📚 الرئيسية

مسابقة دكتوراه 2016Université des Sciences et de la Technologie Houari Boumediène (USTHB) — الموضوع 02

مسابقة تخصص · Probabilités & Statistiques · المدة: 2سا

Concours Doctorat de 3ème cycle en Probabilités, Statistiques et Applications, Épreuve de Statistiques, Faculté de Mathématiques, Département de Probabilités et Statistiques, Université des Sciences et de la Technologie Houari Boumediène (USTHB), durée 2 heures, le 03 octobre 2016.

التمرين 1

Exercice 1 — Loi exponentielle translatée : EMV, UVMB et intervalles de confiance

#maximum-likelihood#umvue#exponential-distribution#confidence-interval#point-estimation

Soit (x1,x2,,xn)(x_1, x_2, \dots, x_n) un nn-échantillon d'une variable aléatoire XX de loi

fX(x)=θ1eθ1(xθ2)1{xθ2},θ1>0.f_X(x) = \theta_1\, e^{-\theta_1 (x - \theta_2)}\,\mathbf{1}_{\{x \geq \theta_2\}}, \qquad \theta_1 \gt 0.

  1. On suppose θ2=0\theta_2 = 0. a. Déterminer l'estimateur du maximum de vraisemblance de θ1\theta_1. Est-il sans biais ? Sinon, calculer son biais et son erreur quadratique moyenne. b. Montrer que θ1\theta_1 admet un estimateur sans biais de variance minimale qu'on déterminera. Est-il efficace ? Justifier. c. Soit KK un nombre réel strictement positif. Déterminer l'estimateur du maximum de vraisemblance de Pθ1(X>K)P_{\theta_1}(X \gt K). Montrer que cette quantité admet un estimateur sans biais optimal, qu'on caractérisera, qui ne peut pas être efficace.

  2. Dans cette question, θ1\theta_1 et θ2\theta_2 sont inconnus. Estimer ces deux paramètres par la méthode des moments puis par la méthode du maximum de vraisemblance. (Indication : on remarquera que Xθ2X - \theta_2 suit une loi exponentielle.)

  3. On suppose θ1=1\theta_1 = 1 et on note θ2^\widehat{\theta_2} l'estimateur du maximum de vraisemblance de θ2\theta_2. a. Déterminer la loi de (θ2^θ2)(\widehat{\theta_2} - \theta_2), et en déduire que

2n(θ2^θ2)n+Lχ222n(\widehat{\theta_2} - \theta_2) \xrightarrow[n \to +\infty]{\mathcal{L}} \chi_2^2

(khi-deux à 22 degrés de liberté). Donner un intervalle de confiance de niveau (1α)(1 - \alpha) pour θ2\theta_2. b. Soient aa et bb deux réels positifs tels que P(a<θ2^θ2<b)=1αP(a \lt \widehat{\theta_2} - \theta_2 \lt b) = 1 - \alpha. En déduire un autre intervalle de confiance de niveau 1α1 - \alpha. En exprimant par exemple bb en fonction de aa, donner l'intervalle de confiance de longueur minimale pour θ2\theta_2.

الحل

1. a.

Avec θ2=0\theta_2 = 0, XX suit la loi exponentielle de densité fX(x)=θ1eθ1x1{x0}f_X(x) = \theta_1 e^{-\theta_1 x}\mathbf{1}_{\{x \geq 0\}}. La vraisemblance est

L(θ1)=i=1nθ1eθ1xi=θ1neθ1i=1nxi.L(\theta_1) = \prod_{i=1}^{n} \theta_1 e^{-\theta_1 x_i} = \theta_1^{\,n}\, e^{-\theta_1 \sum_{i=1}^n x_i}.

La log-vraisemblance (θ1)=nlnθ1θ1xi\ell(\theta_1) = n \ln \theta_1 - \theta_1 \sum x_i vérifie (θ1)=nθ1xi=0\ell'(\theta_1) = \dfrac{n}{\theta_1} - \sum x_i = 0, d'où

θ1^=ni=1nXi=1X.\boxed{\,\widehat{\theta_1} = \frac{n}{\sum_{i=1}^n X_i} = \frac{1}{\overline{X}}.\,}

Posons S=i=1nXiΓ(n,θ1)S = \sum_{i=1}^n X_i \sim \Gamma(n, \theta_1). On a E[S1]=θ1n1\mathbb{E}[S^{-1}] = \dfrac{\theta_1}{n-1} et E[S2]=θ12(n1)(n2)\mathbb{E}[S^{-2}] = \dfrac{\theta_1^2}{(n-1)(n-2)} (pour n3n \geq 3). Donc

E[θ1^]=nE[S1]=nθ1n1θ1: θ1^ est biaiseˊ.\mathbb{E}[\widehat{\theta_1}] = n\,\mathbb{E}[S^{-1}] = \frac{n\,\theta_1}{n-1} \neq \theta_1 : \ \widehat{\theta_1} \text{ est biaisé.}

Biais(θ1^)=nθ1n1θ1=θ1n1.\text{Biais}(\widehat{\theta_1}) = \frac{n \theta_1}{n-1} - \theta_1 = \frac{\theta_1}{n-1}.

Variance : Var(θ1^)=n2(E[S2]E[S1]2)=n2θ12(1(n1)(n2)1(n1)2)=n2θ12(n1)2(n2).\operatorname{Var}(\widehat{\theta_1}) = n^2\left( \mathbb{E}[S^{-2}] - \mathbb{E}[S^{-1}]^2 \right) = n^2 \theta_1^2\left( \frac{1}{(n-1)(n-2)} - \frac{1}{(n-1)^2} \right) = \frac{n^2 \theta_1^2}{(n-1)^2 (n-2)}.

Erreur quadratique moyenne :

EQM(θ1^)=Var(θ1^)+Biais2=n2θ12(n1)2(n2)+θ12(n1)2.\operatorname{EQM}(\widehat{\theta_1}) = \operatorname{Var}(\widehat{\theta_1}) + \text{Biais}^2 = \frac{n^2 \theta_1^2}{(n-1)^2(n-2)} + \frac{\theta_1^2}{(n-1)^2}.

EQM(θ1^)=θ12(n+2)(n1)(n2).\boxed{\,\operatorname{EQM}(\widehat{\theta_1}) = \frac{\theta_1^2 (n+2)}{(n-1)(n-2)}.\,}

1. b.

Comme E ⁣[n1S]=(n1)θ1n1=θ1\mathbb{E}\!\left[\dfrac{n-1}{S}\right] = (n-1)\dfrac{\theta_1}{n-1} = \theta_1, l'estimateur

θ1=n1i=1nXi\theta_1^{*} = \frac{n-1}{\sum_{i=1}^n X_i}

est sans biais. La loi appartient à la famille exponentielle et S=XiS = \sum X_i est une statistique exhaustive et complète. Comme θ1\theta_1^{*} est une fonction de SS sans biais, le théorème de Lehmann-Scheffé assure que c'est l'unique estimateur sans biais de variance minimale (UVMB).

Var(θ1)=(n1)2(θ12(n1)(n2)θ12(n1)2)=θ12n2.\operatorname{Var}(\theta_1^{*}) = (n-1)^2\left( \frac{\theta_1^2}{(n-1)(n-2)} - \frac{\theta_1^2}{(n-1)^2} \right) = \frac{\theta_1^2}{n-2}.

Information de Fisher : pour une observation I1(θ1)=1θ12I_1(\theta_1) = \dfrac{1}{\theta_1^2}, donc In(θ1)=nθ12I_n(\theta_1) = \dfrac{n}{\theta_1^2} et la borne de Fréchet-Darmois-Cramér-Rao vaut 1In(θ1)=θ12n\dfrac{1}{I_n(\theta_1)} = \dfrac{\theta_1^2}{n}. Or

Var(θ1)=θ12n2>θ12n,\operatorname{Var}(\theta_1^{*}) = \frac{\theta_1^2}{n-2} \gt \frac{\theta_1^2}{n},

donc θ1\theta_1^{*} n'est pas efficace (il n'atteint pas la borne FDCR), bien qu'il soit optimal parmi les estimateurs sans biais.

θ1=n1i=1nXi,Var(θ1)=θ12n2>θ12n=borne FDCR.\boxed{\,\theta_1^{*} = \frac{n-1}{\sum_{i=1}^n X_i}, \qquad \operatorname{Var}(\theta_1^{*}) = \frac{\theta_1^2}{n-2} \gt \frac{\theta_1^2}{n} = \text{borne FDCR}.\,}

1. c.

On a Pθ1(X>K)=K+θ1eθ1xdx=eθ1KP_{\theta_1}(X \gt K) = \displaystyle\int_K^{+\infty} \theta_1 e^{-\theta_1 x}\,dx = e^{-\theta_1 K}. Par le principe d'invariance de l'EMV,

Pθ1(X>K)^=eθ1^K=enK/i=1nXi.\widehat{P_{\theta_1}(X \gt K)} = e^{-\widehat{\theta_1} K} = e^{-\,nK / \sum_{i=1}^n X_i}.

Pour l'estimateur sans biais optimal, on part de l'estimateur sans biais élémentaire 1{X1>K}\mathbf{1}_{\{X_1 \gt K\}} (car E[1{X1>K}]=eθ1K\mathbb{E}[\mathbf{1}_{\{X_1 \gt K\}}] = e^{-\theta_1 K}) et on le conditionne par la statistique exhaustive complète S=XiS = \sum X_i (Rao-Blackwell / Lehmann-Scheffé). Pour des exponentielles i.i.d., X1SBeta(1,n1)\dfrac{X_1}{S} \sim \text{Beta}(1, n-1), de densité (n1)(1u)n2(n-1)(1-u)^{n-2} sur (0,1)(0,1). Donc, pour S>KS \gt K,

P(X1>KS)=P ⁣(X1S>KS)=K/S1(n1)(1u)n2du=(1KS)n1.P(X_1 \gt K \mid S) = P\!\left( \frac{X_1}{S} \gt \frac{K}{S} \right) = \int_{K/S}^{1} (n-1)(1-u)^{n-2}\,du = \left( 1 - \frac{K}{S} \right)^{n-1}.

ψ=(1Ki=1nXi)n11{Xi>K} est l’estimateur sans biais optimal (UVMB) de eθ1K.\boxed{\,\psi^{*} = \left( 1 - \frac{K}{\sum_{i=1}^n X_i} \right)^{n-1}\mathbf{1}_{\{\sum X_i \gt K\}} \ \text{est l'estimateur sans biais optimal (UVMB) de } e^{-\theta_1 K}.\,}

Il ne peut pas être efficace : un estimateur sans biais atteint la borne FDCR si et seulement s'il est une fonction affine du score, c'est-à-dire de la forme g(θ1)+c(θ1)(SE[S])g(\theta_1) + c(\theta_1)\big(S - \mathbb{E}[S]\big). Or ψ=(1K/S)n1\psi^{*} = (1 - K/S)^{n-1} est une fonction non affine de SS ; l'égalité de Cauchy-Schwarz dans l'inégalité FDCR ne peut donc pas être atteinte pour eθ1Ke^{-\theta_1 K}.

2.

Car Xθ2E(θ1)X - \theta_2 \sim \mathcal{E}(\theta_1), on a E[X]=θ2+1θ1\mathbb{E}[X] = \theta_2 + \dfrac{1}{\theta_1} et Var(X)=1θ12\operatorname{Var}(X) = \dfrac{1}{\theta_1^2}.

Méthode des moments. On égale les moments théoriques aux moments empiriques :

X=θ2+1θ1,Sn2=1ni=1n(XiX)2=1θ12.\overline{X} = \theta_2 + \frac{1}{\theta_1}, \qquad S_n^2 = \frac{1}{n}\sum_{i=1}^n (X_i - \overline{X})^2 = \frac{1}{\theta_1^2}.

D'où 1θ1=Sn\dfrac{1}{\theta_1} = S_n, puis

θ1~=1Sn,θ2~=XSn.\boxed{\,\widetilde{\theta_1} = \frac{1}{S_n}, \qquad \widetilde{\theta_2} = \overline{X} - S_n.\,}

Méthode du maximum de vraisemblance. La vraisemblance est

L(θ1,θ2)=θ1neθ1i=1n(xiθ2)1{θ2x(1)},x(1)=minixi.L(\theta_1, \theta_2) = \theta_1^{\,n}\, e^{-\theta_1 \sum_{i=1}^n (x_i - \theta_2)}\,\mathbf{1}_{\{\theta_2 \leq x_{(1)}\}}, \qquad x_{(1)} = \min_i x_i.

À θ1>0\theta_1 \gt 0 fixé, LL est croissante en θ2\theta_2 (le facteur enθ1θ2e^{n \theta_1 \theta_2} croît) sous la contrainte θ2x(1)\theta_2 \leq x_{(1)}, donc θ2^=x(1)\widehat{\theta_2} = x_{(1)}. En remplaçant puis en annulant θ1\partial_{\theta_1} \ell :

θ2^=X(1)=miniXi,θ1^=ni=1nXinX(1)=1XX(1).\boxed{\,\widehat{\theta_2} = X_{(1)} = \min_i X_i, \qquad \widehat{\theta_1} = \frac{n}{\sum_{i=1}^n X_i - n X_{(1)}} = \frac{1}{\overline{X} - X_{(1)}}.\,}

3. a.

Avec θ1=1\theta_1 = 1, les Xiθ2E(1)X_i - \theta_2 \sim \mathcal{E}(1) sont i.i.d. et θ2^=X(1)\widehat{\theta_2} = X_{(1)}. Alors

θ2^θ2=mini(Xiθ2)E(n),\widehat{\theta_2} - \theta_2 = \min_i (X_i - \theta_2) \sim \mathcal{E}(n),

car le minimum de nn exponentielles E(1)\mathcal{E}(1) indépendantes est une E(n)\mathcal{E}(n) (densité nent1{t0}n e^{-n t}\mathbf{1}_{\{t \geq 0\}}). Ainsi n(θ2^θ2)E(1)n(\widehat{\theta_2} - \theta_2) \sim \mathcal{E}(1) et

2n(θ2^θ2)E ⁣(12)=χ22(loi exacte, pour tout n).2n(\widehat{\theta_2} - \theta_2) \sim \mathcal{E}\!\left(\tfrac{1}{2}\right) = \chi_2^2 \quad (\text{loi exacte, pour tout } n).

En effet la χ22\chi_2^2 est exactement l'exponentielle de moyenne 22. Soient qα/2q_{\alpha/2} et q1α/2q_{1-\alpha/2} les quantiles d'ordre α/2\alpha/2 et 1α/21 - \alpha/2 de la loi χ22\chi_2^2 ; alors P(qα/22n(θ2^θ2)q1α/2)=1αP\big(q_{\alpha/2} \leq 2n(\widehat{\theta_2} - \theta_2) \leq q_{1-\alpha/2}\big) = 1 - \alpha, ce qui donne (en notant θ2^θ2\widehat{\theta_2} \geq \theta_2) :

θ2[θ2^q1α/22n, θ2^qα/22n].\boxed{\,\theta_2 \in \left[\, \widehat{\theta_2} - \frac{q_{1-\alpha/2}}{2n},\ \widehat{\theta_2} - \frac{q_{\alpha/2}}{2n} \,\right].\,}

Comme qβ=2ln(1β)q_{\beta} = -2\ln(1 - \beta) pour la χ22\chi_2^2, l'intervalle s'écrit explicitement

θ2[θ2^+ln(α/2)n, θ2^+ln(1α/2)n].\theta_2 \in \left[\, \widehat{\theta_2} + \frac{\ln(\alpha/2)}{n},\ \widehat{\theta_2} + \frac{\ln(1 - \alpha/2)}{n} \,\right].

3. b.

Comme θ2^θ2E(n)\widehat{\theta_2} - \theta_2 \sim \mathcal{E}(n), la condition P(a<θ2^θ2<b)=1αP(a \lt \widehat{\theta_2} - \theta_2 \lt b) = 1 - \alpha s'écrit

enaenb=1α,e^{-na} - e^{-nb} = 1 - \alpha,

et l'événement a<θ2^θ2<ba \lt \widehat{\theta_2} - \theta_2 \lt b équivaut à θ2^b<θ2<θ2^a\widehat{\theta_2} - b \lt \theta_2 \lt \widehat{\theta_2} - a. D'où l'intervalle de confiance

θ2(θ2^b, θ2^a),de longueur =ba.\theta_2 \in \big(\, \widehat{\theta_2} - b,\ \widehat{\theta_2} - a \,\big), \qquad \text{de longueur } \ell = b - a.

On exprime bb en fonction de aa : b=1nln ⁣(ena(1α))b = -\dfrac{1}{n}\ln\!\big(e^{-na} - (1 - \alpha)\big). Alors

dda=dbda1=enaena(1α)1=1αena(1α)>0,\frac{d\ell}{da} = \frac{db}{da} - 1 = \frac{e^{-na}}{e^{-na} - (1 - \alpha)} - 1 = \frac{1 - \alpha}{e^{-na} - (1 - \alpha)} \gt 0,

donc \ell est strictement croissante en aa : la longueur est minimale pour la plus petite valeur admissible a=0a = 0, ce qui donne enb=αe^{-nb} = \alpha, soit b=1nln1αb = \dfrac{1}{n}\ln\dfrac{1}{\alpha}.

θ2(θ2^1nln1α, θ2^],longueur minimale min=1nln1α.\boxed{\,\theta_2 \in \left(\, \widehat{\theta_2} - \frac{1}{n}\ln\frac{1}{\alpha},\ \widehat{\theta_2} \,\right], \qquad \text{longueur minimale } \ell_{\min} = \frac{1}{n}\ln\frac{1}{\alpha}.\,}

Cet intervalle unilatéral est naturel puisque θ2θ2^\theta_2 \leq \widehat{\theta_2} presque sûrement.

التمرين 2

Exercice 2 — Modèle gaussien et tests les plus puissants (Neyman-Pearson)

#hypothesis-testing#neyman-pearson#gaussian-model#most-powerful-test#statistical-power

Soit X1,X2,,XnX_1, X_2, \dots, X_n des variables aléatoires indépendantes telles que XiN(iθ,1)X_i \sim \mathcal{N}(i\theta, 1) pour 1in1 \leq i \leq n, où θ\theta est un paramètre réel. On dispose d'une observation xix_i de chacune de ces variables.

  1. Déterminer la loi de i=1niXi\displaystyle\sum_{i=1}^n i X_i.

  2. On veut tester l'hypothèse "θ=θ0\theta = \theta_0" contre l'alternative "θ=θ1\theta = \theta_1", avec θ0<θ1\theta_0 \lt \theta_1. Montrer que les tests les plus puissants ont une région critique de la forme C={i=1nixi>K}C = \left\{ \displaystyle\sum_{i=1}^n i x_i \gt K \right\}. Déterminer alors le test le plus puissant au niveau α\alpha et calculer sa puissance.

  3. Pour tester l'hypothèse "θ=θ0\theta = \theta_0" contre l'alternative "θθ0\theta \neq \theta_0", on considère le test de région critique C1={i=1nixiθ0>K1}C_1 = \left\{ \left| \displaystyle\sum_{i=1}^n i x_i - \theta_0 \right| \gt K_1 \right\}. Déterminer K1K_1 pour qu'il soit de niveau α\alpha.

الحل

On pose T=i=1niXiT = \displaystyle\sum_{i=1}^n i X_i et S2=i=1ni2=n(n+1)(2n+1)6S_2 = \displaystyle\sum_{i=1}^n i^2 = \frac{n(n+1)(2n+1)}{6}.

1.

TT est une combinaison linéaire de variables normales indépendantes, donc reste normale, avec

E[T]=i=1niE[Xi]=i=1niiθ=θi=1ni2=θS2,\mathbb{E}[T] = \sum_{i=1}^n i\,\mathbb{E}[X_i] = \sum_{i=1}^n i \cdot i\theta = \theta \sum_{i=1}^n i^2 = \theta\, S_2,

Var(T)=i=1ni2Var(Xi)=i=1ni2=S2.\operatorname{Var}(T) = \sum_{i=1}^n i^2 \operatorname{Var}(X_i) = \sum_{i=1}^n i^2 = S_2.

T=i=1niXiN(θS2, S2),S2=n(n+1)(2n+1)6.\boxed{\,T = \sum_{i=1}^n i X_i \sim \mathcal{N}\big(\theta\, S_2,\ S_2\big), \qquad S_2 = \frac{n(n+1)(2n+1)}{6}.\,}

2.

La densité conjointe est fθ(x)exp ⁣(12i=1n(xiiθ)2)f_\theta(x) \propto \exp\!\left( -\tfrac{1}{2}\sum_{i=1}^n (x_i - i\theta)^2 \right). Le rapport de vraisemblance de Neyman-Pearson vaut

fθ1(x)fθ0(x)=exp ⁣(12i(xiiθ1)2+12i(xiiθ0)2)=exp ⁣((θ1θ0)iixi12(θ12θ02)S2).\frac{f_{\theta_1}(x)}{f_{\theta_0}(x)} = \exp\!\left( -\tfrac{1}{2}\sum_i (x_i - i\theta_1)^2 + \tfrac{1}{2}\sum_i (x_i - i\theta_0)^2 \right) = \exp\!\left( (\theta_1 - \theta_0)\sum_i i x_i - \tfrac{1}{2}(\theta_1^2 - \theta_0^2) S_2 \right).

Comme θ1θ0>0\theta_1 - \theta_0 \gt 0, ce rapport est une fonction strictement croissante de T=ixiT = \sum i x_i. D'après le lemme de Neyman-Pearson, le test le plus puissant rejette H0H_0 pour les grandes valeurs du rapport, c'est-à-dire

C={i=1nixi>K}.C = \left\{ \sum_{i=1}^n i x_i \gt K \right\}.

Détermination de KK (niveau α\alpha). Sous H0H_0, TN(θ0S2,S2)T \sim \mathcal{N}(\theta_0 S_2, S_2), donc

α=Pθ0(T>K)=1Φ ⁣(Kθ0S2S2)  K=θ0S2+z1αS2,\alpha = P_{\theta_0}(T \gt K) = 1 - \Phi\!\left( \frac{K - \theta_0 S_2}{\sqrt{S_2}} \right) \ \Longrightarrow\ \boxed{\,K = \theta_0 S_2 + z_{1-\alpha}\sqrt{S_2},\,}

z1α=Φ1(1α)z_{1-\alpha} = \Phi^{-1}(1 - \alpha). Le test le plus puissant rejette H0H_0 si Tθ0S2S2>z1α\dfrac{T - \theta_0 S_2}{\sqrt{S_2}} \gt z_{1-\alpha}.

Puissance. Sous H1H_1, TN(θ1S2,S2)T \sim \mathcal{N}(\theta_1 S_2, S_2), donc

π=Pθ1(T>K)=1Φ ⁣(Kθ1S2S2)=1Φ ⁣(z1α(θ1θ0)S2).\pi = P_{\theta_1}(T \gt K) = 1 - \Phi\!\left( \frac{K - \theta_1 S_2}{\sqrt{S_2}} \right) = 1 - \Phi\!\big( z_{1-\alpha} - (\theta_1 - \theta_0)\sqrt{S_2} \big).

π=Φ ⁣((θ1θ0)S2z1α),S2=n(n+1)(2n+1)6.\boxed{\,\pi = \Phi\!\big( (\theta_1 - \theta_0)\sqrt{S_2} - z_{1-\alpha} \big), \qquad \sqrt{S_2} = \sqrt{\tfrac{n(n+1)(2n+1)}{6}}.\,}

Comme S2+S_2 \to +\infty, la puissance tend vers 11 : le test est convergent.

3.

Pour l'alternative bilatérale "θθ0\theta \neq \theta_0", on centre la statistique sur son espérance sous H0H_0, à savoir EH0[T]=θ0S2\mathbb{E}_{H_0}[T] = \theta_0 S_2 (l'énoncé écrit θ0\theta_0 pour θ0S2\theta_0 S_2). La région critique est

C1={i=1nixiθ0S2>K1}.C_1 = \left\{ \left| \sum_{i=1}^n i x_i - \theta_0 S_2 \right| \gt K_1 \right\}.

Sous H0H_0, Tθ0S2S2N(0,1)\dfrac{T - \theta_0 S_2}{\sqrt{S_2}} \sim \mathcal{N}(0, 1), donc pour un niveau α\alpha :

α=Pθ0 ⁣(Tθ0S2>K1)=2(1Φ ⁣(K1S2))  K1S2=z1α/2.\alpha = P_{\theta_0}\!\big( |T - \theta_0 S_2| \gt K_1 \big) = 2\left( 1 - \Phi\!\left( \frac{K_1}{\sqrt{S_2}} \right) \right) \ \Longrightarrow\ \frac{K_1}{\sqrt{S_2}} = z_{1 - \alpha/2}.

K1=z1α/2S2=Φ1 ⁣(1α2)n(n+1)(2n+1)6.\boxed{\,K_1 = z_{1 - \alpha/2}\,\sqrt{S_2} = \Phi^{-1}\!\left( 1 - \tfrac{\alpha}{2} \right)\sqrt{\tfrac{n(n+1)(2n+1)}{6}}.\,}