efficiently without numerical integration. The statistic T n T_n T n can depart from zero in either direction, whereas S n S_n S n is nonnegative and large values indicate a departure from the characterization.
Remark 2. To compare our construction with that of Bhati and Khandelwal [8 ] , note that their empirical discrepancy can be written as
R n ( t ) : = 1 n ∑ j = 1 n [ ( α ^ n + 1 ) e − t − e − t X j t X j − e − t X j ] , t > 0. R_n(t):=\frac{1}{n}\sum_{j=1}^{n}\left[(\widehat{\alpha}_n+1)\frac{e^{-t}-e^{-tX_j}}{tX_j}-e^{-tX_j}\right],\qquad t>0. R n ( t ) := n 1 j = 1 ∑ n [ ( α n + 1 ) t X j e − t − e − t X j − e − t X j ] , t > 0.
The two discrepancies satisfy the exact identity D n ( t ) = t R n ( t ) + r n e − t D_n(t)=tR_n(t)+r_ne^{-t} D n ( t ) = t R n ( t ) + r n e − t , with r n : = α ^ n − α ^ n + 1 n ∑ j = 1 n X j − 1 r_n:=\widehat{\alpha}_n-\frac{\widehat{\alpha}_n+1}{n}\sum_{j=1}^{n}X_j^{-1} r n := α n − n α n + 1 ∑ j = 1 n X j − 1 . The maximum likelihood estimator does not make r n r_n r n identically zero, and the factor t t t changes the weighting across transform arguments. In particular, their signed integral is ∫ 0 ∞ R n ( t ) d t = 1 n ∑ j = 1 n ( ( α ^ n + 1 ) log X j − 1 ) / X j \int_{0}^{\infty}R_n(t)\,\mathrm{d}t=\frac{1}{n}\sum_{j=1}^{n}((\widehat{\alpha}_n+1)\log X_j-1)/X_j ∫ 0 ∞ R n ( t ) d t = n 1 ∑ j = 1 n (( α n + 1 ) log X j − 1 ) / X j , whereas ours is given by (4) . Likewise, their integrated squared discrepancy is ∫ 0 ∞ R n ( t ) 2 d t \int_{0}^{\infty}R_n(t)^2\,\mathrm{d}t ∫ 0 ∞ R n ( t ) 2 d t , rather than (5) .
Asymptotics
For the limit null results, assume that X 1 , … , X n X_1,\ldots,X_n X 1 , … , X n are iid from P ( α 0 ) \mathcal{P}(\alpha_0) P ( α 0 ) , where α 0 > 0 \alpha_0 > 0 α 0 > 0 . The fixed-alternative results state their assumptions separately. Here, a subscript F F F on an expectation, variance, covariance, or probability specifies the sampling distribution. Under the null, log X 1 \log X_1 log X 1 has an exponential distribution with rate α 0 \alpha_0 α 0 . The standard delta method for the reciprocal sample mean ([37 Chapter 3] ) gives the asymptotic linear Bahadur representation
n ( α ^ n − α 0 ) = − α 0 2 n ∑ i = 1 n ( log X i − 1 α 0 ) + o p ( 1 ) . (6) \sqrt{n}(\hat{\alpha}_n-\alpha_0)=-\frac{\alpha_0^2}{\sqrt{n}}\sum_{i=1}^{n}\left(\log X_i-\frac{1}{\alpha_0}\right)+o_p(1). \tag*{(6)} n ( α ^ n − α 0 ) = − n α 0 2 i = 1 ∑ n ( log X i − α 0 1 ) + o p ( 1 ) . (6)
No finite positive moment of X 1 X_1 X 1 is required for this expansion.
Null asymptotics of T n T_n T n
The statistic T n T_n T n depends on the sample only through the means of log X j \log X_j log X j and X j − 2 X_j^{-2} X j − 2 . This reduces its asymptotic analysis to a bivariate central limit theorem and the delta method.
Theorem 2. Under the stated assumptions we have for n → ∞ n \to\infty n → ∞
n T n → D N ( 0 , σ 2 ) , \sqrt{n}T_n \xrightarrow{\mathcal{D}} \mathrm{N}(0,\sigma^2), n T n D N ( 0 , σ 2 ) , where σ 2 : = 16 α 0 / { ( α 0 + 4 ) ( α 0 + 2 ) 2 } \sigma^2 := 16\alpha_0/\{(\alpha_0+4)(\alpha_0+2)^2\} σ 2 := 16 α 0 / {( α 0 + 4 ) ( α 0 + 2 ) 2 } .
Proof. Under the null hypothesis,
E [ log X 1 ] = 1 α 0 , E [ X 1 − 2 ] = α 0 α 0 + 2 , \mathbb{E}[\log X_1]=\frac{1}{\alpha_0},\qquad\mathbb{E}[X_1^{-2}]=\frac{\alpha_0}{\alpha_0+2}, E [ log X 1 ] = α 0 1 , E [ X 1 − 2 ] = α 0 + 2 α 0 , and
V [ log X 1 ] = 1 α 0 2 , V [ X 1 − 2 ] = 4 α 0 ( α 0 + 4 ) ( α 0 + 2 ) 2 , \mathop{\mathrm{V}}[\log X_1]=\frac{1}{\alpha_0^2},\qquad\mathop{\mathrm{V}}[X_1^{-2}]=\frac{4\alpha_0}{(\alpha_0+4)(\alpha_0+2)^2}, V [ log X 1 ] = α 0 2 1 , V [ X 1 − 2 ] = ( α 0 + 4 ) ( α 0 + 2 ) 2 4 α 0 , as well as C o v [ log X 1 , X 1 − 2 ] = − 2 / ( α 0 + 2 ) 2 \mathop{\mathrm{Cov}}[\log X_1,X_1^{-2}]=-2/(\alpha_0+2)^2 Cov [ log X 1 , X 1 − 2 ] = − 2/ ( α 0 + 2 ) 2 . The function h h h defined by h ( x , y ) : = ( 1 − y ) / x − 2 y h(x,y):=(1-y)/x-2y h ( x , y ) := ( 1 − y ) / x − 2 y maps the two sample means to T n T_n T n . The multivariate delta method ([37 Chapter 3] ) therefore gives the centered influence variable
ξ : = − 2 α 0 2 α 0 + 2 ( log X 1 − 1 α 0 ) − ( α 0 + 2 ) ( X 1 − 2 − α 0 α 0 + 2 ) . \xi:=-\frac{2\alpha_0^2}{\alpha_0+2}\left(\log X_1-\frac{1}{\alpha_0}\right)-(\alpha_0+2)\left(X_1^{-2}-\frac{\alpha_0}{\alpha_0+2}\right). ξ := − α 0 + 2 2 α 0 2 ( log X 1 − α 0 1 ) − ( α 0 + 2 ) ( X 1 − 2 − α 0 + 2 α 0 ) . Substitution of these moments gives V [ ξ ] = 16 α 0 / { ( α 0 + 4 ) ( α 0 + 2 ) 2 } \mathop{\mathrm{V}}[\xi]=16\alpha_0/\{(\alpha_0+4)(\alpha_0+2)^2\} V [ ξ ] = 16 α 0 / {( α 0 + 4 ) ( α 0 + 2 ) 2 } and hence the asserted limit. ■
The limiting variance depends on the unknown shape parameter. Substituting its consistent estimate gives a parameter-free normal limit and hence an asymptotic calibration for T n T_n T n .
Corollary 1. Under the assumptions of the preceding theorem, let α ^ n = n / ∑ j = 1 n log X j \widehat{\alpha}_n = n/\sum_{j=1}^{n}\log X_j α n = n / ∑ j = 1 n log X j be the maximum likelihood estimator used in T n T_n T n . Then
α ^ n + 2 4 n α ^ n + 4 α ^ n T n ⟶ D N ( 0 , 1 ) , \frac{\widehat{\alpha}_n+2}{4}\sqrt{n\frac{\widehat{\alpha}_n+4}{\widehat{\alpha}_n}T_n}\mathrel{\overset{\mathcal{D}}{\longrightarrow}}N(0,1), 4 α n + 2 n α n α n + 4 T n ⟶ D N ( 0 , 1 ) ,
as n → ∞ n \to\infty n → ∞ .
The corollary follows from Slutsky’s theorem. A different positive consistent estimator may be used in the studentizing factor, provided that T n T_n T n itself retains the MLE. The simulation study also includes a bootstrap procedure for comparability with the other statistics.
Fixed alternatives and limits of consistency
The same two sample means determine the behavior of T n T_n T n under a fixed alternative. Their population counterparts yield a limit Δ F \Delta_F Δ F , where a nonzero value leads to consistency, whereas cancellation can make the statistic insensitive to a non-Pareto distribution. The following theorem gives both the almost-sure limit and the fluctuations around it.
Theorem 3. Let X 1 , X 2 , … X_1,X_2,\ldots X 1 , X 2 , … be iid from a distribution F F F on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) such that 0 < μ L : = E F [ log X 1 ] < ∞ 0<\mu_L:=\mathbb{E}_F[\log X_1]<\infty 0 < μ L := E F [ log X 1 ] < ∞ , E F [ ( log X 1 ) 2 ] < ∞ \mathbb{E}_F[(\log X_1)^2]<\infty E F [( log X 1 ) 2 ] < ∞ . Define μ Y : = E F [ X 1 − 2 ] \mu_Y:=\mathbb{E}_F[X_1^{-2}] μ Y := E F [ X 1 − 2 ] , and Δ F : = μ L − 1 − ( μ L − 1 + 2 ) μ Y \Delta_F:=\mu_L^{-1}-(\mu_L^{-1}+2)\mu_Y Δ F := μ L − 1 − ( μ L − 1 + 2 ) μ Y . Then
α ^ n ⟶ a . s . μ L − 1 and T n ⟶ a . s . Δ F . \widehat{\alpha}_n \mathrel{\overset{\mathrm{a.s.}}{\longrightarrow}}\mu_L^{-1}
\qquad\text{and}\qquad
T_n \mathrel{\overset{\mathrm{a.s.}}{\longrightarrow}}\Delta_F. α n ⟶ a.s. μ L − 1 and T n ⟶ a.s. Δ F . Moreover,
n ( T n − Δ F ) ⟶ D N ( 0 , τ F 2 ) , \sqrt{n}(T_n-\Delta_F)\mathrel{\overset{\mathcal{D}}{\longrightarrow}}N(0,\tau_F^2), n ( T n − Δ F ) ⟶ D N ( 0 , τ F 2 ) , where
τ F 2 = ( 1 − μ Y ) 2 μ L − 4 V F [ log X 1 ] + ( μ L − 1 + 2 ) 2 V F [ X 1 − 2 ] + 2 ( 1 − μ Y ) μ L − 2 ( μ L − 1 + 2 ) C o v F [ log X 1 , X 1 − 2 ] . \tau_F^2=(1-\mu_Y)^2\mu_L^{-4}\mathop{\mathrm{V}}_F[\log X_1]+(\mu_L^{-1}+2)^2\mathop{\mathrm{V}}_F[X_1^{-2}]+2(1-\mu_Y)\mu_L^{-2}(\mu_L^{-1}+2)\mathop{\mathrm{Cov}}_F[\log X_1,X_1^{-2}]. τ F 2 = ( 1 − μ Y ) 2 μ L − 4 V F [ log X 1 ] + ( μ L − 1 + 2 ) 2 V F [ X 1 − 2 ] + 2 ( 1 − μ Y ) μ L − 2 ( μ L − 1 + 2 ) Cov F [ log X 1 , X 1 − 2 ] . Proof. Let L ‾ n : = n − 1 ∑ j log X j \overline{L}_n:=n^{-1}\sum_j\log X_j L n := n − 1 ∑ j log X j and Y ‾ n : = n − 1 ∑ j X j − 2 \overline{Y}_n:=n^{-1}\sum_jX_j^{-2} Y n := n − 1 ∑ j X j − 2 . Except on the event ∑ j log X j = 0 \sum_j\log X_j=0 ∑ j log X j = 0 , which occurs only finitely often almost surely under the stated assumptions, α ^ n = 1 / L ‾ n \widehat{\alpha}_n=1/\overline{L}_n α n = 1/ L n and T n = h ( L ‾ n , Y ‾ n ) T_n=h(\overline{L}_n,\overline{Y}_n) T n = h ( L n , Y n ) . The strong law and the positivity of μ L \mu_L μ L give both almost-sure limits. For the normal limit, boundedness of X 1 − 2 X_1^{-2} X 1 − 2 and the finite second moment of log X 1 \log X_1 log X 1 allow us to apply the multivariate central limit theorem to ( log X 1 , X 1 − 2 ) ⊤ (\log X_1,X_1^{-2})^\top ( log X 1 , X 1 − 2 ) ⊤ . The delta method [37 Chapter 3] applied at ( μ L , μ Y ) ⊤ (\mu_L,\mu_Y)^\top ( μ L , μ Y ) ⊤ provides τ F 2 : = ∇ h ( μ L , μ Y ) ⊤ Σ F ∇ h ( μ L , μ Y ) \tau_F^2:=\nabla h(\mu_L,\mu_Y)^\top\Sigma_F\nabla h(\mu_L,\mu_Y) τ F 2 := ∇ h ( μ L , μ Y ) ⊤ Σ F ∇ h ( μ L , μ Y ) with h ( ℓ , y ) = ( 1 − y ) / ℓ − 2 y h(\ell,y)=(1-y)/\ell-2y h ( ℓ , y ) = ( 1 − y ) / ℓ − 2 y and covariance matrix
Σ F : = ( V F [ log X 1 ] C o v F [ log X 1 , X 1 − 2 ] C o v F [ log X 1 , X 1 − 2 ] V F [ X 1 − 2 ] ) . \Sigma_F:=\begin{pmatrix}
\mathop{\mathrm{V}}_F[\log X_1] & \mathop{\mathrm{Cov}}_F[\log X_1,X_1^{-2}]\\
\mathop{\mathrm{Cov}}_F[\log X_1,X_1^{-2}] & \mathop{\mathrm{V}}_F[X_1^{-2}]
\end{pmatrix}. Σ F := ( V F [ log X 1 ] Cov F [ log X 1 , X 1 − 2 ] Cov F [ log X 1 , X 1 − 2 ] V F [ X 1 − 2 ] ) . The asserted limit follows by direct calculation using the gradient ∇ h ( ℓ , y ) = ( − ( 1 − y ) / ℓ 2 , − 1 / ℓ − 2 ) ⊤ \nabla h(\ell,y)=(-(1-y)/\ell^2,-1/\ell-2)^\top ∇ h ( ℓ , y ) = ( − ( 1 − y ) / ℓ 2 , − 1/ ℓ − 2 ) ⊤ . Substitution yields the displayed variance formula, with a degenerate normal law understood when τ F 2 = 0 \tau_F^2=0 τ F 2 = 0 . ■
Let Φ \Phi Φ denote the standard normal distribution function and let z p : = Φ − 1 ( p ) z_p:=\Phi^{-1}(p) z p := Φ − 1 ( p ) be its p p p -quantile for 0 < p < 1 0<p<1 0 < p < 1 . Fix a significance level 0 < β < 1 0<\beta<1 0 < β < 1 . The next corollary identifies the alternatives detected consistently by the two-sided normal test.
Corollary 2. Assume the conditions of Theorem 3 . If Δ F ≠ 0 \Delta_F \ne0 Δ F = 0 , then ∣ n T n ∣ → P ∞ |\sqrt{n}T_n| \xrightarrow{\mathbb{P}} \infty ∣ n T n ∣ P ∞ . Define the studentized statistic
W n : = α ^ n + 2 4 n ( α ^ n + 4 ) α ^ n T n . W_n := \frac{\widehat{\alpha}_n+2}{4}\sqrt{\frac{n(\widehat{\alpha}_n+4)}{\widehat{\alpha}_n}}T_n. W n := 4 α n + 2 α n n ( α n + 4 ) T n . The asymptotic normal test based on W n W_n W n is therefore consistent against every fixed alternative F F F such that Δ F ≠ 0 \Delta_F \ne0 Δ F = 0 , that is,
P F [ ∣ W n ∣ > z ~ 1 − β / 2 ] → 1 as n → ∞ . \mathbb{P}_F\left[|W_n|>\widetilde{z}_{1-\beta/2}\right]\to1 \quad\text{as } n\to\infty. P F [ ∣ W n ∣ > z 1 − β /2 ] → 1 as n → ∞.
Proof. By Theorem 3 , T n → Δ F ≠ 0 T_n \to\Delta_F \ne0 T n → Δ F = 0 in probability. The factor multiplying n T n \sqrt{n}T_n n T n in W n W_n W n converges to the finite positive number ( μ L − 1 + 2 ) 1 + 4 μ L / 4 (\mu_L^{-1}+2)\sqrt{1+4\mu_L/4} ( μ L − 1 + 2 ) 1 + 4 μ L /4 . Thus ∣ n T n ∣ → ∞ |\sqrt{n}T_n|\to\infty ∣ n T n ∣ → ∞ and ∣ W n ∣ → ∞ |W_n|\to\infty ∣ W n ∣ → ∞ in probability, which proves the claim. ■
The condition Δ F ≠ 0 \Delta_F \ne0 Δ F = 0 cannot be omitted. The following proposition constructs continuous non-Pareto alternatives with Δ F = 0 \Delta_F=0 Δ F = 0 against which the two-sided normal test is not consistent. To establish this failure, we compute the limiting rejection probability, since the almost-sure convergence of T n T_n T n to zero alone does not determine the test’s power.
Proposition 1. Let λ > 0 \lambda>0 λ > 0 and b > 0 b>0 b > 0 , and define
q λ , b : = 2 λ b − ( λ + 2 ) ( 1 − e − 2 b ) 2 λ b ( 1 − e − 2 b ) . q_{\lambda,b}:=\frac{2\lambda b-(\lambda+2)(1-e^{-2b})}{2\lambda b(1-e^{-2b})}. q λ , b := 2 λb ( 1 − e − 2 b ) 2 λb − ( λ + 2 ) ( 1 − e − 2 b ) . Assume that q λ , b ∈ ( 0 , 1 ) q_{\lambda,b}\in(0,1) q λ , b ∈ ( 0 , 1 ) . Let F λ , b F_{\lambda,b} F λ , b be the distribution on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) with density
f λ , b ( x ) : = λ x − ( λ + 1 ) [ ( 1 − q λ , b ) + q λ , b e λ b 1 { x ≥ e b } ] , x ≥ 1. f_{\lambda,b}(x):=\lambda x^{-(\lambda+1)}\left[(1-q_{\lambda,b})+q_{\lambda,b}e^{\lambda b}1_{\{x\ge e^b\}}\right], \qquad x\ge1. f λ , b ( x ) := λ x − ( λ + 1 ) [ ( 1 − q λ , b ) + q λ , b e λb 1 { x ≥ e b } ] , x ≥ 1. Then F λ , b F_{\lambda,b} F λ , b is absolutely continuous, has support [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) , is not a Pareto distribution, and satisfies
Δ F λ , b = 0. \Delta_{F_{\lambda,b}}=0. Δ F λ , b = 0. Consequently,
T n → a . s . 0 under F λ , b , T_n \xrightarrow{\mathrm{a.s.}}0 \quad\text{under }F_{\lambda,b}, T n a.s. 0 under F λ , b , so the test based on T n T_n T n is not consistent against F λ , b F_{\lambda,b} F λ , b .
Proof. Let E ∼ Exp ( λ ) E\sim\operatorname{Exp}(\lambda) E ∼ Exp ( λ ) and B ∼ Bernoulli ( q λ , b ) B\sim\operatorname{Bernoulli}(q_{\lambda,b}) B ∼ Bernoulli ( q λ , b ) be independent. The random variable X : = e E + b B X:=e^{E+bB} X := e E + b B has the displayed density. The multiplicative factor in this density changes at e b e^b e b , so it is not a Pareto density. Writing q : = q λ , b q:=q_{\lambda,b} q := q λ , b , the moments satisfy
μ L = 1 λ + q b , μ Y = λ λ + 2 { 1 − q ( 1 − e − 2 b ) } = 1 1 + 2 μ L , \mu_L=\frac{1}{\lambda}+qb,\qquad\mu_Y=\frac{\lambda}{\lambda+2}\{1-q(1-e^{-2b})\}=\frac{1}{1+2\mu_L}, μ L = λ 1 + q b , μ Y = λ + 2 λ { 1 − q ( 1 − e − 2 b )} = 1 + 2 μ L 1 ,
where the last identity follows by substitution of q q q . Thus Δ F = 0 \Delta_F=0 Δ F = 0 . Write σ 2 ( a ) : = 16 a / { ( a + 4 ) ( a + 2 ) 2 } \sigma^2(a):=16a/\{(a+4)(a+2)^2\} σ 2 ( a ) := 16 a / {( a + 4 ) ( a + 2 ) 2 } for the null variance at shape a a a . Since log X = E + b B \log X=E+bB log X = E + b B has a finite second moment, Theorem 3 gives T n → 0 T_n\to0 T n → 0 almost surely and
W n → D N ( 0 , τ F 2 σ 2 ( μ L − 1 ) ) . W_n \xrightarrow{\mathcal{D}} N\left(0,\frac{\tau_F^2}{\sigma^2(\mu_L^{-1})}\right). W n D N ( 0 , σ 2 ( μ L − 1 ) τ F 2 ) .
To see that 0 < τ F 2 < ∞ 0 < \tau_F^2 < \infty 0 < τ F 2 < ∞ , consider the influence function ψ F \psi_F ψ F , defined by
ψ F ( y ) : = − ( 1 − μ Y ) μ L − 2 ( y − μ L ) − ( μ L − 1 + 2 ) ( e − 2 y − μ Y ) . \psi_F(y) := -(1-\mu_Y)\mu_L^{-2}(y-\mu_L)-(\mu_L^{-1}+2)(e^{-2y}-\mu_Y). ψ F ( y ) := − ( 1 − μ Y ) μ L − 2 ( y − μ L ) − ( μ L − 1 + 2 ) ( e − 2 y − μ Y ) .
This function is nonconstant on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , where E + b B E+bB E + b B has a positive density, and has a finite second moment. Consequently,
P F [ ∣ W n ∣ > z 1 − β / 2 ] ⟶ 2 [ 1 − Φ ( z 1 − β / 2 σ ( μ L − 1 ) τ F ) ] < 1. \mathbb{P}_F[|W_n|>z_{1-\beta/2}] \longrightarrow2\left[1-\Phi\left(\frac{z_{1-\beta/2}\sigma(\mu_L^{-1})}{\tau_F}\right)\right]<1. P F [ ∣ W n ∣ > z 1 − β /2 ] ⟶ 2 [ 1 − Φ ( τ F z 1 − β /2 σ ( μ L − 1 ) ) ] < 1.
The limiting rejection probability is strictly less than one, so the test is not consistent against this alternative. ■
Null limit of S n S_n S n
The null limit of n S n nS_n n S n follows from a central limit theorem for the process inside its squared integral. To express the statistic as a squared norm, we work in the separable Hilbert space H : = L 2 ( [ 0 , ∞ ) , B , d t ) \mathbb{H}:=L^2([0,\infty),\mathcal{B},dt) H := L 2 ([ 0 , ∞ ) , B , d t ) of equivalence classes of measurable functions f : [ 0 , ∞ ) → R f:[0,\infty)\to\mathbb{R} f : [ 0 , ∞ ) → R satisfying ∫ 0 ∞ ∣ f ( t ) ∣ 2 d t < ∞ \int_0^\infty|f(t)|^2dt<\infty ∫ 0 ∞ ∣ f ( t ) ∣ 2 d t < ∞ , where B \mathcal{B} B denotes the Borel sigma-field on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . Define its scalar product and norm by
⟨ f , g ⟩ H : = ∫ 0 ∞ f ( t ) g ( t ) d t , ∥ f ∥ H : = ⟨ f , f ⟩ H 1 / 2 , f , g ∈ H , \langle f,g\rangle_{\mathbb{H}}:=\int_0^\infty f(t)g(t)\,dt,\qquad\|f\|_{\mathbb{H}}:=\langle f,f\rangle_{\mathbb{H}}^{1/2},\qquad f,g\in\mathbb{H}, ⟨ f , g ⟩ H := ∫ 0 ∞ f ( t ) g ( t ) d t , ∥ f ∥ H := ⟨ f , f ⟩ H 1/2 , f , g ∈ H ,
respectively. The corresponding empirical process is
Z n ( t ) : = n α ^ n exp ( − t ) − 1 n ∑ j = 1 n ( t + α ^ n + 1 X j ) exp ( − t X j ) , t > 0 , Z_n(t):=\sqrt{n}\widehat{\alpha}_n\exp(-t)-\frac{1}{\sqrt{n}}\sum_{j=1}^n\left(t+\frac{\widehat{\alpha}_n+1}{X_j}\right)\exp(-tX_j),\qquad t>0, Z n ( t ) := n α n exp ( − t ) − n 1 j = 1 ∑ n ( t + X j α n + 1 ) exp ( − t X j ) , t > 0 ,
so that n S n = ∫ 0 ∞ Z n 2 ( t ) d t nS_n=\int_0^\infty Z_n^2(t)\,dt n S n = ∫ 0 ∞ Z n 2 ( t ) d t . Separating the parameter-estimation term gives
Z n ( t ) = n α 0 exp ( − t ) − 1 n ∑ j = 1 n ( t + α 0 + 1 X j ) exp ( − t X j ) + n ( α ^ n − α 0 ) ( exp ( − t ) − 1 n ∑ j = 1 n exp ( − t X j ) X j ) , t > 0. \begin{aligned}
Z_n(t)={}&\sqrt{n}\alpha_0\exp(-t)-\frac{1}{\sqrt{n}}\sum_{j=1}^n\left(t+\frac{\alpha_0+1}{X_j}\right)\exp(-tX_j)\\
&+\sqrt{n}(\widehat{\alpha}_n-\alpha_0)\left(\exp(-t)-\frac{1}{n}\sum_{j=1}^n\frac{\exp(-tX_j)}{X_j}\right),\qquad t>0.
\end{aligned} Z n ( t ) = n α 0 exp ( − t ) − n 1 j = 1 ∑ n ( t + X j α 0 + 1 ) exp ( − t X j ) + n ( α n − α 0 ) ( exp ( − t ) − n 1 j = 1 ∑ n X j exp ( − t X j ) ) , t > 0.
The Bahadur representation (6) leads to the approximation process
Z ^ n ( t ) : = n α 0 exp ( − t ) − 1 n ∑ j = 1 n ( t + α 0 + 1 X j ) exp ( − t X j ) − α 0 2 n ∑ i = 1 n ( log X i − 1 α 0 ) ( exp ( − t ) − Ψ α 0 ( t ) ) , t > 0 , \begin{aligned}
\widehat{Z}_n(t):={}&\sqrt{n}\alpha_0\exp(-t)-\frac{1}{\sqrt{n}}\sum_{j=1}^n\left(t+\frac{\alpha_0+1}{X_j}\right)\exp(-tX_j)\\
&-\frac{\alpha_0^2}{\sqrt{n}}\sum_{i=1}^n\left(\log X_i-\frac{1}{\alpha_0}\right)\left(\exp(-t)-\Psi_{\alpha_0}(t)\right),\qquad t>0,
\end{aligned} Z n ( t ) := n α 0 exp ( − t ) − n 1 j = 1 ∑ n ( t + X j α 0 + 1 ) exp ( − t X j ) − n α 0 2 i = 1 ∑ n ( log X i − α 0 1 ) ( exp ( − t ) − Ψ α 0 ( t ) ) , t > 0 ,
where Ψ α 0 ( t ) : = α 0 t α 0 + 1 Γ ( − α 0 − 1 , t ) \Psi_{\alpha_0}(t):=\alpha_0t^{\alpha_0+1}\Gamma(-\alpha_0-1,t) Ψ α 0 ( t ) := α 0 t α 0 + 1 Γ ( − α 0 − 1 , t ) and the upper incomplete gamma function is defined by Γ ( s , t ) : = ∫ t ∞ u s − 1 exp ( − u ) d u \Gamma(s,t):=\int_t^\infty u^{s-1}\exp(-u)\,du Γ ( s , t ) := ∫ t ∞ u s − 1 exp ( − u ) d u . The next theorem identifies the Gaussian limit.
Theorem 4. Under the standing assumptions, there exists a centered Gaussian random element Z Z Z in H \mathbb{H} H with covariance kernel
K α 0 ( s , t ) : = α 0 s t ( s + t ) α 0 Γ ( − α 0 , s + t ) + α 0 ( α 0 + 1 ) ( s + t ) α 0 + 2 Γ ( − α 0 − 1 , s + t ) + α 0 ( α 0 + 1 ) 2 ( s + t ) α 0 + 2 Γ ( − α 0 − 2 , s + t ) − α 0 2 e − s − t − α 0 2 C α 0 ( s ) C α 0 ( t ) , K_{\alpha_0}(s,t) := \alpha_0st(s+t)^{\alpha_0}\Gamma(-\alpha_0,s+t) + \alpha_0(\alpha_0+1)(s+t)^{\alpha_0+2}\Gamma(-\alpha_0-1,s+t) + \alpha_0(\alpha_0+1)^2(s+t)^{\alpha_0+2}\Gamma(-\alpha_0-2,s+t) - \alpha_0^2e^{-s-t} - \alpha_0^2C_{\alpha_0}(s)C_{\alpha_0}(t), K α 0 ( s , t ) := α 0 s t ( s + t ) α 0 Γ ( − α 0 , s + t ) + α 0 ( α 0 + 1 ) ( s + t ) α 0 + 2 Γ ( − α 0 − 1 , s + t ) + α 0 ( α 0 + 1 ) 2 ( s + t ) α 0 + 2 Γ ( − α 0 − 2 , s + t ) − α 0 2 e − s − t − α 0 2 C α 0 ( s ) C α 0 ( t ) , where C α 0 ( t ) : = e − t − Ψ α 0 ( t ) C_{\alpha_0}(t):=e^{-t}-\Psi_{\alpha_0}(t) C α 0 ( t ) := e − t − Ψ α 0 ( t ) , s , t > 0 s,t>0 s , t > 0 , such that
n S n → D ∥ Z ∥ H 2 , as n → ∞ . nS_n \xrightarrow{\mathcal{D}} \lVert Z\rVert_{\mathbb{H}}^2,\qquad\text{as } n\to\infty. n S n D ∥ Z ∥ H 2 , as n → ∞.
Proof. We first show that the approximating process is a sum of centered, square-integrable random elements and that its difference from Z n Z_n Z n is negligible. We then calculate the covariance kernel. Throughout the proof let E α 0 \mathbb{E}_{\alpha_0} E α 0 denote expectation under P ( α 0 ) \mathbb{P}(\alpha_0) P ( α 0 ) , and define Y x ( t ) : = e − t x / x Y_x(t):=e^{-tx}/x Y x ( t ) := e − t x / x , and
h α 0 ( x , t ) : = α 0 e − t − ( t + α 0 + 1 x ) e − t x − α 0 2 ( log x − 1 α 0 ) C α 0 ( t ) . h_{\alpha_0}(x,t):=\alpha_0e^{-t}-\left(t+\frac{\alpha_0+1}{x}\right)e^{-tx}-\alpha_0^2\left(\log x-\frac{1}{\alpha_0}\right)C_{\alpha_0}(t). h α 0 ( x , t ) := α 0 e − t − ( t + x α 0 + 1 ) e − t x − α 0 2 ( log x − α 0 1 ) C α 0 ( t ) .
The characterization gives E α 0 [ h α 0 ( X , t ) ] = 0 \mathbb{E}_{\alpha_0}[h_{\alpha_0}(X,t)]=0 E α 0 [ h α 0 ( X , t )] = 0 . Since x ≥ 1 x\geq1 x ≥ 1 , ∥ Y x ∥ H 2 = 1 / ( 2 x 3 ) ≤ 1 / 2 \lVert Y_x\rVert_{\mathbb{H}}^2=1/(2x^3)\leq1/2 ∥ Y x ∥ H 2 = 1/ ( 2 x 3 ) ≤ 1/2 , 0 ≤ C α 0 ( t ) ≤ e − t 0\leq C_{\alpha_0}(t)\leq e^{-t} 0 ≤ C α 0 ( t ) ≤ e − t , and
∣ h α 0 ( x , t ) ∣ ≤ c α 0 ( 1 + log x ) ( 1 + t ) e − t , \lvert h_{\alpha_0}(x,t)\rvert\leq c_{\alpha_0}(1+\log x)(1+t)e^{-t}, ∣ h α 0 ( x , t )∣ ≤ c α 0 ( 1 + log x ) ( 1 + t ) e − t ,
where c α 0 c_{\alpha_0} c α 0 is a finite constant depending only on α 0 \alpha_0 α 0 . Thus h α 0 ( X , ⋅ ) h_{\alpha_0}(X,\cdot) h α 0 ( X , ⋅ ) is square-integrable in H \mathbb{H} H , because log X ∼ Exp ( α 0 ) \log X\sim\operatorname{Exp}(\alpha_0) log X ∼ Exp ( α 0 ) . The Hilbert-space law of large numbers and central limit theorem [17 Chapter 17] therefore apply. In particular, with H j : = h α 0 ( X j , ⋅ ) H_j:=h_{\alpha_0}(X_j,\cdot) H j := h α 0 ( X j , ⋅ ) and Y ‾ n : = n − 1 ∑ j = 1 n Y X j \overline{Y}_n:=n^{-1}\sum_{j=1}^nY_{X_j} Y n := n − 1 ∑ j = 1 n Y X j , we have ∥ Y ‾ n − E α 0 [ Y X ] ∥ H → 0 \lVert\overline{Y}_n-\mathbb{E}_{\alpha_0}[Y_X]\rVert_{\mathbb{H}}\to0 ∥ Y n − E α 0 [ Y X ] ∥ H → 0 almost surely and n − 1 / 2 ∑ j = 1 n H j → D Z n^{-1/2}\sum_{j=1}^nH_j\xrightarrow{\mathcal{D}}Z n − 1/2 ∑ j = 1 n H j D Z in H \mathbb{H} H . If R n = o P ( 1 ) R_n=o_{\mathbb{P}}(1) R n = o P ( 1 ) is the remainder in (6), then
Z n − 1 n ∑ j = 1 n H j = R n C α 0 − n ( α ^ n − α 0 ) ( Y ‾ n − E α 0 [ Y X ] ) = o P ( 1 ) in H . Z_n-\frac{1}{\sqrt{n}}\sum_{j=1}^nH_j=R_nC_{\alpha_0}-\sqrt{n}(\widehat{\alpha}_n-\alpha_0)(\overline{Y}_n-\mathbb{E}_{\alpha_0}[Y_X])=o_{\mathbb{P}}(1)\quad\text{in }\mathbb{H}. Z n − n 1 j = 1 ∑ n H j = R n C α 0 − n ( α n − α 0 ) ( Y n − E α 0 [ Y X ]) = o P ( 1 ) in H .
Slutsky’s theorem and the continuous mapping theorem give the asserted limit of n S n = ∥ Z n ∥ H 2 nS_n=\lVert Z_n\rVert_{\mathbb{H}}^2 n S n = ∥ Z n ∥ H 2 . It remains to calculate the kernel. Set B t ( x ) : = { t + ( α 0 + 1 ) / x } e − t x B_t(x):=\{t+(\alpha_0+1)/x\}e^{-tx} B t ( x ) := { t + ( α 0 + 1 ) / x } e − t x , suppressing its dependence on α 0 \alpha_0 α 0 in the notation. Differentiating the identity E α 0 [ B t ( X ) ] = α 0 e − t \mathbb{E}_{\alpha_0}[B_t(X)]=\alpha_0e^{-t} E α 0 [ B t ( X )] = α 0 e − t with respect to α 0 \alpha_0 α 0 is justified locally by the exponential law of log X \log X log X and gives
E α 0 [ ( log X − 1 α 0 ) B t ( X ) ] = − C α 0 ( t ) . \mathbb{E}_{\alpha_0}\left[\left(\log X-\frac{1}{\alpha_0}\right)B_t(X)\right]=-C_{\alpha_0}(t). E α 0 [ ( log X − α 0 1 ) B t ( X ) ] = − C α 0 ( t ) .
This differentiation includes the dependence on α 0 \alpha_0 α 0 of both B t B_t B t and the Pareto law. Consequently, after some more calculations, the kernel is as stated in the theorem. ■
Consistency of S n S_n S n
Under the logarithmic moment condition below, S n S_n S n converges to the squared norm of the population discrepancy. The characterization in Theorem 1 makes this limit positive whenever the distribution is not Pareto.
Theorem 5. Let X 1 , X 2 , … X_1,X_2,\ldots X 1 , X 2 , … be iid from a distribution F F F on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) such that 0 < μ L = E F [ log X 1 ] < ∞ 0<\mu_L=\mathbb{E}_F[\log X_1]<\infty 0 < μ L = E F [ log X 1 ] < ∞ , and set α F = 1 / μ L \alpha_F=1/\mu_L α F = 1/ μ L . Define
D F ( t ) : = α F e − t − E F [ ( t + α F + 1 X 1 ) e − t X 1 ] , t > 0 , (7) D_F(t):=\alpha_F e^{-t}-\mathbb{E}_F\left[\left(t+\frac{\alpha_F+1}{X_1}\right)e^{-tX_1}\right],\qquad t>0, \tag*{(7)} D F ( t ) := α F e − t − E F [ ( t + X 1 α F + 1 ) e − t X 1 ] , t > 0 , (7) and S ( F ) : = ∫ 0 ∞ D F 2 ( t ) d t S(F):=\int_0^\infty D_F^2(t)\,\mathrm{d}t S ( F ) := ∫ 0 ∞ D F 2 ( t ) d t . Then
S n → a . s . S ( F ) . (8) S_n \xrightarrow{\mathrm{a.s.}} S(F). \tag*{(8)} S n a.s. S ( F ) . (8) Moreover, S ( F ) = 0 S(F)=0 S ( F ) = 0 if and only if F = P ( α F ) F=\mathcal{P}(\alpha_F) F = P ( α F ) . Hence, for every fixed non-Pareto alternative satisfying the stated moment condition, S ( F ) > 0 S(F)>0 S ( F ) > 0 and n S n → ∞ nS_n\to\infty n S n → ∞ almost surely. Consequently, every right-tailed test based on n S n nS_n n S n with critical values bounded in probability is consistent.
Proof. We prove convergence of the empirical discrepancy in H \mathbb{H} H and then identify the distributions for which its limit vanishes. The reciprocal-mean formula for α ^ n \widehat{\alpha}_n α n holds eventually almost surely under the convention above, so the strong law of large numbers gives α ^ n → α F \widehat{\alpha}_n\to\alpha_F α n → α F almost surely.
For fixed a > 0 a>0 a > 0 , define g a ( x , t ) : = { t + ( a + 1 ) / x } e − t x g_a(x,t):=\{t+(a+1)/x\}e^{-tx} g a ( x , t ) := { t + ( a + 1 ) / x } e − t x . Since x ≥ 1 x\ge1 x ≥ 1 , we have ∫ 0 ∞ g a 2 ( x , t ) d t ≤ c a < ∞ \int_0^\infty g_a^2(x,t)\,\mathrm{d}t\le c_a<\infty ∫ 0 ∞ g a 2 ( x , t ) d t ≤ c a < ∞ for a constant c a c_a c a depending only on a a a . Thus g α F ( X 1 , ⋅ ) g_{\alpha_F}(X_1,\cdot) g α F ( X 1 , ⋅ ) is square-integrable in H \mathbb{H} H , and the Hilbert-space strong law of large numbers ([17 Theorem 17.15] ) yields
∥ n − 1 ∑ j = 1 n g α F ( X j , ⋅ ) − E F [ g α F ( X 1 , ⋅ ) ] ∥ H → 0 almost surely . \left\|n^{-1}\sum_{j=1}^{n}g_{\alpha_F}(X_j,\cdot)-\mathbb{E}_F[g_{\alpha_F}(X_1,\cdot)]\right\|_{\mathbb{H}}\to0\quad\text{almost surely}. n − 1 j = 1 ∑ n g α F ( X j , ⋅ ) − E F [ g α F ( X 1 , ⋅ )] H → 0 almost surely .
To account for parameter estimation, use the bound ∥ g a ( x , ⋅ ) − g b ( x , ⋅ ) ∥ H ≤ ∣ a − b ∣ / 2 \|g_a(x,\cdot)-g_b(x,\cdot)\|_{\mathbb{H}}\le|a-b|/\sqrt{2} ∥ g a ( x , ⋅ ) − g b ( x , ⋅ ) ∥ H ≤ ∣ a − b ∣/ 2 , which holds uniformly in x ≥ 1 x\ge1 x ≥ 1 . Replacing α F \alpha_F α F by α ^ n \widehat{\alpha}_n α n is therefore asymptotically negligible in H \mathbb{H} H . The empirical discrepancy defining S n S_n S n consequently converges almost surely in H \mathbb{H} H to D F D_F D F , and continuity of the squared norm gives (8).
If S ( F ) = 0 S(F)=0 S ( F ) = 0 , then D F ( t ) = 0 D_F(t)=0 D F ( t ) = 0 for Lebesgue-a.e. t > 0 t>0 t > 0 . The function D F D_F D F is continuous in t t t , so D F ( t ) = 0 D_F(t)=0 D F ( t ) = 0 for every $t>0. Theorem 1, applied with α = α F \alpha=\alpha_F α = α F , then implies F = P ( α F ) F=\mathcal{P}(\alpha_F) F = P ( α F ) . The converse is immediate from Theorem 1. If F F F is not Pareto, S ( F ) > 0 S(F)>0 S ( F ) > 0 and n S n → ∞ nS_n\to\infty n S n → ∞ almost surely. ■
Covariance-operator calibration
The null limit in Theorem 4 depends on the unknown parameter α 0 > 0 \alpha_0>0 α 0 > 0 . We therefore use two different calibrations for S n S_n S n . Here, we approximate the covariance eigenvalues and substitute α ^ n \widehat{\alpha}_n α n for α 0 \alpha_0 α 0 , and Section 3.6 treats the parametric bootstrap. Let K α 0 : H → H K_{\alpha_0}:\mathbb{H}\to\mathbb{H} K α 0 : H → H denote the covariance operator defined by
( K α 0 f ) ( s ) : = ∫ 0 ∞ K α 0 ( s , t ) f ( t ) d t . (K_{\alpha_0}f)(s):=\int_0^\infty K_{\alpha_0}(s,t)f(t)\,\mathrm{d}t. ( K α 0 f ) ( s ) := ∫ 0 ∞ K α 0 ( s , t ) f ( t ) d t .
The moment bound in the proof of Theorem 4 shows that this positive covariance operator is trace class, since
tr ( K α 0 ) = E ∥ Z ∥ H 2 = E ∥ h α 0 ( X 1 , ⋅ ) ∥ H 2 < ∞ . \operatorname{tr}(K_{\alpha_0})=\mathbb{E}\|Z\|_{\mathbb{H}}^2=\mathbb{E}\|h_{\alpha_0}(X_1,\cdot)\|_{\mathbb{H}}^2<\infty. tr ( K α 0 ) = E ∥ Z ∥ H 2 = E ∥ h α 0 ( X 1 , ⋅ ) ∥ H 2 < ∞.
Let λ 1 , λ 2 , … \lambda_1,\lambda_2,\ldots λ 1 , λ 2 , … denote its nonnegative eigenvalues, ordered nonincreasingly with multiplicity and extended by zeros if necessary. The trace-class property gives ∑ j λ j < ∞ \sum_j\lambda_j<\infty ∑ j λ j < ∞ . For iid standard normal random variables N 1 , N 2 , … N_1,N_2,\ldots N 1 , N 2 , … , the standard Karhunen–Loève expansion for Gaussian random elements [13 Chapter 3] gives
∥ Z ∥ H 2 = D ∑ j = 1 ∞ λ j N j 2 . \lVert Z\rVert_{\mathbb{H}}^2 \mathrel{\overset{\mathcal{D}}{=}} \sum_{j=1}^{\infty} \lambda_j N_j^2. ∥ Z ∥ H 2 = D j = 1 ∑ ∞ λ j N j 2 .
The nonzero eigenvalues correspond to eigenfunctions f j f_j f j satisfying the integral equation
λ f ( s ) = ( K α 0 f ) ( s ) , s > 0. (9) \lambda f(s) = (\mathcal{K}_{\alpha_0}f)(s), \qquad s>0. \tag*{(9)} λ f ( s ) = ( K α 0 f ) ( s ) , s > 0. (9)
We approximate these eigenvalues using the Nyström method (for details see Nyström [30 ] and Atkinson (1997, Chapter 4)), so truncation and quadrature reduce the integral equation to a finite matrix eigenvalue problem. For a different approach to covariance-operator eigenvalues in goodness-of-fit testing, see the Rayleigh–Ritz method of Ebner et al. [14 ] .
Since the integration domain is unbounded, we first truncate the operator to a sufficiently large interval [ 0 , T ] [0,T] [ 0 , T ] and consider
( K α 0 , T f ) ( s ) : = ∫ 0 T K α 0 ( s , t ) f ( t ) d t , 0 ≤ s ≤ T . (\mathcal{K}_{\alpha_0,T}f)(s) := \int_0^T K_{\alpha_0}(s,t)f(t)\,\mathrm{d}t,\qquad0\le s\le T. ( K α 0 , T f ) ( s ) := ∫ 0 T K α 0 ( s , t ) f ( t ) d t , 0 ≤ s ≤ T .
Next, let t 1 , … , t m ∈ [ 0 , T ] t_1,\ldots,t_m\in[0,T] t 1 , … , t m ∈ [ 0 , T ] be quadrature nodes with positive weights w 1 , … , w m w_1,\ldots,w_m w 1 , … , w m , so that
∫ 0 T g ( t ) d t ≈ ∑ j = 1 m w j g ( t j ) . \int_0^T g(t)\,\mathrm{d}t \approx\sum_{j=1}^{m}w_jg(t_j). ∫ 0 T g ( t ) d t ≈ j = 1 ∑ m w j g ( t j ) .
Then the Nyström approximation of K α 0 , T \mathcal{K}_{\alpha_0,T} K α 0 , T is
( K α 0 , T f ) ( t i ) ≈ ∑ j = 1 m w j K α 0 ( t i , t j ) f ( t j ) , i = 1 , … , m . (\mathcal{K}_{\alpha_0,T}f)(t_i) \approx\sum_{j=1}^{m}w_jK_{\alpha_0}(t_i,t_j)f(t_j),\qquad i=1,\ldots,m. ( K α 0 , T f ) ( t i ) ≈ j = 1 ∑ m w j K α 0 ( t i , t j ) f ( t j ) , i = 1 , … , m .
Writing u j u_j u j for the discrete counterpart of f ( t j ) f(t_j) f ( t j ) , we replace the continuous eigenvalue problem in (9) by
∑ j = 1 m w j K α 0 ( t i , t j ) u j = λ u i , i = 1 , … , m . \sum_{j=1}^{m}w_jK_{\alpha_0}(t_i,t_j)u_j=\lambda u_i,\qquad i=1,\ldots,m. j = 1 ∑ m w j K α 0 ( t i , t j ) u j = λ u i , i = 1 , … , m .
Since the kernel is symmetric, we use the symmetrized system with the column vector v : = ( w 1 u 1 , … , w m u m ) T v:=(\sqrt{w_1}u_1,\ldots,\sqrt{w_m}u_m)^{\mathsf{T}} v := ( w 1 u 1 , … , w m u m ) T :
B v = λ v , B : = ( B i j ) i , j = 1 m , \mathbf{B}v=\lambda v,\qquad\mathbf{B}:=(B_{ij})_{i,j=1}^{m}, B v = λ v , B := ( B ij ) i , j = 1 m ,
where the entries are defined by B i j : = w i K α 0 ( t i , t j ) w j B_{ij}:=\sqrt{w_i}K_{\alpha_0}(t_i,t_j)\sqrt{w_j} B ij := w i K α 0 ( t i , t j ) w j , i , j = 1 , … , m i,j=1,\ldots,m i , j = 1 , … , m . The eigenvalues of B \mathbf{B} B approximate those of K α 0 \mathcal{K}_{\alpha_0} K α 0 . For the fitted model, we substitute α ^ n \widehat{\alpha}_n α n for α 0 \alpha_0 α 0 and denote the resulting eigenvalues by λ ^ j , n \widehat{\lambda}_{j,n} λ j , n after replacing negative numerical values by zero.
Given independent chi-square variables χ 1 , j 2 \chi_{1,j}^2 χ 1 , j 2 with one degree of freedom, we approximate the quantile of ∑ j λ ^ j , n χ 1 , j 2 \sum_j\widehat{\lambda}_{j,n}\chi_{1,j}^2 ∑ j λ j , n χ 1 , j 2 by Monte Carlo simulation.
The next theorem states sufficient conditions for this plug-in method to have the intended asymptotic level. It applies standard weak convergence and quantile convergence [37 Section 21.1] . The key assumption is ℓ 1 \ell^1 ℓ 1 -consistency of the entire ordered eigenvalue sequence, including control of its tail; convergence of finitely many eigenvalues alone is insufficient.
Theorem 6. Assume the null hypothesis H 0 : X 1 , … , X n ∼ i i d P ( α 0 ) H_0: X_1,\ldots,X_n \mathrel{\overset{\mathrm{iid}}{\sim}} \mathcal{P}(\alpha_0) H 0 : X 1 , … , X n ∼ iid P ( α 0 ) with α 0 > 0 \alpha_0>0 α 0 > 0 , and suppose
n S n → D Q α 0 : = ∑ j = 1 ∞ λ j ( α 0 ) χ 1 , j 2 . nS_n \xrightarrow{\mathcal{D}} Q_{\alpha_0}:=\sum_{j=1}^{\infty}\lambda_j(\alpha_0)\chi_{1,j}^2. n S n D Q α 0 := j = 1 ∑ ∞ λ j ( α 0 ) χ 1 , j 2 . where ( χ 1 , j 2 ) j ≥ 1 (\chi_{1,j}^2)_{j\geq1} ( χ 1 , j 2 ) j ≥ 1 are iid chi-square random variables with one degree of freedom and ( λ j ( α 0 ) ) j ≥ 1 (\lambda_j(\alpha_0))_{j\geq1} ( λ j ( α 0 ) ) j ≥ 1 are the eigenvalues of the covariance operator K α 0 \mathcal{K}_{\alpha_0} K α 0 . Let α ^ n \widehat{\alpha}_n α n be a consistent estimator of α 0 \alpha_0 α 0 , and let λ ^ 1 , n ≥ λ ^ 2 , n ≥ ⋯ ≥ 0 \widehat{\lambda}_{1,n}\geq\widehat{\lambda}_{2,n}\geq\cdots\geq0 λ 1 , n ≥ λ 2 , n ≥ ⋯ ≥ 0 be the eigenvalues obtained from a plug-in Nyström approximation of K α ^ n \mathcal{K}_{\widehat{\alpha}_n} K α n , extended by zeros. Assume that
∑ j = 1 ∞ ∣ λ ^ j , n − λ j ( α 0 ) ∣ → P 0. \sum_{j=1}^{\infty}\lvert\widehat{\lambda}_{j,n}-\lambda_j(\alpha_0)\rvert\xrightarrow{\mathbb{P}}0. j = 1 ∑ ∞ ∣ λ j , n − λ j ( α 0 )∣ P 0. Define
Q ^ n : = ∑ j = 1 ∞ λ ^ j , n χ 1 , j 2 . \widehat{Q}_n:=\sum_{j=1}^{\infty}\widehat{\lambda}_{j,n}\chi_{1,j}^2. Q n := j = 1 ∑ ∞ λ j , n χ 1 , j 2 . conditionally on the data, using auxiliary chi-square variables independent of the observations. Let q ^ n , 1 − β \widehat{q}_{n,1-\beta} q n , 1 − β be its conditional lower ( 1 − β ) (1-\beta) ( 1 − β ) -quantile, where 0 < β < 1 0<\beta<1 0 < β < 1 . Write F Q , α 0 F_{Q,\alpha_0} F Q , α 0 for the distribution function of Q α 0 Q_{\alpha_0} Q α 0 and set q : = q 1 − β ( α 0 ) : = inf { x : F Q , α 0 ( x ) ≥ 1 − β } q:=q_{1-\beta}(\alpha_0):=\inf\{x:F_{Q,\alpha_0}(x)\geq1-\beta\} q := q 1 − β ( α 0 ) := inf { x : F Q , α 0 ( x ) ≥ 1 − β } . Assume that F Q , α 0 F_{Q,\alpha_0} F Q , α 0 is continuous at q q q and, for every ε > 0 \varepsilon>0 ε > 0 ,
F Q , α 0 ( q − ε ) < 1 − β < F Q , α 0 ( q + ε ) . F_{Q,\alpha_0}(q-\varepsilon)<1-\beta<F_{Q,\alpha_0}(q+\varepsilon). F Q , α 0 ( q − ε ) < 1 − β < F Q , α 0 ( q + ε ) . Then
q ^ n , 1 − β → P q 1 − β ( α 0 ) , \widehat{q}_{n,1-\beta}\xrightarrow{\mathbb{P}}q_{1-\beta}(\alpha_0), q n , 1 − β P q 1 − β ( α 0 ) , and
lim n → ∞ P H 0 [ n S n > q ^ n , 1 − β ] = β . \lim_{n\to\infty}\mathbb{P}_{H_0}[nS_n>\widehat{q}_{n,1-\beta}]=\beta. n → ∞ lim P H 0 [ n S n > q n , 1 − β ] = β . Proof. The eigenvalue assumption controls the difference between the two weighted chi-square variables under a common coupling. Couple Q ^ n \widehat{Q}_n Q n and Q α 0 Q_{\alpha_0} Q α 0 with the same independent chi-square variables. Conditionally on the data,
E [ ∣ Q ^ n − Q α 0 ∣ ∣ X 1 , … , X n ] ≤ ∑ j = 1 ∞ ∣ λ ^ j , n − λ j ( α 0 ) ∣ → P 0. \mathbb{E}\left[\lvert\widehat{Q}_n-Q_{\alpha_0}\rvert\mid X_1,\ldots,X_n\right]\leq\sum_{j=1}^{\infty}\lvert\widehat{\lambda}_{j,n}-\lambda_j(\alpha_0)\rvert\xrightarrow{\mathbb{P}}0. E [ ∣ Q n − Q α 0 ∣ ∣ X 1 , … , X n ] ≤ j = 1 ∑ ∞ ∣ λ j , n − λ j ( α 0 )∣ P 0. The bound proves conditional weak convergence in probability. The standard criterion for quantile convergence [37 Section 21.1] , applied along almost-surely convergent subsequences, then gives the stated convergence of quantiles. Slutsky’s theorem and continuity of F Q , α 0 F_{Q,\alpha_0} F Q , α 0 at q q q give the asserted level. ■
The quantile condition in Theorem 6 holds for the present covariance operator. Indeed, C α 0 ( t ) > 0 C_{\alpha_0}(t)>0 C α 0 ( t ) > 0 for t > 0 t>0 t > 0 , and h α 0 ( x , t ) h_{\alpha_0}(x,t) h α 0 ( x , t ) has a nonzero logarithmic term, so the operator is nonzero. Its summable weighted chi-square law has a continuous distribution function that is strictly increasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) .
The following proposition gives sufficient conditions for the numerical approximation to satisfy the spectral convergence assumption of Theorem 6.
Proposition 2. Assume H 0 H_0 H 0 and α ^ n → α 0 > 0 \widehat{\alpha}_n \to\alpha_0 > 0 α n → α 0 > 0 in probability. Let ( t i , n , w i , n ) i = 1 m n (t_{i,n},w_{i,n})_{i=1}^{m_n} ( t i , n , w i , n ) i = 1 m n be the Gauss–Legendre rule on [ 0 , τ n ] [0,\tau_n] [ 0 , τ n ] , and let K ~ n \widetilde{K}_n K n be a symmetric numerical approximation to K α ^ n K_{\widehat{\alpha}_n} K α n . Define ϵ n : = max i , j ≤ m n ∣ K ~ n ( t i , n , t j , n ) − K α ^ n ( t i , n , t j , n ) ∣ \epsilon_n := \max_{i,j\le m_n}|\widetilde{K}_n(t_{i,n},t_{j,n})-K_{\widehat{\alpha}_n}(t_{i,n},t_{j,n})| ϵ n := max i , j ≤ m n ∣ K n ( t i , n , t j , n ) − K α n ( t i , n , t j , n ) ∣ . Suppose
τ n ⟶ ∞ , τ n 2 m n ⟶ 0 , m n τ n ϵ n ⟶ P 0. (10) \tau_n \longrightarrow\infty,\qquad\frac{\tau_n^2}{\sqrt{m_n}}\longrightarrow0,\qquad m_n\tau_n\epsilon_n \overset{\mathbb{P}}{\longrightarrow} 0. \tag*{(10)} τ n ⟶ ∞ , m n τ n 2 ⟶ 0 , m n τ n ϵ n ⟶ P 0. (10) Let λ ^ j , n \widehat{\lambda}_{j,n} λ j , n be the ordered eigenvalues of B ~ n = ( w i , n w j , n K ~ n ( t i , n , t j , n ) ) i , j = 1 m n \widetilde{B}_n=(\sqrt{w_{i,n}w_{j,n}}\widetilde{K}_n(t_{i,n},t_{j,n}))_{i,j=1}^{m_n} B n = ( w i , n w j , n K n ( t i , n , t j , n ) ) i , j = 1 m n , with negative eigenvalues replaced by zero and the sequence extended by zeros. Then ∑ j ≥ 1 ∣ λ ^ j , n − λ j ( α 0 ) ∣ → 0 \sum_{j\ge1}|\widehat{\lambda}_{j,n}-\lambda_j(\alpha_0)|\to0 ∑ j ≥ 1 ∣ λ j , n − λ j ( α 0 ) ∣ → 0 in probability. Consequently, calibration using empirical quantiles from M n → ∞ M_n\to\infty M n → ∞ conditionally independent weighted chi-square draws has asymptotic level β \beta β .
Proof. Throughout the proof, ∥ ⋅ ∥ 1 \|\cdot\|_1 ∥ ⋅ ∥ 1 denotes the trace norm and ∥ ⋅ ∥ o p \|\cdot\|_{\mathrm{op}} ∥ ⋅ ∥ op the operator norm. Let G = L 2 ( [ 0 , ∞ ) , e − z d z ) G=L^2([0,\infty),e^{-z}\,dz) G = L 2 ([ 0 , ∞ ) , e − z d z ) and set g a ( t ) ( z ) : = h a ( e z / a , t ) g_a(t)(z):=h_a(e^{z/a},t) g a ( t ) ( z ) := h a ( e z / a , t ) , where h a h_a h a is the influence function in the proof of Theorem 4 . Then K a ( s , t ) = ⟨ g a ( s ) , g a ( t ) ⟩ G K_a(s,t)=\langle g_a(s),g_a(t)\rangle_G K a ( s , t ) = ⟨ g a ( s ) , g a ( t ) ⟩ G . Writing ( v ⊗ v ) f : = ⟨ f , v ⟩ G v (v\otimes v)f:=\langle f,v\rangle_Gv ( v ⊗ v ) f := ⟨ f , v ⟩ G v , define U a ( t ) : = g a ( t ) ⊗ g a ( t ) U_a(t):=g_a(t)\otimes g_a(t) U a ( t ) := g a ( t ) ⊗ g a ( t ) and C a : = ∫ 0 ∞ U a ( t ) d t C_a:=\int_0^\infty U_a(t)\,dt C a := ∫ 0 ∞ U a ( t ) d t . The operators C a C_a C a and K a K_a K a have the same nonzero eigenvalues, as do ∑ i w i , n U a ( t i , n ) \sum_i w_{i,n}U_a(t_{i,n}) ∑ i w i , n U a ( t i , n ) and the exact Nyström matrix B n ( a ) B_n(a) B n ( a ) .
On every compact A ⊂ ( 0 , ∞ ) A\subset(0,\infty) A ⊂ ( 0 , ∞ ) , the expression for h a h_a h a and its derivative in t t t give ∥ g a ( t ) ∥ G + ∥ ∂ t g a ( t ) ∥ G ≤ c A ( 1 + t ) e − t \|g_a(t)\|_G+\|\partial_tg_a(t)\|_G\le c_A(1+t)e^{-t} ∥ g a ( t ) ∥ G + ∥ ∂ t g a ( t ) ∥ G ≤ c A ( 1 + t ) e − t . Thus U a U_a U a is uniformly Lipschitz in trace norm. Moreover, ∥ v ⊗ v − w ⊗ w ∥ 1 ≤ ( ∥ v ∥ + ∥ w ∥ ) ∥ v − w ∥ \|v\otimes v-w\otimes w\|_1\le(\|v\|+\|w\|)\|v-w\| ∥ v ⊗ v − w ⊗ w ∥ 1 ≤ ( ∥ v ∥ + ∥ w ∥ ) ∥ v − w ∥ and dominated convergence show that a ↦ C a a\mapsto C_a a ↦ C a is trace-norm continuous.
Bernstein polynomial approximation of degree m n m_n m n approximates U a U_a U a uniformly on [ 0 , τ n ] [0,\tau_n] [ 0 , τ n ] with trace-norm error at most c A τ n / m n c_A\tau_n/\sqrt{m_n} c A τ n / m n . Positivity and polynomial exactness of the quadrature, together with ∑ i w i , n = τ n \sum_iw_{i,n}=\tau_n ∑ i w i , n = τ n , therefore yield
sup a ∈ A ∥ ∑ i = 1 m n w i , n U a ( t i , n ) − C a ∥ 1 ≤ c A { τ n 2 m n + ( 1 + τ n ) 2 e − 2 τ n } . (11) \sup_{a\in A}\left\|\sum_{i=1}^{m_n}w_{i,n}U_a(t_{i,n})-C_a\right\|_1
\le c_A\left\{\frac{\tau_n^2}{\sqrt{m_n}}+(1+\tau_n)^2e^{-2\tau_n}\right\}. \tag*{(11)} a ∈ A sup i = 1 ∑ m n w i , n U a ( t i , n ) − C a 1 ≤ c A { m n τ n 2 + ( 1 + τ n ) 2 e − 2 τ n } . (11)
Combining (11), (10), and plug-in consistency gives trace-norm convergence to C α 0 C_{\alpha_0} C α 0 in probability. Hence the exact Nyström eigenvalues converge coordinatewise and their sums converge. Nonnegativity then gives convergence in ℓ 1 \ell^1 ℓ 1 .
For the numerically evaluated matrix, ∥ B ~ n − B n ( α ^ n ) ∥ o p ≤ τ n ϵ n \|\widetilde{B}_n-B_n(\widehat{\alpha}_n)\|_{\mathrm{op}}\le\tau_n\epsilon_n ∥ B n − B n ( α n ) ∥ op ≤ τ n ϵ n . The min–max principle bounds the sum of the absolute eigenvalue differences by m n τ n ϵ n m_n\tau_n\epsilon_n m n τ n ϵ n ; replacing negative eigenvalues by zero cannot increase their distance from the nonnegative exact eigenvalues. This proves the asserted spectral convergence.
Finally, conditionally on the data, the variance of the empirical distribution function of the M n M_n M n reference draws at any fixed argument is at most 1 / ( 4 M n ) 1/(4M_n) 1/ ( 4 M n ) . Applying Chebyshev’s inequality at points on either side of the limiting quantile, and using the quantile separation established in Theorem 6 , proves empirical quantile consistency. The asserted asymptotic level follows. ■
Remark 3. With exact kernel evaluation, ϵ n = 0 \epsilon_n=0 ϵ n = 0 . For Gauss–Laguerre evaluation, the substitution X = e z / a X=e^{z/a} X = e z / a expresses the kernel expectations as integrals against e − z d z e^{-z}\,dz e − z d z of continuous functions bounded by c A ( 1 + z ) 2 c_A(1+z)^2 c A ( 1 + z ) 2 on compact parameter sets. The positive Gauss–Laguerre measures converge weakly to this exponential measure by moment exactness and moment determinacy; their exact fourth moments ensure uniform integrability. Consequently, the kernel approximations converge uniformly on compact ( a , s , t ) (a,s,t) ( a , s , t ) -sets.
Writing K a ( L ) K_a^{(L)} K a ( L ) for the approximation with L L L Gauss–Laguerre nodes, choose L n → ∞ L_n \to\infty L n → ∞ so that sup a ∈ [ 1 / n , n ] , s , t ∈ [ 0 , τ n ] ∣ K a ( L n ) ( s , t ) − K a ( s , t ) ∣ ≤ ( n m n τ n ) − 1 \sup_{a\in[1/n,n],\,s,t\in[0,\tau_n]}|K_a^{(L_n)}(s,t)-K_a(s,t)|\le(n m_n \tau_n)^{-1} sup a ∈ [ 1/ n , n ] , s , t ∈ [ 0 , τ n ] ∣ K a ( L n ) ( s , t ) − K a ( s , t ) ∣ ≤ ( n m n τ n ) − 1 . Such a choice is possible by the compact-uniform convergence. Since P ( α ^ n ∈ [ 1 / n , n ] ) → 1 \mathbb{P}(\widehat{\alpha}_n\in[1/n,n])\to1 P ( α n ∈ [ 1/ n , n ]) → 1 , the last condition in (10) follows.
Proposition 2 concerns an increasing-resolution approximation, while the numerical studies below use fixed-resolution implementations with the settings specified in the respective sections.
Parametric-bootstrap calibration
The bootstrap generates samples from the fitted Pareto law and repeats parameter estimation for each sample. Its validity follows from stability of the null limits along convergent sequences of shape parameters. The argument applies to both T n T_n T n and S n S_n S n .
Theorem 7. Let X 1 , … , X n X_1,\ldots,X_n X 1 , … , X n be iid from P ( α 0 ) \mathcal{P}(\alpha_0) P ( α 0 ) , α 0 > 0 \alpha_0>0 α 0 > 0 , and let α ^ n \widehat{\alpha}_n α n be the MLE. For test calibration use exact conditional quantiles, or empirical conditional quantiles from B n → ∞ B_n\to\infty B n → ∞ independent bootstrap samples. Conditionally on the observations, let X 1 ∗ , … , X n ∗ X_1^*,\ldots,X_n^* X 1 ∗ , … , X n ∗ be iid from P ( α ^ n ) \mathcal{P}(\widehat{\alpha}_n) P ( α n ) and compute T n ∗ T_n^* T n ∗ and S n ∗ S_n^* S n ∗ from the bootstrap sample, including re-estimation of α \alpha α . Then the conditional distribution of n S n ∗ nS_n^* n S n ∗ converges in probability to Q α 0 = ∥ Z ∥ H Q_{\alpha_0}=\|Z\|_{\mathbb{H}} Q α 0 = ∥ Z ∥ H . In addition, the conditional distribution of the studentized version of T n ∗ T_n^* T n ∗ converges in probability to the standard normal law. Thus the parametric bootstrap tests based on T n T_n T n and S n S_n S n have asymptotic level β \beta β for every 0 < β < 1 0<\beta<1 0 < β < 1 at continuity points of the corresponding limiting distributions.
Proof. We verify the local-parameter convergence condition in the standard parametric-bootstrap consistency theory [5 Section 2] ; see also [37 Chapter 23] . It suffices to establish the null limits along every deterministic parameter sequence converging to α 0 \alpha_0 α 0 . At parameter a > 0 a>0 a > 0 , the Pareto family with its fixed support is a regular one-parameter exponential family, with score a − 1 − log x a^{-1}-\log x a − 1 − log x and Fisher information a − 2 a^{-2} a − 2 , and its MLE is asymptotically normal by (6).
For any deterministic sequence a n a_n a n converging to a > 0 a>0 a > 0 , couple a row of Pareto observations as X n , j : = exp ( E j / a n ) X_{n,j}:=\exp(E_j/a_n) X n , j := exp ( E j / a n ) , where the E j E_j E j are iid Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) . This representation places all rows on the same probability space. The corresponding MLE is α ~ n : = a n / E ‾ n \widetilde{\alpha}_n:=a_n/\overline{E}_n α n := a n / E n , where E ‾ n : = n − 1 ∑ j = 1 n E j \overline{E}_n:=n^{-1}\sum_{j=1}^n E_j E n := n − 1 ∑ j = 1 n E j , and
n ( α ~ n − a n ) = − a n n ∑ j = 1 n ( E j − 1 ) + a n n ( E ‾ n − 1 ) 2 E ‾ n . \sqrt{n}(\widetilde{\alpha}_n-a_n)=-\frac{a_n}{\sqrt{n}}\sum_{j=1}^n(E_j-1)+\frac{a_n\sqrt{n}(\overline{E}_n-1)^2}{\overline{E}_n}. n ( α n − a n ) = − n a n j = 1 ∑ n ( E j − 1 ) + E n a n n ( E n − 1 ) 2 .
The last term is o P ( 1 ) o_{\mathbb{P}}(1) o P ( 1 ) . Let E E E denote a generic Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) distributed variable. For a n a_n a n in a compact subset of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , the influence functions h a h_a h a from the proof of Theorem 4 satisfy, for a finite constant M M M depending only on that compact set,
∣ h a n ( e E / a n , t ) ∣ ≤ M ( 1 + E ) ( 1 + t ) e − t , E [ ∥ h a n ( e E / a n , ⋅ ) − h a ( e E / a , ⋅ ) ∥ H 2 ] → 0. |h_{a_n}(e^{E/a_n},t)|\le M(1+E)(1+t)e^{-t},\qquad\mathbb{E}\left[\|h_{a_n}(e^{E/a_n},\cdot)-h_a(e^{E/a},\cdot)\|_{\mathbb{H}}^2\right]\to0. ∣ h a n ( e E / a n , t ) ∣ ≤ M ( 1 + E ) ( 1 + t ) e − t , E [ ∥ h a n ( e E / a n , ⋅ ) − h a ( e E / a , ⋅ ) ∥ H 2 ] → 0.
The second statement follows by dominated convergence. Together with the bound ∥ Y x ∥ H 2 ≤ 1 / 2 \|Y_x\|_{\mathbb{H}}^2\le1/2 ∥ Y x ∥ H 2 ≤ 1/2 from the same proof, these bounds verify the negligible remainder, convergence of covariances and their traces, and Lindeberg conditions for the Hilbert-space triangular-array central limit theorem [17 Chapter 17] . Let Q a Q_a Q a denote the limit in Theorem 4 with parameter a a a . It follows that n S n → D Q a nS_n \xrightarrow{\mathcal{D}} Q_a n S n D Q a along every such sequence. The scalar influence function for T n T_n T n is likewise continuous in a a a in L 2 L^2 L 2 under this coupling and bounded by M ( 1 + E ) M(1+E) M ( 1 + E ) ; therefore n T n → D N ( 0 , σ 2 ( a ) ) \sqrt{n}T_n \xrightarrow{\mathcal{D}} N(0,\sigma^2(a)) n T n D N ( 0 , σ 2 ( a )) , where σ 2 ( a ) = 16 a / { ( a + 4 ) ( a + 2 ) 2 } \sigma^2(a)=16a/\{(a+4)(a+2)^2\} σ 2 ( a ) = 16 a / {( a + 4 ) ( a + 2 ) 2 } , and studentization converges to the standard normal law along the same sequences. Applying parametric-bootstrap consistency with a n a_n a n replaced by α ^ n → α 0 \widehat{\alpha}_n \to\alpha_0 α n → α 0 gives both asserted conditional limits. Conditional quantile consistency and the stipulated quantile conditions give the level statements. The scalar unstudentized limit also justifies the equal-tailed bootstrap calibration of T n T_n T n used in the simulations. ■
Simulation study
We use sample sizes n = 20 , 50 , 100 n=20,50,100 n = 20 , 50 , 100 , N = 10,000 N=10{,}000 N = 10 , 000 replications, and nominal significance level η = 0.05 \eta=0.05 η = 0.05 . For the integrated statistic S n S_n S n , we report two calibrations: the procedure S n , K S_{n,K} S n , K uses the plug-in covariance operator described after Theorem 6 , and the procedure S n , B S_{n,B} S n , B uses the parametric bootstrap of Theorem 7. For T n T_n T n , we use a two-sided bootstrap test to facilitate comparison with the competing procedures.
The proposed procedures are implemented as follows.
Draw a random sample X 1 , … , X n X_1,\ldots,X_n X 1 , … , X n from the distribution under study.
Estimate the shape parameter by the MLE α ^ n = n / ∑ i = 1 n log X i \widehat{\alpha}_n=n/\sum_{i=1}^n\log X_i α n = n / ∑ i = 1 n log X i .
Evaluate T n T_n T n and S n S_n S n from (4) and (5) . The statistic T n T_n T n is calibrated as a two-sided statistic and S n S_n S n as a right-tailed statistic.
For the bootstrap calibration, generate B = 500 B=500 B = 500 samples of size n n n from P ( α ^ n ) \mathcal{P}(\widehat{\alpha}_n) P ( α n ) . Re-estimate α \alpha α in each bootstrap sample and compute T n ∗ ( b ) T_n^{*(b)} T n ∗ ( b ) and S n ∗ ( b ) S_n^{*(b)} S n ∗ ( b ) , b = 1 , … , B b=1,\ldots,B b = 1 , … , B . Denote the empirical quantiles of the bootstrap T n T_n T n values at levels η / 2 \eta/2 η /2 and 1 − η / 2 1-\eta/2 1 − η /2 by q n , η / 2 ∗ T q_{n,\eta/2}^{*T} q n , η /2 ∗ T and q n , 1 − η / 2 ∗ T q_{n,1-\eta/2}^{*T} q n , 1 − η /2 ∗ T , respectively. Denote the empirical quantile of the bootstrap S n S_n S n values at level 1 − η 1-\eta 1 − η by q n , 1 − η ∗ S q_{n,1-\eta}^{*S} q n , 1 − η ∗ S .
For the covariance-operator calibration of S n S_n S n , compute the eigenvalues of the plug-in Nyström matrix with α ^ n \widehat{\alpha}_n α n , T = 12 T=12 T = 12 and m = 50 m=50 m = 50 , and approximate the 1 − η 1-\eta 1 − η quantile q n , 1 − η K q_{n,1-\eta}^{K} q n , 1 − η K of ∑ j λ ^ j , n χ 1 , j 2 \sum_j\widehat{\lambda}_{j,n}\chi_{1,j}^2 ∑ j λ j , n χ 1 , j 2 by 20,000 Monte Carlo draws.
Reject with the bootstrap version of T n T_n T n if T n < q n , η / 2 ∗ T T_n<q_{n,\eta/2}^{*T} T n < q n , η /2 ∗ T or T n > q n , 1 − η / 2 ∗ T T_n>q_{n,1-\eta/2}^{*T} T n > q n , 1 − η /2 ∗ T ; reject with S n , B S_{n,B} S n , B if S n > q n , 1 − η ∗ S S_n>q_{n,1-\eta}^{*S} S n > q n , 1 − η ∗ S ; reject with S n , K S_{n,K} S n , K if n S n > q n , 1 − η K nS_n>q_{n,1-\eta}^{K} n S n > q n , 1 − η K .
The competing procedures defined below use the same simulated samples and the same bootstrap sample size B = 500 B=500 B = 500 . After N N N replications, we report the percentage of samples in which each test rejects. If p ^ \widehat{p} p denotes the rejection proportion, the Monte Carlo standard error of the reported percentage 100 p ^ 100\widehat{p} 100 p is 100 { p ^ ( 1 − p ^ ) / N } 1 / 2 100\{\widehat{p}(1-\widehat{p})/N\}^{1/2} 100 { p ( 1 − p ) / N } 1/2 . With N = 10,000 N=10{,}000 N = 10 , 000 , it is about 0.22 percentage points at 5% and at most 0.50 percentage points.
The comparison includes tests based on empirical distribution functions, transforms, and powers and minima of the observations. To fix the common notation, let X ( 1 ) ≤ ⋯ ≤ X ( n ) X_{(1)}\leq\cdots\leq X_{(n)} X ( 1 ) ≤ ⋯ ≤ X ( n ) denote the ordered sample. The empirical distribution function is
F ^ n ( t ) : = 1 n ∑ i = 1 n 1 { X i ≤ t } . \widehat{F}_n(t):=\frac{1}{n}\sum_{i=1}^{n}\mathbf{1}_{\{X_i\leq t\}}. F n ( t ) := n 1 i = 1 ∑ n 1 { X i ≤ t } .
Distribution-function tests. The test of Obradović et al. [31 ] compares the empirical distribution of the observations with that of the maximum ratio within each pair. The latter is
M n ( t ) : = ( n 2 ) − 1 ∑ i = 1 n − 1 ∑ j = i + 1 n 1 { max ( X i / X j , X j / X i ) ≤ t } , t ≥ 1. M_n(t) := \binom{n}{2}^{-1}\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} \mathbf{1}_{\{\max(X_i/X_j,X_j/X_i)\le t\}}, \quad t\ge1. M n ( t ) := ( 2 n ) − 1 i = 1 ∑ n − 1 j = i + 1 ∑ n 1 { m a x ( X i / X j , X j / X i ) ≤ t } , t ≥ 1.
The resulting integrated difference is
O n : = ∫ 1 ∞ ( M n ( t ) − F ^ n ( t ) ) d F ^ n ( t ) . O_n := \int_{1}^{\infty}\left(M_n(t)-\widehat{F}_n(t)\right)\,\mathrm{d}\widehat{F}_n(t). O n := ∫ 1 ∞ ( M n ( t ) − F n ( t ) ) d F n ( t ) .
The three classical empirical-distribution tests compare the sample with the fitted distribution F ^ α ^ n ( ⋅ ) : = F ( ⋅ , α ^ n ) \widehat{F}_{\widehat{\alpha}_n}(\cdot):=F(\cdot,\widehat{\alpha}_n) F α n ( ⋅ ) := F ( ⋅ , α n ) from (1) . The Kolmogorov–Smirnov statistic is defined by
K S n : = max ( max 1 ≤ j ≤ n ( j n − F ^ α ^ n ( X ( j ) ) ) , max 1 ≤ j ≤ n ( F ^ α ^ n ( X ( j ) ) − j − 1 n ) ) . KS_n := \max\left(\max_{1\le j\le n}\left(\frac{j}{n}-\widehat{F}_{\widehat{\alpha}_n}(X_{(j)})\right),\max_{1\le j\le n}\left(\widehat{F}_{\widehat{\alpha}_n}(X_{(j)})-\frac{j-1}{n}\right)\right). K S n := max ( 1 ≤ j ≤ n max ( n j − F α n ( X ( j ) ) ) , 1 ≤ j ≤ n max ( F α n ( X ( j ) ) − n j − 1 ) ) .
The Cramér–von Mises statistic is
C M n : = 1 12 n + ∑ j = 1 n ( F ^ α ^ n ( X ( j ) ) − 2 j − 1 2 n ) 2 , CM_n := \frac{1}{12n}+\sum_{j=1}^{n}\left(\widehat{F}_{\widehat{\alpha}_n}(X_{(j)})-\frac{2j-1}{2n}\right)^2, C M n := 12 n 1 + j = 1 ∑ n ( F α n ( X ( j ) ) − 2 n 2 j − 1 ) 2 ,
whereas the Anderson–Darling statistic is given by
A D n : = − n − 1 n ∑ j = 1 n ( 2 j − 1 ) ( log ( F ^ α ^ n ( X ( j ) ) ) + log ( 1 − F ^ α ^ n ( X ( n + 1 − j ) ) ) ) . AD_n := -n-\frac{1}{n}\sum_{j=1}^{n}(2j-1)\left(\log(\widehat{F}_{\widehat{\alpha}_n}(X_{(j)}))+\log(1-\widehat{F}_{\widehat{\alpha}_n}(X_{(n+1-j)}))\right). A D n := − n − n 1 j = 1 ∑ n ( 2 j − 1 ) ( log ( F α n ( X ( j ) )) + log ( 1 − F α n ( X ( n + 1 − j ) )) ) .
Transform tests. For the Mellin-transform test of Meintanis [24 ] , take a tuning parameter a > 0 a>0 a > 0 and define, for t ≥ 1 t\ge1 t ≥ 1 , the auxiliary functions
I a ( 0 ) ( t ) : = ( a + log t ) − 1 , I a ( 1 ) ( t ) : = 1 − a − log t ( a + log t ) 2 , I_a^{(0)}(t):=(a+\log t)^{-1}, \qquad I_a^{(1)}(t):=\frac{1-a-\log t}{(a+\log t)^2}, I a ( 0 ) ( t ) := ( a + log t ) − 1 , I a ( 1 ) ( t ) := ( a + log t ) 2 1 − a − log t ,
and
I a ( 2 ) ( t ) : = 2 − 2 a + a 2 + 2 ( a − 1 ) log t + log 2 t ( a + log t ) 3 . I_a^{(2)}(t):=\frac{2-2a+a^2+2(a-1)\log t+\log^2 t}{(a+\log t)^3}. I a ( 2 ) ( t ) := ( a + log t ) 3 2 − 2 a + a 2 + 2 ( a − 1 ) log t + log 2 t .
The statistic combines these functions as
G n , a : = 1 n ∑ j , k = 1 n [ ( α ^ n + 1 ) 2 I a ( 0 ) ( X j X k ) + I a ( 2 ) ( X j X k ) + 2 ( α ^ n + 1 ) I a ( 1 ) ( X j X k ) + n α ^ n 2 I a ( 0 ) ( 1 ) ] − 2 α ^ n ∑ j = 1 n [ ( α ^ n + 1 ) I a ( 0 ) ( X j ) + I a ( 1 ) ( X j ) ] . G_{n,a}:=\frac{1}{n}\sum_{j,k=1}^{n}\left[(\widehat{\alpha}_n+1)^2I_a^{(0)}(X_jX_k)+I_a^{(2)}(X_jX_k)+2(\widehat{\alpha}_n+1)I_a^{(1)}(X_jX_k)+n\widehat{\alpha}_n^2I_a^{(0)}(1)\right]-2\widehat{\alpha}_n\sum_{j=1}^{n}\left[(\widehat{\alpha}_n+1)I_a^{(0)}(X_j)+I_a^{(1)}(X_j)\right]. G n , a := n 1 j , k = 1 ∑ n [ ( α n + 1 ) 2 I a ( 0 ) ( X j X k ) + I a ( 2 ) ( X j X k ) + 2 ( α n + 1 ) I a ( 1 ) ( X j X k ) + n α n 2 I a ( 0 ) ( 1 ) ] − 2 α n j = 1 ∑ n [ ( α n + 1 ) I a ( 0 ) ( X j ) + I a ( 1 ) ( X j ) ] .
We use the tuning parameter a = 2 a = 2 a = 2 .
The test of Meintanis et al. [25 ] applies the empirical characteristic function to the fitted probability transforms
U ^ j : = F ^ α ^ n ( X j ) = 1 − X j − α ^ n , j = 1 , 2 , … , n . \widehat{U}_j := \widehat{F}_{\widehat{\alpha}_n}(X_j) = 1 - X_j^{-\widehat{\alpha}_n}, \qquad j = 1,2,\ldots,n. U j := F α n ( X j ) = 1 − X j − α n , j = 1 , 2 , … , n .
In terms of U ^ 1 , … , U ^ n \widehat{U}_1,\ldots,\widehat{U}_n U 1 , … , U n , the statistic is
M n , a : = 1 n ∑ j , k = 1 n 2 a ( U ^ j − U ^ k ) 2 + a 2 + 2 n ( 2 tan − 1 ( 1 a ) − a log ( 1 + 1 a 2 ) ) − 4 ∑ j = 1 n ( tan − 1 ( U ^ j a ) + tan − 1 ( 1 − U ^ j a ) ) . M_{n,a} := \frac{1}{n}\sum_{j,k=1}^{n}\frac{2a}{(\widehat{U}_j-\widehat{U}_k)^2+a^2}+2n\left(2\tan^{-1}\left(\frac{1}{a}\right)-a\log\left(1+\frac{1}{a^2}\right)\right)-4\sum_{j=1}^{n}\left(\tan^{-1}\left(\frac{\widehat{U}_j}{a}\right)+\tan^{-1}\left(\frac{1-\widehat{U}_j}{a}\right)\right). M n , a := n 1 j , k = 1 ∑ n ( U j − U k ) 2 + a 2 2 a + 2 n ( 2 tan − 1 ( a 1 ) − a log ( 1 + a 2 1 ) ) − 4 j = 1 ∑ n ( tan − 1 ( a U j ) + tan − 1 ( a 1 − U j ) ) .
We again use a = 2 a = 2 a = 2 .
Tests based on powers and minima. The test of Allison et al. [1 ] compares empirical distribution functions for powers of the observations and minima of m m m observations. Their difference is
Δ n , m ( x ) : = 1 n ∑ j = 1 n 1 { X j 1 / m ≤ x } − 1 n m ∑ j 1 , j 2 , … , j m = 1 n 1 { min ( X j 1 , X j 2 , … , X j m ) ≤ x } . \Delta_{n,m}(x) := \frac{1}{n}\sum_{j=1}^{n}\mathbf{1}_{\{X_j^{1/m}\le x\}}-\frac{1}{n^m}\sum_{j_1,j_2,\ldots,j_m=1}^{n}\mathbf{1}_{\{\min(X_{j_1},X_{j_2},\ldots,X_{j_m})\le x\}}. Δ n , m ( x ) := n 1 j = 1 ∑ n 1 { X j 1/ m ≤ x } − n m 1 j 1 , j 2 , … , j m = 1 ∑ n 1 { m i n ( X j 1 , X j 2 , … , X j m ) ≤ x } .
Integrating this difference gives the statistic
I n , m : = ∫ 1 ∞ Δ n , m ( x ) d F ^ n ( x ) . I_{n,m} := \int_{1}^{\infty}\Delta_{n,m}(x)\,\mathrm{d}\widehat{F}_n(x). I n , m := ∫ 1 ∞ Δ n , m ( x ) d F n ( x ) .
We use the tuning parameter m = 2 m = 2 m = 2 .
The four statistics from Ndwandwe et al. [27 ] use two sets of rank weights. Define
v j , m : = 1 n m ( ( n − j + 1 ) m − ( n − j ) m ) and u j , m : = ( n − j m − 1 ) . v_{j,m} := \frac{1}{n^m}\left((n-j+1)^m-(n-j)^m\right) \quad\text{and}\quad u_{j,m} := \binom{n-j}{m-1}. v j , m := n m 1 ( ( n − j + 1 ) m − ( n − j ) m ) and u j , m := ( m − 1 n − j ) .
The weights v j , m v_{j,m} v j , m are used for 1 ≤ j ≤ n 1\le j\le n 1 ≤ j ≤ n , and the weights u j , m u_{j,m} u j , m for 1 ≤ j ≤ n − m + 1 1\le j\le n-m+1 1 ≤ j ≤ n − m + 1 . The first pair of statistics uses v j , m v_{j,m} v j , m :
S n , m , a ( 1 ) : = 1 n ∑ j = 1 n ∑ k = 1 n ( 2 a a 2 + ( X ( j ) 1 / m − X ( k ) 1 / m ) 2 − n v j , m 4 a a 2 + ( X ( j ) − X ( k ) 1 / m ) 2 + n 2 v j , m v k , m 2 a a 2 + ( X ( j ) − X ( k ) ) 2 ) , S_{n,m,a}^{(1)} := \frac{1}{n}\sum_{j=1}^{n}\sum_{k=1}^{n}\left(\frac{2a}{a^2+(X_{(j)}^{1/m}-X_{(k)}^{1/m})^2}-nv_{j,m}\frac{4a}{a^2+(X_{(j)}-X_{(k)}^{1/m})^2}+n^2v_{j,m}v_{k,m}\frac{2a}{a^2+(X_{(j)}-X_{(k)})^2}\right), S n , m , a ( 1 ) := n 1 j = 1 ∑ n k = 1 ∑ n a 2 + ( X ( j ) 1/ m − X ( k ) 1/ m ) 2 2 a − n v j , m a 2 + ( X ( j ) − X ( k ) 1/ m ) 2 4 a + n 2 v j , m v k , m a 2 + ( X ( j ) − X ( k ) ) 2 2 a ,
and
S n , m , a ( 2 ) : = 1 n π a ∑ j = 1 n ∑ k = 1 n ( exp ( − ( X ( j ) 1 / m − X ( k ) 1 / m ) 2 4 a ) − 2 n v j , m exp ( − ( X ( j ) − X ( k ) 1 / m ) 2 4 a ) + n 2 v j , m v k , m exp ( − ( X ( j ) − X ( k ) ) 2 4 a ) ) . S_{n,m,a}^{(2)} := \frac{1}{n}\sqrt{\frac{\pi}{a}}\sum_{j=1}^{n}\sum_{k=1}^{n}\left(\exp\left(\frac{-(X_{(j)}^{1/m}-X_{(k)}^{1/m})^2}{4a}\right)-2nv_{j,m}\exp\left(\frac{-(X_{(j)}-X_{(k)}^{1/m})^2}{4a}\right)+n^2v_{j,m}v_{k,m}\exp\left(\frac{-(X_{(j)}-X_{(k)})^2}{4a}\right)\right). S n , m , a ( 2 ) := n 1 a π j = 1 ∑ n k = 1 ∑ n exp 4 a − ( X ( j ) 1/ m − X ( k ) 1/ m ) 2 − 2 n v j , m exp 4 a − ( X ( j ) − X ( k ) 1/ m ) 2 + n 2 v j , m v k , m exp ( 4 a − ( X ( j ) − X ( k ) ) 2 ) .
The second pair uses u j , m u_{j,m} u j , m :
T n , m , a ( 1 ) : = 1 n ∑ j = 1 n ∑ k = 1 n 2 a a 2 + ( X ( j ) 1 / m − X ( k ) 1 / m ) 2 − ∑ j = 1 n − m + 1 ∑ k = 1 n ( n m ) − 1 u j , m 4 a a 2 + ( X ( j ) − X ( k ) 1 / m ) 2 + n ∑ j = 1 n − m + 1 ∑ k = 1 n − m + 1 ( n m ) − 2 u j , m u k , m 2 a a 2 + ( X ( j ) − X ( k ) ) 2 , T^{(1)}_{n,m,a}:=\frac{1}{n}\sum_{j=1}^{n}\sum_{k=1}^{n}\frac{2a}{a^{2}+(X_{(j)}^{1/m}-X_{(k)}^{1/m})^{2}}
-\sum_{j=1}^{n-m+1}\sum_{k=1}^{n}\binom{n}{m}^{-1}u_{j,m}\frac{4a}{a^{2}+(X_{(j)}-X_{(k)}^{1/m})^{2}}
+n\sum_{j=1}^{n-m+1}\sum_{k=1}^{n-m+1}\binom{n}{m}^{-2}u_{j,m}u_{k,m}\frac{2a}{a^{2}+(X_{(j)}-X_{(k)})^{2}}, T n , m , a ( 1 ) := n 1 j = 1 ∑ n k = 1 ∑ n a 2 + ( X ( j ) 1/ m − X ( k ) 1/ m ) 2 2 a − j = 1 ∑ n − m + 1 k = 1 ∑ n ( m n ) − 1 u j , m a 2 + ( X ( j ) − X ( k ) 1/ m ) 2 4 a + n j = 1 ∑ n − m + 1 k = 1 ∑ n − m + 1 ( m n ) − 2 u j , m u k , m a 2 + ( X ( j ) − X ( k ) ) 2 2 a ,
and
T n , m , a ( 2 ) : = π a 1 n ∑ j = 1 n ∑ k = 1 n exp ( − ( X ( j ) 1 / m − X ( k ) 1 / m ) 2 4 a ) − 2 π a ∑ j = 1 n − m + 1 ∑ k = 1 n ( n m ) − 1 u j , m exp ( − ( X ( j ) − X ( k ) 1 / m ) 2 4 a ) + n π a ∑ j = 1 n − m + 1 ∑ k = 1 n − m + 1 ( n m ) − 2 u j , m u k , m exp ( − ( X ( j ) − X ( k ) ) 2 4 a ) . T^{(2)}_{n,m,a}:=\sqrt{\frac{\pi}{a}}\frac{1}{n}\sum_{j=1}^{n}\sum_{k=1}^{n}\exp\left(-\frac{(X_{(j)}^{1/m}-X_{(k)}^{1/m})^{2}}{4a}\right)
-2\sqrt{\frac{\pi}{a}}\sum_{j=1}^{n-m+1}\sum_{k=1}^{n}\binom{n}{m}^{-1}u_{j,m}\exp\left(-\frac{(X_{(j)}-X_{(k)}^{1/m})^{2}}{4a}\right)
+n\sqrt{\frac{\pi}{a}}\sum_{j=1}^{n-m+1}\sum_{k=1}^{n-m+1}\binom{n}{m}^{-2}u_{j,m}u_{k,m}\exp\left(-\frac{(X_{(j)}-X_{(k)})^{2}}{4a}\right). T n , m , a ( 2 ) := a π n 1 j = 1 ∑ n k = 1 ∑ n exp − 4 a ( X ( j ) 1/ m − X ( k ) 1/ m ) 2 − 2 a π j = 1 ∑ n − m + 1 k = 1 ∑ n ( m n ) − 1 u j , m exp − 4 a ( X ( j ) − X ( k ) 1/ m ) 2 + n a π j = 1 ∑ n − m + 1 k = 1 ∑ n − m + 1 ( m n ) − 2 u j , m u k , m exp ( − 4 a ( X ( j ) − X ( k ) ) 2 ) .
Within each pair, superscripts (1) and (2) distinguish the rational and Gaussian terms, respectively. We use m = 3 m=3 m = 3 and a = 2 a=2 a = 2 for all four statistics. Their asymptotic properties under the conditions imposed in the original work are discussed by Ndwandwe et al. [27 ] .
Tests with fitted-shape weights. The two classes proposed by Ngatchou-Wandji et al. [29 ] incorporate the fitted shape parameter through the factors X j − α ^ n ( m − 1 ) X_j^{-\widehat{\alpha}_n(m-1)} X j − α n ( m − 1 ) . To display their common structure, take a > 0 a>0 a > 0 and define the kernels
H 1 , a ( u ) : = 2 a a 2 + u 2 , H 2 , a ( u ) : = π a exp ( − u 2 4 a ) , u ∈ R . H_{1,a}(u):=\frac{2a}{a^{2}+u^{2}},\qquad H_{2,a}(u):=\sqrt{\frac{\pi}{a}}\exp\left(-\frac{u^{2}}{4a}\right),\qquad u\in\mathbb{R}. H 1 , a ( u ) := a 2 + u 2 2 a , H 2 , a ( u ) := a π exp ( − 4 a u 2 ) , u ∈ R .
The two statistics can then be written as
N n , m , a ( r ) : = 1 n ∑ j , k = 1 n [ 1 m 2 H r , a ( X j 1 / m − X k 1 / m ) − 1 m X k − α ^ n ( m − 1 ) H r , a ( X j 1 / m − X k ) − 1 m X j − α ^ n ( m − 1 ) H r , a ( X k 1 / m − X j ) + X j − α ^ n ( m − 1 ) X k − α ^ n ( m − 1 ) H r , a ( X j − X k ) ] , r = 1 , 2. N^{(r)}_{n,m,a}:=\frac{1}{n}\sum_{j,k=1}^{n}\left[\frac{1}{m^{2}}H_{r,a}\left(X_j^{1/m}-X_k^{1/m}\right)-\frac{1}{m}X_k^{-\widehat{\alpha}_n(m-1)}H_{r,a}\left(X_j^{1/m}-X_k\right)-\frac{1}{m}X_j^{-\widehat{\alpha}_n(m-1)}H_{r,a}\left(X_k^{1/m}-X_j\right)+X_j^{-\widehat{\alpha}_n(m-1)}X_k^{-\widehat{\alpha}_n(m-1)}H_{r,a}\left(X_j-X_k\right)\right],\qquad r=1,2. N n , m , a ( r ) := n 1 j , k = 1 ∑ n [ m 2 1 H r , a ( X j 1/ m − X k 1/ m ) − m 1 X k − α n ( m − 1 ) H r , a ( X j 1/ m − X k ) − m 1 X j − α n ( m − 1 ) H r , a ( X k 1/ m − X j ) + X j − α n ( m − 1 ) X k − α n ( m − 1 ) H r , a ( X j − X k ) ] , r = 1 , 2.
Following Ngatchou-Wandji et al. [29 ] , we use ( a , m ) = ( 0.5 , 2 ) (a,m)=(0.5,2) ( a , m ) = ( 0.5 , 2 ) and ( 0.5 , 3 ) (0.5,3) ( 0.5 , 3 ) for each class.
The test based on I n , m I_{n,m} I n , m rejects for large absolute values, and the test based on O n O_n O n uses both tails of its null distribution. The remaining competing tests reject for large values of their statistics.
To compare power, we draw samples from the alternatives in Table 2 , which include distributions used in earlier simulation studies of Pareto goodness-of-fit tests. Distributions originally supported on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) are shifted by one so that all alternatives have support [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) . All densities in Table 2 are zero for x < 1 x<1 x < 1 , and all parameters used in the study are positive. The notation Γ ( θ ) \Gamma(\theta) Γ ( θ ) denotes a gamma distribution with shape θ \theta θ and rate one, shifted by one. The label LogW(θ \theta θ ) is retained to match the simulation files: we generate X : = exp ( Y ) X:=\exp(Y) X := exp ( Y ) , where Y Y Y has a gamma distribution with shape and rate both equal to 1 + θ 1+\theta 1 + θ . This gives the scaled log-gamma density in Table 2 . The additional alternative F 1 , 3 F_{1,3} F 1 , 3 is defined in Proposition 1.
Table 2. Table 2: Alternative distributions and their densities on x > 1 x>1 x > 1 .
Tables 3, 4 and 5 report rejection percentages for the three sample sizes. The first three rows of each table give empirical size under the Pareto model; the remaining rows give power against the alternatives. The final three columns contain the proposed procedures. The results support the following conclusions.
(a) The simulated type-I error rates are generally near the nominal level, with visible finite-sample deviations. For example, under P ( 2 ) \mathcal{P}(2) P ( 2 ) , S n , K S_{n,K} S n , K rejects in approximately 3.9%, 4.2% and 4.8% of samples for n = 20 n=20 n = 20 , 50, 100, respectively. The fixed numerical calibration is therefore somewhat conservative in this case at smaller sample sizes.
(b) No procedure dominates across the alternatives considered.
(c) The proposed tests are competitive for several alternatives. However, T n T_n T n is not omnibus, as Proposition 1 shows. For F 1 , 3 F_{1,3} F 1 , 3 its rejection percentages remain near 5%. Although S n S_n S n is consistent under the stated assumptions, both calibrations also have low power against this alternative at the sample sizes studied.
Direct comparison of bootstrap and covariance estimation methods
For the distributions in Figure 1, we compare S n , B S_{n,B} S n , B and S n , K S_{n,K} S n , K using M = 10,000 M=10{,}000 M = 10 , 000 paired samples at each sample size n = 20 n=20 n = 20 , 35, 50, 75, 100, 150, and 200. At nominal level 5%, we use B = 500 B=500 B = 500 bootstrap samples with parameter re-estimation, 20,000 weighted chi-square draws, and the quadrature settings of Section 3.5. We used 14 parallel workers for the power experiment and measured computation times separately in a single worker. Figure 1 shows similar rejection rates, although S n , K S_{n,K} S n , K is somewhat conservative under P ( 2 ) \mathcal{P}(2) P ( 2 ) at the smaller sample sizes. The largest observed power difference is 1.42 percentage points, in favor of S n , B S_{n,B} S n , B under LogW(0.3) at n = 20 n=20 n = 20 . The computational difference is more pronounced (see Figure 2): over the same sizes studied, median covariance-calibrated test times remain near 0.04–0.05 seconds, whereas bootstrap times rise from about 0.04 to 1.45 seconds. Covariance calibration is faster from n = 35 n=35 n = 35 onward and about 28–37 times faster at n = 200 n=200 n = 200 . Thus, for these settings and alternatives, it substantially reduces computation at larger sample sizes while giving similar power.
Figure 1. Empirical size and power of S n , B S_{n,B} S n , B (blue solid lines) and S n , K S_{n,K} S n , K (orange dashed lines), based on 10,000 paired samples per setting. Upper panels: Pareto null distributions. Lower panels: alternatives from Table 2 ; LogW(0.3) denotes the scaled log-gamma alternative. Shading gives pointwise 95% Wilson Monte Carlo intervals for each rejection probability; dotted horizontal lines mark the nominal level 0.05.
Figure 2. Elapsed time per complete test under P ( α ) \mathcal{P}(\alpha) P ( α ) , measured serially on ten paired data sets per ( n , α ) (n,\alpha) ( n , α ) setting. Lines show medians and shading shows interquartile ranges on a logarithmic time scale. Timings include fitting, the observed statistic, and full calibration, but exclude generation of the observed sample and warm-up calls. Fast calls are batched for at least 0.1 seconds and averaged within each data set.
Data analysis: Danish fire-insurance losses
We consider the Danish fire-insurance losses available as danish in the R package evir ([33 ] ). The data set contains 2,167 claims recorded between 1980 and 1990. The amounts are expressed in millions of Danish kroner (DKK), adjusted to 1985 prices. These data are a standard benchmark in extreme-value analysis; see McNeil [23 ] .
Our aim is to examine how the assessment of an exact Pareto type-I model changes with the lower loss threshold. Let Y Y Y denote a recorded loss and consider the conditional hypothesis H 0 ( u ) : P ( Y > y ∣ Y > u ) = ( y / u ) − α u H_0(u): \mathbb{P}(Y > y \mid Y > u) = (y/u)^{-\alpha_u} H 0 ( u ) : P ( Y > y ∣ Y > u ) = ( y / u ) − α u for y ≥ u y \ge u y ≥ u and some α u > 0 \alpha_u > 0 α u > 0 . For each threshold u u u , we retain losses strictly exceeding u u u and normalize them as X j ( u ) = Y j / u X_j^{(u)} = Y_j/u X j ( u ) = Y j / u . The resulting testing problem has the known lower endpoint 1 1 1 , as assumed in § 2 .
To assess threshold sensitivity, we use the grid u ∈ { 1 , 1.5 , 2 , 3 , 5 , 10 } u \in\{1, 1.5, 2, 3, 5, 10\} u ∈ { 1 , 1.5 , 2 , 3 , 5 , 10 } , measured in millions of DKK, and retain only losses strictly exceeding each threshold. At u = 1 u = 1 u = 1 , excluding the 11 claims recorded exactly at the cutoff leaves n u = 2,156 n_u = 2{,}156 n u = 2 , 156 exceedances. At u = 1.5 u = 1.5 u = 1.5 , six claims equal the cutoff and 1,386 exceed it.
At each threshold, we estimate the Pareto shape by α ^ u = { n u − 1 ∑ j = 1 n u log X j ( u ) } − 1 \widehat{\alpha}_u = \left\{n_u^{-1}\sum_{j=1}^{n_u}\log X_j^{(u)}\right\}^{-1} α u = { n u − 1 ∑ j = 1 n u log X j ( u ) } − 1 . We calculate the signed Stein–Laplace statistic T n u T_{n_u} T n u and the integrated squared statistic n u S n u n_uS_{n_u} n u S n u .
The signed statistic is calibrated by a two-sided parametric bootstrap. For S n u S_{n_u} S n u , we use both the bootstrap calibration and the plug-in covariance-operator calibration developed in § 3.5 . The bootstrap uses B = 4,999 B = 4{,}999 B = 4 , 999 samples of size n u n_u n u from P ( α ^ u ) \mathcal{P}(\widehat{\alpha}_u) P ( α u ) , with shape re-estimation in every sample. The covariance-operator calculation uses the truncation interval [ 0 , 12 ] [0, 12] [ 0 , 12 ] , 50 Gauss–Legendre nodes, 80 Gauss–Laguerre nodes, and M = 200,000 M = 200{,}000 M = 200 , 000 weighted chi-square draws. Both calibrations use the same observed statistic n u S n u n_uS_{n_u} n u S n u .
For comparison, we also consider the Kolmogorov–Smirnov and Anderson–Darling statistics and the three Stein–Laplace procedures of Bhati and Khandelwal [8 ] . In terms of their empirical discrepancy R n u \mathcal{R}_{n_u} R n u , described in § 2 , these are D S 1 = ∫ 0 ∞ R n u ( t ) d t \mathcal{DS}_1 = \int_0^\infty\mathcal{R}_{n_u}(t)\,\mathrm{d}t D S 1 = ∫ 0 ∞ R n u ( t ) d t , D S 2 = sup t > 0 ∣ R n u ( t ) ∣ \mathcal{DS}_2 = \sup_{t>0}|\mathcal{R}_{n_u}(t)| D S 2 = sup t > 0 ∣ R n u ( t ) ∣ , and D S 3 = ∫ 0 ∞ R n u ( t ) 2 d t \mathcal{DS}_3 = \int_0^\infty\mathcal{R}_{n_u}(t)^2\,\mathrm{d}t D S 3 = ∫ 0 ∞ R n u ( t ) 2 d t . The implementation evaluates D S 1 \mathcal{DS}_1 D S 1 in closed form, approximates D S 2 \mathcal{DS}_2 D S 2 on a numerical grid, and evaluates D S 3 \mathcal{DS}_3 D S 3 by quadrature; the same rules are applied to the observed and bootstrap samples.
For a right-tailed statistic with observed value r r r and bootstrap values r 1 ∗ , … , r B ∗ r_1^*,\ldots,r_B^* r 1 ∗ , … , r B ∗ , we report p ^ B = ( 1 + ∑ b = 1 B 1 { r b ∗ ≥ r } ) / ( B + 1 ) \widehat{p}_B = (1+\sum_{b=1}^{B}\mathbf{1}_{\{r_b^*\ge r\}})/(B+1) p B = ( 1 + ∑ b = 1 B 1 { r b ∗ ≥ r } ) / ( B + 1 ) . The signed statistics T n u T_{n_u} T n u and D S 1 \mathcal{DS}_1 D S 1 use twice the smaller plus-one tail probability, capped at one; the remaining statistics use the upper tail. Operator-calibrated p p p -values use the analogous correction with M M M reference draws.
Table 1 reports the numbers of exceedances, estimated shape parameters, test-statistic values, and calibrated p p p -values.
Table 1. Threshold sensitivity of Pareto goodness-of-fit tests for the Danish fire-insurance losses. Thresholds are measured in millions of DKK. Bootstrap p p p -values use B = 4,999 B = 4{,}999 B = 4 , 999 replications; the operator-calibrated p p p -values use M = 200,000 M = 200{,}000 M = 200 , 000 weighted chi-square draws. All analyses use strict exceedances Y > u Y > u Y > u .
At the reporting threshold u = 1 u = 1 u = 1 , the fitted shape parameter is α ^ u = 1.264 \widehat{\alpha}_u = 1.264 α u = 1.264 , and all procedures reject the Pareto model at the 5% level. At u = 1.5 u = 1.5 u = 1.5 , neither calibration of S n u S_{n_u} S n u rejects, although 1,386 exceedances remain. The bootstrap and operator p p p -values are 0.9462 and 0.9475, respectively. The Kolmogorov–Smirnov and Anderson–Darling tests also do not reject, with p p p -values 0.5920 and 0.3634. The three related Stein–Laplace procedures reach the same conclusion.
At the remaining thresholds, neither calibration of S n u S_{n_u} S n u rejects at the nominal 5% level. The fitted shape parameter varies between approximately 1.37 and 1.42 for thresholds from 1.5 to 5 million DKK. At u = 10 u = 10 u = 10 , the estimate increases to 1.61, although the number of exceedances has declined to 109.
Across the six thresholds, bootstrap and covariance-operator calibration yield the same rejection decisions and generally similar p p p -values. Both detect a departure from the Pareto model at the reporting threshold of one million DKK, whereas neither rejects at the five higher thresholds considered. The substantive conclusions of this application are therefore
At u = 1 u = 1 u = 1 , the probabilities 0.0002 and 5.0 ⋅ 10 − 6 5.0 \cdot10^{-6} 5.0 ⋅ 1 0 − 6 are the respective plus-one Monte Carlo reporting minima for B = 4,999 B = 4{,}999 B = 4 , 999 and M = 200,000 M = 200{,}000 M = 200 , 000 ; no reference value exceeded the observed statistic. The minimum for the doubled-tail rule is 0.0004.
unchanged by the choice between the two calibration methods.
Conclusion
The central distinction between the proposed tests is the information they retain from the Stein–Laplace characterization. The signed statistic T n T_n T n reduces the discrepancy to a scalar moment restriction, yielding a simple diagnostic with an explicit studentization and a standard normal limiting null distribution. However, continuous non-Pareto distributions can satisfy the same restriction, so the resulting test is not consistent against every alternative. In contrast, squaring before integration preserves the characterizing property. The statistic S n S_n S n converges to a population discrepancy that is strictly positive for every non-Pareto distribution on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) satisfying 0 < E F [ log X ] < ∞ 0 < E_F[\log X] < \infty 0 < E F [ log X ] < ∞ . Together, these results show why the manner in which a characterization is aggregated is decisive for the consistency of the resulting test.
The covariance structure of the limiting Gaussian process provides both a description of the null law of n S n nS_n n S n and a practical alternative to bootstrap calibration. Approximating the eigenvalues of the fitted covariance operator allows critical values to be obtained from a weighted chi-square distribution without repeatedly generating and refitting Pareto samples. The direct comparison of the two calibrations demonstrates the computational value of this approach. In the reported timing experiment, covariance-operator calibration was approximately 28–37 times faster at n = 200 n = 200 n = 200 , while the paired power experiment showed similar rejection probabilities for the alternatives examined. The parametric bootstrap nevertheless remains useful in small samples, where the operator approximation was conservative in some settings. Its occasionally higher empirical power must therefore be considered together with the corresponding differences in empirical size.
The broader comparison with existing procedures does not support a universal ranking. The proposed tests are competitive against several alternatives, but established procedures are appreciably more powerful in other cases, a pattern consistent with the theoretical findings of Janssen [18 ] . The constructed non-Pareto mixture is particularly informative: it demonstrates the lack of consistency of T n T_n T n , while also exposing low finite-sample power of both calibrations of S n S_n S n .
The Danish fire-insurance application demonstrates the importance of specifying the region over which an exact Pareto model is assessed. The model is strongly rejected at the original reporting threshold of one million DKK, whereas neither calibration of S n S_n S n rejects at any of the five higher thresholds considered. In particular, the nonrejection at 1.5 million DKK occurs with 1,386 exceedances remaining. These findings are compatible with a Pareto approximation farther into the upper tail, but they neither identify a statistically validated threshold nor establish exact Pareto behavior. Developing selection-aware inference for the Stein–Laplace statistics therefore remains a natural extension of the present theory.
Several further questions arise directly from the results. The known-endpoint assumption suggests two distinct extensions: estimation of the support endpoint within an exact Pareto model, and selection of a threshold above which a Pareto approximation is adequate. Goodness-of-fit testing with an estimated endpoint has been studied by Ndwandwe et al. [28 ] , while procedures combining threshold selection and Pareto fitting include those of Grama and Spokoiny [16 ] and Clauset et al. [11 ] . The asymptotic analysis of Drees et al. [12 ] shows that threshold selection can alter the limiting distribution of the subsequent tail-index estimator. For the present Stein–Laplace tests, the corresponding task is to establish the joint limit of the selected threshold, the shape estimator, and the empirical discrepancy, and to determine when covariance-operator or bootstrap calibration remains valid. Ordered testing procedures (Bader et al. [4 ] ) and methods that propagate threshold-selection uncertainty (Murphy et al. [26 ] ) provide relevant starting points for this extension. Another interesting line of further research includes limits under local alternatives and asymptotic efficiency calculations. They would clarify which departures are weakly represented by the unweighted Stein–Laplace discrepancy. Such an analysis could guide the choice of weighted versions of S n S_n S n , with the aim of improving sensitivity while retaining consistency.
Declaration of generative AI and AI-assisted technologies in the manuscript preparation process
OpenAI’s ChatGPT 6 Astra was used as an AI-assisted research tool in the preparation of this work. In particular, the code for the simulation study was partially written by ChatGPT. All suggested proofs were checked and independently verified by the authors. ChatGPT was also used for assistance with exposition, LaTeX formulation, and consistency checks. The authors take full responsibility for all results and conclusions in the paper.
Tables
We provide the tables of alternatives and simulation results of § 4 .
Table 3. Empirical size and power (percentages, rounded to one decimal place) of the proposed and competing tests (n = 20 n = 20 n = 20 , N = 10,000 N = 10{,}000 N = 10 , 000 , B = 500 B = 500 B = 500 , η = 0.05 \eta= 0.05 η = 0.05 ). The distribution F 1 , 3 F_{1,3} F 1 , 3 is the alternative in Proposition 1 with λ = 1 \lambda= 1 λ = 1 and b = 3 b = 3 b = 3 .
Table 4. Empirical size and power (percentages, rounded to one decimal place) of the proposed and competing tests (n = 50 n = 50 n = 50 , N = 10,000 N = 10{,}000 N = 10 , 000 , B = 500 B = 500 B = 500 , η = 0.05 \eta= 0.05 η = 0.05 ). The distribution F 1 , 3 F_{1,3} F 1 , 3 is the alternative in Proposition 1 with λ = 1 \lambda= 1 λ = 1 and b = 3 b = 3 b = 3 .
Table 5. Empirical size and power (percentages, rounded to one decimal place) of the proposed and competing tests (n = 100 n = 100 n = 100 , N = 10,000 N = 10{,}000 N = 10 , 000 , B = 500 B = 500 B = 500 , η = 0.05 \eta= 0.05 η = 0.05 ). The distribution F 1 , 3 F_{1,3} F 1 , 3 is the alternative in Proposition 1 with λ = 1 \lambda= 1 λ = 1 and b = 3 b = 3 b = 3 .