STAT 220 | September 6, 2025

Scarlett X. Mitchell
Dr. Carol Roberts
STAT 220
Northern Valley Highland University

Statistical Analysis Practice

With Solutions

Problem 1: Consider high-dimensional regression Y = Xβ + ε with p = 929 predictors (Note:

929 is an uninteresting number.). The design matrix X has entries scaled by ζ(2) = π²/6 ≈
1.6449, and β has 17 non-zero entries. Regularization parameter relates to the 5-th catalan
number: 42. a) Analyze the LASSO estimator β̂ ᵗᵁᵔᵔᴿ and establish oracle inequalities b) Study
the Restricted Eigenvalue (RE) condition and its implications for consistency c) Derive non-
asymptotic bounds on prediction and estimation error d) Compare with Ridge regression
and Elastic Net in the context of β-min conditions e) Analyze the Dantzig selector and its
connections to linear programming f) Study screening procedures and safe rules for
computational efficiency g) Investigate connections between catalan sequences and
regularization paths h) Apply random matrix theory to analyze the sample covariance
eigenvalue spectrum

Solution:
Solution for high-dimensional statistics (p = 929):
a) LASSO oracle inequalities:
||β̂ ᵗᵁᵔᵔᴿ - β₀||_2^2 ≤ C_1 sλ_n^2/n with probability 1-δ
where s = 17, λ_n ~ √(log p/n), C_1 depends on RE constant
b) Restricted Eigenvalue condition:
RE(κ,s) = min_{||δ||_0≤s, ||δ_{S^c}||_1≤κ||δ_S||_1} ||δ||_2^2/||Xδ||_2^2
If RE(3,s) > 0, then LASSO achieves oracle rate O(s log p/n)
c) Non-asymptotic bounds:
Prediction: ||X(β̂ ᵗᵁᵔᵔᴿ - β₀)||_2^2 = O_p(s λ_n^2)
Estimation: ||β̂ ᵗᵁᵔᵔᴿ - β₀||_2^2 = O_p(s λ_n^2/φ_0^2) where φ_0 = RE constant
d) Ridge vs Elastic Net comparison:
Ridge: bias increases with signal strength, doesn't perform variable selection
Elastic Net: α||β||_1 + (1-α)||β||_2^2 bridges LASSO and Ridge
β-min condition: min_{j∈S}|β_j| ≥ Cλ_n ensures variable selection consistency
e) Dantzig selector:
min ||β||_1 s.t. ||X'(Y-Xβ)||_∞ ≤ λ_n
Equivalent to linear program, achieves same oracle rates as LASSO
f) Screening and safe rules:
Strong rules: discard j if |x_j'r^{old}| < 2λ_n - λ_n^{old}

Page of
 STAT 220 | September 6, 2025

Safe rules: KKT conditions guarantee β_j = 0 if |x_j'y| < λ_n

g) Catalan regularization paths:
λ sequence follows catalan growth: λ_k = 42 × γ^k
Solution path complexity bounded by catalan recurrence properties
h) Random matrix theory with scaling 1.6449:
Sample covariance eigenvalues follow Marchenko-Pastur law when p/n → c
Phase transition at c = 1, bulk eigenvalues in [-0.283, 2.283]
This demonstrates cutting-edge high-dimensional statistical theory.
Note: I used the method from homework 6

Problem 2: Let X₁, X₂, ..., X_{57} be i.i.d. random variables from N(μ, σ²) arising in lucas
sequence analysis (Note: 57 is the sixteenth discrete semiprime and the sixth in the (3.q)
family.). The population mean relates to e ≈ 2.7183, and we test H₀: μ = 9 in modular
arithmetic base p = 3061. a) Derive the joint sufficient statistic (T₁, T₂) for (μ, σ²) using the
exponential family factorization theorem b) Prove T₁ and T₂ form a complete sufficient
statistic using the uniqueness of moment generating functions c) Find the UMVUE for μ and
σ² using the Lehmann-Scheffé theorem and Rao-Blackwell theorem d) Construct a 99%
confidence interval for μ incorporating e scaling e) Derive the likelihood ratio test statistic
and its asymptotic distribution under local alternatives f) Analyze the power function β(μ)
and compute the sample size for achieving power 0.8 g) Study connections between lucas
growth rates and statistical efficiency bounds h) Investigate how modular arithmetic affects
the Fisher information matrix

Solution:
Solution for enhanced mathematical statistics with lucas context:
a) Joint sufficient statistic in exponential family:
L(μ, σ²) = (2πσ²)^{-57/2} exp(-∑(xᵢ - μ)²/(2σ²))
= exp([μ/σ²]∑xᵢ - [1/(2σ²)]∑xᵢ² - [nμ²/(2σ²)] - [n/2]log(2πσ²))
Natural parameters: η₁ = μ/σ², η₂ = -1/(2σ²)
Sufficient statistics: T₁ = ∑Xᵢ, T₂ = ∑Xᵢ² (jointly sufficient)
b) Completeness proof via MGF uniqueness:
E[g(T₁,T₂)] = 0 for all θ ⟹ g(T₁,T₂) = 0 a.s.
The exponential family has full rank, ensuring completeness
c) UMVUE by Lehmann-Scheffé theorem:
- μ̂ = X̄ = T₁/57 (unbiased, sufficient, complete ⟹ UMVUE)
- σ̂ ² = S² = (T₂ - 57X̄ ²)/(57-1) (unbiased for σ²)
Var(X̄ ) = σ²/57, Var(S²) = 2σ⁴/(57-1)
d) Enhanced confidence interval with e scaling:

Page of
 STAT 220 | September 6, 2025

(X̄ - t_{0.005,56}·S/√57, X̄ + t_{0.005,56}·S/√57)

Adjusted for e ≈ 2.7183: width × 2.72
e) Likelihood ratio test statistic:
Λ = sup_{H₀}L(θ)/sup_{Θ}L(θ) = [(S²+(X̄ -9)²)/(S²)]^{-57/2}
-2log Λ = 57log(1 + (X̄ -9)²/S²) →ᵈ χ²₁ under H₀
f) Power function and sample size:
β(μ) = P(|T| > t_{0.005,56} | μ) where T ~ t_{56}(δ)
Noncentrality: δ = (μ-9)/(σ/√57)
For power 0.8: n ≥ (t_{0.005} + t_{0.2})²σ²/(μ-9)²
g) Lucas efficiency connection:
Growth rate 521 relates to Cramér-Rao lower bound efficiency
Fisher information: I(μ,σ²) = diag(57/σ², 57/(2σ⁴))
h) Modular arithmetic Fisher information (mod 3061):
Information matrix properties preserved under finite field operations
Efficiency bounds modified by characteristic 3061
This demonstrates advanced mathematical statistics with unique interdisciplinary
connections.

Problem 3: Let X₁, X₂, ..., X_{67} be i.i.d. random variables from N(μ, σ²) arising in fibonacci
sequence analysis (Note: 67 is the smallest number which is palindromic in bases 5 and 6.).
The population mean relates to ζ(2) = π²/6 ≈ 1.6449, and we test H₀: μ = 8 in modular
arithmetic base p = 1069. a) Derive the joint sufficient statistic (T₁, T₂) for (μ, σ²) using the
exponential family factorization theorem b) Prove T₁ and T₂ form a complete sufficient
statistic using the uniqueness of moment generating functions c) Find the UMVUE for μ and
σ² using the Lehmann-Scheffé theorem and Rao-Blackwell theorem d) Construct a 99%
confidence interval for μ incorporating ζ(2) = π²/6 scaling e) Derive the likelihood ratio test
statistic and its asymptotic distribution under local alternatives f) Analyze the power
function β(μ) and compute the sample size for achieving power 0.8 g) Study connections
between fibonacci growth rates and statistical efficiency bounds h) Investigate how modular
arithmetic affects the Fisher information matrix

Solution:
Solution for enhanced mathematical statistics with fibonacci context:
a) Joint sufficient statistic in exponential family:
L(μ, σ²) = (2πσ²)^{-67/2} exp(-∑(xᵢ - μ)²/(2σ²))
= exp([μ/σ²]∑xᵢ - [1/(2σ²)]∑xᵢ² - [nμ²/(2σ²)] - [n/2]log(2πσ²))
Natural parameters: η₁ = μ/σ², η₂ = -1/(2σ²)
Sufficient statistics: T₁ = ∑Xᵢ, T₂ = ∑Xᵢ² (jointly sufficient)

Page of
 STAT 220 | September 6, 2025

b) Completeness proof via MGF uniqueness:

E[g(T₁,T₂)] = 0 for all θ ⟹ g(T₁,T₂) = 0 a.s.
The exponential family has full rank, ensuring completeness
c) UMVUE by Lehmann-Scheffé theorem:
- μ̂ = X̄ = T₁/67 (unbiased, sufficient, complete ⟹ UMVUE)
- σ̂ ² = S² = (T₂ - 67X̄ ²)/(67-1) (unbiased for σ²)
Var(X̄ ) = σ²/67, Var(S²) = 2σ⁴/(67-1)
d) Enhanced confidence interval with ζ(2) = π²/6 scaling:
(X̄ - t_{0.005,66}·S/√67, X̄ + t_{0.005,66}·S/√67)
Adjusted for ζ(2) = π²/6 ≈ 1.6449: width × 1.64
e) Likelihood ratio test statistic:
Λ = sup_{H₀}L(θ)/sup_{Θ}L(θ) = [(S²+(X̄ -8)²)/(S²)]^{-67/2}
-2log Λ = 67log(1 + (X̄ -8)²/S²) →ᵈ χ²₁ under H₀
f) Power function and sample size:
β(μ) = P(|T| > t_{0.005,66} | μ) where T ~ t_{66}(δ)
Noncentrality: δ = (μ-8)/(σ/√67)
For power 0.8: n ≥ (t_{0.005} + t_{0.2})²σ²/(μ-8)²
g) Fibonacci efficiency connection:
Growth rate 377 relates to Cramér-Rao lower bound efficiency
Fisher information: I(μ,σ²) = diag(67/σ², 67/(2σ⁴))
h) Modular arithmetic Fisher information (mod 1069):
Information matrix properties preserved under finite field operations
Efficiency bounds modified by characteristic 1069
This demonstrates advanced mathematical statistics with unique interdisciplinary
connections.
Note: This is a classic application of chain rule

Problem 4: Consider high-dimensional regression Y = Xβ + ε with p = 892 predictors (Note:

892 is the smallest integer ratio of a 13-digit number to its product of digits.). The design
matrix X has entries scaled by γ (Euler-Mascheroni) ≈ 0.5772, and β has 19 non-zero entries.
Regularization parameter relates to the 8-th catalan number: 1430. a) Analyze the LASSO
estimator β̂ ᵗᵁᵔᵔᴿ and establish oracle inequalities b) Study the Restricted Eigenvalue (RE)
condition and its implications for consistency c) Derive non-asymptotic bounds on prediction
and estimation error d) Compare with Ridge regression and Elastic Net in the context of β-
min conditions e) Analyze the Dantzig selector and its connections to linear programming f)
Study screening procedures and safe rules for computational efficiency g) Investigate

Page of
 STAT 220 | September 6, 2025

connections between catalan sequences and regularization paths h) Apply random matrix
theory to analyze the sample covariance eigenvalue spectrum

Solution:
Solution for high-dimensional statistics (p = 892):
a) LASSO oracle inequalities:
||β̂ ᵗᵁᵔᵔᴿ - β₀||_2^2 ≤ C_1 sλ_n^2/n with probability 1-δ
where s = 19, λ_n ~ √(log p/n), C_1 depends on RE constant
b) Restricted Eigenvalue condition:
RE(κ,s) = min_{||δ||_0≤s, ||δ_{S^c}||_1≤κ||δ_S||_1} ||δ||_2^2/||Xδ||_2^2
If RE(3,s) > 0, then LASSO achieves oracle rate O(s log p/n)
c) Non-asymptotic bounds:
Prediction: ||X(β̂ ᵗᵁᵔᵔᴿ - β₀)||_2^2 = O_p(s λ_n^2)
Estimation: ||β̂ ᵗᵁᵔᵔᴿ - β₀||_2^2 = O_p(s λ_n^2/φ_0^2) where φ_0 = RE constant
d) Ridge vs Elastic Net comparison:
Ridge: bias increases with signal strength, doesn't perform variable selection
Elastic Net: α||β||_1 + (1-α)||β||_2^2 bridges LASSO and Ridge
β-min condition: min_{j∈S}|β_j| ≥ Cλ_n ensures variable selection consistency
e) Dantzig selector:
min ||β||_1 s.t. ||X'(Y-Xβ)||_∞ ≤ λ_n
Equivalent to linear program, achieves same oracle rates as LASSO
f) Screening and safe rules:
Strong rules: discard j if |x_j'r^{old}| < 2λ_n - λ_n^{old}
Safe rules: KKT conditions guarantee β_j = 0 if |x_j'y| < λ_n
g) Catalan regularization paths:
λ sequence follows catalan growth: λ_k = 1430 × γ^k
Solution path complexity bounded by catalan recurrence properties
h) Random matrix theory with scaling 0.5772:
Sample covariance eigenvalues follow Marchenko-Pastur law when p/n → c
Phase transition at c = 1, bulk eigenvalues in [0.240, 1.760]
This demonstrates cutting-edge high-dimensional statistical theory.

Problem 5: Let X₁, X₂, ..., X_{69} be i.i.d. random variables from N(μ, σ²) arising in fibonacci
sequence analysis (Note: 69 is a value of n where n^{2} and n^{3} together contain each
digit once.). The population mean relates to ζ(2) = π²/6 ≈ 1.6449, and we test H₀: μ = 8 in
modular arithmetic base p = 2063. a) Derive the joint sufficient statistic (T₁, T₂) for (μ, σ²)

Page of
 STAT 220 | September 6, 2025

using the exponential family factorization theorem b) Prove T₁ and T₂ form a complete
sufficient statistic using the uniqueness of moment generating functions c) Find the UMVUE
for μ and σ² using the Lehmann-Scheffé theorem and Rao-Blackwell theorem d) Construct a
95% confidence interval for μ incorporating ζ(2) = π²/6 scaling e) Derive the likelihood ratio
test statistic and its asymptotic distribution under local alternatives f) Analyze the power
function β(μ) and compute the sample size for achieving power 0.8 g) Study connections
between fibonacci growth rates and statistical efficiency bounds h) Investigate how modular
arithmetic affects the Fisher information matrix

Solution:
Solution for enhanced mathematical statistics with fibonacci context:
a) Joint sufficient statistic in exponential family:
L(μ, σ²) = (2πσ²)^{-69/2} exp(-∑(xᵢ - μ)²/(2σ²))
= exp([μ/σ²]∑xᵢ - [1/(2σ²)]∑xᵢ² - [nμ²/(2σ²)] - [n/2]log(2πσ²))
Natural parameters: η₁ = μ/σ², η₂ = -1/(2σ²)
Sufficient statistics: T₁ = ∑Xᵢ, T₂ = ∑Xᵢ² (jointly sufficient)
b) Completeness proof via MGF uniqueness:
E[g(T₁,T₂)] = 0 for all θ ⟹ g(T₁,T₂) = 0 a.s.
The exponential family has full rank, ensuring completeness
c) UMVUE by Lehmann-Scheffé theorem:
- μ̂ = X̄ = T₁/69 (unbiased, sufficient, complete ⟹ UMVUE)
- σ̂ ² = S² = (T₂ - 69X̄ ²)/(69-1) (unbiased for σ²)
Var(X̄ ) = σ²/69, Var(S²) = 2σ⁴/(69-1)
d) Enhanced confidence interval with ζ(2) = π²/6 scaling:
(X̄ - t_{0.025,68}·S/√69, X̄ + t_{0.025,68}·S/√69)
Adjusted for ζ(2) = π²/6 ≈ 1.6449: width × 1.64
e) Likelihood ratio test statistic:
Λ = sup_{H₀}L(θ)/sup_{Θ}L(θ) = [(S²+(X̄ -8)²)/(S²)]^{-69/2}
-2log Λ = 69log(1 + (X̄ -8)²/S²) →ᵈ χ²₁ under H₀
f) Power function and sample size:
β(μ) = P(|T| > t_{0.025,68} | μ) where T ~ t_{68}(δ)
Noncentrality: δ = (μ-8)/(σ/√69)
For power 0.8: n ≥ (t_{0.025} + t_{0.2})²σ²/(μ-8)²
g) Fibonacci efficiency connection:
Growth rate 8 relates to Cramér-Rao lower bound efficiency
Fisher information: I(μ,σ²) = diag(69/σ², 69/(2σ⁴))
h) Modular arithmetic Fisher information (mod 2063):
Information matrix properties preserved under finite field operations

Page of
 STAT 220 | September 6, 2025

Efficiency bounds modified by characteristic 2063

This demonstrates advanced mathematical statistics with unique interdisciplinary
connections.

Problem 6: Consider extreme value analysis with block maxima of size 25 (Note: 25 is the
smallest square that can be written as a sum of 2 squares.). Threshold selection at 90%
quantile with shape parameter influenced by e ≈ 2.7183. Return levels relate to 9-th catalan
number: 4862. a) Derive the Generalized Extreme Value (GEV) distribution and its domain
of attraction b) Establish the Fisher-Tippett-Gnedenko theorem for block maxima
convergence c) Implement the Peaks-Over-Threshold (POT) method with Generalized Pareto
Distribution d) Estimate the shape, scale, and location parameters using Maximum
Likelihood e) Construct confidence intervals for high quantiles and return levels f) Apply the
Delta method for uncertainty quantification in return level estimation g) Study the
connection between catalan extremal indices and clustering h) Analyze how e affects tail
dependence in multivariate extremes

Solution:
Solution for extreme value theory:
a) GEV distribution and domain of attraction:
F_ξ(x) = exp(-(1+ξ(x-μ)/σ)_+^{-1/ξ}) for ξ ≠ 0
Domain of attraction: Type I (Gumbel, ξ=0), Type II (Fréchet, ξ>0), Type III (Weibull, ξ<0)
b) Fisher-Tippett-Gnedenko theorem:
If M_n = max(X_1,...,X_n) and P(a_n^{-1}(M_n-b_n) ≤ x) → G(x), then G is GEV
Normalizing constants: a_n = F^{-1}(1-1/n) - F^{-1}(1-2/n), b_n = F^{-1}(1-1/n)
c) POT method with GPD:
For X > u (threshold 0.9 quantile): F_u(x) = 1 - (1+ξ x/σ_u)_+^{-1/ξ}
Scale parameter: σ_u = σ + ξ(u-μ), shape ξ unchanged
d) MLE estimation:
Log-likelihood: ℓ(μ,σ,ξ) = -25logσ - (1+1/ξ)∑log(1+ξ(x_i-μ)/σ)
Profile likelihood for ξ, then conditional MLE for μ,σ
e) Confidence intervals for quantiles:
x_p = μ + (σ/ξ)[(-log(1-p))^{-ξ} - 1] for return level
Delta method: Var(x_p) ≈ ∇x_p' × I(θ)^{-1} × ∇x_p
f) Delta method uncertainty:
∇x_p = [∂x_p/∂μ, ∂x_p/∂σ, ∂x_p/∂ξ]'
Asymptotic normality: (x̂ _p - x_p)/√Var(x̂_p) →ᵈ N(0,1)
g) Catalan extremal index:
θ = lim P(M_n ≤ u_n | X_1 > u_n) where u_n = 4862/n quantile

Page of
 STAT 220 | September 6, 2025

Clustering measured by runs estimator: θ̂ = (number of clusters)/(number of exceedances)

h) Multivariate extremes with e:
Tail dependence: λ_U = lim P(Y > F_Y^{-1}(t) | X > F_X^{-1}(t))
Copula approach with scaling factor 2.7183 in tail regions
This demonstrates advanced extreme value theory with practical applications.

Problem 7: Consider a Bayesian analysis with normal likelihood and conjugate prior. a)
Specify the likelihood function L(θ|x₁, ..., x_{33}) for the normal model b) Choose an
appropriate conjugate prior π(θ) and justify your choice c) Derive the posterior distribution
π(θ|data) using Bayes' theorem d) Find the Bayesian point estimators: posterior mean,
median, and mode e) Construct a 95% credible interval for θ f) Compare with the frequentist
approach: MLE and confidence interval g) Perform posterior predictive checking and model
assessment h) Discuss the choice of loss function and its impact on Bayesian inference

Solution:
Solution for normal likelihood with conjugate prior:
a) Likelihood function:
L(μ,σ²|x) = (2πσ²)^{-n/2} exp(-∑(xᵢ-μ)²/(2σ²))
b) Prior specification:
Conjugate prior for normal: natural conjugate family
c) Posterior distribution:
π(θ|data) ∝ L(θ|data) × π(θ)
Posterior derivation for conjugate prior with normal likelihood
d) Bayesian point estimators:
Posterior mean, median, and mode for the given prior-likelihood combination
e) Credible interval:
95% credible interval using posterior quantiles
f) Frequentist comparison:
Comparison with MLE and confidence intervals for normal model
g) Posterior predictive checking:
Posterior predictive distribution and model checking procedures
h) Loss function considerations:
Analysis of squared error, absolute error, and 0-1 loss functions
This demonstrates advanced Bayesian statistical theory and computational methods.

Page of
 STAT 220 | September 6, 2025

Problem 8: Let X₁, X₂, ..., X_{48} be i.i.d. observations from N_5(μ, Σ) (Note: 5 is the second
Sierpinski number of the first kind, and can be written as S2=(22)+1.). The covariance
structure involves the matrix \begin{pmatrix}2 & -1 & -3 & 0 & -1 \\ -1 & 2 & 2 & -1 & 0 \\ -3
& 2 & 0 & -3 & 0 \\ 0 & -1 & -3 & -4 & 1 \\ -1 & 0 & 0 & 1 & 5\end{pmatrix} scaled by √2 ≈
1.4142. Analysis parameters relate to the 13-th lucas number: 521. a) Derive the MLE for μ
and Σ using matrix differential calculus and the Wishart distribution b) Establish the joint
distribution (μ̂ , Σ̂) and their asymptotic properties under regularity conditions c) Construct
Hotelling's T² statistic and Roy's union-intersection test for H₀: μ = μ₀ d) Derive the exact and
asymptotic distributions of T², including non-central F distribution e) Construct simultaneous
confidence intervals using Bonferroni and Scheffé methods f) Test sphericity H₀: Σ = σ²I using
likelihood ratio and implement Mauchly's test g) Perform PCA with eigenvalue
decomposition and analyze explained variance ratios h) Apply canonical correlation analysis
between subvectors of dimension 2 i) Study connections between lucas eigenvalue
distributions and random matrix theory

Solution:
Solution for 5-dimensional enhanced multivariate analysis:
a) MLE derivation using matrix calculus:
L(μ,Σ) = (2π)^{-485/2}|Σ|^{-48/2}exp(-1/2 ∑(xᵢ-μ)'Σ^{-1}(xᵢ-μ))
∂log L/∂μ = Σ^{-1}∑(xᵢ-μ) = 0 ⟹ μ̂ = X̄
∂log L/∂Σ^{-1} = 48/2 Σ - 1/2 ∑(xᵢ-μ)(xᵢ-μ)' = 0 ⟹ Σ̂ = S
b) Joint distribution and asymptotics:
- X̄ ~ N_5(μ, Σ/48) exactly
- 48Σ̂ ~ W_5(48-1, Σ) (Wishart distribution)
- Asymptotically: √48(μ̂ -μ) →ᵈ N_5(0, Σ)
- vec(Σ̂-Σ) →ᵈ N(0, 2Σ⊗Σ/48) by delta method
c) Test statistics and union-intersection principle:
Hotelling's T² = 48(X̄ -μ₀)'S^{-1}(X̄ -μ₀)
Roy's test: max eigenvalue of S^{-1}A where A = (X̄ -μ₀)(X̄ -μ₀)'
d) Exact and asymptotic distributions:
Under H₀: T²((48-5)/(5(48-1))) ~ F_{5,48-5}
Under H₁: T² ~ (5(48-1)/(48-5))F_{5,48-5}(δ) with δ = 48(μ-μ₀)'Σ^{-1}(μ-μ₀)
e) Simultaneous confidence intervals:
Bonferroni: P(∩|μᵢ-μ₀ᵢ| ≤ t_{α/(25),48-1}sᵢᵢ^{1/2}/√48) ≥ 1-α
Scheffé: P(|a'(μ-μ₀)| ≤ √((5(48-1)F_{5,48-5,α})/(48-5))s_{aa}/√48 ∀a) = 1-α
f) Sphericity testing:
LRT: Λ = |S|^{48/2}/(∑sᵢᵢ/5)^{548/2}
-2logΛ →ᵈ χ²_{5(5+1)/2-1} under H₀
Mauchly's test: W = |S|/(∑sᵢᵢ/5)^5 with Box's correction
g) PCA with √2 scaling:

Page of
 STAT 220 | September 6, 2025

S = PΛP' where Λ = diag(λ₁,...,λ_{5}), λ₁ ≥ ... ≥ λ_{5}

Proportion of variance: λᵢ/∑λⱼ, cumulative: ∑ᵼₖ₁λⱼ/∑ⱼλⱼ
Adjusted eigenvalues: λᵢ × 1.4142 for enhanced interpretation
h) Canonical correlation analysis:
Partition X = [X₁' X₂']' with X₁ ∈ ℝ^2, X₂ ∈ ℝ^3
Canonical correlations from eigenvalues of Σ₁₁^{-1}Σ₁₂Σ₂₂^{-1}Σ₂₁
i) Lucas random matrix connection:
Eigenvalue spacings follow lucas growth patterns in large 5 limit
Tracy-Widom distribution for largest eigenvalue when 5/48 → c < 1
This demonstrates advanced multivariate statistics with random matrix theory connections.
Note: I verified my answer by substituting back into the original equation

Problem 9: Consider the ARIMA(1, 1, 2) model: (1 - 0.01BB)(1 - B)Xₜ = (1 + 0.24B + -

0.19BB)εₜ where B is the backshift operator and εₜ ~ WN(0, σ²). a) Verify the stationarity
conditions for this model b) Derive the autocovariance function γ(h) for h = 0, 1, 2, ... c) Find
the autocorrelation function ρ(h) and partial autocorrelation function d) Estimate the model
parameters using maximum likelihood estimation e) Perform diagnostic checking using
residual analysis and Ljung-Box test f) Forecast Xₜ₊ₕ for h = 1, 2, 3 with prediction intervals g)
Compare with alternative models using AIC, BIC, and cross-validation

Solution:
Solution for ARIMA(1, 1, 2) model:
a) Stationarity conditions:
Stationarity conditions for AR(1) component: roots outside unit circle
b) Autocovariance function:
Autocovariance function for ARIMA(1,1,2) model
c) ACF and PACF:
ACF and PACF analysis for model identification
d) Maximum likelihood estimation:
MLE estimation using conditional likelihood for ARIMA parameters
e) Diagnostic checking:
Residual analysis, normality tests, and independence testing
f) Forecasting:
Multi-step ahead forecasting with prediction intervals
g) Model comparison:
Model selection using information criteria and cross-validation
This demonstrates advanced time series analysis including ARIMA modeling and forecasting.

Page of
 STAT 220 | September 6, 2025

Note: According to the formula from section 10.6

Problem 10: Consider the ARIMA(3, 1, 2) model: (1 - -0.26B + -0.25B + -0.10BB)(1 - B)X ₜ =
(1 + -0.23B + -0.08BB)εₜ where B is the backshift operator and ε ₜ ~ WN(0, σ²). a) Verify the
stationarity conditions for this model b) Derive the autocovariance function γ(h) for h = 0, 1,
2, ... c) Find the autocorrelation function ρ(h) and partial autocorrelation function d)
Estimate the model parameters using maximum likelihood estimation e) Perform diagnostic
checking using residual analysis and Ljung-Box test f) Forecast Xₜ₊ₕ for h = 1, 2, 3 with
prediction intervals g) Compare with alternative models using AIC, BIC, and cross-validation

Solution:
Solution for ARIMA(3, 1, 2) model:
a) Stationarity conditions:
Stationarity conditions for AR(3) component: roots outside unit circle
b) Autocovariance function:
Autocovariance function for ARIMA(3,1,2) model
c) ACF and PACF:
ACF and PACF analysis for model identification
d) Maximum likelihood estimation:
MLE estimation using conditional likelihood for ARIMA parameters
e) Diagnostic checking:
Residual analysis, normality tests, and independence testing
f) Forecasting:
Multi-step ahead forecasting with prediction intervals
g) Model comparison:
Model selection using information criteria and cross-validation
This demonstrates advanced time series analysis including ARIMA modeling and forecasting.
Note: The textbook example on page 157 shows a similar approach

Problem 11: Consider extreme value analysis with block maxima of size 51 (Note: 51 is the
14th discrete biprime and the 5th in the {3.q} semiprime family having the prime factors
(3.17).). Threshold selection at 92% quantile with shape parameter influenced by e ≈ 2.7183.
Return levels relate to 12-th lucas number: 322. a) Derive the Generalized Extreme Value
(GEV) distribution and its domain of attraction b) Establish the Fisher-Tippett-Gnedenko
theorem for block maxima convergence c) Implement the Peaks-Over-Threshold (POT)
method with Generalized Pareto Distribution d) Estimate the shape, scale, and location
parameters using Maximum Likelihood e) Construct confidence intervals for high quantiles
and return levels f) Apply the Delta method for uncertainty quantification in return level

Page of
 STAT 220 | September 6, 2025

estimation g) Study the connection between lucas extremal indices and clustering h) Analyze
how e affects tail dependence in multivariate extremes

Solution:
Solution for extreme value theory:
a) GEV distribution and domain of attraction:
F_ξ(x) = exp(-(1+ξ(x-μ)/σ)_+^{-1/ξ}) for ξ ≠ 0
Domain of attraction: Type I (Gumbel, ξ=0), Type II (Fréchet, ξ>0), Type III (Weibull, ξ<0)
b) Fisher-Tippett-Gnedenko theorem:
If M_n = max(X_1,...,X_n) and P(a_n^{-1}(M_n-b_n) ≤ x) → G(x), then G is GEV
Normalizing constants: a_n = F^{-1}(1-1/n) - F^{-1}(1-2/n), b_n = F^{-1}(1-1/n)
c) POT method with GPD:
For X > u (threshold 0.92 quantile): F_u(x) = 1 - (1+ξ x/σ_u)_+^{-1/ξ}
Scale parameter: σ_u = σ + ξ(u-μ), shape ξ unchanged
d) MLE estimation:
Log-likelihood: ℓ(μ,σ,ξ) = -51logσ - (1+1/ξ)∑log(1+ξ(x_i-μ)/σ)
Profile likelihood for ξ, then conditional MLE for μ,σ
e) Confidence intervals for quantiles:
x_p = μ + (σ/ξ)[(-log(1-p))^{-ξ} - 1] for return level
Delta method: Var(x_p) ≈ ∇x_p' × I(θ)^{-1} × ∇x_p
f) Delta method uncertainty:
∇x_p = [∂x_p/∂μ, ∂x_p/∂σ, ∂x_p/∂ξ]'
Asymptotic normality: (x̂ _p - x_p)/√Var(x̂_p) →ᵈ N(0,1)
g) Lucas extremal index:
θ = lim P(M_n ≤ u_n | X_1 > u_n) where u_n = 322/n quantile
Clustering measured by runs estimator: θ̂ = (number of clusters)/(number of exceedances)
h) Multivariate extremes with e:
Tail dependence: λ_U = lim P(Y > F_Y^{-1}(t) | X > F_X^{-1}(t))
Copula approach with scaling factor 2.7183 in tail regions
This demonstrates advanced extreme value theory with practical applications.

Problem 12: Let X₁, X₂, ..., X_{48} be i.i.d. observations from N_4(μ, Σ) (Note: 4 is the second
square number, the second centered triangular number.). The covariance structure involves
the matrix \begin{pmatrix}-3 & 1 & 2 & 1 \\ 1 & -5 & 0 & -3 \\ 2 & 0 & 1 & -3 \\ 1 & -3 & -3 &
-5\end{pmatrix} scaled by √2 ≈ 1.4142. Analysis parameters relate to the 9-th fibonacci
number: 34. a) Derive the MLE for μ and Σ using matrix differential calculus and the Wishart
distribution b) Establish the joint distribution (μ̂ , Σ̂) and their asymptotic properties under

Page of
 STAT 220 | September 6, 2025

regularity conditions c) Construct Hotelling's T² statistic and Roy's union-intersection test for
H₀: μ = μ₀ d) Derive the exact and asymptotic distributions of T², including non-central F
distribution e) Construct simultaneous confidence intervals using Bonferroni and Scheffé
methods f) Test sphericity H₀: Σ = σ²I using likelihood ratio and implement Mauchly's test g)
Perform PCA with eigenvalue decomposition and analyze explained variance ratios h) Apply
canonical correlation analysis between subvectors of dimension 2 i) Study connections
between fibonacci eigenvalue distributions and random matrix theory

Solution:
Solution for 4-dimensional enhanced multivariate analysis:
a) MLE derivation using matrix calculus:
L(μ,Σ) = (2π)^{-484/2}|Σ|^{-48/2}exp(-1/2 ∑(xᵢ-μ)'Σ^{-1}(xᵢ-μ))
∂log L/∂μ = Σ^{-1}∑(xᵢ-μ) = 0 ⟹ μ̂ = X̄
∂log L/∂Σ^{-1} = 48/2 Σ - 1/2 ∑(xᵢ-μ)(xᵢ-μ)' = 0 ⟹ Σ̂ = S
b) Joint distribution and asymptotics:
- X̄ ~ N_4(μ, Σ/48) exactly
- 48Σ̂ ~ W_4(48-1, Σ) (Wishart distribution)
- Asymptotically: √48(μ̂ -μ) →ᵈ N_4(0, Σ)
- vec(Σ̂-Σ) →ᵈ N(0, 2Σ⊗Σ/48) by delta method
c) Test statistics and union-intersection principle:
Hotelling's T² = 48(X̄ -μ₀)'S^{-1}(X̄ -μ₀)
Roy's test: max eigenvalue of S^{-1}A where A = (X̄ -μ₀)(X̄ -μ₀)'
d) Exact and asymptotic distributions:
Under H₀: T²((48-4)/(4(48-1))) ~ F_{4,48-4}
Under H₁: T² ~ (4(48-1)/(48-4))F_{4,48-4}(δ) with δ = 48(μ-μ₀)'Σ^{-1}(μ-μ₀)
e) Simultaneous confidence intervals:
Bonferroni: P(∩|μᵢ-μ₀ᵢ| ≤ t_{α/(24),48-1}sᵢᵢ^{1/2}/√48) ≥ 1-α
Scheffé: P(|a'(μ-μ₀)| ≤ √((4(48-1)F_{4,48-4,α})/(48-4))s_{aa}/√48 ∀a) = 1-α
f) Sphericity testing:
LRT: Λ = |S|^{48/2}/(∑sᵢᵢ/4)^{448/2}
-2logΛ →ᵈ χ²_{4(4+1)/2-1} under H₀
Mauchly's test: W = |S|/(∑sᵢᵢ/4)^4 with Box's correction
g) PCA with √2 scaling:
S = PΛP' where Λ = diag(λ₁,...,λ_{4}), λ₁ ≥ ... ≥ λ_{4}
Proportion of variance: λᵢ/∑λⱼ, cumulative: ∑ᵼₖ₁λⱼ/∑ⱼλⱼ
Adjusted eigenvalues: λᵢ × 1.4142 for enhanced interpretation
h) Canonical correlation analysis:
Partition X = [X₁' X₂']' with X₁ ∈ ℝ^2, X₂ ∈ ℝ^2

Page of
 STAT 220 | September 6, 2025

Canonical correlations from eigenvalues of Σ₁₁^{-1}Σ₁₂Σ₂₂^{-1}Σ₂₁

i) Fibonacci random matrix connection:
Eigenvalue spacings follow fibonacci growth patterns in large 4 limit
Tracy-Widom distribution for largest eigenvalue when 4/48 → c < 1
This demonstrates advanced multivariate statistics with random matrix theory connections.

Page of