JEL Classification Numbers: C12, C31, C38
Corresponding author and proofs: Chihwa Kao, Department of Economics, University of Connecticut, Storrs, CT 06269, USA; e-mail: chih-hwa.kao@uconn.edu.
Introduction
The LM test of Breusch and Pagan (1980) aggregates squared sample correlations and has a chi-squared null limit when the cross-sectional dimension is fixed and the time dimension grows. In high dimensions, Schott (2005) derives a Gaussian limit for the sum of squared correlations when both dimensions increase proportionally under temporally independent Gaussian sampling. Baltagi, Feng, and Kao (2012) extend high-dimensional LM testing to fixed-effects panel regressions, and Huang, Li, and Yao (2025) develop unified and robust LM-type tests for large panels; both assume serially independent errors.
The existing literature does not establish the centering and scaling of the quadratic LM statistic under serial correlation when the cross-sectional and time dimensions grow proportionally. Baltagi, Kao, and Peng (2016) document size distortions of existing cross-sectional correlation tests under serial dependence and develop a modified CD test; Hounyo, Kao, and Kim (2025) also study CD-type statistics. Those procedures aggregate correlations linearly, whereas the LM statistic aggregates their squares. Chen and Liu (2018) consider a separable Gaussian model but test independence in the temporal dimension. Thus, these results do not directly provide a high-dimensional null distribution for the quadratic LM statistic with serially correlated observations.
We study an n × T n \times T n × T Gaussian panel whose cross-sectional units are independent under the null and share a stationary, short-memory Toeplitz temporal covariance matrix Γ \Gamma Γ , normalized by Tr ( Γ ) / T = 1 \operatorname{Tr}(\Gamma)/T = 1 Tr ( Γ ) / T = 1 . Write R = ( r i j ) R = (r_{ij}) R = ( r ij ) for the sample correlation matrix formed using the maintained zero population means. Let λ = T / n \lambda= T/n λ = T / n denote the dimension ratio, with λ → λ 0 ∈ ( 0 , ∞ ) \lambda\to\lambda_0 \in(0,\infty) λ → λ 0 ∈ ( 0 , ∞ ) . We consider the statistic
Tr ( R 2 ) = n + 2 ∑ i < j r i j 2 . \operatorname{Tr}(R^2) = n + 2\sum_{i<j} r_{ij}^2. Tr ( R 2 ) = n + 2 i < j ∑ r ij 2 .
Although this is a linear spectral statistic (LSS), existing CLTs for high-dimensional sample correlation matrices (Gao, Han, Pan, and Yang, 2017; Yin, Zheng, and Zou, 2023; Li, Pan, Xie, and Zhou, 2024) address temporally independent sampling and cannot be applied directly to our setting.
The contribution of this paper is to show how to center and scale Tr ( R 2 ) \operatorname{Tr}(R^{2}) Tr ( R 2 ) under serial correlation. Let τ k , T \tau_{k,T} τ k , T denote the normalized temporal spectral moments of Γ \Gamma Γ , and let K T K_{T} K T be the function of τ 2 , T , τ 3 , T , τ 4 , T \tau_{2,T},\tau_{3,T},\tau_{4,T} τ 2 , T , τ 3 , T , τ 4 , T defined in § 2.1 . We show that
E { Tr ( R 2 ) } = n + n − 1 λ τ 2 , T + 4 ( n − 1 ) n λ 2 ( τ 2 , T 2 − τ 3 , T ) + O ( n − 1 ) , \mathbb{E}\{\operatorname{Tr}(R^{2})\}=n+\frac{n-1}{\lambda}\tau_{2,T}+\frac{4(n-1)}{n\lambda^{2}}(\tau_{2,T}^{2}-\tau_{3,T})+O(n^{-1}), E { Tr ( R 2 )} = n + λ n − 1 τ 2 , T + n λ 2 4 ( n − 1 ) ( τ 2 , T 2 − τ 3 , T ) + O ( n − 1 ) ,
Var { Tr ( R 2 ) } = 4 τ 2 , T 2 λ 2 + 8 K T λ 3 + O ( n − 1 ) , \operatorname{Var}\{\operatorname{Tr}(R^{2})\}=\frac{4\tau_{2,T}^{2}}{\lambda^{2}}+\frac{8K_{T}}{\lambda^{3}}+O(n^{-1}), Var { Tr ( R 2 )} = λ 2 4 τ 2 , T 2 + λ 3 8 K T + O ( n − 1 ) ,
and that Tr ( R 2 ) \operatorname{Tr}(R^{2}) Tr ( R 2 ) , centered and scaled accordingly, is asymptotically standard normal. The centering and scaling depend on the temporal dependence only through τ 2 , T \tau_{2,T} τ 2 , T , τ 3 , T \tau_{3,T} τ 3 , T , and τ 4 , T \tau_{4,T} τ 4 , T , and reduce to the known-mean version of Schott (2005) when Γ = I T \Gamma=I_{T} Γ = I T . Whereas the LSS literature relies on random matrix theory, we obtain these moments directly by Wick’s theorem. This is the key step: once the centering and scaling are known, asymptotic normality is established by a martingale central limit theorem. Details are given in § 3 .
The null CLT concerns the oracle statistic, which uses the population spectral moments τ k , T \tau_{k,T} τ k , T ; feasible versions based on estimated moments are evaluated in the Monte Carlo experiments, which also illustrate the size distortions that arise when serial correlation is ignored.
We also characterize the behavior of the test under alternatives. We establish consistency against fixed alternatives and a Gaussian shift under local alternatives, with aggregate signal ∥ R ∗ − I n ∥ F 2 \lVert R_{*}-I_{n}\rVert_{F}^{2} ∥ R ∗ − I n ∥ F 2 , where R ∗ R_{*} R ∗ is the population cross-sectional correlation matrix. This yields an explicit local-power expression and a detection boundary for the statistic under the alternatives considered here; it is not asserted to be a minimax boundary over all tests.
§ 2 presents the model and statistic, § 3 establishes the null expansions and CLT, § 4 studies power, and § 5 reports simulations. § 6 concludes. Proofs of the null results (Lemmas 2 and 3 and Theorem 1) are in the Appendix; the remaining proofs, including those of the power results, are in the Online Supplement.
Model and Test Statistic
Model Specification
Let X = ( x i t ) ∈ R n × T X=(x_{it})\in\mathbb{R}^{n\times T} X = ( x i t ) ∈ R n × T denote observations on n n n cross-sectional units over T T T periods.
Assume the separable Gaussian model
X = U 1 / 2 Z Γ 1 / 2 , X=U^{1/2}Z\Gamma^{1/2}, X = U 1/2 Z Γ 1/2 ,
where U ∈ R n × n U\in\mathbb{R}^{n\times n} U ∈ R n × n is a positive-definite cross-sectional covariance matrix, Γ ∈ R T × T \Gamma\in\mathbb{R}^{T\times T} Γ ∈ R T × T is a temporal covariance matrix, and Z Z Z has independent N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) entries. We use symmetric matrix square roots. Then vec ( X ) ∼ N ( 0 , Γ ⊗ U ) \operatorname{vec}(X)\sim N(0,\Gamma\otimes U) vec ( X ) ∼ N ( 0 , Γ ⊗ U ) under column-wise vectorization, and Cov ( x i t , x j s ) = U i j Γ t s \operatorname{Cov}(x_{it},x_{js})=U_{ij}\Gamma_{ts} Cov ( x i t , x j s ) = U ij Γ t s . We write λ = T / n \lambda=T/n λ = T / n and impose proportional growth in Assumption 3.
The temporal covariance matrix Γ = Γ T = ( γ ∣ t − s ∣ ) t , s = 1 T \Gamma=\Gamma_T=(\gamma_{\lvert t-s\rvert})_{t,s=1}^T Γ = Γ T = ( γ ∣ t − s ∣ ) t , s = 1 T , with γ 0 = 1 \gamma_0=1 γ 0 = 1 , is stationary and short-memory; the precise conditions are given in Assumption 2. We suppress the subscript T T T except where the dependence on T T T matters. Define the normalized temporal spectral moments by
τ k , T = 1 T Tr ( Γ k ) = 1 T ∑ t = 1 T ν t , T k , k = 1 , 2 , … , \tau_{k,T}=\frac{1}{T}\operatorname{Tr}(\Gamma^k)=\frac{1}{T}\sum_{t=1}^{T}\nu_{t,T}^{k},\qquad k=1,2,\ldots, τ k , T = T 1 Tr ( Γ k ) = T 1 t = 1 ∑ T ν t , T k , k = 1 , 2 , … ,
where ν 1 , T , … , ν T , T \nu_{1,T},\ldots,\nu_{T,T} ν 1 , T , … , ν T , T are the eigenvalues of Γ \Gamma Γ . Since γ 0 = 1 \gamma_0=1 γ 0 = 1 , τ 1 , T = 1 \tau_{1,T}=1 τ 1 , T = 1 . The normalization γ 0 = 1 \gamma_0=1 γ 0 = 1 is without loss of generality for the sample correlation matrix, which is invariant to positive scalar rescaling of Γ \Gamma Γ . Define
K T = τ 4 , T − 2 τ 2 , T τ 3 , T + τ 2 , T 3 , K = lim T → ∞ K T = τ 4 − 2 τ 2 τ 3 + τ 2 3 , K_T=\tau_{4,T}-2\tau_{2,T}\tau_{3,T}+\tau_{2,T}^{3},\qquad K=\lim_{T\to\infty}K_T=\tau_4-2\tau_2\tau_3+\tau_2^3, K T = τ 4 , T − 2 τ 2 , T τ 3 , T + τ 2 , T 3 , K = T → ∞ lim K T = τ 4 − 2 τ 2 τ 3 + τ 2 3 ,
where τ k = lim T → ∞ τ k , T \tau_k=\lim_{T\to\infty}\tau_{k,T} τ k = lim T → ∞ τ k , T .
Hypotheses and Test Statistic
Let x i = ( x i 1 , … , x i T ) ′ ∈ R T x_i=(x_{i1},\ldots,x_{iT})^\prime\in\mathbb{R}^T x i = ( x i 1 , … , x i T ) ′ ∈ R T be the i i i th row of X X X written as a column vector. We test cross-sectional independence. Under Gaussianity, this is equivalent to U U U being diagonal, or to the population correlation matrix R ∗ = diag ( U ) − 1 / 2 U diag ( U ) − 1 / 2 R_*=\operatorname{diag}(U)^{-1/2}U\operatorname{diag}(U)^{-1/2} R ∗ = diag ( U ) − 1/2 U diag ( U ) − 1/2 being I n I_n I n . Because the sample correlation matrix is invariant to positive row-wise rescaling, we use the normalization U = I n U = I_n U = I n when deriving the null distribution. The null hypothesis is therefore H 0 : R ∗ = I n H_0 : R_* = I_n H 0 : R ∗ = I n , with U = I n U = I_n U = I n as its normalized representation.
With the population mean specified as zero, define the uncentered sample covariance and correlation matrices by
S = 1 T X X ′ , R = diag ( S ) − 1 / 2 S diag ( S ) − 1 / 2 . S = \frac{1}{T}XX^{\prime}, \qquad R = \operatorname{diag}(S)^{-1/2} S \operatorname{diag}(S)^{-1/2}. S = T 1 X X ′ , R = diag ( S ) − 1/2 S diag ( S ) − 1/2 .
In particular,
r i j = ∑ t = 1 T x i t x j t ( ∑ t = 1 T x i t 2 ) 1 / 2 ( ∑ t = 1 T x j t 2 ) 1 / 2 , Tr ( R 2 ) = n + 2 ∑ i < j r i j 2 . r_{ij} = \frac{\sum_{t=1}^{T} x_{it}x_{jt}}{\left(\sum_{t=1}^{T} x_{it}^{2}\right)^{1/2}\left(\sum_{t=1}^{T} x_{jt}^{2}\right)^{1/2}}, \qquad\operatorname{Tr}(R^{2}) = n + 2\sum_{i<j} r_{ij}^{2}. r ij = ( ∑ t = 1 T x i t 2 ) 1/2 ( ∑ t = 1 T x j t 2 ) 1/2 ∑ t = 1 T x i t x j t , Tr ( R 2 ) = n + 2 i < j ∑ r ij 2 .
This identity expresses the quadratic spectral statistic as a sum of squared off-diagonal sample correlations, the basis of the LM-type test.
Under H 0 H_0 H 0 and Γ = I T \Gamma= I_T Γ = I T , the known-mean version of the standardized statistic of Schott (2005) is
L M S = ∑ i < j r i j 2 − n ( n − 1 ) 2 T n ( n − 1 ) ( T − 1 ) T 2 ( T + 2 ) = Tr ( R 2 ) − n − n ( n − 1 ) T 2 n ( n − 1 ) ( T − 1 ) T 2 ( T + 2 ) . LM_S = \frac{\sum_{i<j} r_{ij}^{2} - \frac{n(n-1)}{2T}}{\sqrt{\frac{n(n-1)(T-1)}{T^{2}(T+2)}}} = \frac{\operatorname{Tr}(R^{2}) - n - \frac{n(n-1)}{T}}{2\sqrt{\frac{n(n-1)(T-1)}{T^{2}(T+2)}}}. L M S = T 2 ( T + 2 ) n ( n − 1 ) ( T − 1 ) ∑ i < j r ij 2 − 2 T n ( n − 1 ) = 2 T 2 ( T + 2 ) n ( n − 1 ) ( T − 1 ) Tr ( R 2 ) − n − T n ( n − 1 ) .
Assumptions
The null results below use Assumptions 1–3. The analysis of alternatives additionally invokes Assumption 4.
Assumption 1 The entries Z i t Z_{it} Z i t are independent and identically distributed N ( 0 , 1 ) \mathrm{N}(0,1) N ( 0 , 1 ) random variables.
Assumption 2 (Stationary short-memory temporal covariance ) For each T T T , Γ T = ( γ ∣ t − s ∣ ) t , s = 1 T \Gamma_T = (\gamma_{\lvert t-s\rvert})_{t,s=1}^{T} Γ T = ( γ ∣ t − s ∣ ) t , s = 1 T is a deterministic symmetric positive semidefinite Toeplitz matrix, with γ 0 = 1 \gamma_0 = 1 γ 0 = 1 and ∑ h = 0 ∞ ∣ γ h ∣ < ∞ \sum_{h=0}^{\infty} \lvert\gamma_h\rvert< \infty ∑ h = 0 ∞ ∣ γ h ∣ < ∞ . In particular, T − 1 Tr ( Γ T ) = 1 T^{-1}\operatorname{Tr}(\Gamma_T) = 1 T − 1 Tr ( Γ T ) = 1 and sup T ∥ Γ T ∥ o p < ∞ \sup_T \lVert\Gamma_T\rVert_{\mathrm{op}} < \infty sup T ∥ Γ T ∥ op < ∞ .
Assumption 3 (Proportional growth ) As n , T → ∞ n,T \to\infty n , T → ∞ , T / n → λ 0 ∈ ( 0 , ∞ ) T/n \to\lambda_0 \in(0,\infty) T / n → λ 0 ∈ ( 0 , ∞ ) .
Gaussianity permits exact higher-order moment calculations using Wick’s theorem (Isserlis’ formula). The short-memory assumption controls τ k , T \tau_{k,T} τ k , T for each fixed k k k , while proportional growth allows n n n and T T T to increase together.
Main Results
Null mean and variance
We expand each r i j 2 = S i j 2 / ( S i i S j j ) r_{ij}^{2}=S_{ij}^{2}/(S_{ii}S_{jj}) r ij 2 = S ij 2 / ( S ii S j j ) around S i i = S j j = 1 S_{ii}=S_{jj}=1 S ii = S j j = 1 , which requires the sample variances to be uniformly close to one.
Lemma 1 (Diagonal Concentration). Under Assumptions 1–3 and H 0 H_{0} H 0 ,
∥ diag ( S ) − I n ∥ op = O a . s . ( log n T ) . \lVert\operatorname{diag}(S)-I_{n}\rVert_{\operatorname{op}}=O_{\mathrm{a.s.}}\left(\sqrt{\frac{\log n}{T}}\right). ∥ diag ( S ) − I n ∥ op = O a.s. ( T log n ) .
The bound follows from Gaussian quadratic-form concentration and a union bound; see Section S.A.3 of the Online Supplement. Although the diagonal elements of S S S converge uniformly to one, this alone is insufficient to establish the limiting distribution of Tr ( R 2 ) \operatorname{Tr}(R^{2}) Tr ( R 2 ) . The randomness of the diagonal normalization must be accounted for in the asymptotic mean and variance expansions.
Lemma 2 (Null Mean). Under Assumptions 1–3 and H 0 H_{0} H 0 ,
E { Tr ( R 2 ) } = n + n − 1 λ τ 2 , T + 4 ( n − 1 ) n λ 2 ( τ 2 , T 2 − τ 3 , T ) + O ( n − 1 ) . \mathbb{E}\{\operatorname{Tr}(R^{2})\}=n+\frac{n-1}{\lambda}\tau_{2,T}+\frac{4(n-1)}{n\lambda^{2}}(\tau_{2,T}^{2}-\tau_{3,T})+O(n^{-1}). E { Tr ( R 2 )} = n + λ n − 1 τ 2 , T + n λ 2 4 ( n − 1 ) ( τ 2 , T 2 − τ 3 , T ) + O ( n − 1 ) .
Remark 3 (Interpretation of Lemma 2) Under H 0 H_{0} H 0 , all ordered off-diagonal pairs are identically distributed, so
E { Tr ( R 2 ) } = n + n ( n − 1 ) E ( r 12 2 ) . \mathbb{E}\{\operatorname{Tr}(R^{2})\}=n+n(n-1)\mathbb{E}(r_{12}^{2}). E { Tr ( R 2 )} = n + n ( n − 1 ) E ( r 12 2 ) .
The pairwise expansion in Appendix § 10 is
E ( r 12 2 ) = τ 2 , T T + 4 ( τ 2 , T 2 − τ 3 , T ) T 2 + O ( T − 3 ) . \mathbb{E}(r_{12}^{2})=\frac{\tau_{2,T}}{T}+\frac{4(\tau_{2,T}^{2}-\tau_{3,T})}{T^{2}}+O(T^{-3}). E ( r 12 2 ) = T τ 2 , T + T 2 4 ( τ 2 , T 2 − τ 3 , T ) + O ( T − 3 ) .
*The leading term τ 2 , T / T = E ( S 12 2 ) \tau_{2,T}/T=\mathbb{E}(S_{12}^{2}) τ 2 , T / T = E ( S 12 2 ) captures temporal correlation. Since τ 2 , T ≥ 1 \tau_{2,T}\ge1 τ 2 , T ≥ 1 , with equality only when Γ = I T \Gamma=I_{T} Γ = I T , centering as if observations were temporally independent ($\tau_{2,T}=
incurs the order-n n n error ( n − 1 ) ( τ 2 , T − 1 ) / λ (n-1)(\tau_{2,T}-1)/\lambda ( n − 1 ) ( τ 2 , T − 1 ) / λ . The second term is the correction from the diagonal normalization: the linear denominator terms contribute − 4 τ 3 , T / T 2 -4\tau_{3,T}/T^2 − 4 τ 3 , T / T 2 , while the quadratic terms contribute 4 τ 2 , T 2 / T 2 + O ( T − 3 ) 4\tau_{2,T}^2/T^2+O(T^{-3}) 4 τ 2 , T 2 / T 2 + O ( T − 3 ) . By Cauchy–Schwarz and τ 1 , T = 1 \tau_{1,T}=1 τ 1 , T = 1 , τ 2 , T 2 ≤ τ 3 , T \tau_{2,T}^2\leq\tau_{3,T} τ 2 , T 2 ≤ τ 3 , T ; hence this second-order correction is nonpositive. The O ( T − 3 ) O(T^{-3}) O ( T − 3 ) pairwise remainder sums to O ( n − 1 ) O(n^{-1}) O ( n − 1 ) when T ≍ n T\asymp n T ≍ n .
For Γ = I T \Gamma=I_T Γ = I T , all τ k , T = 1 \tau_{k,T}=1 τ k , T = 1 and r 12 2 ∼ B e t a ( 1 / 2 , ( T − 1 ) / 2 ) r_{12}^2\sim Beta(1/2,(T-1)/2) r 12 2 ∼ B e t a ( 1/2 , ( T − 1 ) /2 ) . Thus E ( r 12 2 ) = 1 / T \mathbb{E}(r_{12}^2)=1/T E ( r 12 2 ) = 1/ T exactly, recovering E { T r ( R 2 ) } = n + n ( n − 1 ) / T \mathbb{E}\{Tr(R^2)\}=n+n(n-1)/T E { T r ( R 2 )} = n + n ( n − 1 ) / T .
The variance has two components: the variance of individual squared correlations and the covariance between squared correlations that share a unit. Disjoint pairs are independent under the null. The shared-unit covariance vanishes in the temporally independent benchmark; under temporal dependence, its aggregate contribution can remain of order O ( 1 ) O(1) O ( 1 ) as n n n and T T T grow proportionally.
Lemma 3 (Null Variance) Under Assumptions 1–3 and H 0 H_0 H 0 ,
Var { T r ( R 2 ) } = 4 τ 2 , T 2 λ 2 + 8 K T λ 3 + O ( n − 1 ) . \operatorname{Var}\{Tr(R^2)\}=\frac{4\tau_{2,T}^2}{\lambda^2}+\frac{8K_T}{\lambda^3}+O(n^{-1}). Var { T r ( R 2 )} = λ 2 4 τ 2 , T 2 + λ 3 8 K T + O ( n − 1 ) .
By conditional independence, b T = V a r { E ( r 12 2 ∣ Z 1 ) } b_T = Var\{E(r_{12}^{2} \mid Z_1)\} b T = V a r { E ( r 12 2 ∣ Z 1 )} . With q ( z ) = z ′ Γ z q(z) = z^{\prime}\Gamma z q ( z ) = z ′ Γ z and B T = E { Z Z ′ / q ( Z ) } B_T = E\{ZZ^{\prime}/q(Z)\} B T = E { Z Z ′ / q ( Z )} for Z ∼ N ( 0 , I T ) Z \sim N(0,I_T) Z ∼ N ( 0 , I T ) ,
E ( r 12 2 ∣ Z 1 = z ) = z ′ Γ B T Γ z q ( z ) . E(r_{12}^{2} \mid Z_1 = z) = \frac{z^{\prime}\Gamma B_T\Gamma z}{q(z)}. E ( r 12 2 ∣ Z 1 = z ) = q ( z ) z ′ Γ B T Γ z .
Its centered fluctuation is governed at leading order by
W T ( Z 1 ) = Z 1 ′ Γ 2 Z 1 T Z 1 ′ Z 1 , D 1 = Z 1 ′ ( Γ 2 − τ 2 , T Γ ) Z 1 . W_T(Z_1) = \frac{Z_1^{\prime}\Gamma^{2}Z_1}{T Z_1^{\prime}Z_1}, \qquad D_1 = Z_1^{\prime}(\Gamma^{2}-\tau_{2,T}\Gamma)Z_1. W T ( Z 1 ) = T Z 1 ′ Z 1 Z 1 ′ Γ 2 Z 1 , D 1 = Z 1 ′ ( Γ 2 − τ 2 , T Γ ) Z 1 .
for which V a r ( D 1 ) = 2 T K T Var(D_1) = 2TK_T V a r ( D 1 ) = 2 T K T . The leading term of W T − τ 2 , T / T W_T-\tau_{2,T}/T W T − τ 2 , T / T is D 1 / T 2 D_1/T^{2} D 1 / T 2 , and V a r ( W T ) = V a r ( D 1 ) / T 4 + O ( T − 4 ) = 2 K T / T 3 + O ( T − 4 ) Var(W_T) = Var(D_1)/T^{4}+O(T^{-4}) = 2K_T/T^{3}+O(T^{-4}) V a r ( W T ) = V a r ( D 1 ) / T 4 + O ( T − 4 ) = 2 K T / T 3 + O ( T − 4 ) . The exact conditional expectation differs from W T W_T W T by a constant plus a centered term of order O P ( T − 5 / 2 ) O_P(T^{-5/2}) O P ( T − 5/2 ) , so it has the same variance to this order (Appendix B, Steps 3–4). In terms of the eigenvalues of Γ \Gamma Γ ,
K T = τ 4 , T − 2 τ 2 , T τ 3 , T + τ 2 , T 3 = 1 T ∑ t = 1 T ν t , T 2 ( ν t , T − τ 2 , T ) 2 ≥ 0 , K_T = \tau_{4,T}-2\tau_{2,T}\tau_{3,T}+\tau_{2,T}^{3}
= \frac{1}{T}\sum_{t=1}^{T}\nu_{t,T}^{2}(\nu_{t,T}-\tau_{2,T})^{2} \ge0, K T = τ 4 , T − 2 τ 2 , T τ 3 , T + τ 2 , T 3 = T 1 t = 1 ∑ T ν t , T 2 ( ν t , T − τ 2 , T ) 2 ≥ 0 ,
a dispersion measure of the temporal eigenvalues weighted by their squares. For Γ = I T \Gamma=I_T Γ = I T , K T = 0 K_T=0 K T = 0 , the shared-unit covariance vanishes exactly, and
V a r { T r ( R 2 ) } = 4 n ( n − 1 ) ( T − 1 ) T 2 ( T + 2 ) . Var\{Tr(R^{2})\} = \frac{4n(n-1)(T-1)}{T^{2}(T+2)}. V a r { T r ( R 2 )} = T 2 ( T + 2 ) 4 n ( n − 1 ) ( T − 1 ) .
which is the variance in the denominator of L M S LMS L M S . Since τ k , T → τ k \tau_{k,T} \to\tau_k τ k , T → τ k and K T → K K_T \to K K T → K , replacing τ 2 , T \tau_{2,T} τ 2 , T and K T K_T K T by their limits changes the leading variance by o ( 1 ) o(1) o ( 1 ) .
Carrying the expansions of Appendix B one order further gives
a T = 2 τ 2 , T 2 T 2 + 28 τ 2 , T 3 − 40 τ 2 , T τ 3 , T + 6 τ 4 , T T 3 + O ( T − 4 ) , a_T = \frac{2\tau_{2,T}^{2}}{T^{2}}+\frac{28\tau_{2,T}^{3}-40\tau_{2,T}\tau_{3,T}+6\tau_{4,T}}{T^{3}}+O(T^{-4}), a T = T 2 2 τ 2 , T 2 + T 3 28 τ 2 , T 3 − 40 τ 2 , T τ 3 , T + 6 τ 4 , T + O ( T − 4 ) ,
b T = 2 K T T 3 + L T T 4 + O ( T − 5 ) , b_T = \frac{2K_T}{T^{3}}+\frac{L_T}{T^{4}}+O(T^{-5}), b T = T 3 2 K T + T 4 L T + O ( T − 5 ) ,
where
L T = 40 τ 2 , T 4 − 104 τ 2 , T 2 τ 3 , T + 60 τ 2 , T τ 4 , T + 28 τ 3 , T 2 − 24 τ 5 , T . L_T = 40\tau_{2,T}^{4}-104\tau_{2,T}^{2}\tau_{3,T}+60\tau_{2,T}\tau_{4,T}+28\tau_{3,T}^{2}-24\tau_{5,T}. L T = 40 τ 2 , T 4 − 104 τ 2 , T 2 τ 3 , T + 60 τ 2 , T τ 4 , T + 28 τ 3 , T 2 − 24 τ 5 , T .
Consequently, with T = n λ T = n\lambda T = nλ ,
Var { Tr ( R 2 ) } = 4 τ 2 , T 2 λ 2 + 8 K T λ 3 + c T n + O ( n − 2 ) , \operatorname{Var}\{\operatorname{Tr}(R^2)\} = \frac{4\tau_{2,T}^2}{\lambda^2} + \frac{8K_T}{\lambda^3} + \frac{c_T}{n} + O(n^{-2}), Var { Tr ( R 2 )} = λ 2 4 τ 2 , T 2 + λ 3 8 K T + n c T + O ( n − 2 ) ,
c T = − 4 τ 2 , T 2 λ 2 + 4 ( 8 τ 2 , T 3 − 8 τ 2 , T τ 3 , T − 3 τ 4 , T ) λ 3 + 4 L T λ 4 . c_T = -\frac{4\tau_{2,T}^2}{\lambda^2} + \frac{4(8\tau_{2,T}^3 - 8\tau_{2,T}\tau_{3,T} - 3\tau_{4,T})}{\lambda^3} + \frac{4L_T}{\lambda^4}. c T = − λ 2 4 τ 2 , T 2 + λ 3 4 ( 8 τ 2 , T 3 − 8 τ 2 , T τ 3 , T − 3 τ 4 , T ) + λ 4 4 L T .
For Γ = I T \Gamma= I_T Γ = I T , L T = 0 L_T = 0 L T = 0 and c T = − 4 / λ 2 − 12 / λ 3 c_T = -4/\lambda^2 - 12/\lambda^3 c T = − 4/ λ 2 − 12/ λ 3 , which is the n − 1 n^{-1} n − 1 term of the exact variance above. In the AR(1) designs of § 5 with ρ > 0 \rho> 0 ρ > 0 , c T c_T c T is negative and large: for ρ = 0.4 \rho= 0.4 ρ = 0.4 and n = T = 100 n = T = 100 n = T = 100 , c T ≈ − 128 c_T \approx-128 c T ≈ − 128 , and the leading variance exceeds the exact variance by about 10%; for ρ = 0.7 \rho= 0.7 ρ = 0.7 , by about 26%. The O ( 1 ) O(1) O ( 1 ) terms are unaffected, so Theorem 1 is unchanged; the n − 1 n^{-1} n − 1 term accounts for the finite-sample conservativeness of the oracle test reported in § 5 .
Relative to the temporally independent benchmark, temporal dependence changes the leading mean and variance, adds an O ( 1 ) O(1) O ( 1 ) mean correction from the diagonal normalization, and introduces the O ( 1 ) O(1) O ( 1 ) shared-unit variance correction. The latter two corrections are of the same order as the null fluctuations and cannot generally be omitted.
Central limit theorem
Let μ n \mu_n μ n and σ n 2 \sigma_n^2 σ n 2 denote the centering and variance expressions evaluated at the finite-T T T spectral moments,
μ n = n + n − 1 λ τ 2 , T + 4 ( n − 1 ) n λ 2 ( τ 2 , T 2 − τ 3 , T ) , σ n 2 = 4 τ 2 , T 2 λ 2 + 8 K T λ 3 . (2) \mu_n = n + \frac{n-1}{\lambda}\tau_{2,T} + \frac{4(n-1)}{n\lambda^2}(\tau_{2,T}^2-\tau_{3,T}), \qquad\sigma_n^2 = \frac{4\tau_{2,T}^2}{\lambda^2} + \frac{8K_T}{\lambda^3}. \tag*{(2)} μ n = n + λ n − 1 τ 2 , T + n λ 2 4 ( n − 1 ) ( τ 2 , T 2 − τ 3 , T ) , σ n 2 = λ 2 4 τ 2 , T 2 + λ 3 8 K T . (2)
By Lemmas 2 and 3, E { Tr ( R 2 ) } = μ n + O ( n − 1 ) \mathbb{E}\{\operatorname{Tr}(R^2)\} = \mu_n + O(n^{-1}) E { Tr ( R 2 )} = μ n + O ( n − 1 ) and Var { Tr ( R 2 ) } = σ n 2 + O ( n − 1 ) \operatorname{Var}\{\operatorname{Tr}(R^2)\} = \sigma_n^2 + O(n^{-1}) Var { Tr ( R 2 )} = σ n 2 + O ( n − 1 ) . Replacing τ k , T \tau_{k,T} τ k , T and K T K_T K T by their limits τ k \tau_k τ k and K K K in the O ( 1 ) O(1) O ( 1 ) terms changes μ n \mu_n μ n and σ n 2 \sigma_n^2 σ n 2 by o ( 1 ) o(1) o ( 1 ) (Remarks 1 and 4). Define the LM statistic
L M = Tr ( R 2 ) − μ n σ n . LM = \frac{\operatorname{Tr}(R^2)-\mu_n}{\sigma_n}. L M = σ n Tr ( R 2 ) − μ n .
The test rejects H 0 H_0 H 0 at level α \alpha α when ∣ L M ∣ > z α / 2 |LM| > z_{\alpha/2} ∣ L M ∣ > z α /2 , where z α / 2 z_{\alpha/2} z α /2 is the ( 1 − α / 2 ) (1-\alpha/2) ( 1 − α /2 ) standard normal quantile.
Theorem 1 (Null CLT). Under Assumptions 1–3 and H 0 H_0 H 0 , L M → d N ( 0 , 1 ) LM \xrightarrow{d} N(0,1) L M d N ( 0 , 1 ) .
Since τ 2 , T ≥ 1 \tau_{2,T} \ge1 τ 2 , T ≥ 1 and K T ≥ 0 K_T \ge0 K T ≥ 0 (Remark 2), σ n 2 ≥ 4 / λ 2 \sigma_n^2 \ge4/\lambda^2 σ n 2 ≥ 4/ λ 2 is bounded away from zero. Theorem 1 continues to hold with the O ( 1 ) O(1) O ( 1 ) terms of μ n \mu_n μ n and σ n 2 \sigma_n^2 σ n 2 evaluated at the limits τ k \tau_k τ k and K K K ; only the leading mean term must retain τ 2 , T \tau_{2,T} τ 2 , T . When Γ = I T \Gamma= I_T Γ = I T , μ n \mu_n μ n is the centering of L M S LM_S L M S and σ n 2 \sigma_n^2 σ n 2 equals its variance up to O ( n − 1 ) O(n^{-1}) O ( n − 1 ) , so Theorem 1 recovers the known-mean version of Schott (2005). The test is an oracle test because it uses the population spectral moments of Γ \Gamma Γ ; feasible versions are investigated numerically in § 5 .
so they are not generally martingale differences. The Hoeffding–Doob increments in Appendix § 10.2 instead account for both projections:
Y k = 2 ( n − 1 ) ξ k + 2 ∑ i < k h i k o Y_k = 2(n-1)\xi_k + 2\sum_{i<k} h^{o}_{ik} Y k = 2 ( n − 1 ) ξ k + 2 i < k ∑ h ik o
where h i k o = r i k 2 − θ T − ξ i − ξ k h^{o}_{ik}=r_{ik}^{2}-\theta_T-\xi_i-\xi_k h ik o = r ik 2 − θ T − ξ i − ξ k is the degenerate component. The nondegenerate first-order projection yields the additional 8 K T / λ 3 8K_T/\lambda^3 8 K T / λ 3 variance component. Thus, replacing the dependence-adjusted centering and variance by their temporally independent benchmarks can lead to substantial size distortions.
Power of the LM Test
This section examines the LM statistic under cross-sectional dependence. We derive its mean and variance under alternatives, establish consistency when the aggregate signal ∑ i ≠ j ( ρ i j ∗ ) 2 \sum_{i\ne j}(\rho_{ij}^{*})^2 ∑ i = j ( ρ ij ∗ ) 2 diverges, and characterize local asymptotic power and the detection boundary. The proofs of Lemmas 5 and 6, Theorem 2, and Corollary 1 are in Sections S.B–S.E of the Online Supplement.
Setup Under the Alternative
Under the alternative, the data model remains X = U 1 / 2 Z Γ 1 / 2 X=U^{1/2}Z\Gamma^{1/2} X = U 1/2 Z Γ 1/2 with Z Z Z having i.i.d. N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) entries, but now U ≠ I n U\ne I_n U = I n . We parameterize the cross-sectional dependence through the population correlation matrix
R ∗ = ( diag ( U ) ) − 1 / 2 U ( diag ( U ) ) − 1 / 2 R_*=(\operatorname{diag}(U))^{-1/2}U(\operatorname{diag}(U))^{-1/2} R ∗ = ( diag ( U ) ) − 1/2 U ( diag ( U ) ) − 1/2
with ( i , j ) (i,j) ( i , j ) entry ρ i j ∗ = U i j / U i i U j j \rho_{ij}^{*}=U_{ij}/\sqrt{U_{ii}U_{jj}} ρ ij ∗ = U ij / U ii U j j . The hypotheses are H 0 : R ∗ = I n H_0:R_*=I_n H 0 : R ∗ = I n and H A : R ∗ ≠ I n H_A:R_*\ne I_n H A : R ∗ = I n . We write E A E_A E A , V a r A Var_A V a r A , and P A P_A P A for moments and probabilities under the alternative, and E 0 E_0 E 0 , V a r 0 Var_0 V a r 0 for those under H 0 H_0 H 0 with the same Γ \Gamma Γ ; thus E 0 E_0 E 0 and V a r 0 Var_0 V a r 0 are the moments in Lemmas 2 and 3. We define the cross-sectional spectral moments of R ∗ R_\ast R ∗ by
ϕ k = 1 n T r ( R ∗ k ) , k = 1 , 2 , 3 , 4. \phi_k = \frac{1}{n}Tr(R_\ast^k), \qquad k = 1,2,3,4. ϕ k = n 1 T r ( R ∗ k ) , k = 1 , 2 , 3 , 4.
By construction, ϕ 1 = 1 \phi_1 = 1 ϕ 1 = 1 . Under H 0 H_0 H 0 , ϕ k = 1 \phi_k = 1 ϕ k = 1 for all k k k . Under H A H_A H A , ϕ 2 > 1 \phi_2 > 1 ϕ 2 > 1 and the aggregate signal is
T r ( R ∗ 2 ) − n = n ( ϕ 2 − 1 ) = ∑ i ≠ j ( ρ i j ∗ ) 2 > 0. Tr(R_\ast^2) - n = n(\phi_2 - 1) = \sum_{i \ne j}(\rho_{ij}^\ast)^2 > 0. T r ( R ∗ 2 ) − n = n ( ϕ 2 − 1 ) = i = j ∑ ( ρ ij ∗ ) 2 > 0.
We call an alternative sequence local when n ( ϕ 2 − 1 ) = O ( 1 ) n(\phi_2 - 1) = O(1) n ( ϕ 2 − 1 ) = O ( 1 ) . By Assumptions 2–3 and Remark 2, σ n \sigma_n σ n is bounded above and away from zero, so under local alternatives the aggregate signal is at most of the order of the null standard deviation. We add one assumption governing the alternative.
Assumption 4 (Cross-Sectional Covariance). The matrix U U U is symmetric positive definite with c 1 ≤ min i U i i ≤ ∥ U ∥ o p ≤ c 2 c_1 \le\min_i U_{ii} \le\lVert U\rVert_{\mathrm{op}} \le c_2 c 1 ≤ min i U ii ≤ ∥ U ∥ op ≤ c 2 for constants 0 < c 1 ≤ c 2 < ∞ 0 < c_1 \le c_2 < \infty 0 < c 1 ≤ c 2 < ∞ independent of n n n .
Assumption 4 is condition (1.1) of Heiny (2022): the lower bound keeps each cross-sectional variance away from zero, and the operator-norm bound controls the largest eigenvalue of S S S . Under this condition Heiny (2022, Thm. 1.2)Thm. 1.2 shows that diag ( S ) \operatorname{diag}(S) diag ( S ) concentrates around the population diagonal for temporally independent data. Hounyo, Kao, and Kim [2 ] extend this result to temporal dependence, under finite fourth moments (Thm. 3) and, with the explicit rate O ( log n / T ) O(\sqrt{\log n/T}) O ( log n / T ) , under Gaussianity (Thm. 5); Lemma 4 below is the special case used here. Assumption 4 implies (i) ∥ R ∗ ∥ o p ≤ c 2 / c 1 < ∞ \lVert R_\ast\rVert_{\mathrm{op}} \le c_2/c_1 < \infty ∥ R ∗ ∥ op ≤ c 2 / c 1 < ∞ ; (ii) ϕ k = O ( 1 ) \phi_k = O(1) ϕ k = O ( 1 ) for each k k k ; and (iii) T r ( R ∗ 2 ) − n = O ( n ) Tr(R_\ast^2) - n = O(n) T r ( R ∗ 2 ) − n = O ( n ) .
Moments Under the Alternative
The first step extends the null diagonal-concentration result to general U U U . Under H A H_A H A , E [ S i i ] = U i i \mathbb{E}[S_{ii}] = U_{ii} E [ S ii ] = U ii , and the diagonal concentrates around diag ( U ) \operatorname{diag}(U) diag ( U ) .
Lemma 4 (Diagonal Concentration Under the Alternative). Under Assumptions 1–4,
∥ diag ( S ) − diag ( U ) ∥ o p = O a . s . ( log n T ) . \lVert\operatorname{diag}(S) - \operatorname{diag}(U)\rVert_{op} = O_{\mathrm{a.s.}}\left(\sqrt{\frac{\log n}{T}}\right). ∥ diag ( S ) − diag ( U ) ∥ o p = O a.s. ( T log n ) . The proof is in Section S.A.3 of the Online Supplement.
Under the separable Gaussian model, for each i i i ,
S i i U i i = d 1 T z i ′ Γ z i , z i ∼ N ( 0 , I T ) , \frac{S_{ii}}{U_{ii}} \mathrel{\overset{d}{=}} \frac{1}{T}z_i^\prime\Gamma z_i,\qquad z_i \sim N(0,I_T), U ii S ii = d T 1 z i ′ Γ z i , z i ∼ N ( 0 , I T ) ,
where = d \mathrel{\overset{d}{=}} = d denotes equality in distribution. Thus the standardized diagonal elements have the same marginal distribution as under the null. The uniform bounds on U i i U_{ii} U ii in Assumption 4 yield the same concentration rate as in Lemma 1.
Lemma 5 (Alternative Mean). Under Assumptions 1–4, the expectation of Tr ( R 2 ) \operatorname{Tr}(R^2) Tr ( R 2 ) under H A H_A H A satisfies
E A [ Tr ( R 2 ) ] = Tr ( R ∗ 2 ) + n − 1 λ τ 2 , T + 4 ( n − 1 ) λ 2 n ( τ 2 , T 2 − τ 3 , T ) − 3 τ 2 , T ( ϕ 2 − 1 ) λ + 2 τ 2 , T T ∑ i ≠ j ( ρ i j ∗ ) 4 + O ( n − 1 ) . (3) \mathbb{E}_A[\operatorname{Tr}(R^2)] = \operatorname{Tr}(R_*^2) + \frac{n-1}{\lambda}\tau_{2,T} + \frac{4(n-1)}{\lambda^2 n}\left(\tau_{2,T}^2-\tau_{3,T}\right) - \frac{3\tau_{2,T}(\phi_2-1)}{\lambda} + \frac{2\tau_{2,T}}{T}\sum_{i\ne j}(\rho_{ij}^*)^4 + O(n^{-1}). \tag*{(3)} E A [ Tr ( R 2 )] = Tr ( R ∗ 2 ) + λ n − 1 τ 2 , T + λ 2 n 4 ( n − 1 ) ( τ 2 , T 2 − τ 3 , T ) − λ 3 τ 2 , T ( ϕ 2 − 1 ) + T 2 τ 2 , T i = j ∑ ( ρ ij ∗ ) 4 + O ( n − 1 ) . (3)
Remark 7 In terms of μ n \mu_n μ n in (2), Lemma 5 reads
E A [ Tr ( R 2 ) ] = μ n ⏟ null mean + n ( ϕ 2 − 1 ) ⏟ leading signal − 3 τ 2 , T ( ϕ 2 − 1 ) λ + 2 τ 2 , T T ∑ i ≠ j ( ρ i j ∗ ) 4 + O ( n − 1 ) . \mathbb{E}_A[\operatorname{Tr}(R^2)] = \underbrace{\mu_n}_{\text{null mean}} + \underbrace{n(\phi_2-1)}_{\text{leading signal}} - \frac{3\tau_{2,T}(\phi_2-1)}{\lambda} + \frac{2\tau_{2,T}}{T}\sum_{i\ne j}(\rho_{ij}^*)^4 + O(n^{-1}). E A [ Tr ( R 2 )] = null mean μ n + leading signal n ( ϕ 2 − 1 ) − λ 3 τ 2 , T ( ϕ 2 − 1 ) + T 2 τ 2 , T i = j ∑ ( ρ ij ∗ ) 4 + O ( n − 1 ) .
Since ∑ i ≠ j ( ρ i j ∗ ) 4 ≤ n ( ϕ 2 − 1 ) \sum_{i\ne j}(\rho_{ij}^*)^4 \le n(\phi_2-1) ∑ i = j ( ρ ij ∗ ) 4 ≤ n ( ϕ 2 − 1 ) , the interaction terms vanish under H 0 H_0 H 0 and are O ( n − 1 ) O(n^{-1}) O ( n − 1 ) under local alternatives with n ( ϕ 2 − 1 ) = O ( 1 ) n(\phi_2-1)=O(1) n ( ϕ 2 − 1 ) = O ( 1 ) . Pair by pair, $\mathbb{E}_A(r_{ij}^2)=(\rho_{ij}^)^2+\tau_{2,T}(1-(\rho_{ij}^ )^2)(12 ( ρ i j ∗ ) 2 / T + O ( T − 2 ) 2(\rho^{*}_{ij})^2/T+O(T^{-2}) 2 ( ρ ij ∗ ) 2 / T + O ( T − 2 ) , which for Γ = I T \Gamma=I_T Γ = I T is the standard expansion for a bivariate Gaussian sample with known means. Since ( 1 − ρ 2 ) ( 1 − 2 ρ 2 ) = 1 − 3 ρ 2 + 2 ρ 4 (1-\rho^2)(1-2\rho^2)=1-3\rho^2+2\rho^4 ( 1 − ρ 2 ) ( 1 − 2 ρ 2 ) = 1 − 3 ρ 2 + 2 ρ 4 , summing the T − 1 T^{-1} T − 1 term over the n ( n − 1 ) n(n-1) n ( n − 1 ) ordered pairs gives ( n − 1 ) τ 2 , T / λ (n-1)\tau_{2,T}/\lambda ( n − 1 ) τ 2 , T / λ and the two interaction terms. The O ( T − 2 ) O(T^{-2}) O ( T − 2 ) term equals 4 ( τ 2 , T 2 − τ 3 , T ) / T 2 + O ( ( ρ i j ∗ ) 2 T − 2 ) + O ( T − 3 ) 4(\tau_{2,T}^2-\tau_{3,T})/T^2+O((\rho^{*}_{ij})^2T^{-2})+O(T^{-3}) 4 ( τ 2 , T 2 − τ 3 , T ) / T 2 + O (( ρ ij ∗ ) 2 T − 2 ) + O ( T − 3 ) , and summation over pairs gives the remainder O ( n − 1 ) O(n^{-1}) O ( n − 1 ) .
Remark 8. By Lemma 5, the standardized mean shift of L M LM L M is
E A [ Tr ( R 2 ) ] − μ n σ n = Δ n + O ( ϕ 2 − 1 ) + O ( n − 1 ) , Δ n = n ( ϕ 2 − 1 ) σ n = λ n ( ϕ 2 − 1 ) 2 τ 2 , T 2 + 2 K T / λ . \frac{\mathbb{E}_A[\operatorname{Tr}(R^2)]-\mu_n}{\sigma_n}
=\Delta_n+O(\phi_2-1)+O(n^{-1}),\qquad
\Delta_n=\frac{n(\phi_2-1)}{\sigma_n}
=\frac{\lambda n(\phi_2-1)}{2\sqrt{\tau_{2,T}^2+2K_T/\lambda}}. σ n E A [ Tr ( R 2 )] − μ n = Δ n + O ( ϕ 2 − 1 ) + O ( n − 1 ) , Δ n = σ n n ( ϕ 2 − 1 ) = 2 τ 2 , T 2 + 2 K T / λ λn ( ϕ 2 − 1 ) .
and Theorem 2 shows that Δ n \Delta_n Δ n is the noncentrality parameter of L M LM L M under local alternatives. Temporal dependence reduces power by inflating σ n \sigma_n σ n through both τ 2 , T \tau_{2,T} τ 2 , T and K T K_T K T . Using only the first variance term, i.e. 2 τ 2 , T / λ 2\tau_{2,T}/\lambda 2 τ 2 , T / λ as the standard deviation, would overstate Δ n \Delta_n Δ n by the factor 1 + 2 K T / ( λ τ 2 , T 2 ) \sqrt{1+2K_T/(\lambda\tau_{2,T}^2)} 1 + 2 K T / ( λ τ 2 , T 2 ) . For an A R ( 1 ) AR(1) A R ( 1 ) process with persistence ρ \rho ρ , τ 2 , T → τ 2 = ( 1 + ρ 2 ) / ( 1 − ρ 2 ) \tau_{2,T}\to\tau_2=(1+\rho^2)/(1-\rho^2) τ 2 , T → τ 2 = ( 1 + ρ 2 ) / ( 1 − ρ 2 ) (see (1) ) and K T → K = 2 ρ 2 ( 1 + ρ 2 ) / ( 1 − ρ 2 ) 3 K_T\to K=2\rho^2(1+\rho^2)/(1-\rho^2)^3 K T → K = 2 ρ 2 ( 1 + ρ 2 ) / ( 1 − ρ 2 ) 3 , so that
σ n 2 → 4 ( 1 + ρ 2 ) 2 ( 1 − ρ 2 ) 2 λ 2 + 16 ρ 2 ( 1 + ρ 2 ) ( 1 − ρ 2 ) 3 λ 3 . \sigma_n^2\to\frac{4(1+\rho^2)^2}{(1-\rho^2)^2\lambda^2}+\frac{16\rho^2(1+\rho^2)}{(1-\rho^2)^3\lambda^3}. σ n 2 → ( 1 − ρ 2 ) 2 λ 2 4 ( 1 + ρ 2 ) 2 + ( 1 − ρ 2 ) 3 λ 3 16 ρ 2 ( 1 + ρ 2 ) .
At λ = 1 \lambda=1 λ = 1 , the limiting factor 1 + 2 K / ( λ τ 2 2 ) \sqrt{1+2K/(\lambda\tau_2^2)} 1 + 2 K / ( λ τ 2 2 ) is about 1.29 for ρ = 0.4 \rho=0.4 ρ = 0.4 and 1.89 for ρ = 0.7 \rho=0.7 ρ = 0.7 (about 1.72 and 2.96 at λ = 1 / 3 \lambda=1/3 λ = 1/3 ). Within the local-alternative regime, the leading power depends on the alternative through the aggregate signal n ( ϕ 2 − 1 ) = ∑ i ≠ j ( ρ i j ∗ ) 2 n(\phi_2-1)=\sum_{i\ne j}(\rho^{*}_{ij})^2 n ( ϕ 2 − 1 ) = ∑ i = j ( ρ ij ∗ ) 2 . Many weak correlations are therefore detected with power tending to one once the aggregate signal diverges, whereas a fixed number of pairs with fixed correlations gives n ( ϕ 2 − 1 ) = O ( 1 ) n(\phi_2-1)=O(1) n ( ϕ 2 − 1 ) = O ( 1 ) and power bounded away from one.
Lemma 6 (Alternative Variance). Under Assumptions 1–4, Var A ( Tr ( R 2 ) ) = O ( 1 ) \operatorname{Var}_A(\operatorname{Tr}(R^2))=O(1) Var A ( Tr ( R 2 )) = O ( 1 ) . If in addition n ( ϕ 2 − 1 ) = O ( 1 ) n(\phi_2-1)=O(1) n ( ϕ 2 − 1 ) = O ( 1 ) , then
Var A ( Tr ( R 2 ) ) = 4 τ 2 , T 2 λ 2 + 8 K T λ 3 + O ( n − 1 ) = σ n 2 + O ( n − 1 ) . (4) \operatorname{Var}_A(\operatorname{Tr}(R^2))
=\frac{4\tau_{2,T}^2}{\lambda^2}+\frac{8K_T}{\lambda^3}+O(n^{-1})
=\sigma_n^2+O(n^{-1}). \tag*{(4)} Var A ( Tr ( R 2 )) = λ 2 4 τ 2 , T 2 + λ 3 8 K T + O ( n − 1 ) = σ n 2 + O ( n − 1 ) . (4) where σ n 2 \sigma_n^2 σ n 2 is the null variance in (2) .
Remark 9 The first statement holds uniformly over alternatives satisfying the stated assumptions and is sufficient, together with a diverging mean shift, for consistency; it follows from the Gaussian Poincaré inequality (Lemma S.7 in the Online Supplement). The sharper local variance statement follows from the pairwise covariance expansion in Lemma S.9 in the Online Supplement, combined with Lemma 3.
Remark 10 (Persistence of the null variance) Under local alternatives, the aggregate signal n ( ϕ 2 − 1 ) = ∑ i ≠ j ( ρ i j ∗ ) 2 n(\phi_2-1)=\sum_{i\ne j}(\rho_{ij}^{*})^2 n ( ϕ 2 − 1 ) = ∑ i = j ( ρ ij ∗ ) 2 is O ( 1 ) O(1) O ( 1 ) . It shifts the mean of Tr ( R 2 ) \operatorname{Tr}(R^2) Tr ( R 2 ) by an O ( 1 ) O(1) O ( 1 ) amount, whereas Lemma S.9 shows that its variance changes by only O ( n − 1 ) O(n^{-1}) O ( n − 1 ) :
Var A { Tr ( R 2 ) } − Var 0 { Tr ( R 2 ) } = O ( n − 1 ) . \operatorname{Var}_{A}\{\operatorname{Tr}(R^2)\}-\operatorname{Var}_{0}\{\operatorname{Tr}(R^2)\}=O(n^{-1}). Var A { Tr ( R 2 )} − Var 0 { Tr ( R 2 )} = O ( n − 1 ) .
Thus both components of the null variance, 4 τ 2 , T 2 / λ 2 4\tau_{2,T}^2/\lambda^2 4 τ 2 , T 2 / λ 2 and 8 K T / λ 3 8K_T/\lambda^3 8 K T / λ 3 , remain the leading components under local alternatives.
The distinction reflects how cross-sectional correlations enter the two moments. The leading mean shift sums ( ρ i j ∗ ) 2 (\rho_{ij}^{*})^2 ( ρ ij ∗ ) 2 directly. By contrast, variance corrections involve covariances of squared sample correlations, including pairs that share a unit and pairs that are disjoint. Sign invariance restricts the correlation monomials in these covariances: each index must have even total degree, although this condition alone does not determine their asymptotic order. The uniform coefficient estimates and graph summation in Lemma S.9 are what establish the aggregate O ( n − 1 ) O(n^{-1}) O ( n − 1 ) bound.
A single pair shows the mechanism. Each ( ρ i j ∗ ) 2 (\rho_{ij}^{*})^2 ( ρ ij ∗ ) 2 enters E A ( r i j 2 ) \mathbb{E}_{A}(r_{ij}^2) E A ( r ij 2 ) with coefficient one, but the variance of r i j 2 r_{ij}^2 r ij 2 is already of order T − 1 T^{-1} T − 1 . A delta-method expansion gives Var A ( r i j ) = ( 1 − ( ρ i j ∗ ) 2 ) τ 2 , T / T + O ( T − 2 ) \operatorname{Var}_{A}(r_{ij})=(1-(\rho_{ij}^{*})^2)\tau_{2,T}/T+O(T^{-2}) Var A ( r ij ) = ( 1 − ( ρ ij ∗ ) 2 ) τ 2 , T / T + O ( T − 2 ) , so the leading term of Var A ( r i j 2 ) − Var 0 ( r i j 2 ) \operatorname{Var}_{A}(r_{ij}^2)-\operatorname{Var}_{0}(r_{ij}^2) Var A ( r ij 2 ) − Var 0 ( r ij 2 ) is 4 ( ρ i j ∗ ) 2 ( 1 − ( ρ i j ∗ ) 2 ) τ 2 , T / T 4(\rho_{ij}^{*})^2(1-(\rho_{ij}^{*})^2)\tau_{2,T}/T 4 ( ρ ij ∗ ) 2 ( 1 − ( ρ ij ∗ ) 2 ) τ 2 , T / T . Summed over pairs, these terms contribute at most of order τ 2 , T n ( ϕ 2 − 1 ) / T \tau_{2,T}n(\phi_2-1)/T τ 2 , T n ( ϕ 2 − 1 ) / T to Var A { Tr ( R 2 ) } \operatorname{Var}_{A}\{\operatorname{Tr}(R^2)\} Var A { Tr ( R 2 )} , which is O ( n − 1 ) O(n^{-1}) O ( n − 1 ) under local alternatives, whereas the same signal shifts the mean by n ( ϕ 2 − 1 ) n(\phi_2-1) n ( ϕ 2 − 1 ) . Equivalently, the coupling in Lemma S.8 writes the alternative statistic as the null statistic plus the constant n ( ϕ 2 − 1 ) n(\phi_2-1) n ( ϕ 2 − 1 ) and an L 2 L^2 L 2 -negligible remainder.
For fixed alternatives with ϕ 2 − 1 \phi_2-1 ϕ 2 − 1 bounded away from zero, ∑ i ≠ j ( ρ i j ∗ ) 2 \sum_{i\ne j}(\rho_{ij}^{*})^2 ∑ i = j ( ρ ij ∗ ) 2 is of order n n n , so the local variance comparison need not apply. The local-power approximation below is therefore stated under n ( ϕ 2 − 1 ) = O ( 1 ) n(\phi_{2}-1)=O(1) n ( ϕ 2 − 1 ) = O ( 1 ) .
Consistency and Local Power
Proposition 1 (Consistency Under Fixed Alternatives). Under Assumptions 1–4, if ϕ 2 − 1 \phi_{2}-1 ϕ 2 − 1 is bounded away from zero, then P A ( ∣ L M ∣ > z α / 2 ) → 1 P_{A}(|LM|>z_{\alpha/2})\to1 P A ( ∣ L M ∣ > z α /2 ) → 1 .
Lemma 5 gives a mean shift of order n n n under the stated fixed-signal condition, while Lemma 6 bounds the alternative variance by O ( 1 ) O(1) O ( 1 ) . Chebyshev’s inequality therefore establishes the result. The following theorem treats local alternatives, for which the mean shift remains bounded.
Theorem 2 (CLT Under Local Alternatives). Under Assumptions 1–4, suppose that n ( ϕ 2 − 1 ) = O ( 1 ) n(\phi_{2}-1)=O(1) n ( ϕ 2 − 1 ) = O ( 1 ) . Then
L M − Δ n → d N ( 0 , 1 ) , Δ n = n ( ϕ 2 − 1 ) σ n . LM-\Delta_{n}\xrightarrow{d}N(0,1), \qquad\Delta_{n}=\frac{n(\phi_{2}-1)}{\sigma_{n}}. L M − Δ n d N ( 0 , 1 ) , Δ n = σ n n ( ϕ 2 − 1 ) .
The next result, proved in Section S.E of the Online Supplement, gives the signal scale n ( ϕ 2 − 1 ) n(\phi_2-1) n ( ϕ 2 − 1 ) at which the LM test has nontrivial asymptotic power.
Corollary 1 (Detection Boundary). Under Assumptions 1–4:
(a) the LM test has power tending to one if and only if n ( ϕ 2 − 1 ) → ∞ n(\phi_2-1) \to\infty n ( ϕ 2 − 1 ) → ∞ , equivalently,
Tr ( R ∗ 2 ) − n σ n → ∞ ; \frac{\operatorname{Tr}(R_*^2)-n}{\sigma_n} \to\infty; σ n Tr ( R ∗ 2 ) − n → ∞ ; (b) the test has asymptotic power equal to α \alpha α if and only if n ( ϕ 2 − 1 ) → 0 n(\phi_2-1) \to0 n ( ϕ 2 − 1 ) → 0 ;
(c) if n ( ϕ 2 − 1 ) / σ n → Δ ∈ ( 0 , ∞ ) n(\phi_2-1)/\sigma_n \to\Delta\in(0,\infty) n ( ϕ 2 − 1 ) / σ n → Δ ∈ ( 0 , ∞ ) , the power converges to Φ ( Δ − z α / 2 ) + Φ ( − Δ − z α / 2 ) ∈ ( α , 1 ) \Phi(\Delta-z_{\alpha/2})+\Phi(-\Delta-z_{\alpha/2}) \in(\alpha,1) Φ ( Δ − z α /2 ) + Φ ( − Δ − z α /2 ) ∈ ( α , 1 ) .
Section 5 reports the finite-sample power of the test under factor and spatial alternatives (Table 3 and Table 4 ).
Table 3. Size-unadjusted empirical power at the 5% nominal level (ρ = 0.0 \rho= 0.0 ρ = 0.0 )
Table 4. Size-unadjusted empirical power at the 5% nominal level (ρ = 0.4 \rho= 0.4 ρ = 0.4 ). Notes: L M i n f LM_{\mathrm{inf}} L M inf denotes the oracle LM test. L M U T LM_{UT} L M U T and L M T LM_T L M T denote the feasible LM tests using untapered and tapered autocovariance estimators, respectively. The parameters κ \kappa κ and ψ \psi ψ control the strength of cross-sectional dependence in the factor and SAR(1) models, respectively. Schott’s test is excluded because it is not robust to temporal dependence.
Monte Carlo Simulation
This section examines the finite-sample size and power of the proposed test and compares it with Schott’s (2005) test L M S LM_S L M S . All experiments use 5,000 Monte Carlo replications and the 5% nominal level.
Design and implementation
Under the null hypothesis, the data are generated as
x i t = ρ x i t − 1 + 1 − ρ 2 v i t , x_{it} = \rho x_{it-1} + \sqrt{1-\rho^2}v_{it}, x i t = ρ x i t − 1 + 1 − ρ 2 v i t ,
where x i 0 x_{i0} x i 0 and v i t v_{it} v i t are i.i.d. draws from one of the following distributions: (i) N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) , (ii) χ 2 ( 1 ) − 1 \chi^2(1)-1 χ 2 ( 1 ) − 1 , and (iii) t ( 5 ) t(5) t ( 5 ) . Under the null hypothesis, the rows x 1 , … , x n x_1,\ldots,x_n x 1 , … , x n are cross-sectionally independent AR(1) processes with autoregressive parameter ρ ∈ { 0 , 0.4 , 0.7 } \rho\in\{0,0.4,0.7\} ρ ∈ { 0 , 0.4 , 0.7 } , and the temporal covariance matrix is the Toeplitz matrix Γ t s = ρ ∣ t − s ∣ \Gamma_{ts}=\rho^{\lvert t-s\rvert} Γ t s = ρ ∣ t − s ∣ , whose limiting spectral moments are given in (1) . Since x i 0 x_{i0} x i 0 and v i t v_{it} v i t have a common variance, each row has covariance matrix proportional to Γ \Gamma Γ ; because R R R is invariant to rescaling the rows, the innovations need not be standardized. We consider n , T ∈ { 100 , 200 , 300 } n,T \in\{100,200,300\} n , T ∈ { 100 , 200 , 300 } .
We compute three versions of the proposed test. The oracle test L M i n f LM_{\mathrm{inf}} L M inf uses μ n \mu_n μ n and σ n 2 \sigma_n^2 σ n 2 in (2) , evaluated at the finite-T T T spectral moments τ k , T \tau_{k,T} τ k , T , k = 2 , 3 , 4 k=2,3,4 k = 2 , 3 , 4 , of the data-generating Γ \Gamma Γ . The feasible tests L M U T LM_{UT} L M U T and L M T LM_T L M T replace τ k , T \tau_{k,T} τ k , T by plug-in estimates based on the pooled untapered and tapered autocovariance estimators of Hounyo, Kao, and Kim [2 ] , normalized as in § 2 .
The tapered estimator is defined as
Γ ^ K = 1 n ∑ i = 1 n Γ ^ i , K \widehat{\Gamma}_{K}=\frac{1}{n}\sum_{i=1}^{n}\widehat{\Gamma}_{i,K} Γ K = n 1 i = 1 ∑ n Γ i , K
where Γ ^ i , K \widehat{\Gamma}_{i,K} Γ i , K is the T × T T \times T T × T Toeplitz matrix whose h h h -th diagonal elements, for h = 0 , … , T − 1 h=0,\ldots,T-1 h = 0 , … , T − 1 , are given by
K ( h m T ) 1 T ∑ t = 1 T − h x i t x i , t + h . \mathcal{K}\left(\frac{h}{m_T}\right)\frac{1}{T}\sum_{t=1}^{T-h}x_{it}x_{i,t+h}. K ( m T h ) T 1 t = 1 ∑ T − h x i t x i , t + h .
Here K ( ⋅ ) \mathcal{K}(\cdot) K ( ⋅ ) is a kernel function and m T m_T m T is the bandwidth parameter. In our simulations, we employ the rectangular kernel, K ( z ) = 1 { ∣ z ∣ ≤ 1 } \mathcal{K}(z)=1\{|z|\le1\} K ( z ) = 1 { ∣ z ∣ ≤ 1 } , which exhibited substantially better size control than the Bartlett and Parzen kernels in additional simulations not reported here.
The untapered estimator corresponds to the full-bandwidth rectangular-kernel case with m T = T m_T=T m T = T .
For bandwidth selection in the tapered estimator, we adopt the following data-driven rule from Hounyo et al. [2 ] :
m ^ T = max { 1 , min { T − 1 , ⌊ m ~ T ⌋ } } . \widehat{m}_T=\max\{1,\min\{T-1,\lfloor\widetilde{m}_T\rfloor\}\}. m T = max { 1 , min { T − 1 , ⌊ m T ⌋}} .
where
m ~ T = 0.8 { α ^ 2 n T log ( n T ) } 1 / 3 α ^ = 2 ∑ h = 1 l h ∣ γ ^ ( h ) ∣ ∣ γ ^ ( 0 ) ∣ + 2 ∑ h = 1 l ∣ γ ^ ( h ) ∣ γ ^ ( h ) = 1 n T ∑ i = 1 n ∑ t = 1 T − h x i t x i , t + h . \widetilde{m}_T=0.8\left\{\frac{\widehat{\alpha}^{2}nT}{\log(nT)}\right\}^{1/3}
\qquad
\widehat{\alpha}=\frac{2\sum_{h=1}^{l}h|\widehat{\gamma}(h)|}{|\widehat{\gamma}(0)|+2\sum_{h=1}^{l}|\widehat{\gamma}(h)|}
\qquad
\widehat{\gamma}(h)=\frac{1}{nT}\sum_{i=1}^{n}\sum_{t=1}^{T-h}x_{it}x_{i,t+h}. m T = 0.8 { log ( n T ) α 2 n T } 1/3 α = ∣ γ ( 0 ) ∣ + 2 ∑ h = 1 l ∣ γ ( h ) ∣ 2 ∑ h = 1 l h ∣ γ ( h ) ∣ γ ( h ) = n T 1 i = 1 ∑ n t = 1 ∑ T − h x i t x i , t + h .
with l = min { T − 1 , ⌊ ( n T ) 1 / 5 ⌋ } l=\min\{T-1,\lfloor(nT)^{1/5}\rfloor\} l = min { T − 1 , ⌊( n T ) 1/5 ⌋} . This rule was developed for the Bartlett kernel; we use it for the rectangular kernel as a practical choice.
Size
Table 1 reports the empirical sizes. We summarize four findings.
Schott’s test. L M S LM_S L M S is accurately sized when ρ = 0 \rho=0 ρ = 0 , but its empirical size is 1.000 for ρ ∈ { 0.4 , 0.7 } \rho\in\{0.4,0.7\} ρ ∈ { 0.4 , 0.7 } in every design and error distribution.
Oracle test. The L M inf LM_{\inf} L M i n f test provides accurate empirical sizes with or without temporal dependence, as Theorem 1 predicts. Its slight under-rejection under strong persistence reflects the O ( n − 1 ) O(n^{-1}) O ( n − 1 ) term in Remark 4: both variance components fall short of their leading approximations by a relative amount of order T − 1 T^{-1} T − 1 . For ρ = 0.4 \rho=0.4 ρ = 0.4 , the second-order terms of a T a_T a T and b T b_T b T in Remark 4 are − 7.7 % -7.7\% − 7.7% for a T a_T a T and − 9.6 % -9.6\% − 9.6% for b T b_T b T at T = 100 T=100 T = 100 , and − 3.9 % -3.9\% − 3.9% and − 4.8 % -4.8\% − 4.8% at T = 200 T = 200 T = 200 ; for ρ = 0.7 \rho= 0.7 ρ = 0.7 they are about − 23 % -23\% − 23% at T = 100 T = 100 T = 100 . The leading formula therefore overstates Var ( Tr ( R 2 ) ) \operatorname{Var}(\operatorname{Tr}(R^{2})) Var ( Tr ( R 2 )) in small samples. Under normal innovations, the largest deviations from the nominal level are at ρ = 0.7 \rho= 0.7 ρ = 0.7 and ( n , T ) = ( 200 , 100 ) (n,T) = (200, 100) ( n , T ) = ( 200 , 100 ) and ( 100 , 100 ) (100, 100) ( 100 , 100 ) , where the empirical sizes are 0.030 and 0.031. When ρ = 0 \rho= 0 ρ = 0 , L M i n f LM_{\mathrm{inf}} L M inf and L M S LM_{S} L M S have nearly identical sizes.
Non-Gaussian innovations. Theorem 1 relies on Gaussianity. Under the t ( 5 ) t(5) t ( 5 ) distribution, the empirical sizes of L M i n f LM_{\mathrm{inf}} L M inf remain close to those under normality. Under the χ 2 ( 1 ) − 1 \chi^{2}(1)^{-1} χ 2 ( 1 ) − 1 distribution, however, both L M i n f LM_{\mathrm{inf}} L M inf and L M S LM_{S} L M S over-reject even without temporal dependence (empirical sizes between 0.059 and 0.092 at ρ = 0 \rho= 0 ρ = 0 ), consistent with the fourth-cumulant term that enters the variance of Tr ( R 2 ) \operatorname{Tr}(R^{2}) Tr ( R 2 ) for non-Gaussian data (Gao et al., 2017). These distortions become larger when temporal dependence is strong and T T T is relatively small. For example, when ρ = 0.7 \rho= 0.7 ρ = 0.7 and ( n , T ) = ( 300 , 100 ) (n,T) = (300, 100) ( n , T ) = ( 300 , 100 ) , the empirical size of L M i n f LM_{\mathrm{inf}} L M inf reaches 0.135 under the χ 2 ( 1 ) − 1 \chi^{2}(1)^{-1} χ 2 ( 1 ) − 1 distribution, compared with 0.034 under normality. These distortions diminish as T T T increases.
Feasible tests. When ρ = 0 \rho= 0 ρ = 0 , L M U T LM_{UT} L M U T and L M T LM_{T} L M T have empirical sizes close to the nominal level. In the presence of temporal dependence and under normal or t ( 5 ) t(5) t ( 5 ) innovations, both tests are markedly conservative: across the designs with ρ ∈ { 0.4 , 0.7 } \rho\in\{0.4, 0.7\} ρ ∈ { 0.4 , 0.7 } their empirical sizes range from 0.003 to 0.035, except at ( n , T ) = ( 300 , 100 ) (n,T) = (300, 100) ( n , T ) = ( 300 , 100 ) with ρ = 0.7 \rho= 0.7 ρ = 0.7 , where they lie between 0.048 and 0.103. This conservativeness does not diminish when n n n and T T T grow together. For example, when ρ = 0.4 \rho= 0.4 ρ = 0.4 and the innovations are normally distributed, the empirical sizes of L M U T LM_{UT} L M U T and L M T LM_{T} L M T are 0.008 and 0.014 at ( n , T ) = ( 100 , 100 ) (n,T) = (100, 100) ( n , T ) = ( 100 , 100 ) and 0.012 and 0.018 at ( n , T ) = ( 300 , 300 ) (n,T) = (300, 300) ( n , T ) = ( 300 , 300 ) . It diminishes, however, as T T T grows with n n n fixed. The tapered test is generally less conservative than the untapered test. Under the χ 2 ( 1 ) − 1 \chi^{2}(1)^{-1} χ 2 ( 1 ) − 1 distribution with strong temporal dependence and T T T small relative to n n n , the feasible tests instead over-reject substantially: when ρ = 0.7 \rho= 0.7 ρ = 0.7 and ( n , T ) = ( 300 , 100 ) (n,T) = (300, 100) ( n , T ) = ( 300 , 100 ) , the empirical sizes of L M U T LM_{UT} L M U T and L M T LM_{T} L M T are 0.540 and 0.583, respectively; these over-rejections diminish as T T T increases.
Table 2 separates the effects of estimating the centering and the scaling terms, for ρ = 0.4 \rho= 0.4 ρ = 0.4 and normal innovations. The hybrid tests L M T , μ LM_{T,\mu} L M T , μ and L M U T , μ LM_{UT,\mu} L M U T , μ , which use the true centering and estimated scaling, have empirical sizes close to those of L M inf LM_{\inf} L M i n f . The hybrid tests L M T , σ LM_{T,\sigma} L M T , σ and L M U T , σ LM_{UT,\sigma} L M U T , σ , which use the estimated centering and true scaling, remain conservative, with rejection frequencies similar to those of L M T LM_T L M T and L M U T LM_{UT} L M U T . Estimation of the centering term, rather than the scaling term, is therefore primarily responsible for the under-rejection.
Table 2. Empirical size of hybrid LM tests at the 5% nominal level (ρ = 0.4 \rho= 0.4 ρ = 0.4 , N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) )
A first-order expansion of the normalized spectral-moment estimator gives
τ ^ 2 , T − τ 2 , T ≈ 2 n T ∑ i = 1 n D i , D i = Z i ′ ( Γ 2 − τ 2 , T Γ ) Z i . \widehat{\tau}_{2,T}-\tau_{2,T}\approx\frac{2}{nT}\sum_{i=1}^{n}D_i,\qquad D_i=Z_i^{\prime}(\Gamma^2-\tau_{2,T}\Gamma)Z_i. τ 2 , T − τ 2 , T ≈ n T 2 i = 1 ∑ n D i , D i = Z i ′ ( Γ 2 − τ 2 , T Γ ) Z i .
Because the leading null centering contains ( n − 1 ) τ 2 , T / λ (n-1)\tau_{2,T}/\lambda ( n − 1 ) τ 2 , T / λ , its plug-in estimation error is approximately
n − 1 λ ( τ ^ 2 , T − τ 2 , T ) ≈ 2 ( n − 1 ) T 2 ∑ i = 1 n D i . \frac{n-1}{\lambda}(\widehat{\tau}_{2,T}-\tau_{2,T})\approx\frac{2(n-1)}{T^2}\sum_{i=1}^{n}D_i. λ n − 1 ( τ 2 , T − τ 2 , T ) ≈ T 2 2 ( n − 1 ) i = 1 ∑ n D i .
This matches, to first order, the first-order Hoeffding projection of T r ( R 2 ) Tr(R^2) T r ( R 2 ) that generates the additional oracle variance 8 K T / λ 3 8K_T/\lambda^3 8 K T / λ 3 . Thus, estimating τ 2 , T \tau_{2,T} τ 2 , T from the same observations may absorb that projection, leaving a plug-in-centered fluctuation with variance closer to the degenerate component 4 τ 2 , T 2 / λ 2 4\tau_{2,T}^2/\lambda^2 4 τ 2 , T 2 / λ 2 than to the full oracle variance. Standardizing by the full oracle variance would then make the feasible tests conservative. This argument is heuristic.
Power
We consider two alternative specifications of cross-sectional dependence: a factor model and a spatial autoregressive (SAR(1)) model.
For the factor model, the data are generated as
x i t = α i f t + ϵ i t , ϵ i t = ρ ϵ i , t − 1 + 1 − ρ 2 v i t . x_{it}=\alpha_i f_t+\epsilon_{it},\qquad\epsilon_{it}=\rho\epsilon_{i,t-1}+\sqrt{1-\rho^2}v_{it}. x i t = α i f t + ϵ i t , ϵ i t = ρ ϵ i , t − 1 + 1 − ρ 2 v i t .
where ϵ i 0 \epsilon_{i0} ϵ i 0 and v i t v_{it} v i t are i.i.d. N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) , α i ∼ i i d κ U ( 0 , 1 ) \alpha_i\mathrel{\overset{\mathrm{iid}}{\sim}}\kappa U(0,1) α i ∼ iid κ U ( 0 , 1 ) , and f t ∼ i i d N ( 0 , 1 ) f_t\mathrel{\overset{\mathrm{iid}}{\sim}}N(0,1) f t ∼ iid N ( 0 , 1 ) , with all random variables mutually independent. Thus, κ \kappa κ determines the strength of cross-sectional dependence induced by the common factor. Conditional on the factor loadings, the covariance matrix is Var ( vec ( X ) ) = I T ⊗ α α ′ + Γ ⊗ I n \operatorname{Var}(\operatorname{vec}(X)) = I_T \otimes\alpha\alpha^{\prime} + \Gamma\otimes I_n Var ( vec ( X )) = I T ⊗ α α ′ + Γ ⊗ I n , where Γ \Gamma Γ is the AR(1) temporal covariance matrix. This covariance structure is generally not separable when ρ ≠ 0 \rho\ne0 ρ = 0 . Moreover, the factor model generates a spiked cross-sectional covariance matrix, with ∥ U ∥ o p ≍ n \lVert U\rVert_{\mathrm{op}} \asymp n ∥ U ∥ op ≍ n for fixed κ > 0 \kappa> 0 κ > 0 , and therefore lies outside Assumption 4.
For the SAR(1) model, the data are generated as
x t = ψ W n x t + ε t , x t = ( x 1 t , … , x n t ) ′ , ε t = ( ε 1 t , … , ε n t ) ′ , ε i t = ρ ε i t − 1 + 1 − ρ 2 v i t , ε i 0 , v i t ∼ i i d N ( 0 , 1 ) , W n = [ w i j ] i j = 1 n is an ( n × n ) spatial contiguity matrix, \begin{aligned}
x_t &= \psi W_n x_t + \varepsilon_t,\qquad x_t = (x_{1t},\ldots,x_{nt})^{\prime},\qquad\varepsilon_t = (\varepsilon_{1t},\ldots,\varepsilon_{nt})^{\prime},\\
\varepsilon_{it} &= \rho\varepsilon_{it-1} + \sqrt{1-\rho^2}v_{it},\qquad\varepsilon_{i0},v_{it} \mathrel{\sim^{\mathrm{iid}}} N(0,1),\\
W_n &= [w_{ij}]_{ij=1}^{n}\text{ is an }(n\times n)\text{ spatial contiguity matrix,}
\end{aligned} x t ε i t W n = ψ W n x t + ε t , x t = ( x 1 t , … , x n t ) ′ , ε t = ( ε 1 t , … , ε n t ) ′ , = ρ ε i t − 1 + 1 − ρ 2 v i t , ε i 0 , v i t ∼ iid N ( 0 , 1 ) , = [ w ij ] ij = 1 n is an ( n × n ) spatial contiguity matrix,
where the spatial coefficient ψ \psi ψ determines the degree of cross-sectional dependence. In the SAR(1) DGP, the units are placed on a finite square integer lattice containing n n n sites. W n W_n W n is a contiguity matrix in which individuals i i i and j j j are considered to be neighbors if their L 1 L_1 L 1 distance equals 1. The matrix is row-standardized and has zero diagonal entries. This design satisfies Assumption 4 and the separable Gaussian representation.
No temporal dependence. Table 3 reports power for ρ = 0 \rho= 0 ρ = 0 . Since all four tests generally exhibit empirical sizes close to the nominal level under normality when ρ = 0 \rho= 0 ρ = 0 , we compare their rejection probabilities without size adjustment. Under the factor model, L M i n f LM_{\mathrm{inf}} L M inf exhibits power comparable to L M S LM_S L M S , with rejection probabilities increasing toward one as the sample sizes and/or the strength of cross-sectional dependence increase. The feasible tapered test L M T LM_T L M T performs almost as well as L M i n f LM_{\mathrm{inf}} L M inf , whereas the untapered test L M U T LM_{UT} L M U T exhibits power losses, particularly when cross-sectional dependence is relatively weak. Similar patterns emerge under the SAR(1) model: L M T LM_T L M T achieves power close to that of L M i n f LM_{\mathrm{inf}} L M inf and L M S LM_S L M S , while L M U T LM_{UT} L M U T tends to be less powerful.
Temporal dependence. Table 4 reports power for ρ = 0.4 \rho= 0.4 ρ = 0.4 . Because L M S LM_S L M S is invalid under temporal dependence, we report only the proposed tests. The feasible tests are conservative in this design (Table 1), so their unadjusted power is understated relative to tests with exact size. The power of L M i n f LM_{\mathrm{inf}} L M inf is generally lower than that reported in Table 3 , consistent with Remark 8: temporal dependence inflates σ n \sigma_n σ n and lowers Δ n \Delta_n Δ n . The rejection probabilities approach one as the strength of cross-sectional dependence increases and, in many configurations, as the sample sizes increase. The tapered test L M T LM_T L M T tends to reject more often than the untapered test L M U T LM_{UT} L M U T under both alternatives. Moreover, when the cross-sectional dependence is weaker (κ = 0.2 \kappa= 0.2 κ = 0.2 or ψ = 0.2 \psi= 0.2 ψ = 0.2 ), L M T LM_T L M T rejects more often than L M inf LM_{\inf} L M i n f in every configuration, although it is conservative under the null.
Accuracy of the local-alternative approximation. Theorem 2 rests on the coupling in Lemma S.8: under a local alternative, Tr ( R 2 ) = Tr ( R 0 2 ) + n ( ϕ 2 − 1 ) + D n \operatorname{Tr}(R^2) = \operatorname{Tr}(R_0^2) + n(\phi_2 - 1) + D_n Tr ( R 2 ) = Tr ( R 0 2 ) + n ( ϕ 2 − 1 ) + D n , where R 0 R_0 R 0 is the null sample correlation matrix built from the same innovations and E ( D n 2 ) → 0 \mathbb{E}(D_n^2) \to0 E ( D n 2 ) → 0 . To gauge the size of D n D_n D n , we generate X = R ∗ 1 / 2 Z Γ 1 / 2 X = R_*^{1/2}Z\Gamma^{1/2} X = R ∗ 1/2 Z Γ 1/2 and X 0 = Z Γ 1 / 2 X_0 = Z\Gamma^{1/2} X 0 = Z Γ 1/2 from the same matrix Z Z Z of i.i.d. N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) entries, where Γ \Gamma Γ is the AR(1) temporal covariance matrix with ρ = 0.4 \rho= 0.4 ρ = 0.4 and R ∗ = I n + E R_* = I_n + E R ∗ = I n + E is equicorrelated, with E i j = e n = 2 / n ( n − 1 ) E_{ij} = e_n = 2/\sqrt{n(n-1)} E ij = e n = 2/ n ( n − 1 ) for i ≠ j i \ne j i = j , so that n ( ϕ 2 − 1 ) = ∥ E ∥ F 2 = 4 n(\phi_2 - 1) = \lVert E\rVert_F^2 = 4 n ( ϕ 2 − 1 ) = ∥ E ∥ F 2 = 4 for every n n n . We take n = T ∈ { 100 , 200 , 400 } n = T \in\{100, 200, 400\} n = T ∈ { 100 , 200 , 400 } . The Monte Carlo mean of D n D_n D n is at most 0.21 in absolute value, and its standard deviations are 2.19, 1.63, and 1.16, respectively. The standard deviation falls at the rate n − 1 / 2 n^{-1/2} n − 1/2 (n \sqrt{n} n times the standard deviation is 21.9, 23.1, and 23.1), as the bound E ( D n 2 ) = O ( T − 1 ) \mathbb{E}(D_n^2) = O(T^{-1}) E ( D n 2 ) = O ( T − 1 ) in the proof of Lemma S.8 predicts. Relative to the null standard deviation σ n ≈ 3.5 \sigma_n \approx3.5 σ n ≈ 3.5 in this design, the approximation error is not negligible at n = T = 100 n = T = 100 n = T = 100 but shrinks steadily with the sample size.
Conclusion and Discussion
We develop a high-dimensional LM test of cross-sectional independence for Gaussian panels with a common stationary, short-memory temporal covariance matrix. Under proportional growth of the cross-sectional and time dimensions, we derive explicit expansions for the null mean and variance of Tr ( R 2 ) \operatorname{Tr}(R^2) Tr ( R 2 ) and establish its asymptotic normality. The results identify two distinct consequences of serial correlation: random diagonal normalization produces an O ( 1 ) O(1) O ( 1 ) centering correction, and squared sample correlations sharing a cross-sectional unit generate an additional O ( 1 ) O(1) O ( 1 ) variance component. Both vanish under temporal independence.
The proofs combine uniform concentration of sample variances, Gaussian quadraticform expansions, a Hoeffding decomposition, and a martingale CLT. Under alternatives, we establish consistency, a local Gaussian-shift limit, and a detection boundary for the proposed statistic. The local-power analysis separates the cross-sectional signal n ( ϕ 2 − 1 ) n(\phi_2 - 1) n ( ϕ 2 − 1 ) from the temporal spectral quantities determining the null variance.
The theoretical results are established under Gaussian innovations and a common stationary, short-memory temporal covariance structure. Extending the asymptotic theory to non-Gaussian innovations and heterogeneous temporal covariance structures, and establishing the limiting distribution of feasible statistics based on estimated temporal spectral moments, are natural directions for future research. The extension to fixed-effects panel regressions with serially correlated disturbances also requires accounting for the effects of within transformation and slope estimation.
Acknowledgments
During the preparation of this manuscript, the authors used Claude (Anthropic) and Chat-GPT (OpenAI) to assist with manuscript drafting and proofreading. No AI tool was used to write the simulation code or to produce the numerical results. The authors reviewed and edited all generated text and remain fully responsible for the final version.
Competing interests
The authors declare none.
Funding statement
This research received no specific grant from any funding agency, commercial or not-for-profit sectors.
Table 1. Empirical size at the 5% nominal level. Panel A: N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) and χ 2 ( 1 ) − 1 \chi^2(1)-1 χ 2 ( 1 ) − 1 innovations.
Table 1 (continued). Empirical size at the 5% nominal level. Panel B: t ( 5 ) t(5) t ( 5 ) innovations. Notes: L M i n f LM_{\mathrm{inf}} L M inf denotes the oracle LM test using the true spectral moments. L M U T LM_{UT} L M U T and L M T LM_T L M T denote the feasible LM tests using untapered and tapered autocovariance estimators, respectively. L M S LM_S L M S denotes Schott’s (2005) test. All results are based on 5,000 Monte Carlo replications.
Notes: L M i n f LM_{\mathrm{inf}} L M inf denotes the oracle LM test using the true spectral moments. L M U T LM_{UT} L M U T and L M T LM_T L M T denote the feasible LM tests using untapered and tapered autocovariance estimators, respectively. The hybrid tests L M T , μ LM_{T,\mu} L M T , μ and L M U T , μ LM_{UT,\mu} L M U T , μ use the true centering term and the estimated scaling term, whereas L M T , σ LM_{T,\sigma} L M T , σ and L M U T , σ LM_{UT,\sigma} L M U T , σ use the estimated centering term and the true scaling term. The subscripts T T T and U T UT U T indicate tapered and untapered estimation, respectively.
Notes: L M i n f LM_{\mathrm{inf}} L M inf denotes the oracle LM test. L M U T LM_{UT} L M U T and L M T LM_T L M T denote the feasible LM tests using untapered and tapered autocovariance estimators, respectively. L M S LM_S L M S denotes Schott’s (2005) test. The parameters κ \kappa κ and ψ \psi ψ control the strength of cross-sectional dependence in the factor and SAR(1) models, respectively.
This Appendix proves the null results: Lemma 2 (Appendix A), Lemma 3 (Appendix B), and Theorem 1 (Appendix C). The model, notation, Gaussian moment identities, and preliminary lemmas used here (Lemmas S.1–S.5) are in Section S.A of the Online Supplement, which also proves Lemmas 1 and 4. The proofs of the power results (Lemmas 5 and 6, Theorem 2, and Corollary 1) and the auxiliary inequalities are in the Online Supplement, whose sections, lemmas, and equations carry the prefix “S.”.
Proof of Lemma 2
Under H 0 H_0 H 0 the rows of Z Z Z are i.i.d., so all off-diagonal pairs are identically distributed; since r i i = 1 r_{ii}=1 r ii = 1 ,
E [ Tr ( R 2 ) ] = n + n ( n − 1 ) E ( r 12 2 ) . (5) \mathbb{E}\left[\operatorname{Tr}(R^2)\right] = n+n(n-1)\mathbb{E}(r_{12}^2). \tag*{(5)} E [ Tr ( R 2 ) ] = n + n ( n − 1 ) E ( r 12 2 ) . (5)
Throughout, S 12 = T − 1 Z 1 ′ Γ Z 2 S_{12}=T^{-1}Z_1^\prime\Gamma Z_2 S 12 = T − 1 Z 1 ′ Γ Z 2 , δ i = S i i − 1 \delta_i=S_{ii}-1 δ i = S ii − 1 , P 1 = Z 1 ′ Γ 2 Z 1 P_1=Z_1^\prime\Gamma^2Z_1 P 1 = Z 1 ′ Γ 2 Z 1 , d i = T δ i = Z i ′ Γ Z i − T d_i=T\delta_i=Z_i^\prime\Gamma Z_i-T d i = T δ i = Z i ′ Γ Z i − T , and t k = Tr ( Γ k ) = T τ k , T t_k=\operatorname{Tr}(\Gamma^k)=T\tau_{k,T} t k = Tr ( Γ k ) = T τ k , T , so that t k = O ( T ) t_k=O(T) t k = O ( T ) for each fixed k k k . Let E 12 = { ∣ δ 1 ∣ ≤ 1 / 2 , ∣ δ 2 ∣ ≤ 1 / 2 } \mathcal{E}_{12}=\{|\delta_1|\le1/2,|\delta_2|\le1/2\} E 12 = { ∣ δ 1 ∣ ≤ 1/2 , ∣ δ 2 ∣ ≤ 1/2 } .
Step 1: Expansion. On E 12 \mathcal{E}_{12} E 12 , Taylor’s formula for ( 1 + x ) − 1 (1+x)^{-1} ( 1 + x ) − 1 gives
1 ( 1 + δ 1 ) ( 1 + δ 2 ) = Π 3 ( δ 1 , δ 2 ) + Q 4 , ∣ Q 4 ∣ ≤ C ( ∣ δ 1 ∣ + ∣ δ 2 ∣ ) 4 . \frac{1}{(1+\delta_1)(1+\delta_2)}=\Pi_3(\delta_1,\delta_2)+Q_4,\qquad|Q_4|\le C(|\delta_1|+|\delta_2|)^4. ( 1 + δ 1 ) ( 1 + δ 2 ) 1 = Π 3 ( δ 1 , δ 2 ) + Q 4 , ∣ Q 4 ∣ ≤ C ( ∣ δ 1 ∣ + ∣ δ 2 ∣ ) 4 .
where
Π 3 ( δ 1 , δ 2 ) = 1 − ( δ 1 + δ 2 ) + ( δ 1 2 + δ 1 δ 2 + δ 2 2 ) − ( δ 1 3 + δ 1 2 δ 2 + δ 1 δ 2 2 + δ 2 3 ) . \Pi_3(\delta_1,\delta_2)=1-(\delta_1+\delta_2)+(\delta_1^2+\delta_1\delta_2+\delta_2^2)-(\delta_1^3+\delta_1^2\delta_2+\delta_1\delta_2^2+\delta_2^3). Π 3 ( δ 1 , δ 2 ) = 1 − ( δ 1 + δ 2 ) + ( δ 1 2 + δ 1 δ 2 + δ 2 2 ) − ( δ 1 3 + δ 1 2 δ 2 + δ 1 δ 2 2 + δ 2 3 ) .
We retain the cubic terms because bounding them in absolute value gives only O ( T − 5 / 2 ) O(T^{-5/2}) O ( T − 5/2 ) after multiplication by S 12 2 S_{12}^2 S 12 2 , whereas their expectations are O ( T − 3 ) O(T^{-3}) O ( T − 3 ) (Step 2). Since r 12 2 = S 12 2 Π 3 + S 12 2 Q 4 r_{12}^2=S_{12}^2\Pi_3+S_{12}^2Q_4 r 12 2 = S 12 2 Π 3 + S 12 2 Q 4 on E 12 \mathcal{E}_{12} E 12 ,
E ( r 12 2 ) = E ( S 12 2 Π 3 ) − E ( S 12 2 Π 3 1 E 12 c ) + E ( S 12 2 Q 4 1 E 12 ) + E ( r 12 2 1 E 12 c ) . \mathbb{E}(r_{12}^2)=\mathbb{E}(S_{12}^2\Pi_3)-\mathbb{E}(S_{12}^2\Pi_3 1_{\mathcal{E}_{12}^{c}})+\mathbb{E}(S_{12}^2Q_4 1_{\mathcal{E}_{12}})+\mathbb{E}(r_{12}^2 1_{\mathcal{E}_{12}^{c}}). E ( r 12 2 ) = E ( S 12 2 Π 3 ) − E ( S 12 2 Π 3 1 E 12 c ) + E ( S 12 2 Q 4 1 E 12 ) + E ( r 12 2 1 E 12 c ) .
By Hölder’s inequality and Lemma S.1, E { S 12 2 ( ∣ δ 1 ∣ + ∣ δ 2 ∣ ) 4 } ≤ C ∥ S 12 ∥ 2 2 ( ∥ δ 1 ∥ 8 + ∥ δ 2 ∥ 8 ) 4 = O ( T − 3 ) \mathbb{E}\{S_{12}^2(|\delta_1|+|\delta_2|)^4\}\le C\|S_{12}\|_2^2(\|\delta_1\|_8+\|\delta_2\|_8)^4=O(T^{-3}) E { S 12 2 ( ∣ δ 1 ∣ + ∣ δ 2 ∣ ) 4 } ≤ C ∥ S 12 ∥ 2 2 ( ∥ δ 1 ∥ 8 + ∥ δ 2 ∥ 8 ) 4 = O ( T − 3 ) . By the per-unit tail bound following the proof of Lemma 4, P ( E 12 c ) ≤ C e − c T P(\mathcal{E}_{12}^{c})\le Ce^{-cT} P ( E 12 c ) ≤ C e − c T . Since 0 ≤ r 12 2 ≤ 1 0 \le r_{12}^{2} \le1 0 ≤ r 12 2 ≤ 1 , E ( r 12 2 1 ε 12 c ) = O ( e − c T ) \mathbb{E}(r_{12}^{2}1_{\varepsilon_{12}^{c}})=O(e^{-cT}) E ( r 12 2 1 ε 12 c ) = O ( e − c T ) ; by Cauchy–Schwarz and ∥ S 12 2 Π 3 ∥ 2 = O ( T − 1 ) \lVert S_{12}^{2}\Pi_{3}\rVert_{2}=O(T^{-1}) ∥ S 12 2 Π 3 ∥ 2 = O ( T − 1 ) , E ( S 12 2 Π 3 1 ε 12 c ) = O ( e − c T ) \mathbb{E}(S_{12}^{2}\Pi_{3}1_{\varepsilon_{12}^{c}})=O(e^{-cT}) E ( S 12 2 Π 3 1 ε 12 c ) = O ( e − c T ) as well. Consequently,
E ( r 12 2 ) = E ( S 12 2 Π 3 ) + O ( T − 3 ) . (6) \mathbb{E}(r_{12}^{2})=\mathbb{E}(S_{12}^{2}\Pi_{3})+O(T^{-3}). \tag*{(6)} E ( r 12 2 ) = E ( S 12 2 Π 3 ) + O ( T − 3 ) . (6)
Step 2: Moments. Since all matrices involved are powers of Γ \Gamma Γ , the joint cumulants of Gaussian quadratic forms are κ ( Z ′ A 1 Z , … , Z ′ A r Z ) = 2 r − 1 ( r − 1 ) ! Tr ( A 1 ⋯ A r ) \kappa(Z^{\prime}A_{1}Z,\ldots,Z^{\prime}A_{r}Z)=2^{r-1}(r-1)!\operatorname{Tr}(A_{1}\cdots A_{r}) κ ( Z ′ A 1 Z , … , Z ′ A r Z ) = 2 r − 1 ( r − 1 )! Tr ( A 1 ⋯ A r ) ; for r = 2 , 3 r=2,3 r = 2 , 3 these follow from (S.A.3)–(S.A.4). Conditioning on Z 1 Z_{1} Z 1 , Lemma S.4 gives E ( S 12 2 ∣ Z 1 ) = P 1 / T 2 \mathbb{E}(S_{12}^{2}\mid Z_{1})=P_{1}/T^{2} E ( S 12 2 ∣ Z 1 ) = P 1 / T 2 , so the monomials in δ 1 \delta_{1} δ 1 alone reduce to E ( S 12 2 δ 1 k ) = E ( P 1 d 1 k ) / T k + 2 \mathbb{E}(S_{12}^{2}\delta_{1}^{k})=\mathbb{E}(P_{1}d_{1}^{k})/T^{k+2} E ( S 12 2 δ 1 k ) = E ( P 1 d 1 k ) / T k + 2 . Since d 1 d_{1} d 1 is centered, the moment–cumulant formula gives E ( P 1 d 1 ) = Cov ( P 1 , d 1 ) = 2 t 3 \mathbb{E}(P_{1}d_{1})=\operatorname{Cov}(P_{1},d_{1})=2t_{3} E ( P 1 d 1 ) = Cov ( P 1 , d 1 ) = 2 t 3 ,
E ( P 1 d 1 2 ) = E ( P 1 ) Var ( d 1 ) + κ ( P 1 , d 1 , d 1 ) = 2 t 2 2 + 8 t 4 , \mathbb{E}(P_{1}d_{1}^{2})=\mathbb{E}(P_{1})\operatorname{Var}(d_{1})+\kappa(P_{1},d_{1},d_{1})=2t_{2}^{2}+8t_{4}, E ( P 1 d 1 2 ) = E ( P 1 ) Var ( d 1 ) + κ ( P 1 , d 1 , d 1 ) = 2 t 2 2 + 8 t 4 ,
and
E ( P 1 d 1 3 ) = E ( P 1 ) κ 3 ( d 1 ) + 3 Cov ( P 1 , d 1 ) Var ( d 1 ) + κ ( P 1 , d 1 , d 1 , d 1 ) = 8 t 2 t 3 + 12 t 2 t 3 + 48 t 5 = 20 t 2 t 3 + 48 t 5 . \begin{aligned}
\mathbb{E}(P_{1}d_{1}^{3})&=\mathbb{E}(P_{1})\kappa_{3}(d_{1})+3\operatorname{Cov}(P_{1},d_{1})\operatorname{Var}(d_{1})+\kappa(P_{1},d_{1},d_{1},d_{1})\\
&=8t_{2}t_{3}+12t_{2}t_{3}+48t_{5}=20t_{2}t_{3}+48t_{5}.
\end{aligned} E ( P 1 d 1 3 ) = E ( P 1 ) κ 3 ( d 1 ) + 3 Cov ( P 1 , d 1 ) Var ( d 1 ) + κ ( P 1 , d 1 , d 1 , d 1 ) = 8 t 2 t 3 + 12 t 2 t 3 + 48 t 5 = 20 t 2 t 3 + 48 t 5 .
For the mixed monomials, conditionally on Z 1 Z_{1} Z 1 , Cov { ( Z 2 ′ Γ Z 1 ) 2 , Z 2 ′ Γ Z 2 ∣ Z 1 } = 2 Z 1 ′ Γ 3 Z 1 \operatorname{Cov}\{(Z_{2}^{\prime}\Gamma Z_{1})^{2},Z_{2}^{\prime}\Gamma Z_{2}\mid Z_{1}\}=2Z_{1}^{\prime}\Gamma^{3}Z_{1} Cov {( Z 2 ′ Γ Z 1 ) 2 , Z 2 ′ Γ Z 2 ∣ Z 1 } = 2 Z 1 ′ Γ 3 Z 1 , so E ( S 12 2 d 2 ∣ Z 1 ) = 2 Z 1 ′ Γ 3 Z 1 / T 2 \mathbb{E}(S_{12}^{2}d_{2}\mid Z_{1})=2Z_{1}^{\prime}\Gamma^{3}Z_{1}/T^{2} E ( S 12 2 d 2 ∣ Z 1 ) = 2 Z 1 ′ Γ 3 Z 1 / T 2 ; hence E ( S 12 2 δ 1 δ 2 ) = 2 E ( Z 1 ′ Γ 3 Z 1 d 1 ) / T 4 \mathbb{E}(S_{12}^{2}\delta_{1}\delta_{2})=2\mathbb{E}(Z_{1}^{\prime}\Gamma^{3}Z_{1}d_{1})/T^{4} E ( S 12 2 δ 1 δ 2 ) = 2 E ( Z 1 ′ Γ 3 Z 1 d 1 ) / T 4 and E ( S 12 2 δ 1 2 δ 2 ) = 2 E ( Z 1 ′ Γ 3 Z 1 d 1 2 ) / T 5 \mathbb{E}(S_{12}^{2}\delta_{1}^{2}\delta_{2})=2\mathbb{E}(Z_{1}^{\prime}\Gamma^{3}Z_{1}d_{1}^{2})/T^{5} E ( S 12 2 δ 1 2 δ 2 ) = 2 E ( Z 1 ′ Γ 3 Z 1 d 1 2 ) / T 5 . By symmetry in the two rows,
E ( S 12 2 ) = t 2 T 2 , E ( S 12 2 δ i ) = 2 t 3 T 3 , E ( S 12 2 δ i 2 ) = 2 t 2 2 + 8 t 4 T 4 , E ( S 12 2 δ 1 δ 2 ) = 4 t 4 T 4 , E ( S 12 2 δ i 3 ) = 20 t 2 t 3 + 48 t 5 T 5 , E ( S 12 2 δ 1 2 δ 2 ) = E ( S 12 2 δ 1 δ 2 2 ) = 4 t 2 t 3 + 16 t 5 T 5 . \begin{aligned}
\mathbb{E}(S_{12}^{2})&=\frac{t_{2}}{T^{2}}, &\qquad\mathbb{E}(S_{12}^{2}\delta_{i})&=\frac{2t_{3}}{T^{3}},\\
\mathbb{E}(S_{12}^{2}\delta_{i}^{2})&=\frac{2t_{2}^{2}+8t_{4}}{T^{4}}, &\qquad\mathbb{E}(S_{12}^{2}\delta_{1}\delta_{2})&=\frac{4t_{4}}{T^{4}},\\
\mathbb{E}(S_{12}^{2}\delta_{i}^{3})&=\frac{20t_{2}t_{3}+48t_{5}}{T^{5}}, &\qquad\mathbb{E}(S_{12}^{2}\delta_{1}^{2}\delta_{2})=\mathbb{E}(S_{12}^{2}\delta_{1}\delta_{2}^{2})&=\frac{4t_{2}t_{3}+16t_{5}}{T^{5}}.
\end{aligned} E ( S 12 2 ) E ( S 12 2 δ i 2 ) E ( S 12 2 δ i 3 ) = T 2 t 2 , = T 4 2 t 2 2 + 8 t 4 , = T 5 20 t 2 t 3 + 48 t 5 , E ( S 12 2 δ i ) E ( S 12 2 δ 1 δ 2 ) E ( S 12 2 δ 1 2 δ 2 ) = E ( S 12 2 δ 1 δ 2 2 ) = T 3 2 t 3 , = T 4 4 t 4 , = T 5 4 t 2 t 3 + 16 t 5 .
Since t k = O ( T ) t_{k}=O(T) t k = O ( T ) , every cubic monomial has expectation O ( T − 3 ) O(T^{-3}) O ( T − 3 ) uniformly.
Step 3: Assembly. Collecting the monomials of Π 3 \Pi_3 Π 3 by symmetry,
E ( S 12 2 Π 3 ) = E ( S 12 2 ) − 2 E ( S 12 2 δ 1 ) + { 2 E ( S 12 2 δ 1 2 ) + E ( S 12 2 δ 1 δ 2 ) } − { 2 E ( S 12 2 δ 1 3 ) + 2 E ( S 12 2 δ 1 2 δ 2 ) } = t 2 T 2 − 4 t 3 T 3 + 4 t 2 2 + 20 t 4 T 4 − 48 t 2 t 3 + 128 t 5 T 5 . \mathbb{E}(S_{12}^{2}\Pi_3)=\mathbb{E}(S_{12}^{2})-2\mathbb{E}(S_{12}^{2}\delta_1)+\{2\mathbb{E}(S_{12}^{2}\delta_1^{2})+\mathbb{E}(S_{12}^{2}\delta_1\delta_2)\}
-\{2\mathbb{E}(S_{12}^{2}\delta_1^{3})+2\mathbb{E}(S_{12}^{2}\delta_1^{2}\delta_2)\}
=\frac{t_2}{T^2}-\frac{4t_3}{T^3}+\frac{4t_2^2+20t_4}{T^4}-\frac{48t_2t_3+128t_5}{T^5}. E ( S 12 2 Π 3 ) = E ( S 12 2 ) − 2 E ( S 12 2 δ 1 ) + { 2 E ( S 12 2 δ 1 2 ) + E ( S 12 2 δ 1 δ 2 )} − { 2 E ( S 12 2 δ 1 3 ) + 2 E ( S 12 2 δ 1 2 δ 2 )} = T 2 t 2 − T 3 4 t 3 + T 4 4 t 2 2 + 20 t 4 − T 5 48 t 2 t 3 + 128 t 5 .
Substituting into (A.2),
E ( r 12 2 ) = t 2 T 2 − 4 t 3 T 3 + 4 t 2 2 + 20 t 4 T 4 + O ( T − 3 ) = τ 2 , T T + 4 ( τ 2 , T 2 − τ 3 , T ) T 2 + O ( T − 3 ) , \mathbb{E}(r_{12}^{2})=\frac{t_2}{T^2}-\frac{4t_3}{T^3}+\frac{4t_2^2+20t_4}{T^4}+O(T^{-3})=\frac{\tau_{2,T}}{T}+\frac{4(\tau_{2,T}^{2}-\tau_{3,T})}{T^2}+O(T^{-3}), E ( r 12 2 ) = T 2 t 2 − T 3 4 t 3 + T 4 4 t 2 2 + 20 t 4 + O ( T − 3 ) = T τ 2 , T + T 2 4 ( τ 2 , T 2 − τ 3 , T ) + O ( T − 3 ) ,
because 20 t 4 / T 4 20t_4/T^4 20 t 4 / T 4 and the cubic contribution are O ( T − 3 ) O(T^{-3}) O ( T − 3 ) . Insert this into (A.1) with T = n λ T=n\lambda T = nλ :
n ( n − 1 ) τ 2 , T / T = ( n − 1 ) τ 2 , T / λ n(n-1)\tau_{2,T}/T=(n-1)\tau_{2,T}/\lambda n ( n − 1 ) τ 2 , T / T = ( n − 1 ) τ 2 , T / λ , 4 n ( n − 1 ) / T 2 = 4 ( n − 1 ) / ( n λ 2 ) 4n(n-1)/T^2=4(n-1)/(n\lambda^2) 4 n ( n − 1 ) / T 2 = 4 ( n − 1 ) / ( n λ 2 ) , and n ( n − 1 ) O ( T − 3 ) = O ( n − 1 ) n(n-1)O(T^{-3})=O(n^{-1}) n ( n − 1 ) O ( T − 3 ) = O ( n − 1 ) . All constants are uniform under Assumptions 1–3.
Proof of Lemma 3
With t k t_k t k , d i d_i d i , and c 12 c_{12} c 12 as in Appendix A, write h i j = r i j 2 h_{ij}=r_{ij}^{2} h ij = r ij 2 , q ( z ) = z ′ Γ z q(z)=z'\Gamma z q ( z ) = z ′ Γ z , P ( z ) = z ′ Γ 2 z P(z)=z'\Gamma^2z P ( z ) = z ′ Γ 2 z (so P i = P ( Z i ) P_i=P(Z_i) P i = P ( Z i ) ), and g ( z ) = E ( h 12 ∣ Z 1 = z ) g(z)=\mathbb{E}(h_{12}\mid Z_1=z) g ( z ) = E ( h 12 ∣ Z 1 = z ) . For a generic Z ∼ N ( 0 , I T ) Z\sim N(0,I_T) Z ∼ N ( 0 , I T ) , put d = q ( Z ) − T d=q(Z)-T d = q ( Z ) − T .
Step 1: Exact decomposition. Since Tr ( R 2 ) = n + 2 ∑ i < j h i j \operatorname{Tr}(R^2)=n+2\sum_{i<j}h_{ij} Tr ( R 2 ) = n + 2 ∑ i < j h ij , Var { Tr ( R 2 ) } = 4 Var ( ∑ i < j h i j ) \operatorname{Var}\{\operatorname{Tr}(R^2)\}=4\operatorname{Var}(\sum_{i<j}h_{ij}) Var { Tr ( R 2 )} = 4 Var ( ∑ i < j h ij ) . Under H 0 H_0 H 0 , disjoint pairs are independent, and two pairs sharing one index are conditionally independent given the shared row, so that Cov ( h 12 , h 13 ) = Var { g ( Z 1 ) } \operatorname{Cov}(h_{12},h_{13})=\operatorname{Var}\{g(Z_1)\} Cov ( h 12 , h 13 ) = Var { g ( Z 1 )} . There are n ( n − 1 ) / 2 n(n-1)/2 n ( n − 1 ) /2 unordered pairs and n ( n − 1 ) ( n − 2 ) n(n-1)(n-2) n ( n − 1 ) ( n − 2 ) ordered pairs of distinct unordered pairs with exactly one common index. Hence
Var { Tr ( R 2 ) } = 2 n ( n − 1 ) a T + 4 n ( n − 1 ) ( n − 2 ) b T , a T = Var ( h 12 ) , b T = Var { g ( Z 1 ) } . (7) \operatorname{Var}\{\operatorname{Tr}(R^2)\}=2n(n-1)a_T+4n(n-1)(n-2)b_T,\qquad a_T=\operatorname{Var}(h_{12}),\quad b_T=\operatorname{Var}\{g(Z_1)\}. \tag*{(7)} Var { Tr ( R 2 )} = 2 n ( n − 1 ) a T + 4 n ( n − 1 ) ( n − 2 ) b T , a T = Var ( h 12 ) , b T = Var { g ( Z 1 )} . (7)
We show a T = 2 τ 2 , T 2 / T 2 + O ( T − 3 ) a_T=2\tau_{2,T}^{2}/T^2+O(T^{-3}) a T = 2 τ 2 , T 2 / T 2 + O ( T − 3 ) and b T = 2 K T / T 3 + O ( T − 4 ) b_T=2K_T/T^3+O(T^{-4}) b T = 2 K T / T 3 + O ( T − 4 ) .
Step 2: Self-pair term. Put A = Z 1 ′ Γ Z 2 = T S 12 A=Z_1'\Gamma Z_2=TS_{12} A = Z 1 ′ Γ Z 2 = T S 12 , so that h 12 2 = T − 4 A 4 ( 1 + δ 1 ) − 2 ( 1 + δ 2 ) − 2 h_{12}^{2}=T^{-4}A^4(1+\delta_1)^{-2}(1+\delta_2)^{-2} h 12 2 = T − 4 A 4 ( 1 + δ 1 ) − 2 ( 1 + δ 2 ) − 2 . On E 12 \mathcal{E}_{12} E 12 ,
( 1 + δ 1 ) − 2 ( 1 + δ 2 ) − 2 = 1 − 2 δ 1 − 2 δ 2 + R 2 , ∣ R 2 ∣ ≤ C ( δ 1 2 + δ 2 2 ) . (1+\delta_1)^{-2}(1+\delta_2)^{-2}=1-2\delta_1-2\delta_2+R_2,\qquad|R_2|\le C(\delta_1^2+\delta_2^2). ( 1 + δ 1 ) − 2 ( 1 + δ 2 ) − 2 = 1 − 2 δ 1 − 2 δ 2 + R 2 , ∣ R 2 ∣ ≤ C ( δ 1 2 + δ 2 2 ) .
Since E ( A 4 ∣ Z 1 ) = 3 P 1 2 \mathbb{E}(A^4\mid Z_1)=3P_1^2 E ( A 4 ∣ Z 1 ) = 3 P 1 2 , Lemma S.2 and the moment–cumulant formula give, for i = 1 , 2 i=1,2 i = 1 , 2 (by symmetry in the two rows),
E ( A 4 ) = 3 ( t 2 2 + 2 t 4 ) , E ( A 4 d i ) = 3 E ( P 1 2 d 1 ) = 3 ( 4 t 2 t 3 + 8 t 5 ) = O ( T 2 ) , \mathbb{E}(A^4)=3(t_2^2+2t_4),\qquad\mathbb{E}(A^4d_i)=3\mathbb{E}(P_1^2d_1)=3(4t_2t_3+8t_5)=O(T^2), E ( A 4 ) = 3 ( t 2 2 + 2 t 4 ) , E ( A 4 d i ) = 3 E ( P 1 2 d 1 ) = 3 ( 4 t 2 t 3 + 8 t 5 ) = O ( T 2 ) ,
so the linear terms contribute T − 5 E ( A 4 d i ) = O ( T − 3 ) T^{-5}\mathbb{E}(A^4d_i)=O(T^{-3}) T − 5 E ( A 4 d i ) = O ( T − 3 ) . By Hölder’s inequality and Lemma S.1, E ( A 4 δ i 2 ) ≤ ∥ A ∥ 8 4 ∥ δ i ∥ 4 2 = O ( T ) \mathbb{E}(A^4\delta_i^2)\le\lVert A\rVert_8^4\lVert\delta_i\rVert_4^2=O(T) E ( A 4 δ i 2 ) ≤ ∥ A ∥ 8 4 ∥ δ i ∥ 4 2 = O ( T ) , so the remainder contributes O ( T − 3 ) O(T^{-3}) O ( T − 3 ) . The exceptional event E 12 c \mathcal{E}_{12}^c E 12 c contributes O ( e − c T ) O(e^{-cT}) O ( e − c T ) , as in Appendix A, using 0 ≤ h 12 2 ≤ 1 0\le h_{12}^2\le1 0 ≤ h 12 2 ≤ 1 . Hence E ( h 12 2 ) = ( 3 t 2 2 + 6 t 4 ) / T 4 + O ( T − 3 ) = 3 τ 2 , T 2 / T 2 + O ( T − 3 ) \mathbb{E}(h_{12}^2)=(3t_2^2+6t_4)/T^4+O(T^{-3})=3\tau_{2,T}^2/T^2+O(T^{-3}) E ( h 12 2 ) = ( 3 t 2 2 + 6 t 4 ) / T 4 + O ( T − 3 ) = 3 τ 2 , T 2 / T 2 + O ( T − 3 ) . Lemma 2 gives E ( h 12 ) = τ 2 , T / T + O ( T − 2 ) \mathbb{E}(h_{12})=\tau_{2,T}/T+O(T^{-2}) E ( h 12 ) = τ 2 , T / T + O ( T − 2 ) , so a T = 2 τ 2 , T 2 / T 2 + O ( T − 3 ) a_T=2\tau_{2,T}^2/T^2+O(T^{-3}) a T = 2 τ 2 , T 2 / T 2 + O ( T − 3 ) .
Step 3: Representation of g g g . Let ν 1 , T , … , ν T , T \nu_{1,T},\ldots,\nu_{T,T} ν 1 , T , … , ν T , T be the eigenvalues of Γ \Gamma Γ and m T = # { t : ν t , T ≥ 1 / 2 } m_T=\#\{t:\nu_{t,T}\ge1/2\} m T = # { t : ν t , T ≥ 1/2 } . Since T tr Γ = T T\operatorname{tr}\Gamma=T T tr Γ = T and ∥ Γ ∥ o p ≤ K Γ \lVert\Gamma\rVert_{\mathrm{op}}\le K_\Gamma ∥ Γ ∥ op ≤ K Γ , T ≤ T / 2 + K Γ m T T\le T/2+K_\Gamma m_T T ≤ T /2 + K Γ m T , so m T ≥ T / ( 2 K Γ ) m_T\ge T/(2K_\Gamma) m T ≥ T / ( 2 K Γ ) . Hence q ( Z ) q(Z) q ( Z ) stochastically dominates χ m T 2 / 2 \chi^2_{m_T}/2 χ m T 2 /2 , and for each fixed p ≥ 1 p\ge1 p ≥ 1 and all T T T with m T > 2 p m_T>2p m T > 2 p ,
E q ( Z ) − p = O ( T − p ) . (8) \mathbb{E}q(Z)^{-p}=O(T^{-p}). \tag*{(8)} E q ( Z ) − p = O ( T − p ) . (8)
Let B T = E { Z Z ′ / q ( Z ) } B_T=\mathbb{E}\{ZZ^\prime/q(Z)\} B T = E { Z Z ′ / q ( Z )} . Independence of Z 1 Z_1 Z 1 and Z 2 Z_2 Z 2 gives the exact representation g ( z ) = z ′ Γ B T Γ z / q ( z ) g(z)=z^\prime\Gamma B_T\Gamma z/q(z) g ( z ) = z ′ Γ B T Γ z / q ( z ) . The identity T / q = 1 − d / T + d 2 / ( T q ) T/q=1-d/T+d^2/(Tq) T / q = 1 − d / T + d 2 / ( T q ) and E ( Z Z ′ d ) = 2 Γ \mathbb{E}(ZZ^\prime d)=2\Gamma E ( Z Z ′ d ) = 2Γ yield
T B T − I T = − 2 Γ T + 1 T E { Z Z ′ d 2 q ( Z ) } . TB_T-I_T=-\frac{2\Gamma}{T}+\frac{1}{T}\mathbb{E}\left\{\frac{ZZ^\prime d^2}{q(Z)}\right\}. T B T − I T = − T 2Γ + T 1 E { q ( Z ) Z Z ′ d 2 } .
For a unit vector u u u , Hölder’s inequality and (B.2) give
E { ( u ′ Z ) 2 d 2 q ( Z ) } ≤ ∥ u ′ Z ∥ 8 2 ∥ d ∥ 8 2 ∥ q ( Z ) − 1 ∥ 2 = O ( 1 ) . \begin{aligned}
\mathbb{E}\left\{\frac{(u^\prime Z)^2d^2}{q(Z)}\right\}\le\lVert u^\prime Z\rVert_8^2\lVert d\rVert_8^2\lVert q(Z)^{-1}\rVert_2 \\
&=O(1).
\end{aligned} E { q ( Z ) ( u ′ Z ) 2 d 2 } ≤ ∥ u ′ Z ∥ 8 2 ∥ d ∥ 8 2 ∥ q ( Z ) − 1 ∥ 2 = O ( 1 ) .
Since E { Z Z ′ d 2 / q ( Z ) } \mathbb{E}\{ZZ^\prime d^2/q(Z)\} E { Z Z ′ d 2 / q ( Z )} is positive semidefinite, its operator norm is the supremum of this quantity over unit u u u , and is therefore O ( 1 ) O(1) O ( 1 ) . Hence
∥ B T − T − 1 I T ∥ op = O ( T − 2 ) , in particular ∥ B T ∥ op = O ( T − 1 ) . (9) \lVert B_T-T^{-1}I_T\rVert_{\operatorname{op}}=O(T^{-2}),\qquad\text{in particular}\qquad\lVert B_T\rVert_{\operatorname{op}}=O(T^{-1}). \tag*{(9)} ∥ B T − T − 1 I T ∥ op = O ( T − 2 ) , in particular ∥ B T ∥ op = O ( T − 1 ) . (9)
Set N T = Γ ( B T − T − 1 I T ) Γ N_T=\Gamma(B_T-T^{-1}I_T)\Gamma N T = Γ ( B T − T − 1 I T ) Γ and ϖ T = Tr ( N T ) / T \varpi_T=\operatorname{Tr}(N_T)/T ϖ T = Tr ( N T ) / T . Then ∥ N T ∥ op = O ( T − 2 ) \lVert N_T\rVert_{\operatorname{op}}=O(T^{-2}) ∥ N T ∥ op = O ( T − 2 ) , ∣ ϖ T ∣ ≤ ∥ N T ∥ op |\varpi_T|\leq\lVert N_T\rVert_{\operatorname{op}} ∣ ϖ T ∣ ≤ ∥ N T ∥ op , Tr ( N T − ϖ T I ) = 0 \operatorname{Tr}(N_T-\varpi_T I)=0 Tr ( N T − ϖ T I ) = 0 , and ∥ N T − ϖ T Γ ∥ F ≤ T 1 / 2 ∥ N T − ϖ T Γ ∥ op = O ( T − 3 / 2 ) \lVert N_T-\varpi_T\Gamma\rVert_F\leq T^{1/2}\lVert N_T-\varpi_T\Gamma\rVert_{\operatorname{op}}=O(T^{-3/2}) ∥ N T − ϖ T Γ ∥ F ≤ T 1/2 ∥ N T − ϖ T Γ ∥ op = O ( T − 3/2 ) . With W = P ( Z ) / { T q ( Z ) } W=P(Z)/\{Tq(Z)\} W = P ( Z ) / { T q ( Z )} ,
g ( Z ) = W + ϖ T + Z ′ ( N T − ϖ T Γ ) Z q ( Z ) . g(Z)=W+\varpi_T+\frac{Z^{\prime}(N_T-\varpi_T\Gamma)Z}{q(Z)}. g ( Z ) = W + ϖ T + q ( Z ) Z ′ ( N T − ϖ T Γ ) Z .
For a trace-zero symmetric H H H , E ( Z ′ H Z ) 4 ≤ C ∥ H ∥ F 4 \mathbb{E}(Z^{\prime}HZ)^4\leq C\lVert H\rVert_F^4 E ( Z ′ H Z ) 4 ≤ C ∥ H ∥ F 4 ; Cauchy–Schwarz and (B.2) therefore bound the L 2 L^2 L 2 norm of the last term by O ( T − 3 / 2 ) ⋅ O ( T − 1 ) = O ( T − 5 / 2 ) O(T^{-3/2})\cdot O(T^{-1})=O(T^{-5/2}) O ( T − 3/2 ) ⋅ O ( T − 1 ) = O ( T − 5/2 ) . Thus g ( Z ) − E g ( Z ) = W − E W + e T g(Z)-\mathbb{E}g(Z)=W-\mathbb{E}W+e_T g ( Z ) − E g ( Z ) = W − E W + e T with ∥ e T ∥ 2 = O ( T − 5 / 2 ) \lVert e_T\rVert_2=O(T^{-5/2}) ∥ e T ∥ 2 = O ( T − 5/2 ) .
Step 4: Variance of g g g . Put D = P ( Z ) − τ 2 , T q ( Z ) = Z ′ H D Z D=P(Z)-\tau_{2,T}q(Z)=Z^{\prime}H_DZ D = P ( Z ) − τ 2 , T q ( Z ) = Z ′ H D Z with H D = Γ 2 − τ 2 , T Γ H_D=\Gamma^2-\tau_{2,T}\Gamma H D = Γ 2 − τ 2 , T Γ . Then Tr ( H D ) = 0 \operatorname{Tr}(H_D)=0 Tr ( H D ) = 0 and Tr ( H D 2 ) = T K T \operatorname{Tr}(H_D^2)=TK_T Tr ( H D 2 ) = T K T . The exact identity
W − τ 2 , T T = D T 2 − D d T 3 + D d 2 T 3 q ( Z ) W-\frac{\tau_{2,T}}{T}=\frac{D}{T^2}-\frac{Dd}{T^3}+\frac{Dd^2}{T^3q(Z)} W − T τ 2 , T = T 2 D − T 3 D d + T 3 q ( Z ) D d 2
holds, and by Hölder’s inequality and (B.2) the three terms on the right have L 2 L^2 L 2 norms O ( T − 3 / 2 ) O(T^{-3/2}) O ( T − 3/2 ) , O ( T − 2 ) O(T^{-2}) O ( T − 2 ) , and O ( T − 5 / 2 ) O(T^{-5/2}) O ( T − 5/2 ) . Since D D D and d d d are centered, (S.A.3) and (S.A.4) give
E D 2 = 2 T K T , E ( D 2 d ) = 8 Tr ( H D 2 Γ ) = 8 ( t 5 − 2 t 2 t 4 T + t 2 2 t 3 T 2 ) = O ( T ) . \mathbb{E}D^2=2TK_T,\qquad\mathbb{E}(D^2d)=8\operatorname{Tr}(H_D^2\Gamma)=8\left(t_5-\frac{2t_2t_4}{T}+\frac{t_2^2t_3}{T^2}\right)=O(T). E D 2 = 2 T K T , E ( D 2 d ) = 8 Tr ( H D 2 Γ ) = 8 ( t 5 − T 2 t 2 t 4 + T 2 t 2 2 t 3 ) = O ( T ) .
The first term therefore has variance 2 K T / T 3 2K_T/T^3 2 K T / T 3 ; its covariance with the second term is − E ( D 2 d ) / T 5 = O ( T − 4 ) -\mathbb{E}(D^2d)/T^5=O(T^{-4}) − E ( D 2 d ) / T 5 = O ( T − 4 ) ; and all remaining variance and covariance terms are O ( T − 4 ) O(T^{-4}) O ( T − 4 ) by Cauchy–Schwarz. Hence Var ( W ) = 2 K T / T 3 + O ( T − 4 ) \operatorname{Var}(W)=2K_T/T^3+O(T^{-4}) Var ( W ) = 2 K T / T 3 + O ( T − 4 ) . Finally, b T = Var ( W ) + 2 Cov ( W , e T ) + E e T 2 b_T=\operatorname{Var}(W)+2\operatorname{Cov}(W,e_T)+\mathbb{E}e_T^2 b T = Var ( W ) + 2 Cov ( W , e T ) + E e T 2 and ∥ W − E W ∥ 2 = O ( T − 3 / 2 ) \lVert W-\mathbb{E}W\rVert_2=O(T^{-3/2}) ∥ W − E W ∥ 2 = O ( T − 3/2 ) , so
b T = 2 K T T 3 + O ( T − 4 ) . (10) b_T=\frac{2K_T}{T^3}+O(T^{-4}). \tag*{(10)} b T = T 3 2 K T + O ( T − 4 ) . (10)
These bounds are absolute and do not require K T > 0 K_T>0 K T > 0 .
Step 5: Assembly. Insert both expansions into (B.1). Because T = n λ T=n\lambda T = nλ , the aggregate errors are O ( n 2 T − 3 + n 3 T − 4 ) = O ( n − 1 ) O(n^2T^{-3}+n^3T^{-4})=O(n^{-1}) O ( n 2 T − 3 + n 3 T − 4 ) = O ( n − 1 ) . Replacing n ( n − 1 ) n(n-1) n ( n − 1 ) and n ( n − 1 ) ( n − 2 ) n(n-1)(n-2) n ( n − 1 ) ( n − 2 ) by n 2 n^2 n 2 and n 3 n^3 n 3 costs O ( n − 1 ) O(n^{-1}) O ( n − 1 ) , since τ 2 , T \tau_{2,T} τ 2 , T and K T K_T K T are uniformly bounded. Hence
Var { Tr ( R 2 ) } = 4 τ 2 , T 2 λ 2 + 8 K T λ 3 + O ( n − 1 ) . \operatorname{Var}\{\operatorname{Tr}(R^2)\}=\frac{4\tau_{2,T}^2}{\lambda^2}+\frac{8K_T}{\lambda^3}+O(n^{-1}). Var { Tr ( R 2 )} = λ 2 4 τ 2 , T 2 + λ 3 8 K T + O ( n − 1 ) .
Second-order terms. The second-order expansions in Remark 4 follow from the same argument with longer Taylor polynomials. For a T a_T a T , expand ( 1 + δ 1 ) − 2 ( 1 + δ 2 ) − 2 (1+\delta_1)^{-2}(1+\delta_2)^{-2} ( 1 + δ 1 ) − 2 ( 1 + δ 2 ) − 2 through total degree three in ( δ 1 , δ 2 ) (\delta_1,\delta_2) ( δ 1 , δ 2 ) ; by Hölder’s inequality the remainder contributes O ( ∥ A ∥ 8 4 ∥ δ ∥ 8 4 T − 4 ) = O ( T − 4 ) O(\lVert A\rVert_8^4\lVert\delta\rVert_8^4T^{-4})=O(T^{-4}) O (∥ A ∥ 8 4 ∥ δ ∥ 8 4 T − 4 ) = O ( T − 4 ) . For b T b_T b T , write b T = E ( h 12 h 13 ) − ( E h 12 ) 2 b_T=\mathbb{E}(h_{12}h_{13})-(\mathbb{E}h_{12})^2 b T = E ( h 12 h 13 ) − ( E h 12 ) 2 , expand ( 1 + δ 1 ) − 2 ( 1 + δ 2 ) − 1 ( 1 + δ 3 ) − 1 (1+\delta_1)^{-2}(1+\delta_2)^{-1}(1+\delta_3)^{-1} ( 1 + δ 1 ) − 2 ( 1 + δ 2 ) − 1 ( 1 + δ 3 ) − 1 through degree five and ( 1 + δ 1 ) − 1 ( 1 + δ 2 ) − 1 (1+\delta_1)^{-1}(1+\delta_2)^{-1} ( 1 + δ 1 ) − 1 ( 1 + δ 2 ) − 1 through degree seven; the remainders are O ( T − 5 ) O(T^{-5}) O ( T − 5 ) , and the event { max a ≤ 3 ∣ δ a ∣ > 1 / 2 } \{\max_{a\leq3}|\delta_a|>1/2\} { max a ≤ 3 ∣ δ a ∣ > 1/2 } contributes O ( e − c T ) O(e^{-cT}) O ( e − c T ) as before. The retained terms are products of centered Gaussian quadratic forms in ( Z 1 , Z 2 , Z 3 ) (Z_1,Z_2,Z_3) ( Z 1 , Z 2 , Z 3 ) . In an eigenbasis of Γ \Gamma Γ these forms are T − 1 ∑ ν ν t , T ζ t T^{-1}\sum_\nu\nu_{t,T}\zeta_t T − 1 ∑ ν ν t , T ζ t with i.i.d. copies ζ t \zeta_t ζ t of a vector of products of independent standard normals, so every k k k th joint cumulant equals T 1 − k τ k , T T^{1-k}\tau_{k,T} T 1 − k τ k , T times a rational constant. The moment–cumulant formula therefore gives each expectation as an exact polynomial in T − 1 T^{-1} T − 1 with coefficients polynomial in τ 2 , T , … , τ 5 , T \tau_{2,T},\ldots,\tau_{5,T} τ 2 , T , … , τ 5 , T . An expectation of a product of d d d centered forms is O ( T − ⌈ d / 2 ⌉ ) O(T^{-\lceil d/2\rceil}) O ( T − ⌈ d /2 ⌉ ) , and retaining those up to T − 3 T^{-3} T − 3 for a T a_T a T and T − 4 T^{-4} T − 4 for b T b_T b T yields the second-order expansions of a T a_T a T and b T b_T b T . The same computation also gives E ( r 12 2 ) = τ 2 , T / T + 4 ( τ 2 , T 2 − τ 3 , T ) / T 2 + ( 28 τ 2 , T 3 − 48 τ 2 , T τ 3 , T + 20 τ 4 , T ) / T 3 + O ( T − 4 ) \mathbb{E}(r_{12}^2)=\tau_{2,T}/T+4(\tau_{2,T}^2-\tau_{3,T})/T^2+(28\tau_{2,T}^3-48\tau_{2,T}\tau_{3,T}+20\tau_{4,T})/T^3+O(T^{-4}) E ( r 12 2 ) = τ 2 , T / T + 4 ( τ 2 , T 2 − τ 3 , T ) / T 2 + ( 28 τ 2 , T 3 − 48 τ 2 , T τ 3 , T + 20 τ 4 , T ) / T 3 + O ( T − 4 ) . In b T b_T b T the T − 2 T^{-2} T − 2 terms of E ( h 12 h 13 ) \mathbb{E}(h_{12}h_{13}) E ( h 12 h 13 ) and ( E h 12 ) 2 (\mathbb{E}h_{12})^2 ( E h 12 ) 2 cancel exactly, in agreement with Step 4.
Proof of Theorem 1
We first establish the exact Hoeffding decomposition of Tr ( R 2 ) \operatorname{Tr}(R^2) Tr ( R 2 ) and moment bounds for its Doob increments (Lemma C.1), and then apply the martingale CLT. Put
q i = Z i ′ Γ Z i , u i = T 1 / 2 Z i q i 1 / 2 , h i j = h T ( Z i , Z j ) = ( u i ′ u j ) 2 = r i j 2 . q_i=Z_i'\Gamma Z_i,\qquad u_i=\frac{T^{1/2}Z_i}{q_i^{1/2}},\qquad h_{ij}=h_T(Z_i,Z_j)=(u_i'u_j)^2=r_{ij}^2. q i = Z i ′ Γ Z i , u i = q i 1/2 T 1/2 Z i , h ij = h T ( Z i , Z j ) = ( u i ′ u j ) 2 = r ij 2 .
The vectors u 1 , … , u n u_1,\ldots,u_n u 1 , … , u n are i.i.d. and ∥ u i ∥ = 1 \lVert u_i\rVert=1 ∥ u i ∥ = 1 . Define
M T = E ( u 1 u 1 ′ ) , θ T = E ( h 12 ) = Tr ( M T 2 ) , M_T=\mathbb{E}(u_1u_1'),\qquad\theta_T=\mathbb{E}(h_{12})=\operatorname{Tr}(M_T^2), M T = E ( u 1 u 1 ′ ) , θ T = E ( h 12 ) = Tr ( M T 2 ) ,
and the first-order and second-order degenerate Hoeffding projections
ξ i = E ( h i j ∣ Z i ) − θ T = u i ′ M T u i − θ T , \xi_i=\mathbb{E}(h_{ij}\mid Z_i)-\theta_T=u_i'M_Tu_i-\theta_T, ξ i = E ( h ij ∣ Z i ) − θ T = u i ′ M T u i − θ T ,
h i j ∘ = h i j − θ T − ξ i − ξ j . h_{ij}^{\circ}=h_{ij}-\theta_T-\xi_i-\xi_j. h ij ∘ = h ij − θ T − ξ i − ξ j .
Then
E ξ i = 0 , E ( h i j ∘ ∣ Z i ) = E ( h i j ∘ ∣ Z j ) = 0. \mathbb{E}\xi_i=0,\qquad\mathbb{E}(h_{ij}^{\circ}\mid Z_i)=\mathbb{E}(h_{ij}^{\circ}\mid Z_j)=0. E ξ i = 0 , E ( h ij ∘ ∣ Z i ) = E ( h ij ∘ ∣ Z j ) = 0.
Lemma C.1 (Hoeffding-projection bounds) Under Assumptions 1–3 and H 0 H_0 H 0 ,
E ξ 1 2 = 2 K T T 3 + O ( T − 4 ) , (11) \mathbb{E}\xi_1^2=\frac{2K_T}{T^3}+O(T^{-4}), \tag*{(11)} E ξ 1 2 = T 3 2 K T + O ( T − 4 ) , (11) E [ ( h 12 ∘ ) 2 ] = 2 τ T 2 T 2 + O ( T − 3 ) , (12) \mathbb{E}[(h_{12}^{\circ})^2]=\frac{2\tau_T^2}{T^2}+O(T^{-3}), \tag*{(12)} E [( h 12 ∘ ) 2 ] = T 2 2 τ T 2 + O ( T − 3 ) , (12) E ξ 1 4 = O ( T − 6 ) , E [ ( h 12 ∘ ) 4 ] = O ( T − 4 ) . (13) \mathbb{E}\xi_1^4=O(T^{-6}),\qquad\mathbb{E}[(h_{12}^{\circ})^4]=O(T^{-4}). \tag*{(13)} E ξ 1 4 = O ( T − 6 ) , E [( h 12 ∘ ) 4 ] = O ( T − 4 ) . (13)
For F k = σ ( Z 1 , … , Z k ) \mathcal{F}_k=\sigma(Z_1,\ldots,Z_k) F k = σ ( Z 1 , … , Z k ) , the Doob increment of Tr ( R 2 ) \operatorname{Tr}(R^2) Tr ( R 2 ) is
Y k = 2 ( n − 1 ) ξ k + 2 ∑ i < k h i k ∘ . (14) Y_k=2(n-1)\xi_k+2\sum_{i<k}h_{ik}^{\circ}. \tag*{(14)} Y k = 2 ( n − 1 ) ξ k + 2 i < k ∑ h ik ∘ . (14)
Moreover,
∑ k = 1 n E ( Y k 4 ) = o ( 1 ) , (15) \sum_{k=1}^{n}\mathbb{E}(Y_k^4)=o(1), \tag*{(15)} k = 1 ∑ n E ( Y k 4 ) = o ( 1 ) , (15)
∑ k = 1 n E ( Y k 2 ∣ F k − 1 ) − ∑ k = 1 n E ( Y k 2 ) → p 0. (16) \sum_{k=1}^{n}\mathbb{E}(Y_k^2\mid\mathcal{F}_{k-1})-\sum_{k=1}^{n}\mathbb{E}(Y_k^2)\xrightarrow{p}0. \tag*{(16)} k = 1 ∑ n E ( Y k 2 ∣ F k − 1 ) − k = 1 ∑ n E ( Y k 2 ) p 0. (16)
Consequently,
max 1 ≤ k ≤ n ∣ Y k ∣ → p 0 , max 1 ≤ k ≤ n ∣ Y k ∣ σ n → p 0 , \max_{1\leq k\leq n}|Y_k|\xrightarrow{p}0,\qquad\max_{1\leq k\leq n}\frac{|Y_k|}{\sigma_n}\xrightarrow{p}0, 1 ≤ k ≤ n max ∣ Y k ∣ p 0 , 1 ≤ k ≤ n max σ n ∣ Y k ∣ p 0 ,
where σ n 2 \sigma_n^2 σ n 2 is given in (2).
Proof. Second moments. For Z ∼ N ( 0 , I T ) Z \sim N(0,I_T) Z ∼ N ( 0 , I T ) let q = Z ′ Γ Z q = Z^{\prime}\Gamma Z q = Z ′ Γ Z and Ψ = Γ 1 / 2 Z Z ′ Γ 1 / 2 / q \Psi= \Gamma^{1/2}ZZ^{\prime}\Gamma^{1/2}/q Ψ = Γ 1/2 Z Z ′ Γ 1/2 / q , so that Ψ i = u i u i ′ \Psi_i = u_i u_i^{\prime} Ψ i = u i u i ′ and M T = E Ψ = Γ 1 / 2 B T Γ 1 / 2 M_T = \mathbb{E}\Psi= \Gamma^{1/2}B_T\Gamma^{1/2} M T = E Ψ = Γ 1/2 B T Γ 1/2 with B T B_T B T as in Appendix B. Since ξ 1 = ⟨ Ψ 1 , M T ⟩ F − θ T \xi_1 = \langle\Psi_1,M_T\rangle_F-\theta_T ξ 1 = ⟨ Ψ 1 , M T ⟩ F − θ T is the centered conditional mean g ( Z 1 ) − E g ( Z 1 ) g(Z_1)-\mathbb{E}g(Z_1) g ( Z 1 ) − E g ( Z 1 ) in the proof of Lemma 3 (Appendix B), E ξ 1 2 = b T \mathbb{E}\xi_1^2=b_T E ξ 1 2 = b T , and (B.4) gives (C.1). Orthogonality of the Hoeffding components gives E [ ( h 12 ∘ ) 2 ] = Var ( h 12 ) − 2 E ξ 1 2 = a T − 2 b T \mathbb{E}[(h_{12}^{\circ})^2]=\operatorname{Var}(h_{12})-2\mathbb{E}\xi_1^2=a_T-2b_T E [( h 12 ∘ ) 2 ] = Var ( h 12 ) − 2 E ξ 1 2 = a T − 2 b T , which is (C.2) by Lemma 3.
A uniform fourth-moment bound. For a deterministic symmetric H H H , put Q H = Γ 1 / 2 H Γ 1 / 2 Q_H=\Gamma^{1/2}H\Gamma^{1/2} Q H = Γ 1/2 H Γ 1/2 , c H = Tr ( Q H ) / T c_H=\operatorname{Tr}(Q_H)/T c H = Tr ( Q H ) / T , and B H = Q H − c H Γ B_H=Q_H-c_H\Gamma B H = Q H − c H Γ . Then
⟨ Ψ , H ⟩ F − c H = Z ′ B H Z q , Tr ( B H ) = 0 , ∥ B H ∥ F ≤ C ∥ H ∥ F . \langle\Psi,H\rangle_F-c_H=\frac{Z^{\prime}B_HZ}{q},\qquad\operatorname{Tr}(B_H)=0,\qquad\lVert B_H\rVert_F\le C\lVert H\rVert_F. ⟨ Ψ , H ⟩ F − c H = q Z ′ B H Z , Tr ( B H ) = 0 , ∥ B H ∥ F ≤ C ∥ H ∥ F .
because ∣ c H ∣ ≤ ∥ H ∥ F ∥ Γ ∥ F / T |c_H|\le\lVert H\rVert_F\lVert\Gamma\rVert_F/T ∣ c H ∣ ≤ ∥ H ∥ F ∥ Γ ∥ F / T and ∥ Γ ∥ F ≤ K Γ T 1 / 2 \lVert\Gamma\rVert_F\le K_\Gamma T^{1/2} ∥ Γ ∥ F ≤ K Γ T 1/2 . Gaussian quadratic-form moments give E ( Z ′ B H Z ) 8 ≤ C ∥ B H ∥ F 8 \mathbb{E}(Z^{\prime}B_HZ)^8\le C\lVert B_H\rVert_F^8 E ( Z ′ B H Z ) 8 ≤ C ∥ B H ∥ F 8 , and (B.2) gives E q − 8 = O ( T − 8 ) \mathbb{E}q^{-8}=O(T^{-8}) E q − 8 = O ( T − 8 ) . By Cauchy–Schwarz, and since centering at the mean instead of c H c_H c H costs at most a factor 2 4 2^4 2 4 ,
E ∣ ⟨ Ψ , H ⟩ F − E ⟨ Ψ , H ⟩ F ∣ 4 ≤ C ∥ B H ∥ F 4 T − 4 . (17) \mathbb{E}\left|\langle\Psi,H\rangle_F-\mathbb{E}\langle\Psi,H\rangle_F\right|^4\le C\lVert B_H\rVert_F^4T^{-4}. \tag*{(17)} E ∣ ⟨ Ψ , H ⟩ F − E ⟨ Ψ , H ⟩ F ∣ 4 ≤ C ∥ B H ∥ F 4 T − 4 . (17)
In particular, by Lyapunov’s inequality, Var ⟨ Ψ , H ⟩ F ≤ C T − 2 ∥ H ∥ F 2 \operatorname{Var}\langle\Psi,H\rangle_F\le CT^{-2}\lVert H\rVert_F^2 Var ⟨ Ψ , H ⟩ F ≤ C T − 2 ∥ H ∥ F 2 .
Fourth moments. By (B.3), ∥ M T ∥ op = O ( T − 1 ) \lVert M_T\rVert_{\operatorname{op}}=O(T^{-1}) ∥ M T ∥ op = O ( T − 1 ) , so ∥ Q M T ∥ F = O ( T − 1 / 2 ) \lVert Q_{M_T}\rVert_F=O(T^{-1/2}) ∥ Q M T ∥ F = O ( T − 1/2 ) , ∣ c M T ∣ = O ( T − 1 ) |c_{M_T}|=O(T^{-1}) ∣ c M T ∣ = O ( T − 1 ) , and ∥ B M T ∥ F = O ( T − 1 / 2 ) \lVert B_{M_T}\rVert_F=O(T^{-1/2}) ∥ B M T ∥ F = O ( T − 1/2 ) . Applying (C.7) with H = M T H=M_T H = M T , now with ∥ B M T ∥ F 4 = O ( T − 2 ) \lVert B_{M_T}\rVert_F^4=O(T^{-2}) ∥ B M T ∥ F 4 = O ( T − 2 ) , gives E ξ 1 4 = O ( T − 6 ) \mathbb{E}\xi_1^4=O(T^{-6}) E ξ 1 4 = O ( T − 6 ) . Next, write h 12 = ℓ 2 / q 2 h_{12}=\ell^2/q_2 h 12 = ℓ 2 / q 2 with ℓ = u 1 ′ Γ 1 / 2 Z 2 \ell=u_1^{\prime}\Gamma^{1/2}Z_2 ℓ = u 1 ′ Γ 1/2 Z 2 . Conditionally on u 1 u_1 u 1 , ℓ \ell ℓ is centered normal with variance u 1 ′ Γ u 1 ≤ K Γ u_1^{\prime}\Gamma u_1\le K_\Gamma u 1 ′ Γ u 1 ≤ K Γ , so E ( ℓ 16 ∣ u 1 ) ≤ C \mathbb{E}(\ell^{16}\mid u_1)\le C E ( ℓ 16 ∣ u 1 ) ≤ C , and Cauchy–Schwarz with E q 2 − 8 = O ( T − 8 ) \mathbb{E}q_2^{-8}=O(T^{-8}) E q 2 − 8 = O ( T − 8 ) gives
E ( h 12 4 ∣ u 1 ) ≤ { E ( ℓ 16 ∣ u 1 ) } 1 / 2 { E q 2 − 8 } 1 / 2 = O ( T − 4 ) . \mathbb{E}(h_{12}^4\mid u_1)\le\{\mathbb{E}(\ell^{16}\mid u_1)\}^{1/2}\{\mathbb{E}q_2^{-8}\}^{1/2}=O(T^{-4}). E ( h 12 4 ∣ u 1 ) ≤ { E ( ℓ 16 ∣ u 1 ) } 1/2 { E q 2 − 8 } 1/2 = O ( T − 4 ) .
Since h 12 ∘ = h 12 − θ T − ξ 1 − ξ 2 h_{12}^{\circ}=h_{12}-\theta_T-\xi_1-\xi_2 h 12 ∘ = h 12 − θ T − ξ 1 − ξ 2 and θ T = O ( T − 1 ) \theta_T=O(T^{-1}) θ T = O ( T − 1 ) , E ( h 12 ∘ ) 4 = O ( T − 4 ) \mathbb{E}(h_{12}^{\circ})^4=O(T^{-4}) E ( h 12 ∘ ) 4 = O ( T − 4 ) . This proves (C.3).
Doob increments. The exact Hoeffding identity is
2 ∑ i < j ( h i j − θ T ) = 2 ( n − 1 ) ∑ i = 1 n ξ i + 2 ∑ i < j h i j ∘ . (18) 2\sum_{i<j}(h_{ij}-\theta_T)=2(n-1)\sum_{i=1}^{n}\xi_i+2\sum_{i<j}h_{ij}^{\circ}. \tag*{(18)} 2 i < j ∑ ( h ij − θ T ) = 2 ( n − 1 ) i = 1 ∑ n ξ i + 2 i < j ∑ h ij ∘ . (18)
Its left side is T r ( R 2 ) − E T r ( R 2 ) Tr(R^2)-ETr(R^2) T r ( R 2 ) − E T r ( R 2 ) . Taking the difference of its conditional expectations given F k \mathcal{F}_k F k and F k − 1 \mathcal{F}_{k-1} F k − 1 , the linear part contributes 2 ( n − 1 ) ξ k 2(n-1)\xi_k 2 ( n − 1 ) ξ k , and, by degeneracy, the only surviving degenerate terms are 2 ∑ i < k h i k ∘ 2\sum_{i<k}h_{ik}^{\circ} 2 ∑ i < k h ik ∘ . This proves (C.4).
Write Ψ i = u i u i ′ \Psi_i=u_i u_i^{\prime} Ψ i = u i u i ′ , A i = Ψ i − M T A_i=\Psi_i-M_T A i = Ψ i − M T , and M ‾ = ( n − 1 ) M T \overline{M}=(n-1)M_T M = ( n − 1 ) M T . Since ξ k = ⟨ M T , A k ⟩ F \xi_k=\langle M_T,A_k\rangle_F ξ k = ⟨ M T , A k ⟩ F and h i k ∘ = ⟨ A i , A k ⟩ F h_{ik}^{\circ}=\langle A_i,A_k\rangle_F h ik ∘ = ⟨ A i , A k ⟩ F , the increment (C.4) is Y k = 2 ⟨ M ‾ + ∑ i < k A i , A k ⟩ F Y_k=2\langle\overline{M}+\sum_{i<k}A_i,A_k\rangle_F Y k = 2 ⟨ M + ∑ i < k A i , A k ⟩ F . Conditional on A k A_k A k , the random variables X i k = ⟨ A i , A k ⟩ F X_{ik}=\langle A_i,A_k\rangle_F X ik = ⟨ A i , A k ⟩ F ( i < k ) (i<k) ( i < k ) are independent and centered. Uniformly in u k u_k u k ,
E ( X i k 2 ∣ A k ) ≤ C T − 2 , E ( X i k 4 ∣ A k ) ≤ C T − 4 . \mathbb{E}(X_{ik}^{2}\mid A_k)\le CT^{-2},\qquad\mathbb{E}(X_{ik}^{4}\mid A_k)\le CT^{-4}. E ( X ik 2 ∣ A k ) ≤ C T − 2 , E ( X ik 4 ∣ A k ) ≤ C T − 4 .
These bounds hold because, conditional on u k = v u_k=v u k = v , X i k = ⟨ Ψ i , H v ⟩ F − E ⟨ Ψ i , H v ⟩ F X_{ik}=\langle\Psi_i,H_v\rangle_F-\mathbb{E}\langle\Psi_i,H_v\rangle_F X ik = ⟨ Ψ i , H v ⟩ F − E ⟨ Ψ i , H v ⟩ F with H v = v v ′ − M T H_v=vv^{\prime}-M_T H v = v v ′ − M T , and ∥ H v ∥ F ≤ 2 \lVert H_v\rVert_F\le2 ∥ H v ∥ F ≤ 2 uniformly in v v v , so (C.7) applies. Moreover, ⟨ M ‾ , A k ⟩ F = ( n − 1 ) ξ k \langle\overline{M},A_k\rangle_F=(n-1)\xi_k ⟨ M , A k ⟩ F = ( n − 1 ) ξ k and E ξ k 4 = O ( T − 6 ) \mathbb{E}\xi_k^4=O(T^{-6}) E ξ k 4 = O ( T − 6 ) by (C.3). Since E { ( ∑ i < k X i k ) 4 ∣ A k } = ∑ i < k E ( X i k 4 ∣ A k ) + 3 ∑ i ≠ j < k E ( X i k 2 ∣ A k ) E ( X j k 2 ∣ A k ) \mathbb{E}\{(\sum_{i<k}X_{ik})^4\mid A_k\}=\sum_{i<k}\mathbb{E}(X_{ik}^4\mid A_k)+3\sum_{i\ne j<k}\mathbb{E}(X_{ik}^2\mid A_k)\mathbb{E}(X_{jk}^2\mid A_k) E {( ∑ i < k X ik ) 4 ∣ A k } = ∑ i < k E ( X ik 4 ∣ A k ) + 3 ∑ i = j < k E ( X ik 2 ∣ A k ) E ( X j k 2 ∣ A k ) and ( x + y ) 4 ≤ 8 ( x 4 + y 4 ) (x+y)^4\le8(x^4+y^4) ( x + y ) 4 ≤ 8 ( x 4 + y 4 ) ,
E Y k 4 ≤ C { n 4 T − 6 + k T − 4 + k 2 T − 4 } ≤ C { n 4 T − 6 + n 2 T − 4 } = O ( n − 2 ) . \mathbb{E}Y_k^4\le C\{n^4T^{-6}+kT^{-4}+k^2T^{-4}\}\le C\{n^4T^{-6}+n^2T^{-4}\}=O(n^{-2}). E Y k 4 ≤ C { n 4 T − 6 + k T − 4 + k 2 T − 4 } ≤ C { n 4 T − 6 + n 2 T − 4 } = O ( n − 2 ) .
Summing over k ≤ n k\le n k ≤ n proves (C.5).
We next verify predictable quadratic-variation convergence. Define the covariance operator C T ( H ) = E { ⟨ A 1 , H ⟩ F A 1 } C_T(H)=\mathbb{E}\{\langle A_1,H\rangle_F A_1\} C T ( H ) = E {⟨ A 1 , H ⟩ F A 1 } on the real Hilbert space of symmetric T × T T\times T T × T matrices with the Frobenius inner product; operator traces T r o p Tr_{op} T r o p are taken on this space. The operator C T C_T C T is positive and self-adjoint, and its Rayleigh quotient at H H H is V a r ⟨ Ψ , H ⟩ F Var\langle\Psi,H\rangle_F V a r ⟨ Ψ , H ⟩ F . By (C.7) and T r o p C T = E ∥ A 1 ∥ F 2 = 1 − T r ( M T 2 ) ≤ 1 Tr_{op}C_T=\mathbb{E}\lVert A_1\rVert_F^2=1-Tr(M_T^2)\le1 T r o p C T = E ∥ A 1 ∥ F 2 = 1 − T r ( M T 2 ) ≤ 1 , the eigenvalues d j ≥ 0 d_j\ge0 d j ≥ 0 of C T C_T C T satisfy $\max_j d_j\le CT^{-2}】【。 and ∑ j d j ≤ 1 \sum_j d_j \le1 ∑ j d j ≤ 1 . Hence
∥ C T ∥ op ≤ C T − 2 , \lVert C_T\rVert_{\operatorname{op}} \le CT^{-2}, ∥ C T ∥ op ≤ C T − 2 ,
Tr op ( C T 2 ) ≤ ( max j d j ) ∑ j d j ≤ C T − 2 , Tr op ( C T 4 ) ≤ ( max j d j ) 3 ∑ j d j ≤ C T − 6 . \operatorname{Tr}_{\operatorname{op}}(C_T^2) \le\left(\max_j d_j\right)\sum_j d_j \le CT^{-2}, \qquad\operatorname{Tr}_{\operatorname{op}}(C_T^4) \le\left(\max_j d_j\right)^3\sum_j d_j \le CT^{-6}. Tr op ( C T 2 ) ≤ ( j max d j ) j ∑ d j ≤ C T − 2 , Tr op ( C T 4 ) ≤ ( j max d j ) 3 j ∑ d j ≤ C T − 6 .
Since Y k = 2 ⟨ M ˉ + ∑ i < k A i , A k ⟩ F Y_k = 2\langle\bar{M}+\sum_{i<k} A_i,A_k\rangle_F Y k = 2 ⟨ M ˉ + ∑ i < k A i , A k ⟩ F and A k A_k A k is independent of F k − 1 \mathcal{F}_{k-1} F k − 1 ,
E ( Y k 2 ∣ F k − 1 ) = 4 ⟨ M ˉ + ∑ i < k A i , C T ( M ˉ + ∑ i < k A i ) ⟩ F . \mathbb{E}(Y_k^2\mid\mathcal{F}_{k-1})=4\left\langle\bar{M}+\sum_{i<k}A_i,C_T\left(\bar{M}+\sum_{i<k}A_i\right)\right\rangle_F. E ( Y k 2 ∣ F k − 1 ) = 4 ⟨ M ˉ + i < k ∑ A i , C T ( M ˉ + i < k ∑ A i ) ⟩ F .
After summing over k k k and subtracting expectations, the fluctuation V n = ∑ k { E ( Y k 2 ∣ F k − 1 ) − E Y k 2 } \mathcal{V}_n=\sum_k\{\mathbb{E}(Y_k^2\mid\mathcal{F}_{k-1})-\mathbb{E}Y_k^2\} V n = ∑ k { E ( Y k 2 ∣ F k − 1 ) − E Y k 2 } is exactly
V n = 8 ∑ i = 1 n − 1 ( n − i ) ⟨ M ˉ , C T A i ⟩ F + 4 ∑ i = 1 n − 1 ( n − i ) { ⟨ A i , C T A i ⟩ F − E ⟨ A i , C T A i ⟩ F } + 8 ∑ i < j ≤ n − 1 ( n − j ) ⟨ A i , C T A j ⟩ F . \begin{aligned}
\mathcal{V}_n=8\sum_{i=1}^{n-1}(n-i)\langle\bar{M},C_TA_i\rangle_F \\
&+4\sum_{i=1}^{n-1}(n-i)\{\langle A_i,C_TA_i\rangle_F-\mathbb{E}\langle A_i,C_TA_i\rangle_F\} \\
&+8\sum_{i<j\le n-1}(n-j)\langle A_i,C_TA_j\rangle_F.
\end{aligned} V n = 8 i = 1 ∑ n − 1 ( n − i ) ⟨ M ˉ , C T A i ⟩ F + 4 i = 1 ∑ n − 1 ( n − i ) {⟨ A i , C T A i ⟩ F − E ⟨ A i , C T A i ⟩ F } + 8 i < j ≤ n − 1 ∑ ( n − j ) ⟨ A i , C T A j ⟩ F .
All three sums are centered. The third is orthogonal in L 2 L^2 L 2 to the first two, and its distinct unordered pairs are mutually orthogonal, including pairs sharing one index, by conditioning on the unshared index. Each of the first two sums has independent summands. Hence E V n 2 \mathbb{E}\mathcal{V}_n^2 E V n 2 is at most a constant times the sum of the three variances. Since ∥ M T ∥ F 2 = Tr ( M T 2 ) = O ( T − 1 ) \lVert M_T\rVert_F^2=\operatorname{Tr}(M_T^2)=O(T^{-1}) ∥ M T ∥ F 2 = Tr ( M T 2 ) = O ( T − 1 ) and E ⟨ M ˉ , C T A i ⟩ F 2 = ⟨ C T M ˉ , C T ( C T M ˉ ) ⟩ F = ⟨ M ˉ , C T 3 M ˉ ⟩ F \mathbb{E}\langle\bar{M},C_TA_i\rangle_F^2=\langle C_T\bar{M},C_T(C_T\bar{M})\rangle_F=\langle\bar{M},C_T^3\bar{M}\rangle_F E ⟨ M ˉ , C T A i ⟩ F 2 = ⟨ C T M ˉ , C T ( C T M ˉ ) ⟩ F = ⟨ M ˉ , C T 3 M ˉ ⟩ F , the variance of the first sum is at most
C n 3 ⟨ M ˉ , C T 3 M ˉ ⟩ F ≤ C n 3 ∥ C T ∥ op ∥ M ˉ ∥ F 2 = O ( T − 2 ) . Cn^3\langle\bar{M},C_T^3\bar{M}\rangle_F\le Cn^3\lVert C_T\rVert_{\operatorname{op}}\lVert\bar{M}\rVert_F^2=O(T^{-2}). C n 3 ⟨ M ˉ , C T 3 M ˉ ⟩ F ≤ C n 3 ∥ C T ∥ op ∥ M ˉ ∥ F 2 = O ( T − 2 ) .
For the second sum, ∥ A i ∥ F ≤ 2 \lVert A_i\rVert_F\le2 ∥ A i ∥ F ≤ 2 and ∣ ⟨ A i , C T A i ⟩ F ∣ ≤ 4 ∥ C T ∥ op |\langle A_i,C_TA_i\rangle_F|\le4\lVert C_T\rVert_{\operatorname{op}} ∣ ⟨ A i , C T A i ⟩ F ∣ ≤ 4 ∥ C T ∥ op , giving variance at most C n 3 T − 4 = O ( T − 1 ) Cn^3T^{-4}=O(T^{-1}) C n 3 T − 4 = O ( T − 1 ) . For the third, distinct unordered pairs are orthogonal and
E ⟨ A i , C T A j ⟩ F 2 = Tr op ( C T 4 ) = O ( T − 6 ) , \mathbb{E}\langle A_i,C_TA_j\rangle_F^2=\operatorname{Tr}_{\operatorname{op}}(C_T^4)=O(T^{-6}), E ⟨ A i , C T A j ⟩ F 2 = Tr op ( C T 4 ) = O ( T − 6 ) ,
so its variance is bounded by C n 4 T − 6 = O ( T − 2 ) Cn^4T^{-6}=O(T^{-2}) C n 4 T − 6 = O ( T − 2 ) . Consequently,
E ν n 2 = O ( T − 1 ) = o ( 1 ) , \mathbb{E}\nu_n^2=O(T^{-1})=o(1), E ν n 2 = O ( T − 1 ) = o ( 1 ) ,
which proves (C.6).
The fourth-moment bound also gives the maximal-increment assertion:
P ( max k ≤ n ∣ Y k ∣ > ε ) ≤ ε − 4 ∑ k = 1 n E ( Y k 4 ) = o ( 1 ) . P\left(\max_{k\le n}|Y_k|>\varepsilon\right)\le\varepsilon^{-4}\sum_{k=1}^{n}\mathbb{E}(Y_k^4)=o(1). P ( k ≤ n max ∣ Y k ∣ > ε ) ≤ ε − 4 k = 1 ∑ n E ( Y k 4 ) = o ( 1 ) .
The normalized assertion follows because σ n 2 ≥ 4 / λ 2 \sigma_n^2\ge4/\lambda^2 σ n 2 ≥ 4/ λ 2 is bounded away from zero as λ → λ 0 ∈ ( 0 , ∞ ) \lambda\to\lambda_0\in(0,\infty) λ → λ 0 ∈ ( 0 , ∞ ) . ■
Proof of Theorem 1. By (C.8),
T r ( R 2 ) − E [ T r ( R 2 ) ] = 2 ( n − 1 ) ∑ i = 1 n ξ i + 2 ∑ i < j h i j ∘ = ∑ k = 1 n Y k . Tr(R^2)-\mathbb{E}[Tr(R^2)]=2(n-1)\sum_{i=1}^{n}\xi_i+2\sum_{i<j}h_{ij}^{\circ}=\sum_{k=1}^{n}Y_k. T r ( R 2 ) − E [ T r ( R 2 )] = 2 ( n − 1 ) i = 1 ∑ n ξ i + 2 i < j ∑ h ij ∘ = k = 1 ∑ n Y k .
Orthogonality of the Hoeffding components and (C.1)–(C.2) give
∑ k = 1 n E ( Y k 2 ) = 4 n ( n − 1 ) 2 E ξ 1 2 + 4 ( n 2 ) E [ ( h 12 ∘ ) 2 ] = 4 n ( n − 1 ) τ 2 , T 2 T 2 + 8 n ( n − 1 ) 2 K T T 3 + O ( n − 1 ) = 4 τ 2 , T 2 λ 2 + 8 K T λ 3 + O ( n − 1 ) = σ n 2 + O ( n − 1 ) , (19) \begin{aligned}
\sum_{k=1}^{n}\mathbb{E}(Y_k^2)
&=4n(n-1)^2\mathbb{E}\xi_1^2+4\binom{n}{2}\mathbb{E}[(h_{12}^{\circ})^2]\\
&=\frac{4n(n-1)\tau_{2,T}^2}{T^2}+\frac{8n(n-1)^2K_T}{T^3}+O(n^{-1})\\
&=\frac{4\tau_{2,T}^2}{\lambda^2}+\frac{8K_T}{\lambda^3}+O(n^{-1})=\sigma_n^2+O(n^{-1}), \tag*{(19)}
\end{aligned} k = 1 ∑ n E ( Y k 2 ) = 4 n ( n − 1 ) 2 E ξ 1 2 + 4 ( 2 n ) E [( h 12 ∘ ) 2 ] = T 2 4 n ( n − 1 ) τ 2 , T 2 + T 3 8 n ( n − 1 ) 2 K T + O ( n − 1 ) = λ 2 4 τ 2 , T 2 + λ 3 8 K T + O ( n − 1 ) = σ n 2 + O ( n − 1 ) , (19)
since the projection remainders contribute O ( n 3 T − 4 + n 2 T − 3 ) = O ( n − 1 ) O(n^3T^{-4}+n^2T^{-3})=O(n^{-1}) O ( n 3 T − 4 + n 2 T − 3 ) = O ( n − 1 ) , and replacing n ( n − 1 ) 2 n(n-1)^2 n ( n − 1 ) 2 and n ( n − 1 ) n(n-1) n ( n − 1 ) by n 3 n^3 n 3 and n 2 n^2 n 2 costs O ( n − 1 ) O(n^{-1}) O ( n − 1 ) . (The first line equals 2 n ( n − 1 ) a T + 4 n ( n − 1 ) ( n − 2 ) b T 2n(n-1)a_T+4n(n-1)(n-2)b_T 2 n ( n − 1 ) a T + 4 n ( n − 1 ) ( n − 2 ) b T , in agreement with (B.1).) Equations (C.6) and (C.9) imply
1 σ n 2 ∑ k = 1 n E ( Y k 2 ∣ F k − 1 ) → p 1. \frac{1}{\sigma_n^2}\sum_{k=1}^{n}\mathbb{E}(Y_k^2\mid\mathcal{F}_{k-1})\xrightarrow{p}1. σ n 2 1 k = 1 ∑ n E ( Y k 2 ∣ F k − 1 ) p 1.
For every ε > 0 \varepsilon> 0 ε > 0 ,
∑ k = 1 n E [ Y k 2 σ n 2 1 { ∣ Y k ∣ > ε σ n } ∣ F k − 1 ] ≤ 1 ε 2 σ n 4 ∑ k = 1 n E ( Y k 4 ∣ F k − 1 ) ⟶ p 0 , \sum_{k=1}^{n} \mathbb{E}\left[\frac{Y_k^2}{\sigma_n^2}\mathbf{1}_{\{|Y_k|>\varepsilon\sigma_n\}}\mid\mathcal{F}_{k-1}\right]\leq\frac{1}{\varepsilon^2\sigma_n^4}\sum_{k=1}^{n}\mathbb{E}\left(Y_k^4\mid\mathcal{F}_{k-1}\right)\overset{p}{\longrightarrow}0, k = 1 ∑ n E [ σ n 2 Y k 2 1 { ∣ Y k ∣ > ε σ n } ∣ F k − 1 ] ≤ ε 2 σ n 4 1 k = 1 ∑ n E ( Y k 4 ∣ F k − 1 ) ⟶ p 0 ,
because the expectation of the last sum tends to zero by (C.5). Thus the conditional Lindeberg condition holds in probability. Together with predictable quadratic-variation convergence to the constant 1, the Hall and Heyde (1980, Cor. 3.1) martingale CLT yields ∑ k Y k / σ n → d N ( 0 , 1 ) \sum_k Y_k/\sigma_n \xrightarrow{d} N(0,1) ∑ k Y k / σ n d N ( 0 , 1 ) , that is, ( Tr ( R 2 ) − E [ Tr ( R 2 ) ] ) / σ n → d N ( 0 , 1 ) (\operatorname{Tr}(R^2)-\mathbb{E}[\operatorname{Tr}(R^2)])/ \sigma_n \xrightarrow{d} N(0,1) ( Tr ( R 2 ) − E [ Tr ( R 2 )]) / σ n d N ( 0 , 1 ) . Because the limit of the conditional variance is constant, the nesting condition on the σ \sigma σ -fields in that corollary is not needed, so the array filtrations F k = F n , k \mathcal{F}_k=\mathcal{F}_{n,k} F k = F n , k are admissible. Finally, σ n ≥ 2 / λ \sigma_n\geq2/\lambda σ n ≥ 2/ λ and Lemma 2 give ( E [ Tr ( R 2 ) ] − μ n ) / σ n = O ( n − 1 ) (\mathbb{E}[\operatorname{Tr}(R^2)]-\mu_n)/\sigma_n=O(n^{-1}) ( E [ Tr ( R 2 )] − μ n ) / σ n = O ( n − 1 ) , and Slutsky’s theorem yields L M → d N ( 0 , 1 ) LM\xrightarrow{d}N(0,1) L M d N ( 0 , 1 ) .
Ulrich Hounyo; Chihwa Kao; Min Seong Kim
Online Supplement to
This Online Supplement contains the proofs not given in the Appendix of the main paper. Section S.A collects the model, notation, Gaussian moment identities, and the expansion of the diagonal normalization used throughout, including in the Appendix of the main paper; Section S.A.3 also proves Lemmas 1 and 4, which are special cases of results in Hounyo, Kao, and Kim [2 ] . Sections S.B–S.E prove the power results: Lemma 5 (Section S.B), Lemma 6 (Section S.C), Theorem 2 (Section S.D), and Corollary 1 (Section S.E). Section S.F collects the auxiliary Gaussian and operator inequalities. Unprefixed theorem, lemma, assumption, section, and equation numbers refer to the main paper.
Setup, Notation, and Preliminary Results
Model, normalization, and notation
Throughout the theoretical analysis, except in Sections S.B–S.E, which consider alternatives and use separate notation, we work under the null hypothesis H 0 : U = I n H_0 : U = I_n H 0 : U = I n . Under H 0 H_0 H 0 , the separable Gaussian model reduces to X = Z Γ 1 / 2 X = Z\Gamma^{1/2} X = Z Γ 1/2 , where Z ∈ R n × T Z \in\mathbb{R}^{n \times T} Z ∈ R n × T has i.i.d. N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) entries and Γ \Gamma Γ satisfies Assumption 2, with the normalization τ 1 , T = T − 1 Tr ( Γ ) = 1 \tau_{1,T} = T^{-1}\operatorname{Tr}(\Gamma) = 1 τ 1 , T = T − 1 Tr ( Γ ) = 1 imposed in Section 2. Let Z i ∈ R T Z_i \in\mathbb{R}^T Z i ∈ R T denote the transpose of the ith row of Z Z Z . The sample covariance matrix and its entries are
S = 1 T Z Z ′ , S i j = 1 T Z i ′ Γ Z j , S = \frac{1}{T}ZZ', \qquad S_{ij} = \frac{1}{T}Z_i'\Gamma Z_j, S = T 1 Z Z ′ , S ij = T 1 Z i ′ Γ Z j ,
and the corresponding sample correlations satisfy
r i j = S i j S i i S j j = Z i ′ Z j ( Z i ′ Z i ) ( Z j ′ Z j ) . r_{ij}=\frac{S_{ij}}{\sqrt{S_{ii}S_{jj}}}=\frac{Z_i^{\prime}Z_j}{\sqrt{(Z_i^{\prime}Z_i)(Z_j^{\prime}Z_j)}}. r ij = S ii S j j S ij = ( Z i ′ Z i ) ( Z j ′ Z j ) Z i ′ Z j .
By the normalization of Γ \Gamma Γ , E ( S i i ) = T − 1 Tr ( Γ ) = 1 \mathbb{E}(S_{ii})=T^{-1}\operatorname{Tr}(\Gamma)=1 E ( S ii ) = T − 1 Tr ( Γ ) = 1 . Under the null, the cross-sectional rows are independent, while temporal dependence is retained through Γ \Gamma Γ . We further define
δ i = S i i − 1 , ∥ δ ∥ ∞ = max 1 ≤ i ≤ n ∣ δ i ∣ , P i = Z i ′ Γ 2 Z i . \delta_i=S_{ii}-1,\qquad\|\delta\|_{\infty}=\max_{1\le i\le n}|\delta_i|,\qquad P_i=Z_i^{\prime}\Gamma^2Z_i. δ i = S ii − 1 , ∥ δ ∥ ∞ = 1 ≤ i ≤ n max ∣ δ i ∣ , P i = Z i ′ Γ 2 Z i .
Thus P i P_i P i is distributed as a weighted sum of independent χ 1 2 \chi_1^2 χ 1 2 variables, with weights given by the eigenvalues of Γ 2 \Gamma^2 Γ 2 . As in § 2 , τ k , T = O ( 1 ) \tau_{k,T}=O(1) τ k , T = O ( 1 ) for each fixed k k k . A symbol with two indices, such as δ t s \delta_{ts} δ t s , denotes the Kronecker delta; with a single index, δ i \delta_i δ i always denotes the diagonal deviation above.
For a matrix H H H , ∥ H ∥ o p \|H\|_{op} ∥ H ∥ o p and ∥ H ∥ F \|H\|_{F} ∥ H ∥ F denote the operator and Frobenius norms, and ⟨ H , G ⟩ F = Tr ( H ′ G ) \langle H,G\rangle_F=\operatorname{Tr}(H^{\prime}G) ⟨ H , G ⟩ F = Tr ( H ′ G ) . For a random variable Y Y Y and p ≥ 1 p\ge1 p ≥ 1 , ∥ Y ∥ p = ( E ∣ Y ∣ p ) 1 / p \|Y\|_p=(\mathbb{E}|Y|^p)^{1/p} ∥ Y ∥ p = ( E ∣ Y ∣ p ) 1/ p . Throughout, C C C and c c c denote positive constants that do not depend on ( n , T ) (n,T) ( n , T ) and may change from line to line.
Uniformity convention and Gaussian moment bounds. For a fixed data-generating temporal covariance sequence satisfying Assumption 2, put K Γ = sup T ∥ Γ T ∥ o p < ∞ K_{\Gamma}=\sup_T\|\Gamma_T\|_{op}<\infty K Γ = sup T ∥ Γ T ∥ o p < ∞ . All bounds below, including the constants implicit in O ( ⋅ ) O(\cdot) O ( ⋅ ) , are uniform in the indices and in n , T n,T n , T for this sequence, with constants allowed to depend on K Γ K_{\Gamma} K Γ and on the fixed moment order. A bound uniform over a class of temporal covariance sequences requires the additional common-envelope restriction sup { Γ T } ∈ G , T ∥ Γ T ∥ o p < ∞ \sup_{\{\Gamma_T\}\in\mathcal{G},\,T}\|\Gamma_T\|_{op}<\infty sup { Γ T } ∈ G , T ∥ Γ T ∥ o p < ∞ ; Assumption 2 alone does not supply one common constant over every admissible sequence.
Lemma S.1 (Uniform Gaussian moment bounds). For each fixed integer m ≥ 1 m\ge1 m ≥ 1 , under Assumptions 1–2 and H 0 H_0 H 0 ,
E ∣ S i i − 1 ∣ 2 m ≤ C m T − m , E ∣ S i j ∣ 2 m ≤ C m T − m ( i ≠ j ) , \mathbb{E}|S_{ii}-1|^{2m}\le C_mT^{-m},\qquad\mathbb{E}|S_{ij}|^{2m}\le C_mT^{-m}\quad(i\ne j), E ∣ S ii − 1 ∣ 2 m ≤ C m T − m , E ∣ S ij ∣ 2 m ≤ C m T − m ( i = j ) ,
where C m C_m C m depends only on m m m and K Γ K_\Gamma K Γ . In particular, ∥ δ i ∥ 2 m = O ( T − 1 / 2 ) \lVert\delta_i\rVert_{2m}=O(T^{-1/2}) ∥ δ i ∥ 2 m = O ( T − 1/2 ) and ∥ S i j ∥ 2 m = O ( T − 1 / 2 ) \lVert S_{ij}\rVert_{2m}=O(T^{-1/2}) ∥ S ij ∥ 2 m = O ( T − 1/2 ) , uniformly in the indices.
Proof. Write δ i = T − 1 ( Z i ′ Γ Z i − T Tr Γ ) \delta_i=T^{-1}(Z_i^\prime\Gamma Z_i-T\operatorname{Tr}\Gamma) δ i = T − 1 ( Z i ′ Γ Z i − T Tr Γ ) . The centered Gaussian quadratic-form moment inequality gives E ∣ Z i ′ Γ Z i − T Tr Γ ∣ 2 m ≤ C m [ T Tr ( Γ 2 ) ] m \mathbb{E}\lvert Z_i^\prime\Gamma Z_i-T\operatorname{Tr}\Gamma\rvert^{2m}\le C_m[T\operatorname{Tr}(\Gamma^2)]^m E ∣ Z i ′ Γ Z i − T Tr Γ ∣ 2 m ≤ C m [ T Tr ( Γ 2 ) ] m . Since Tr ( Γ 2 ) ≤ K Γ T \operatorname{Tr}(\Gamma^2)\le K_\Gamma T Tr ( Γ 2 ) ≤ K Γ T , the first bound follows. For i ≠ j i\ne j i = j , conditional on Z i Z_i Z i , S i j S_{ij} S ij is centered Gaussian with variance T − 2 Z i ′ Γ 2 Z i T^{-2}Z_i^\prime\Gamma^2Z_i T − 2 Z i ′ Γ 2 Z i . Hence
E ∣ S i j ∣ 2 m = ( 2 m − 1 ) ! ! T − 2 m E ( Z i ′ Γ 2 Z i ) m . \mathbb{E}\lvert S_{ij}\rvert^{2m}=(2m-1)!!\,T^{-2m}\mathbb{E}(Z_i^\prime\Gamma^2Z_i)^m. E ∣ S ij ∣ 2 m = ( 2 m − 1 )!! T − 2 m E ( Z i ′ Γ 2 Z i ) m .
Minkowski’s inequality applied to the nonnegative eigenvalue-weighted chi-square sum gives ∥ Z i ′ Γ 2 Z i ∥ m ≤ ∥ χ 1 2 ∥ m Tr ( Γ 2 ) ≤ C m T \lVert Z_i^\prime\Gamma^2Z_i\rVert_m\le\lVert\chi^2_1\rVert_m\operatorname{Tr}(\Gamma^2)\le C_mT ∥ Z i ′ Γ 2 Z i ∥ m ≤ ∥ χ 1 2 ∥ m Tr ( Γ 2 ) ≤ C m T , proving the second bound. ■
Isserlis’ (Wick’s) Theorem
The moment calculations used throughout the proofs rely on the following standard Gaussian identities. Let Z ∼ N ( 0 , I T ) Z\sim N(0,I_T) Z ∼ N ( 0 , I T ) and let A A A , B B B , and C C C be deterministic symmetric T × T T\times T T × T matrices.
Lemma S.2 (Isserlis identities; Isserlis (1918))
E [ Z ′ A Z ] = Tr ( A ) , (20) \mathbb{E}[Z^\prime A Z]=\operatorname{Tr}(A), \tag*{(20)} E [ Z ′ A Z ] = Tr ( A ) , (20) E [ ( Z ′ A Z ) 2 ] = [ Tr ( A ) ] 2 + 2 Tr ( A 2 ) , (21) \mathbb{E}[(Z^\prime A Z)^2]=[\operatorname{Tr}(A)]^2+2\operatorname{Tr}(A^2), \tag*{(21)} E [( Z ′ A Z ) 2 ] = [ Tr ( A ) ] 2 + 2 Tr ( A 2 ) , (21) E [ ( Z ′ A Z ) ( Z ′ B Z ) ] = Tr ( A ) Tr ( B ) + 2 Tr ( A B ) , (22) \mathbb{E}[(Z^\prime A Z)(Z^\prime B Z)]=\operatorname{Tr}(A)\operatorname{Tr}(B)+2\operatorname{Tr}(AB), \tag*{(22)} E [( Z ′ A Z ) ( Z ′ B Z )] = Tr ( A ) Tr ( B ) + 2 Tr ( A B ) , (22) E [ ( Z ′ A Z ) ( Z ′ B Z ) ( Z ′ C Z ) ] = Tr ( A ) Tr ( B ) Tr ( C ) + 2 Tr ( A ) Tr ( B C ) + 2 Tr ( B ) Tr ( A C ) + 2 Tr ( C ) Tr ( A B ) + 8 Tr ( A B C ) . (23) \begin{aligned}
\mathbb{E}[(Z^\prime A Z)(Z^\prime B Z)(Z^\prime C Z)]
&=\operatorname{Tr}(A)\operatorname{Tr}(B)\operatorname{Tr}(C)+2\operatorname{Tr}(A)\operatorname{Tr}(BC)+2\operatorname{Tr}(B)\operatorname{Tr}(AC)\\
&\quad+2\operatorname{Tr}(C)\operatorname{Tr}(AB)+8\operatorname{Tr}(ABC). \tag*{(23)}
\end{aligned} E [( Z ′ A Z ) ( Z ′ B Z ) ( Z ′ C Z )] = Tr ( A ) Tr ( B ) Tr ( C ) + 2 Tr ( A ) Tr ( B C ) + 2 Tr ( B ) Tr ( A C ) + 2 Tr ( C ) Tr ( A B ) + 8 Tr ( A B C ) . (23)
These identities follow from Wick’s theorem [3 ] by summing over the pairings of the Gaussian coordinates; (S.A.1) is immediate, and (S.A.2) is the case B = A B=A B = A of (S.A.3). The following corollary collects the special cases used repeatedly in the proofs.
Corollary S.1. Under Z i ∼ N ( 0 , I T ) Z_i \sim N(0, I_T) Z i ∼ N ( 0 , I T ) independent across i i i :
E [ P i ] = E [ Z i ⊤ Γ 2 Z i ] = T τ 2 , T , (24) \mathbb{E}[P_i] = \mathbb{E}[Z_i^\top\Gamma^2 Z_i] = T\tau_{2,T}, \tag*{(24)} E [ P i ] = E [ Z i ⊤ Γ 2 Z i ] = T τ 2 , T , (24) E [ P i ⋅ Z i ⊤ Γ Z i ] = ( T τ 2 , T ) ( T ) + 2 T τ 3 , T = T 2 τ 2 , T + 2 T τ 3 , T , (25) \mathbb{E}[P_i \cdot Z_i^\top\Gamma Z_i] = (T\tau_{2,T})(T) + 2T\tau_{3,T} = T^2\tau_{2,T} + 2T\tau_{3,T}, \tag*{(25)} E [ P i ⋅ Z i ⊤ Γ Z i ] = ( T τ 2 , T ) ( T ) + 2 T τ 3 , T = T 2 τ 2 , T + 2 T τ 3 , T , (25) E [ P i 2 ] = E [ ( Z i ⊤ Γ 2 Z i ) 2 ] = ( T τ 2 , T ) 2 + 2 T τ 4 , T = T 2 τ 2 , T 2 + 2 T τ 4 , T , (26) \mathbb{E}[P_i^2] = \mathbb{E}[(Z_i^\top\Gamma^2 Z_i)^2] = (T\tau_{2,T})^2 + 2T\tau_{4,T} = T^2\tau_{2,T}^2 + 2T\tau_{4,T}, \tag*{(26)} E [ P i 2 ] = E [( Z i ⊤ Γ 2 Z i ) 2 ] = ( T τ 2 , T ) 2 + 2 T τ 4 , T = T 2 τ 2 , T 2 + 2 T τ 4 , T , (26) E [ P i ( Z i ⊤ Γ Z i ) 2 ] = T 3 τ 2 , T + 2 T 2 τ 2 , T 2 + 4 T 2 τ 3 , T + 8 T τ 4 , T . (27) \mathbb{E}[P_i(Z_i^\top\Gamma Z_i)^2] = T^3\tau_{2,T} + 2T^2\tau_{2,T}^2 + 4T^2\tau_{3,T} + 8T\tau_{4,T}. \tag*{(27)} E [ P i ( Z i ⊤ Γ Z i ) 2 ] = T 3 τ 2 , T + 2 T 2 τ 2 , T 2 + 4 T 2 τ 3 , T + 8 T τ 4 , T . (27)
Identities (S.A.6), (S.A.7) and (S.A.8) follow from (S.A.3), (S.A.2) and (S.A.4) with ( A , B ) = ( Γ 2 , Γ ) (A, B) = (\Gamma^2, \Gamma) ( A , B ) = ( Γ 2 , Γ ) , A = Γ 2 A = \Gamma^2 A = Γ 2 and ( A , B , C ) = ( Γ 2 , Γ , Γ ) (A, B, C) = (\Gamma^2, \Gamma, \Gamma) ( A , B , C ) = ( Γ 2 , Γ , Γ ) , respectively.
Lemma S.3 (Rate for the spectral moments). Let Γ T \Gamma_T Γ T satisfy Assumption 2, write γ h = γ ∣ h ∣ \gamma_h = \gamma_{|h|} γ h = γ ∣ h ∣ for h ∈ Z h \in\mathbb{Z} h ∈ Z , and put A = ∑ h ∈ Z ∣ γ h ∣ A = \sum_{h\in\mathbb{Z}}|\gamma_h| A = ∑ h ∈ Z ∣ γ h ∣ and B = ∑ h ∈ Z ∣ h ∣ ∣ γ h ∣ B = \sum_{h\in\mathbb{Z}}|h||\gamma_h| B = ∑ h ∈ Z ∣ h ∣∣ γ h ∣ . If B < ∞ B < \infty B < ∞ , then for each fixed k ≥ 2 k \ge2 k ≥ 2 and every T T T ,
∣ τ k , T − τ k ∣ ≤ ( k − 1 ) B A k − 2 T . |\tau_{k,T} - \tau_k| \le\frac{(k-1)BA^{k-2}}{T}. ∣ τ k , T − τ k ∣ ≤ T ( k − 1 ) B A k − 2 . Consequently K T − K = O ( T − 1 ) K_T - K = O(T^{-1}) K T − K = O ( T − 1 ) .
Proof. Since f ( ω ) = ∑ h γ h e i h ω f(\omega) = \sum_h \gamma_h e^{ih\omega} f ( ω ) = ∑ h γ h e ihω with A < ∞ A < \infty A < ∞ ,
τ k = ( 2 π ) − 1 ∫ − π π f k = ∑ h 1 + ⋯ + h k = 0 h j ∈ Z ∏ j = 1 k γ h j . \tau_k = (2\pi)^{-1}\int_{-\pi}^{\pi} f^k = \sum_{\substack{h_1+\cdots+h_k=0\\h_j\in\mathbb{Z}}}\prod_{j=1}^{k}\gamma_{h_j}. τ k = ( 2 π ) − 1 ∫ − π π f k = h 1 + ⋯ + h k = 0 h j ∈ Z ∑ j = 1 ∏ k γ h j .
an absolutely convergent sum. With t k + 1 = t 1 t_{k+1}=t_1 t k + 1 = t 1 , T r ( Γ T k ) = ∑ t 1 , … , t k = 1 T ∏ j = 1 k γ t j − t j + 1 Tr(\Gamma_T^k) = \sum_{t_1,\ldots,t_k=1}^{T}\prod_{j=1}^{k}\gamma_{t_j-t_{j+1}} T r ( Γ T k ) = ∑ t 1 , … , t k = 1 T ∏ j = 1 k γ t j − t j + 1 .
Parametrize by t 1 t_1 t 1 and the lags h j = t j − t j + 1 h_j=t_j-t_{j+1} h j = t j − t j + 1 , j < k j<k j < k , so that h k = − ( h 1 + ⋯ + h k − 1 ) h_k=-(h_1+\cdots+h_{k-1}) h k = − ( h 1 + ⋯ + h k − 1 ) . For fixed lags, the points t j = t 1 − ( h 1 + ⋯ + h j − 1 ) t_j=t_1-(h_1+\cdots+h_{j-1}) t j = t 1 − ( h 1 + ⋯ + h j − 1 ) all lie in { 1 , … , T } \{1,\ldots,T\} { 1 , … , T } for exactly ( T − w ) + (T-w)_+ ( T − w ) + values of t 1 t_1 t 1 , where w w w is the range of the partial sums and satisfies w ≤ ∑ j < k ∣ h j ∣ w\le\sum_{j<k}|h_j| w ≤ ∑ j < k ∣ h j ∣ . Hence, using ∣ γ h k ∣ ≤ γ 0 = 1 |\gamma_{h_k}| \le\gamma_0 = 1 ∣ γ h k ∣ ≤ γ 0 = 1 ,
∣ T τ k − Tr ( Γ k r ) ∣ ≤ ∑ h 1 , … , h k − 1 ∏ j < k ∣ γ h j ∣ min ( T , w ) ≤ ∑ h 1 , … , h k − 1 ∏ j < k ∣ γ h j ∣ ∑ j < k ∣ h j ∣ = ( k − 1 ) B A k − 2 . \begin{aligned}
|T\tau_k-\operatorname{Tr}(\Gamma_k^r)| \le\sum_{h_1,\ldots,h_{k-1}} \prod_{j<k}|\gamma_{h_j}|\min(T,w) \\
&\le\sum_{h_1,\ldots,h_{k-1}} \prod_{j<k}|\gamma_{h_j}|\sum_{j<k}|h_j|=(k-1)BA^{k-2}.
\end{aligned} ∣ T τ k − Tr ( Γ k r ) ∣ ≤ h 1 , … , h k − 1 ∑ j < k ∏ ∣ γ h j ∣ min ( T , w ) ≤ h 1 , … , h k − 1 ∑ j < k ∏ ∣ γ h j ∣ j < k ∑ ∣ h j ∣ = ( k − 1 ) B A k − 2 .
Dividing by T T T gives the bound. Since K T K_T K T and K K K are the same polynomial in ( τ 2 , T , τ 3 , T , τ 4 , T ) (\tau_{2,T},\tau_{3,T},\tau_{4,T}) ( τ 2 , T , τ 3 , T , τ 4 , T ) and ( τ 2 , τ 3 , τ 4 ) (\tau_2,\tau_3,\tau_4) ( τ 2 , τ 3 , τ 4 ) , with bounded arguments, K T − K = O ( T − 1 ) K_T-K=O(T^{-1}) K T − K = O ( T − 1 ) . ■
For i ≠ j i\ne j i = j , the following conditional moments hold.
Lemma S.4 (Conditional distribution of S i j S_{ij} S ij ) Given Z i Z_i Z i , the random variable S i j = 1 T Z j ′ Γ Z i S_{ij}=\frac{1}{T}Z_j' \Gamma Z_i S ij = T 1 Z j ′ Γ Z i is Gaussian with
E [ S i j ∣ Z i ] = 0 , Var [ S i j ∣ Z i ] = Z i ′ Γ 2 Z i T 2 = P i T 2 . \mathbb{E}[S_{ij}\mid Z_i]=0,\qquad\operatorname{Var}[S_{ij}\mid Z_i]=\frac{Z_i'\Gamma^2Z_i}{T^2}=\frac{P_i}{T^2}. E [ S ij ∣ Z i ] = 0 , Var [ S ij ∣ Z i ] = T 2 Z i ′ Γ 2 Z i = T 2 P i . In particular, E [ S i j 2 ] = τ 2 , T / T \mathbb{E}[S_{ij}^2]=\tau_{2,T}/T E [ S ij 2 ] = τ 2 , T / T by (S.A.5).
This holds because, conditional on Z i Z_i Z i , S i j = Z j ′ ( Γ Z i / T ) S_{ij}=Z_j'(\Gamma Z_i/T) S ij = Z j ′ ( Γ Z i / T ) is a linear form in Z j ∼ N ( 0 , I T ) Z_j\sim N(0,I_T) Z j ∼ N ( 0 , I T ) , which is independent of Z i Z_i Z i .
Diagonal concentration: proof of Lemmas 1 and 4
Lemmas 1 and 4 are special cases of Hounyo et al. [2 ] . We give the proof for completeness; this subsection allows a general U U U satisfying Assumption 4. Since Lemma 1 is the case U = I n U=I_n U = I n of Lemma 4, it suffices to prove Lemma 4.
Proof of Lemma 4. Let X = U 1 / 2 Z Γ 1 / 2 X=U^{1/2}Z\Gamma^{1/2} X = U 1/2 Z Γ 1/2 and a i = U 1 / 2 e i a_i=U^{1/2}e_i a i = U 1/2 e i , so that ∥ a i ∥ 2 = U i i \lVert a_i\rVert^2=U_{ii} ∥ a i ∥ 2 = U ii and
S i i = 1 T a i ′ Z Γ Z ′ a i = vec ( Z ) ′ B i vec ( Z ) , B i = Γ T ⊗ ( a i a i ′ ) . S_{ii}=\frac{1}{T}a_i'Z\Gamma Z'a_i=\operatorname{vec}(Z)'B_i\operatorname{vec}(Z),\qquad B_i=\frac{\Gamma}{T}\otimes(a_ia_i'). S ii = T 1 a i ′ Z Γ Z ′ a i = vec ( Z ) ′ B i vec ( Z ) , B i = T Γ ⊗ ( a i a i ′ ) .
Since diag ( S ) − diag ( U ) \operatorname{diag}(S)-\operatorname{diag}(U) diag ( S ) − diag ( U ) is diagonal, its operator norm is max i ∣ S i i − U i i ∣ \max_i|S_{ii}-U_{ii}| max i ∣ S ii − U ii ∣ . Because Tr ( Γ ) = T \operatorname{Tr}(\Gamma)=T Tr ( Γ ) = T , we have E ( S i i ) = Tr ( B i ) = U i i \mathbb{E}(S_{ii}) = \operatorname{Tr}(B_i) = U_{ii} E ( S ii ) = Tr ( B i ) = U ii . Moreover,
∥ B i ∥ F 2 = Tr ( Γ 2 ) U i i 2 T 2 = τ 2 , T U i i 2 T ≤ τ ˉ 2 c 2 2 T , ∥ B i ∥ op = ∥ Γ ∥ op U i i T ≤ K Γ c 2 T . \lVert B_i\rVert_F^2 = \frac{\operatorname{Tr}(\Gamma^2)U_{ii}^2}{T^2} = \frac{\tau_{2,T}U_{ii}^2}{T} \le\frac{\bar{\tau}_2c_2^2}{T}, \qquad\lVert B_i\rVert_{\operatorname{op}} = \frac{\lVert\Gamma\rVert_{\operatorname{op}}U_{ii}}{T} \le\frac{K_\Gamma c_2}{T}. ∥ B i ∥ F 2 = T 2 Tr ( Γ 2 ) U ii 2 = T τ 2 , T U ii 2 ≤ T τ ˉ 2 c 2 2 , ∥ B i ∥ op = T ∥ Γ ∥ op U ii ≤ T K Γ c 2 .
where τ ˉ 2 = sup T τ 2 , T \bar{\tau}_2 = \sup_T \tau_{2,T} τ ˉ 2 = sup T τ 2 , T and K Γ = sup T ∥ Γ ∥ op K_\Gamma= \sup_T \lVert\Gamma\rVert_{\operatorname{op}} K Γ = sup T ∥ Γ ∥ op are finite by Assumption 2, and U i i ≤ c 2 U_{ii} \le c_2 U ii ≤ c 2 by Assumption 4. Since vec ( Z ) \operatorname{vec}(Z) vec ( Z ) has i.i.d. N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) entries, the Hanson–Wright inequality [4 Thm. 1.1] gives, for every ε > 0 \varepsilon> 0 ε > 0 ,
P ( ∣ S i i − U i i ∣ > ε ) ≤ 2 exp { − c min ( ε 2 ∥ B i ∥ F 2 , ε ∥ B i ∥ op ) } . P\left(\lvert S_{ii}-U_{ii}\rvert>\varepsilon\right)\leq2\exp\left\{-c\min\left(\frac{\varepsilon^2}{\lVert B_i\rVert_F^2},\frac{\varepsilon}{\lVert B_i\rVert_{\operatorname{op}}}\right)\right\}. P ( ∣ S ii − U ii ∣ > ε ) ≤ 2 exp { − c min ( ∥ B i ∥ F 2 ε 2 , ∥ B i ∥ op ε ) } .
where c > 0 c > 0 c > 0 is a universal constant independent of i i i , n n n , and T T T . Set ε n = M log n / T \varepsilon_n = M\sqrt{\log n/T} ε n = M log n / T . Then
ε n 2 ∥ B i ∥ F 2 ≥ M 2 log n τ ˉ 2 c 2 2 , ε n ∥ B i ∥ op ≥ M T log n K Γ c 2 . \frac{\varepsilon_n^2}{\lVert B_i\rVert_F^2}\geq\frac{M^2\log n}{\bar{\tau}_2c_2^2}, \qquad\frac{\varepsilon_n}{\lVert B_i\rVert_{\operatorname{op}}}\geq\frac{M\sqrt{T\log n}}{K_\Gamma c_2}. ∥ B i ∥ F 2 ε n 2 ≥ τ ˉ 2 c 2 2 M 2 log n , ∥ B i ∥ op ε n ≥ K Γ c 2 M T log n .
Under Assumption 3, n → ∞ n\to\infty n → ∞ , T → ∞ T\to\infty T → ∞ , and T / log n → ∞ T/\log n\to\infty T / log n → ∞ , so the first term is the smaller one for all large n n n , and
P ( ∣ S i i − U i i ∣ > M log n T ) ≤ 2 n − c M 2 / ( τ ˉ 2 c 2 2 ) P\left(\lvert S_{ii}-U_{ii}\rvert>M\sqrt{\frac{\log n}{T}}\right)\leq2n^{-cM^2/(\bar{\tau}_2c_2^2)} P ( ∣ S ii − U ii ∣ > M T log n ) ≤ 2 n − c M 2 / ( τ ˉ 2 c 2 2 )
uniformly in i i i . A union bound over i = 1 , … , n i=1,\ldots,n i = 1 , … , n gives
P ( max i ∣ S i i − U i i ∣ > M log n T ) ≤ 2 n 1 − c M 2 / ( τ ˉ 2 c 2 2 ) . P\left(\max_i\lvert S_{ii}-U_{ii}\rvert>M\sqrt{\frac{\log n}{T}}\right)\leq2n^{1-cM^2/(\bar{\tau}_2c_2^2)}. P ( i max ∣ S ii − U ii ∣ > M T log n ) ≤ 2 n 1 − c M 2 / ( τ ˉ 2 c 2 2 ) .
Choosing M > ( 2 τ ˉ 2 c 2 2 / c ) 1 / 2 M>(2\bar{\tau}_2c_2^2/c)^{1/2} M > ( 2 τ ˉ 2 c 2 2 / c ) 1/2 makes the exponent smaller than − 1 -1 − 1 , so the right-hand side is summable in n n n . By the Borel–Cantelli lemma, max i ∣ S i i − U i i ∣ ≤ M log n / T \max_i\lvert S_{ii}-U_{ii}\rvert\leq M\sqrt{\log n/T} max i ∣ S ii − U ii ∣ ≤ M log n / T for all sufficiently large n n n , almost surely. This proves Lemma 4, and Lemma 1 follows with U = I n U=I_n U = I n and c 2 = 1 c_2=1 c 2 = 1 . ▪
For every fixed ε > 0 \varepsilon>0 ε > 0 , the same exponential bound yields P ( ∣ S i i − U i i ∣ > ε ) ≤ C ε e − c ε T P(\lvert S_{ii}-U_{ii}\rvert>\varepsilon)\leq C_\varepsilon e^{-c_\varepsilon T} P (∣ S ii − U ii ∣ > ε ) ≤ C ε e − c ε T uniformly in i i i , where C ε , c ε > 0 C_\varepsilon,c_\varepsilon>0 C ε , c ε > 0 may depend on ε \varepsilon ε but not on i i i , n n n , or T T T . This per-unit tail bound is the form used in the proofs of Lemmas 2, 3, and 5.
Diagonal normalization
Write R = D S − 1 / 2 S D S − 1 / 2 R=D_S^{-1/2}SD_S^{-1/2} R = D S − 1/2 S D S − 1/2 , where D S = diag ( S 11 , … , S n n ) D_S=\operatorname{diag}(S_{11},\ldots,S_{nn}) D S = diag ( S 11 , … , S nn ) , and let E n = { ∥ δ ∥ ∞ ≤ 1 / 2 } \mathcal{E}_n=\{\lVert\delta\rVert_\infty\le1/2\} E n = {∥ δ ∥ ∞ ≤ 1/2 } .
Lemma S.5 (Expansion of the diagonal normalization) Under Assumptions 1–3 and H 0 H_0 H 0 , P ( E n ) → 1 P(\mathcal{E}_n)\to1 P ( E n ) → 1 , and on E n \mathcal{E}_n E n the following hold uniformly in i ≠ j i\ne j i = j :
(a) S j j − 1 = 1 − δ j + O ( δ j 2 ) S_{jj}^{-1}=1-\delta_j+O(\delta_j^2) S j j − 1 = 1 − δ j + O ( δ j 2 ) ;
(b) ( S i i S j j ) − 1 = 1 − ( δ i + δ j ) + O ( ∥ δ ∥ ∞ 2 ) (S_{ii}S_{jj})^{-1}=1-(\delta_i+\delta_j)+O(\lVert\delta\rVert_\infty^2) ( S ii S j j ) − 1 = 1 − ( δ i + δ j ) + O (∥ δ ∥ ∞ 2 ) ;
(c) r i j 2 = S i j 2 − S i j 2 ( δ i + δ j ) + O ( S i j 2 ∥ δ ∥ ∞ 2 ) r_{ij}^2=S_{ij}^2-S_{ij}^2(\delta_i+\delta_j)+O(S_{ij}^2\lVert\delta\rVert_\infty^2) r ij 2 = S ij 2 − S ij 2 ( δ i + δ j ) + O ( S ij 2 ∥ δ ∥ ∞ 2 ) .
Proof. By Lemma 1, ∥ δ ∥ ∞ = ∥ diag ( S ) − I n ∥ o p = O ( log n / T ) \lVert\delta\rVert_\infty=\lVert\operatorname{diag}(S)-I_n\rVert_{\mathrm{op}}=O(\sqrt{\log n/T}) ∥ δ ∥ ∞ = ∥ diag ( S ) − I n ∥ op = O ( log n / T ) almost surely, so ∥ δ ∥ ∞ → 0 \lVert\delta\rVert_\infty\to0 ∥ δ ∥ ∞ → 0 almost surely and P ( E n ) → 1 P(\mathcal{E}_n)\to1 P ( E n ) → 1 . On E n \mathcal{E}_n E n , every S j j ∈ [ 1 / 2 , 3 / 2 ] S_{jj}\in[1/2,3/2] S j j ∈ [ 1/2 , 3/2 ] , and Taylor’s formula gives S j j − 1 = ( 1 + δ j ) − 1 = 1 − δ j + ε j S_{jj}^{-1}=(1+\delta_j)^{-1}=1-\delta_j+\varepsilon_j S j j − 1 = ( 1 + δ j ) − 1 = 1 − δ j + ε j with ∣ ε j ∣ ≤ C δ j 2 |\varepsilon_j|\le C\delta_j^2 ∣ ε j ∣ ≤ C δ j 2 uniformly in j j j . This is (a); (b) follows by multiplying the expansions for i i i and j j j , and (c) follows from (b) since r i j 2 = S i j 2 / ( S i i S j j ) r_{ij}^2=S_{ij}^2/(S_{ii}S_{jj}) r ij 2 = S ij 2 / ( S ii S j j ) .
The pairwise bounds in Lemma S.5 do not by themselves make the aggregate remainder negligible. The displayed big-O O O statements are pointwise deterministic inequalities on E n \mathcal{E}_n E n , with constants independent of i , j , n , T i,j,n,T i , j , n , T once K Γ K_\Gamma K Γ is fixed. Summed over all pairs, S i j 2 ∥ δ ∥ ∞ 2 S_{ij}^2\lVert\delta\rVert_\infty^2 S ij 2 ∥ δ ∥ ∞ 2 is only O ( log n ) O(\log n) O ( log n ) , and the first-order terms S i j 2 ( δ i + δ j ) S_{ij}^2(\delta_i+\delta_j) S ij 2 ( δ i + δ j ) contribute at the CLT scale to both the mean and the variance. The proofs of Lemmas 2 and 3 therefore retain these terms explicitly and control the remainders in expectation after summation over pairs.
Proof of Lemma 5
We establish a uniform expansion for a single squared sample correlation and then sum it.
Fix i ≠ j i\ne j i = j and write ρ = ρ i j ∗ \rho=\rho_{ij}^* ρ = ρ ij ∗ . Put v a = Z ′ U 1 / 2 e a / U a a 1 / 2 v_a=Z'U^{1/2}e_a/U_{aa}^{1/2} v a = Z ′ U 1/2 e a / U aa 1/2 for a = i , j a=i,j a = i , j , so that v a ∼ N ( 0 , I T ) v_a\sim N(0,I_T) v a ∼ N ( 0 , I T ) and Cov ( v i , v j ) = ρ I T \operatorname{Cov}(v_i,v_j)=\rho I_T Cov ( v i , v j ) = ρ I T . Then r i j = s / ( q i q j ) 1 / 2 r_{ij}=s/(q_iq_j)^{1/2} r ij = s / ( q i q j ) 1/2 with
s = T − 1 v i ′ Γ v j = S i j U i i U j j , q a = T − 1 v a ′ Γ v a = S a a U a a . s=T^{-1}v_i'\Gamma v_j=\frac{S_{ij}}{\sqrt{U_{ii}U_{jj}}},\qquad q_a=T^{-1}v_a'\Gamma v_a=\frac{S_{aa}}{U_{aa}}. s = T − 1 v i ′ Γ v j = U ii U j j S ij , q a = T − 1 v a ′ Γ v a = U aa S aa .
Lemma S.1 (Uniform pairwise mean expansion) Under Assumptions 1–4, uniformly in i ≠ j i \ne j i = j and all sufficiently large T T T ,
E A ( r i j 2 ) = ρ 2 + τ 2 , T T ( 1 − 3 ρ 2 + 2 ρ 4 ) + 4 ( τ 2 , T 2 − τ 3 , T ) T 2 + ϵ i j , ∣ ϵ i j ∣ ≤ C ( ρ 2 T 2 + 1 T 3 ) . (28) \mathbb{E}_{A}(r_{ij}^{2})=\rho^{2}+\frac{\tau_{2,T}}{T}(1-3\rho^{2}+2\rho^{4})+\frac{4(\tau_{2,T}^{2}-\tau_{3,T})}{T^{2}}+\epsilon_{ij}, \qquad\lvert\epsilon_{ij}\rvert\le C\left(\frac{\rho^{2}}{T^{2}}+\frac{1}{T^{3}}\right). \tag*{(28)} E A ( r ij 2 ) = ρ 2 + T τ 2 , T ( 1 − 3 ρ 2 + 2 ρ 4 ) + T 2 4 ( τ 2 , T 2 − τ 3 , T ) + ϵ ij , ∣ ϵ ij ∣ ≤ C ( T 2 ρ 2 + T 3 1 ) . (28) The constant C C C depends only on K Γ K_{\Gamma} K Γ , not on i , j , n , T i,j,n,T i , j , n , T or ρ \rho ρ .
Proof. The argument does not divide by 1 − ρ 2 1-\rho^{2} 1 − ρ 2 , so the bounds are uniform as ∣ ρ ∣ → 1 \lvert\rho\rvert\to1 ∣ ρ ∣ → 1 . In an eigenbasis of Γ \Gamma Γ ,
( s , q i , q j ) = T − 1 ∑ t = 1 T ν t , T ( X t Y t , X t 2 , Y t 2 ) , Y t = ρ X t + 1 − ρ 2 Z t , (s,q_i,q_j)=T^{-1}\sum_{t=1}^{T}\nu_{t,T}(X_tY_t,X_t^{2},Y_t^{2}), \qquad Y_t=\rho X_t+\sqrt{1-\rho^{2}}Z_t, ( s , q i , q j ) = T − 1 t = 1 ∑ T ν t , T ( X t Y t , X t 2 , Y t 2 ) , Y t = ρ X t + 1 − ρ 2 Z t , where the X t , Z t X_t,Z_t X t , Z t are independent standard normals. Put η = ( η 1 , η 2 , η 3 ) = ( s − ρ , q i − 1 , q j − 1 ) \eta=(\eta_1,\eta_2,\eta_3)=(s-\rho,q_i-1,q_j-1) η = ( η 1 , η 2 , η 3 ) = ( s − ρ , q i − 1 , q j − 1 ) and h ( x , y , z ) = x 2 / ( y z ) h(x,y,z)=x^{2}/(yz) h ( x , y , z ) = x 2 / ( y z ) , so that r i j 2 = h ( ( ρ , 1 , 1 ) + η ) r_{ij}^{2}=h((\rho,1,1)+\eta) r ij 2 = h (( ρ , 1 , 1 ) + η ) and 0 ≤ r i j 2 ≤ 1 0\le r_{ij}^{2}\le1 0 ≤ r ij 2 ≤ 1 . Let ζ = ( X Y − ρ , X 2 − 1 , Y 2 − 1 ) \zeta=(XY-\rho,X^{2}-1,Y^{2}-1) ζ = ( X Y − ρ , X 2 − 1 , Y 2 − 1 ) for a single bivariate normal pair with unit variances and correlation ρ \rho ρ , and let C a b = E ( ζ a ζ b ) C_{ab}=\mathbb{E}(\zeta_a\zeta_b) C ab = E ( ζ a ζ b ) and H a b c = E ( ζ a ζ b ζ c ) H_{abc}=\mathbb{E}(\zeta_a\zeta_b\zeta_c) H ab c = E ( ζ a ζ b ζ c ) ; explicitly,
C = ( 1 + ρ 2 2 ρ 2 ρ 2 ρ 2 2 ρ 2 2 ρ 2 ρ 2 2 ) . C=\begin{pmatrix}
1+\rho^{2} & 2\rho& 2\rho\\
2\rho& 2 & 2\rho^{2}\\
2\rho& 2\rho^{2} & 2
\end{pmatrix}. C = 1 + ρ 2 2 ρ 2 ρ 2 ρ 2 2 ρ 2 2 ρ 2 ρ 2 2 . Since η = T − 1 ∑ t ν t , T ζ t \eta=T^{-1}\sum_t\nu_{t,T}\zeta_t η = T − 1 ∑ t ν t , T ζ t with i.i.d. copies ζ t \zeta_t ζ t of ζ \zeta ζ , the k k k th joint cumulant of coordinates of η \eta η equals T − k T r ( Γ k ) T^{-k}T r(\Gamma^{k}) T − k T r ( Γ k ) times the corresponding cumulant of ζ \zeta ζ , hence is O ( T 1 − k ) O(T^{1-k}) O ( T 1 − k ) uniformly in ∣ ρ ∣ ≤ 1 \lvert\rho\rvert\le1 ∣ ρ ∣ ≤ 1 , and is a polynomial in ρ \rho ρ . In particular,
E ( η a η b ) = τ 2 , T T C a b , E ( η a η b η c ) = τ 3 , T T 2 H a b c , \mathbb{E}(\eta_a\eta_b)=\frac{\tau_{2,T}}{T}C_{ab}, \qquad\mathbb{E}(\eta_a\eta_b\eta_c)=\frac{\tau_{3,T}}{T^{2}}H_{abc}, E ( η a η b ) = T τ 2 , T C ab , E ( η a η b η c ) = T 2 τ 3 , T H ab c , E ( η a η b η c η d ) = τ 2 , T 2 T 2 ( C a b C c d + C a c C b d + C a d C b c ) + O ( T − 3 ) . (29) \mathbb{E}(\eta_a\eta_b\eta_c\eta_d)=\frac{\tau_{2,T}^{2}}{T^{2}}(C_{ab}C_{cd}+C_{ac}C_{bd}+C_{ad}C_{bc})+O(T^{-3}). \tag*{(29)} E ( η a η b η c η d ) = T 2 τ 2 , T 2 ( C ab C c d + C a c C b d + C a d C b c ) + O ( T − 3 ) . (29) where the first two identities are exact. By the moment–cumulant formula, a product of d d d coordinates of η \eta η has expectation O ( T − ⌈ d / 2 ⌉ ) O(T^{-\lceil d/2\rceil}) O ( T − ⌈ d /2 ⌉ ) ; in particular E ∥ η ∥ 2 m ≤ C m T − m \mathbb{E}\lVert\eta\rVert^{2m}\le C_mT^{-m} E ∥ η ∥ 2 m ≤ C m T − m for each
fixed m m m , and degrees five and six contribute O ( T − 3 ) O(T^{-3}) O ( T − 3 ) . Moreover, since each q a q_a q a has the null distribution of S a a S_{aa} S aa , P { min ( q i , q j ) < 1 / 2 } ≤ C e − c T P\{\min(q_i,q_j)<1/2\}\le Ce^{-cT} P { min ( q i , q j ) < 1/2 } ≤ C e − c T by the per-unit tail bound following the proof of Lemma 4.
On { min ( q i , q j ) ≥ 1 / 2 } \{\min(q_i,q_j)\ge1/2\} { min ( q i , q j ) ≥ 1/2 } , apply Taylor’s theorem to h h h at ( ρ , 1 , 1 ) (\rho,1,1) ( ρ , 1 , 1 ) through total degree five. The segment joining ( ρ , 1 , 1 ) (\rho,1,1) ( ρ , 1 , 1 ) and ( s , q i , q j ) (s,q_i,q_j) ( s , q i , q j ) lies in { ( x , y , z ) : y , z ≥ 1 / 2 , x 2 ≤ y z } \{(x,y,z):y,z\ge1/2,\ x^2\le yz\} {( x , y , z ) : y , z ≥ 1/2 , x 2 ≤ y z } , because both endpoints are entries of 2 × 2 2\times2 2 × 2 positive semidefinite matrices with diagonal entries at least 1 / 2 1/2 1/2 and that set is convex. Every sixth-order derivative of h h h is a finite sum of terms x k y − a z − b x^k y^{-a}z^{-b} x k y − a z − b with k ≤ 2 k\le2 k ≤ 2 and a , b ≥ 1 a,b\ge1 a , b ≥ 1 ; since ∣ x ∣ k ≤ ( y z ) k / 2 |x|^k\le(yz)^{k/2} ∣ x ∣ k ≤ ( y z ) k /2 and y , z ≥ 1 / 2 y,z\ge1/2 y , z ≥ 1/2 there, it is bounded on this set. The expected sixth-order remainder is therefore O ( E ∥ η ∥ 6 ) = O ( T − 3 ) O(\mathbb{E}\|\eta\|^6)=O(T^{-3}) O ( E ∥ η ∥ 6 ) = O ( T − 3 ) . The complementary event contributes O ( e − c T ) O(e^{-cT}) O ( e − c T ) , using 0 ≤ r i j 2 ≤ 1 0\le r_{ij}^2\le1 0 ≤ r ij 2 ≤ 1 for the correlation and Cauchy–Schwarz for the Taylor polynomial. Writing P d P_d P d for the homogeneous degree-d d d part of h ( ( ρ , 1 , 1 ) + η ) h((\rho,1,1)+\eta) h (( ρ , 1 , 1 ) + η ) ,
P d ( η ) = ∑ k = 0 min ( 2 , d ) ( 2 k ) ρ 2 − k η 1 k ( − 1 ) d − k ∑ l = 0 d − k η 2 l η 3 d − k − l P_d(\eta)=\sum_{k=0}^{\min(2,d)}\binom{2}{k}\rho^{2-k}\eta_1^k(-1)^{d-k}\sum_{l=0}^{d-k}\eta_2^l\eta_3^{d-k-l} P d ( η ) = k = 0 ∑ m i n ( 2 , d ) ( k 2 ) ρ 2 − k η 1 k ( − 1 ) d − k l = 0 ∑ d − k η 2 l η 3 d − k − l
we have E P 1 ( η ) = 0 \mathbb{E}P_1(\eta)=0 E P 1 ( η ) = 0 , and (S.B.2) gives
E A ( r i j 2 ) = ρ 2 + T − 1 A T ( ρ ) + T − 2 B T ( ρ ) + O ( T − 3 ) , \mathbb{E}_A(r_{ij}^2)=\rho^2+T^{-1}A_T(\rho)+T^{-2}B_T(\rho)+O(T^{-3}), E A ( r ij 2 ) = ρ 2 + T − 1 A T ( ρ ) + T − 2 B T ( ρ ) + O ( T − 3 ) ,
uniformly in ∣ ρ ∣ ≤ 1 |\rho|\le1 ∣ ρ ∣ ≤ 1 , where A T A_T A T collects the C C C -contraction of P 2 P_2 P 2 , and B T B_T B T collects the H H H -contraction of P 3 P_3 P 3 and the pairwise C C C -contractions of P 4 P_4 P 4 . For example, P 2 ( η ) = η 1 2 − 2 ρ η 1 ( η 2 + η 3 ) + ρ 2 ( η 2 2 + η 2 η 3 + η 3 2 ) P_2(\eta)=\eta_1^2-2\rho\eta_1(\eta_2+\eta_3)+\rho^2(\eta_2^2+\eta_2\eta_3+\eta_3^2) P 2 ( η ) = η 1 2 − 2 ρ η 1 ( η 2 + η 3 ) + ρ 2 ( η 2 2 + η 2 η 3 + η 3 2 ) , whose C C C -contraction is ( 1 + ρ 2 ) − 8 ρ 2 + ρ 2 ( 4 + 2 ρ 2 ) = 1 − 3 ρ 2 + 2 ρ 4 (1+\rho^2)-8\rho^2+\rho^2(4+2\rho^2)=1-3\rho^2+2\rho^4 ( 1 + ρ 2 ) − 8 ρ 2 + ρ 2 ( 4 + 2 ρ 2 ) = 1 − 3 ρ 2 + 2 ρ 4 . Carrying out the contractions,
A T ( ρ ) = τ 2 , T ( 1 − 3 ρ 2 + 2 ρ 4 ) , B T ( ρ ) = 2 ( ρ 2 − 1 ) [ ( 4 ρ 4 + ρ 2 − 2 ) τ 2 , T 2 + ( 2 − 4 ρ 2 ) τ 3 , T ] . A_T(\rho)=\tau_{2,T}(1-3\rho^2+2\rho^4),\qquad B_T(\rho)=2(\rho^2-1)\left[(4\rho^4+\rho^2-2)\tau_{2,T}^2+(2-4\rho^2)\tau_{3,T}\right]. A T ( ρ ) = τ 2 , T ( 1 − 3 ρ 2 + 2 ρ 4 ) , B T ( ρ ) = 2 ( ρ 2 − 1 ) [ ( 4 ρ 4 + ρ 2 − 2 ) τ 2 , T 2 + ( 2 − 4 ρ 2 ) τ 3 , T ] .
At ρ = 0 \rho=0 ρ = 0 , B T ( 0 ) = 4 ( τ 2 , T 2 − τ 3 , T ) B_T(0)=4(\tau_{2,T}^2-\tau_{3,T}) B T ( 0 ) = 4 ( τ 2 , T 2 − τ 3 , T ) , in agreement with the null expansion in Appendix A of the main paper. Moreover,
B T ( ρ ) − B T ( 0 ) = 2 ρ 2 [ ( 4 ρ 4 − 3 ρ 2 − 3 ) τ 2 , T 2 + ( 6 − 4 ρ 2 ) τ 3 , T ] . \mathcal{B}_T(\rho)-\mathcal{B}_T(0)=2\rho^2\left[(4\rho^4-3\rho^2-3)\tau_{2,T}^2+(6-4\rho^2)\tau_{3,T}\right]. B T ( ρ ) − B T ( 0 ) = 2 ρ 2 [ ( 4 ρ 4 − 3 ρ 2 − 3 ) τ 2 , T 2 + ( 6 − 4 ρ 2 ) τ 3 , T ] .
so ∣ B T ( ρ ) − B T ( 0 ) ∣ ≤ C ρ 2 |\mathcal{B}_T(\rho)-\mathcal{B}_T(0)|\le C\rho^2 ∣ B T ( ρ ) − B T ( 0 ) ∣ ≤ C ρ 2 for ∣ ρ ∣ ≤ 1 |\rho|\le1 ∣ ρ ∣ ≤ 1 , since τ 2 , T \tau_{2,T} τ 2 , T and τ 3 , T \tau_{3,T} τ 3 , T are bounded. This proves (S.B.1).
The case ∣ ρ ∣ = 1 |\rho|=1 ∣ ρ ∣ = 1 is included: there r i j 2 = 1 r_{ij}^2=1 r ij 2 = 1 and A T ( ± 1 ) = B T ( ± 1 ) = 0 \mathcal{A}_T(\pm1)=\mathcal{B}_T(\pm1)=0 A T ( ± 1 ) = B T ( ± 1 ) = 0 .
We now sum (S.B.1) over the n ( n − 1 ) n(n-1) n ( n − 1 ) ordered pairs. Since ∑ i ≠ j ( ρ i j ∗ ) 2 = n ( ϕ 2 − 1 ) = O ( n ) \sum_{i\ne j}(\rho_{ij}^*)^2=n(\phi_2-1)=O(n) ∑ i = j ( ρ ij ∗ ) 2 = n ( ϕ 2 − 1 ) = O ( n ) by Assumption 4 and T = n λ T=n\lambda T = nλ ,
∑ i ≠ j ∣ ϵ i j ∣ ≤ C { n ( ϕ 2 − 1 ) T 2 + n ( n − 1 ) T 3 } = O ( n − 1 ) . \sum_{i\ne j}|\epsilon_{ij}|\le C\left\{\frac{n(\phi_2-1)}{T^2}+\frac{n(n-1)}{T^3}\right\}=O(n^{-1}). i = j ∑ ∣ ϵ ij ∣ ≤ C { T 2 n ( ϕ 2 − 1 ) + T 3 n ( n − 1 ) } = O ( n − 1 ) .
Hence, using T r ( R 2 ) = n + ∑ i ≠ j r i j 2 Tr(R^2)=n+\sum_{i\ne j}r_{ij}^2 T r ( R 2 ) = n + ∑ i = j r ij 2 , n + n ( ϕ 2 − 1 ) = T r ( R ∗ 2 ) n+n(\phi_2-1)=Tr(R_*^2) n + n ( ϕ 2 − 1 ) = T r ( R ∗ 2 ) , ∑ i ≠ j ( 1 − 3 ( ρ i j ∗ ) 2 + 2 ( ρ i j ∗ ) 4 ) = n ( n − 1 ) − 3 n ( ϕ 2 − 1 ) + 2 ∑ i ≠ j ( ρ i j ∗ ) 4 \sum_{i\ne j}(1-3(\rho_{ij}^*)^2+2(\rho_{ij}^*)^4)=n(n-1)-3n(\phi_2-1)+2\sum_{i\ne j}(\rho_{ij}^*)^4 ∑ i = j ( 1 − 3 ( ρ ij ∗ ) 2 + 2 ( ρ ij ∗ ) 4 ) = n ( n − 1 ) − 3 n ( ϕ 2 − 1 ) + 2 ∑ i = j ( ρ ij ∗ ) 4 , and n / T = 1 / λ n/T=1/\lambda n / T = 1/ λ ,
E A [ T r ( R 2 ) ] = T r ( R ∗ 2 ) + ( n − 1 ) τ 2 , T λ + 4 ( n − 1 ) ( τ 2 , T 2 − τ 3 , T ) n λ 2 − 3 τ 2 , T ( ϕ 2 − 1 ) λ + 2 τ 2 , T T ∑ i ≠ j ( ρ i j ∗ ) 4 + O ( n − 1 ) , \mathbb{E}_{A}[Tr(R^2)]=Tr(R_*^2)+\frac{(n-1)\tau_{2,T}}{\lambda}+\frac{4(n-1)(\tau_{2,T}^2-\tau_{3,T})}{n\lambda^2}-\frac{3\tau_{2,T}(\phi_2-1)}{\lambda}+\frac{2\tau_{2,T}}{T}\sum_{i\ne j}(\rho_{ij}^*)^4+O(n^{-1}), E A [ T r ( R 2 )] = T r ( R ∗ 2 ) + λ ( n − 1 ) τ 2 , T + n λ 2 4 ( n − 1 ) ( τ 2 , T 2 − τ 3 , T ) − λ 3 τ 2 , T ( ϕ 2 − 1 ) + T 2 τ 2 , T i = j ∑ ( ρ ij ∗ ) 4 + O ( n − 1 ) ,
which is (3). All constants are uniform under Assumptions 1–4.
Proof of Lemma 6
We establish the two assertions of Lemma 6 through separate auxiliary results: a uniform Gaussian Poincaré bound and a sharper covariance comparison under local alternatives. We also prove here the coupling result used for Theorem 2 (Lemma S.2).
Throughout, R ∗ = I n + E R_* = I_n+E R ∗ = I n + E with E E E symmetric and zero-diagonal, so that ∥ E ∥ F 2 = ∑ i ≠ j ( ρ i j ∗ ) 2 = n ( ϕ 2 − 1 ) \|E\|_F^2=\sum_{i\ne j}(\rho_{ij}^*)^2=n(\phi_2-1) ∥ E ∥ F 2 = ∑ i = j ( ρ ij ∗ ) 2 = n ( ϕ 2 − 1 ) and, by Assumption 4, ∥ R ∗ ∥ o p ≤ c 2 / c 1 \|R_*\|_{\mathrm{op}}\le c_2/c_1 ∥ R ∗ ∥ op ≤ c 2 / c 1 . Because R R R is invariant to rescaling the rows of X X X , its law depends on U U U only through R ∗ R_* R ∗ ; without loss of generality we take
X = W Z Γ 1 / 2 , W = R ∗ 1 / 2 = I n + Θ , (30) X=WZ\Gamma^{1/2},\qquad W=R_*^{1/2}=I_n+\Theta, \tag*{(30)} X = W Z Γ 1/2 , W = R ∗ 1/2 = I n + Θ , (30)
with W W W the symmetric square root. Since W 2 = I n + E W^2=I_n+E W 2 = I n + E and T r ( E ) = 0 Tr(E)=0 T r ( E ) = 0 ,
2 T r ( Θ ) + T r ( Θ 2 ) = 0 exactly , ∥ Θ ∥ F ≤ ∥ E ∥ F , ∥ Θ ∥ o p ≤ C , (31) 2Tr(\Theta)+Tr(\Theta^2)=0 \quad\text{exactly}, \qquad\|\Theta\|_F\leq\|E\|_F, \qquad\|\Theta\|_{\mathrm{op}}\leq C, \tag*{(31)} 2 T r ( Θ ) + T r ( Θ 2 ) = 0 exactly , ∥Θ ∥ F ≤ ∥ E ∥ F , ∥Θ ∥ op ≤ C , (31)
where the Frobenius bound follows from ∣ μ − 1 ∣ ≤ ∣ μ − 1 ∣ |\sqrt{\mu}-1|\leq|\mu-1| ∣ μ − 1∣ ≤ ∣ μ − 1∣ applied to the eigenvalues μ \mu μ of R ∗ R_\ast R ∗ .
Operator-norm moments. Put c + = sup n / T < ∞ c_+=\sup n/T<\infty c + = sup n / T < ∞ and recall K Γ = sup T ∥ Γ T ∥ o p K_\Gamma=\sup_T\|\Gamma_T\|_{\mathrm{op}} K Γ = sup T ∥ Γ T ∥ op ; for p ≥ 2 p\geq2 p ≥ 2 let
z p = 1 + c + + 4 p + 2 , m p = 1 + K Γ z p 2 ; z_p=1+\sqrt{c_+}+\sqrt{4p+2}, \qquad m_p=1+K_\Gamma z_p^2; z p = 1 + c + + 4 p + 2 , m p = 1 + K Γ z p 2 ;
these constants depend only on c + c_+ c + and K Γ K_\Gamma K Γ , not on n , T n,T n , T . The Gaussian operator-norm tail inequality Pr { ∥ Z ∥ o p > T + n + t } ≤ e − t 2 / 2 \Pr\{\|Z\|_{\mathrm{op}}>\sqrt{T}+\sqrt{n}+t\}\leq e^{-t^2/2} Pr { ∥ Z ∥ op > T + n + t } ≤ e − t 2 /2 for t ≥ 0 t\geq0 t ≥ 0 ([1 Thm. II.13] ), followed by integration of the tail, gives, for fixed p ≥ 2 p\geq2 p ≥ 2 and all sufficiently large T T T ,
∥ ∥ Z Γ ∥ o p ∥ p ≤ K Γ z p T , ∥ ∥ N ∥ o p ∥ p ≤ m p , N = T − 1 Z Γ Z ′ − I n . (32) \left\|\|Z\Gamma\|_{\mathrm{op}}\right\|_p\leq K_\Gamma z_p\sqrt{T}, \qquad\left\|\|N\|_{\mathrm{op}}\right\|_p\leq m_p, \qquad N=T^{-1}Z\Gamma Z'-I_n. \tag*{(32)} ∥ ∥ Z Γ ∥ op ∥ p ≤ K Γ z p T , ∥ ∥ N ∥ op ∥ p ≤ m p , N = T − 1 Z Γ Z ′ − I n . (32)
To verify these bounds, let ξ = ∥ Z ∥ o p − T − n \xi=\|Z\|_{\mathrm{op}}-\sqrt{T}-\sqrt{n} ξ = ∥ Z ∥ op − T − n . Tail integration gives ∥ ξ + ∥ p ≤ 4 p + 2 \|\xi_+\|_p\leq\sqrt{4p+2} ∥ ξ + ∥ p ≤ 4 p + 2 , so ∥ ∥ Z ∥ o p ∥ p ≤ z p T \left\|\|Z\|_{\mathrm{op}}\right\|_p\leq z_p\sqrt{T} ∥ ∥ Z ∥ op ∥ p ≤ z p T ; the first bound then follows from ∥ Z Γ ∥ o p ≤ K Γ ∥ Z ∥ o p \|Z\Gamma\|_{\mathrm{op}}\leq K_\Gamma\|Z\|_{\mathrm{op}} ∥ Z Γ ∥ op ≤ K Γ ∥ Z ∥ op , and the second from ∥ N ∥ o p ≤ 1 + K Γ ∥ Z ∥ o p 2 / T \|N\|_{\mathrm{op}}\leq1+K_\Gamma\|Z\|_{\mathrm{op}}^2/T ∥ N ∥ op ≤ 1 + K Γ ∥ Z ∥ op 2 / T and ∥ ∥ Z ∥ o p 2 ∥ p = ∥ ∥ Z ∥ o p ∥ 2 p 2 \left\|\|Z\|_{\mathrm{op}}^2\right\|_p=\left\|\|Z\|_{\mathrm{op}}\right\|_{2p}^2 ∥ Z ∥ op 2 p = ∥ ∥ Z ∥ op ∥ 2 p 2 . No invertibility of Γ \Gamma Γ is needed.
Lemma S.1 (Variance bound). Under Assumptions 1–4, V a r ( T r ( R 2 ) ) ≤ C Var(Tr(R^2))\leq C V a r ( T r ( R 2 )) ≤ C for a constant C C C not depending on ( n , T ) (n,T) ( n , T ) .
Proof. Let x i x_i x i denote the rows of X X X and y i = x i / ∥ x i ∥ y_i=x_i/\|x_i\| y i = x i /∥ x i ∥ , so that f = T r ( R 2 ) = ∑ i , j ( y i ′ y j ) 2 f=Tr(R^2)=\sum_{i,j}(y_i'y_j)^2 f = T r ( R 2 ) = ∑ i , j ( y i ′ y j ) 2 . On the full-measure set where all x i ≠ 0 x_i\ne0 x i = 0 , f f f is smooth and
∂ f ∂ x i = 4 ∥ x i ∥ ( I T − y i y i ′ ) ∑ j r i j y j . \frac{\partial f}{\partial x_i}=\frac{4}{\|x_i\|}(I_T-y_i y_i')\sum_j r_{ij}y_j. ∂ x i ∂ f = ∥ x i ∥ 4 ( I T − y i y i ′ ) j ∑ r ij y j .
Since ∥ ∑ j r i j y j ∥ 2 = ( R 3 ) i i ≤ ∥ R ∥ o p ( R 2 ) i i \left\|\sum_j r_{ij}y_j\right\|^2 = (R^3)_{ii} \le\|R\|_{\mathrm{op}}(R^2)_{ii} ∑ j r ij y j 2 = ( R 3 ) ii ≤ ∥ R ∥ op ( R 2 ) ii and Tr ( R 2 ) ≤ n ∥ R ∥ o p \operatorname{Tr}(R^2) \le n\|R\|_{\mathrm{op}} Tr ( R 2 ) ≤ n ∥ R ∥ op ,
∑ i = 1 n ∥ ∂ f ∂ x i ∥ 2 ≤ 16 ∥ R ∥ o p Tr ( R 2 ) min 1 ≤ i ≤ n ∥ x i ∥ 2 ≤ 16 n ∥ R ∥ o p 2 T min 1 ≤ i ≤ n S i i . \sum_{i=1}^{n}\left\|\frac{\partial f}{\partial x_i}\right\|^2 \le\frac{16\|R\|_{\mathrm{op}}\operatorname{Tr}(R^2)}{\min_{1\le i\le n}\|x_i\|^2} \le\frac{16n\|R\|_{\mathrm{op}}^2}{T\min_{1\le i\le n}S_{ii}}. i = 1 ∑ n ∂ x i ∂ f 2 ≤ min 1 ≤ i ≤ n ∥ x i ∥ 2 16∥ R ∥ op Tr ( R 2 ) ≤ T min 1 ≤ i ≤ n S ii 16 n ∥ R ∥ op 2 .
The map Z ↦ X Z \mapsto X Z ↦ X is linear (in Frobenius norm) with operator norm at most ∥ W ∥ o p ∥ Γ 1 / 2 ∥ o p ≤ C \|W\|_{\mathrm{op}}\|\Gamma^{1/2}\|_{\mathrm{op}} \le C ∥ W ∥ op ∥ Γ 1/2 ∥ op ≤ C , and ∥ R ∥ o p ≤ ∥ S ∥ o p / min i S i i ≤ C ∥ Z Z ′ / T ∥ o p / min i S i i \|R\|_{\mathrm{op}} \le\|S\|_{\mathrm{op}}/\min_i S_{ii} \le C\|ZZ'/T\|_{\mathrm{op}}/\min_i S_{ii} ∥ R ∥ op ≤ ∥ S ∥ op / min i S ii ≤ C ∥ Z Z ′ / T ∥ op / min i S ii . Hence
∥ ∇ Z f ∥ 2 ≤ C ( n / T ) ∥ Z Z ′ / T ∥ o p 2 ( min i S i i ) − 3 . \|\nabla_Z f\|^2 \le C(n/T)\|ZZ'/T\|_{\mathrm{op}}^2\left(\min_i S_{ii}\right)^{-3}. ∥ ∇ Z f ∥ 2 ≤ C ( n / T ) ∥ Z Z ′ / T ∥ op 2 ( i min S ii ) − 3 .
Under Assumption 3, E ∥ Z Z ′ / T ∥ o p 4 = O ( 1 ) \mathbb{E}\|ZZ'/T\|_{\mathrm{op}}^4 = O(1) E ∥ Z Z ′ / T ∥ op 4 = O ( 1 ) by (S.C.2) with Γ = I T \Gamma= I_T Γ = I T . Each S i i S_{ii} S ii is distributed as v ′ Γ v / T v'\Gamma v/T v ′ Γ v / T with v ∼ N ( 0 , I T ) v \sim N(0,I_T) v ∼ N ( 0 , I T ) , which has bounded inverse moments of any fixed order by (B.2) and satisfies P ( S i i < 1 / 2 ) ≤ C e − c T P(S_{ii}<1/2) \le Ce^{-cT} P ( S ii < 1/2 ) ≤ C e − c T ; therefore E ( min i S i i ) − 6 ≤ 2 6 + ∑ i ( E S i i − 12 ) 1 / 2 P ( S i i < 1 / 2 ) 1 / 2 = O ( 1 ) \mathbb{E}(\min_i S_{ii})^{-6} \le2^6+\sum_i(\mathbb{E}S_{ii}^{-12})^{1/2}P(S_{ii}<1/2)^{1/2}=O(1) E ( min i S ii ) − 6 ≤ 2 6 + ∑ i ( E S ii − 12 ) 1/2 P ( S ii < 1/2 ) 1/2 = O ( 1 ) . By Hölder’s inequality, E ∥ ∇ Z f ∥ 2 = O ( n / T ) = O ( 1 ) \mathbb{E}\|\nabla_Z f\|^2 = O(n/T) = O(1) E ∥ ∇ Z f ∥ 2 = O ( n / T ) = O ( 1 ) , and the Gaussian Poincaré inequality gives Var ( f ) ≤ E ∥ ∇ Z f ∥ 2 ≤ C \operatorname{Var}(f) \le\mathbb{E}\|\nabla_Z f\|^2 \le C Var ( f ) ≤ E ∥ ∇ Z f ∥ 2 ≤ C . (The singular set where some x i = 0 x_i=0 x i = 0 has codimension rank ( Γ ) ≥ 3 \operatorname{rank}(\Gamma) \ge3 rank ( Γ ) ≥ 3 for large T T T , so f f f , which is bounded by n 2 n^2 n 2 , belongs to the Gaussian Sobolev space and the inequality applies.) ■
Lemma S.2 (Coupling under local alternatives) Under Assumptions 1–4, suppose n ( ϕ 2 − 1 ) = O ( 1 ) n(\phi_2-1)=O(1) n ( ϕ 2 − 1 ) = O ( 1 ) . Let R 0 R_0 R 0 be the sample correlation matrix of X 0 = Z Γ 1 / 2 X_0=Z\Gamma^{1/2} X 0 = Z Γ 1/2 , built from the same Z Z Z , so that R 0 R_0 R 0 has the null distribution. Then
D n = Tr ( R 2 ) − Tr ( R 0 2 ) − n ( ϕ 2 − 1 ) satisfies E ( D n 2 ) = O ( n − 1 ) . D_n=\operatorname{Tr}(R^2)-\operatorname{Tr}(R_0^2)-n(\phi_2-1)\quad\text{satisfies}\quad\mathbb{E}(D_n^2)=O(n^{-1}). D n = Tr ( R 2 ) − Tr ( R 0 2 ) − n ( ϕ 2 − 1 ) satisfies E ( D n 2 ) = O ( n − 1 ) .
Proof. Write S 0 = X 0 X 0 ′ / T = I n + N S_0=X_0X_0'/T=I_n+N S 0 = X 0 X 0 ′ / T = I n + N . Then S = W S 0 W = R ∗ + G S=WS_0W=R_*+G S = W S 0 W = R ∗ + G with
G = W N W = N + H , H = Θ N + N Θ + Θ N Θ . G=WNW=N+H,\qquad H=\Theta N+N\Theta+\Theta N\Theta. G = W N W = N + H , H = Θ N + N Θ + Θ N Θ.
The entries of N N N form a centered Gaussian chaos of order two with the exact covariance
Cov ( N a b , N c d ) = ( δ a c δ b d + δ a d δ b c ) τ 2 , T T , (33) \operatorname{Cov}(N_{ab},N_{cd})=(\delta_{ac}\delta_{bd}+\delta_{ad}\delta_{bc})\frac{\tau_{2,T}}{T}, \tag*{(33)} Cov ( N ab , N c d ) = ( δ a c δ b d + δ a d δ b c ) T τ 2 , T , (33)
so that, for a deterministic symmetric matrix A A A , Var ( Tr ( A N ) ) = 2 τ 2 , T ∥ A ∥ F 2 / T \operatorname{Var}(\operatorname{Tr}(AN)) = 2\tau_{2,T}\lVert A\rVert_F^2/T Var ( Tr ( A N )) = 2 τ 2 , T ∥ A ∥ F 2 / T . Joint cumulants of order m m m of entries of N N N admit trace-contraction formulas with scale O ( T − ( m − 1 ) ) O(T^{-(m-1)}) O ( T − ( m − 1 ) ) for fixed m m m under the uniform spectral-norm bound. Moreover, by Gaussian hypercontractivity ∥ Q ∥ p ≤ C p ∥ Q ∥ 2 \lVert Q\rVert_p \le C_p\lVert Q\rVert_2 ∥ Q ∥ p ≤ C p ∥ Q ∥ 2 for any polynomial Q Q Q of bounded degree in the entries of N N N .
Put a i = S i i = 1 + G i i a_i=S_{ii}=1+G_{ii} a i = S ii = 1 + G ii and b i = ( S 0 ) i i = 1 + N i i b_i=(S_0)_{ii}=1+N_{ii} b i = ( S 0 ) ii = 1 + N ii . On E = { a i , b i ∈ [ 1 / 2 , 2 ] ∀ i } \mathcal{E}=\{a_i,b_i\in[1/2,2]\ \forall i\} E = { a i , b i ∈ [ 1/2 , 2 ] ∀ i } , whose complement has probability at most C n e − c T Cne^{-cT} C n e − c T , the functions ( a i , a j ) ↦ 1 / ( a i a j ) (a_i,a_j)\mapsto1/(a_i a_j) ( a i , a j ) ↦ 1/ ( a i a j ) have bounded derivatives of all orders; the probability bound holds because each a i a_i a i and b i b_i b i is a unit-mean Gaussian quadratic form with eigenvalue weights bounded by K Γ K_\Gamma K Γ . Since n ≤ Tr ( R 2 ) n\le\operatorname{Tr}(R^2) n ≤ Tr ( R 2 ) , Tr ( R 0 2 ) ≤ n 2 \operatorname{Tr}(R_0^2)\le n^2 Tr ( R 0 2 ) ≤ n 2 and ∥ E ∥ F 2 ≤ n ( n − 1 ) \lVert E\rVert_F^2\le n(n-1) ∥ E ∥ F 2 ≤ n ( n − 1 ) , ∣ D n ∣ ≤ 3 n 2 \lvert D_n\rvert\le3n^2 ∣ D n ∣ ≤ 3 n 2 always, so E c \mathcal{E}^c E c contributes O ( n 5 e − c T ) O(n^5e^{-cT}) O ( n 5 e − c T ) to E ( D n 2 ) \mathbb{E}(D_n^2) E ( D n 2 ) , and we work on E \mathcal{E} E ; below, ∥ T j ∥ 2 \lVert T_j\rVert_2 ∥ T j ∥ 2 denotes ∥ T j 1 E ∥ 2 \lVert T_j1_{\mathcal{E}}\rVert_2 ∥ T j 1 E ∥ 2 . Expanding S i j 2 = ( E i j + N i j + H i j ) 2 S_{ij}^2=(E_{ij}+N_{ij}+H_{ij})^2 S ij 2 = ( E ij + N ij + H ij ) 2 and using ∑ i ≠ j E i j 2 = n ( ϕ 2 − 1 ) \sum_{i\ne j}E_{ij}^2=n(\phi_2-1) ∑ i = j E ij 2 = n ( ϕ 2 − 1 ) ,
D n = T 1 + T 2 + T 3 + T 4 , D_n=T_1+T_2+T_3+T_4, D n = T 1 + T 2 + T 3 + T 4 ,
where
T 1 = ∑ i ≠ j E i j 2 ( 1 a i a j − 1 ) , T 2 = 2 ∑ i ≠ j E i j G i j a i a j , T 3 = ∑ i ≠ j 2 N i j H i j + H i j 2 a i a j , T 4 = ∑ i ≠ j N i j 2 ( 1 a i a j − 1 b i b j ) . \begin{aligned}
T_1&=\sum_{i\ne j}E_{ij}^2\left(\frac{1}{a_i a_j}-1\right),&
T_2&=2\sum_{i\ne j}\frac{E_{ij}G_{ij}}{a_i a_j},\\
T_3&=\sum_{i\ne j}\frac{2N_{ij}H_{ij}+H_{ij}^2}{a_i a_j},&
T_4&=\sum_{i\ne j}N_{ij}^2\left(\frac{1}{a_i a_j}-\frac{1}{b_i b_j}\right).
\end{aligned} T 1 T 3 = i = j ∑ E ij 2 ( a i a j 1 − 1 ) , = i = j ∑ a i a j 2 N ij H ij + H ij 2 , T 2 T 4 = 2 i = j ∑ a i a j E ij G ij , = i = j ∑ N ij 2 ( a i a j 1 − b i b j 1 ) .
Let w i = W e i w_i=We_i w i = W e i , so that ∥ w i ∥ 2 = ( R ⋆ ) i i = 1 \lVert w_i\rVert^2=(R_\star)_{ii}=1 ∥ w i ∥ 2 = ( R ⋆ ) ii = 1 , and let f i = ∥ Θ i ⋅ ∥ 2 f_i=\lVert\Theta_{i\cdot}\rVert_2 f i = ∥ Θ i ⋅ ∥ 2 , where Θ i ⋅ \Theta_{i\cdot} Θ i ⋅ is the i i i th row of Θ \Theta Θ ; thus ∑ i f i ≤ n ∥ Θ ∥ F = O ( n ) \sum_i f_i\le\sqrt{n}\lVert\Theta\rVert_F=O(\sqrt{n}) ∑ i f i ≤ n ∥ Θ ∥ F = O ( n ) . Since G a b = w a ′ N w b G_{ab}=w_a^\prime Nw_b G ab = w a ′ N w b and H i i = Tr { ( w i w i ′ − e i e i ′ ) N } H_{ii}=\operatorname{Tr}\{(w_iw_i^\prime-e_ie_i^\prime)N\} H ii = Tr {( w i w i ′ − e i e i ′ ) N } with ∥ w i w i ′ − e i e i ′ ∥ F ≤ 2 ∥ w i − e i ∥ = 2 f i \lVert w_iw_i^\prime-e_ie_i^\prime\rVert_F\le\sqrt{2}\lVert w_i-e_i\rVert=\sqrt{2}f_i ∥ w i w i ′ − e i e i ′ ∥ F ≤ 2 ∥ w i − e i ∥ = 2 f i (both vectors have unit length), (S.C.3) and hypercontractivity give, for each fixed p ≥ 2 p\ge2 p ≥ 2 and uniformly in the indices,
∥ G a b ∥ p + ∥ N a b ∥ p ≤ C p T − 1 / 2 , ∥ H i i ∥ p ≤ C p T − 1 / 2 f i . \lVert G_{ab}\rVert_p+\lVert N_{ab}\rVert_p\le C_pT^{-1/2},\qquad\lVert H_{ii}\rVert_p\le C_pT^{-1/2}f_i. ∥ G ab ∥ p + ∥ N ab ∥ p ≤ C p T − 1/2 , ∥ H ii ∥ p ≤ C p T − 1/2 f i .
Term T 1 T_1 T 1 . On E \mathcal{E} E , ∣ 1 / ( a i a j ) − 1 ∣ ≤ C ( ∣ G i i ∣ + ∣ G j j ∣ ) |1/(a_i a_j)-1|\le C(|G_{ii}|+|G_{jj}|) ∣1/ ( a i a j ) − 1∣ ≤ C ( ∣ G ii ∣ + ∣ G j j ∣ ) , so Minkowski’s inequality gives
∥ T 1 1 E ∥ 2 ≤ C ∑ i ≠ j E i j 2 ( ∥ G i i ∥ 2 + ∥ G j j ∥ 2 ) ≤ C T − 1 / 2 ∥ E ∥ F 2 = O ( T − 1 / 2 ) . \|T_1\mathbb{1}_{\mathcal{E}}\|_2\le C\sum_{i\ne j}E_{ij}^2(\|G_{ii}\|_2+\|G_{jj}\|_2)\le CT^{-1/2}\|E\|_F^2=O(T^{-1/2}). ∥ T 1 1 E ∥ 2 ≤ C i = j ∑ E ij 2 ( ∥ G ii ∥ 2 + ∥ G j j ∥ 2 ) ≤ C T − 1/2 ∥ E ∥ F 2 = O ( T − 1/2 ) .
Thus ∥ T 1 ∥ 2 = O ( T − 1 / 2 ) \|T_1\|_2=O(T^{-1/2}) ∥ T 1 ∥ 2 = O ( T − 1/2 ) .
Term T 2 T_2 T 2 . Taylor expansion on E \mathcal{E} E yields
T 2 = L 2 + Q 2 + R 2 , L 2 = 2 T Tr ( E G ) , Q 2 = − 4 ∑ i ≠ j E i j G i j G i i . T_2=L_2+Q_2+R_2,\qquad L_2=2T\operatorname{Tr}(EG),\qquad Q_2=-4\sum_{i\ne j}E_{ij}G_{ij}G_{ii}. T 2 = L 2 + Q 2 + R 2 , L 2 = 2 T Tr ( E G ) , Q 2 = − 4 i = j ∑ E ij G ij G ii .
where ∣ R 2 ∣ ≤ C ∑ i ≠ j ∣ E i j G i j ∣ ( G i i 2 + G j j 2 ) |R_2|\le C\sum_{i\ne j}|E_{ij}G_{ij}|(G_{ii}^2+G_{jj}^2) ∣ R 2 ∣ ≤ C ∑ i = j ∣ E ij G ij ∣ ( G ii 2 + G j j 2 ) . The linear term is centered and, by the exact covariance identity,
E L 2 2 = 8 τ 2 , T T ∥ W E W ∥ F 2 = O ( T − 1 ) . \mathbb{E}L_2^2=\frac{8\tau_{2,T}}{T}\|WEW\|_F^2=O(T^{-1}). E L 2 2 = T 8 τ 2 , T ∥ W E W ∥ F 2 = O ( T − 1 ) .
For the quadratic term set D = diag ( G ) D=\operatorname{diag}(G) D = diag ( G ) and J 2 = Tr ( D E G ) J_2=\operatorname{Tr}(DEG) J 2 = Tr ( D E G ) , so Q 2 = − 4 J 2 Q_2=-4J_2 Q 2 = − 4 J 2 (because E i i = 0 E_{ii}=0 E ii = 0 ). Wick’s second-moment identity gives
E J 2 = 2 τ 2 , T T ∑ i ≠ j E i j ( R ∗ ) i j = 2 τ 2 , T T ∥ E ∥ F 2 = O ( T − 1 ) . \mathbb{E}J_2=\frac{2\tau_{2,T}}{T}\sum_{i\ne j}E_{ij}(R_*)_{ij}=\frac{2\tau_{2,T}}{T}\|E\|_F^2=O(T^{-1}). E J 2 = T 2 τ 2 , T i = j ∑ E ij ( R ∗ ) ij = T 2 τ 2 , T ∥ E ∥ F 2 = O ( T − 1 ) .
Differentiation with respect to the symmetric matrix G G G gives
d J 2 = Tr ( B G d G ) , B G = diag ( diag ( E G ) ) + 1 2 ( E D + D E ) . dJ_2=\operatorname{Tr}(B_G\,dG),\qquad B_G=\operatorname{diag}(\operatorname{diag}(EG))+\frac{1}{2}(ED+DE). d J 2 = Tr ( B G d G ) , B G = diag ( diag ( E G )) + 2 1 ( E D + D E ) .
Here G = W N W G=WNW G = W N W , so B N = W B G W B_N=WB_GW B N = W B G W and ∇ Z J 2 = 2 T − 1 B N Z Γ \nabla_ZJ_2=2T^{-1}B_NZ\Gamma ∇ Z J 2 = 2 T − 1 B N Z Γ . For every i i i ,
E ∣ ( E G ) i i ∣ 2 = τ 2 , T T [ ( E R ∗ E ) i i ( R ∗ ) i i + ( E R ∗ ) i i 2 ] ≤ C T ∥ E i ⋅ ∥ 2 2 . \mathbb{E}|(EG)_{ii}|^2=\frac{\tau_{2,T}}{T}\left[(ER_*E)_{ii}(R_*)_{ii}+(ER_*)_{ii}^2\right]\le\frac{C}{T}\|E_{i\cdot}\|_2^2. E ∣ ( E G ) ii ∣ 2 = T τ 2 , T [ ( E R ∗ E ) ii ( R ∗ ) ii + ( E R ∗ ) ii 2 ] ≤ T C ∥ E i ⋅ ∥ 2 2 .
Also E G i i 2 = 2 τ 2 , T / T \mathbb{E}G_{ii}^2=2\tau_{2,T}/T E G ii 2 = 2 τ 2 , T / T , whence E ∥ B G ∥ F 2 ≤ C T − 1 ∥ E ∥ F 2 = O ( T − 1 ) \mathbb{E}\|B_G\|_F^2\le CT^{-1}\|E\|_F^2=O(T^{-1}) E ∥ B G ∥ F 2 ≤ C T − 1 ∥ E ∥ F 2 = O ( T − 1 ) . Since B G B_G B G is linear in the Gaussian quadratic chaos N N N , hypercontractivity yields E ∥ B N ∥ F 4 = O ( T − 2 ) \mathbb{E}\|B_N\|_F^4=O(T^{-2}) E ∥ B N ∥ F 4 = O ( T − 2 ) ; for example, apply the fourth-moment inequality entrywise and then Minkowski’s inequality to the Frobenius sum. The standard Gaussian operator-norm moment bound E ∥ Z Γ ∥ o p 4 = O ( T 2 ) \mathbb{E}\|Z\Gamma\|_{\mathrm{op}}^4=O(T^2) E ∥ Z Γ ∥ op 4 = O ( T 2 ) now gives, by Gaussian Poincaré,
Var ( J 2 ) ≤ 4 T 2 E ∥ B N Z Γ ∥ F 2 ≤ 4 T 2 ( E ∥ B N ∥ F 4 ) 1 / 2 ( E ∥ Z Γ ∥ o p 4 ) 1 / 2 = O ( T − 2 ) . \operatorname{Var}(J_2) \le\frac{4}{T^2}\mathbb{E}\lVert B_N Z\Gamma\rVert_F^2 \le\frac{4}{T^2}\left(\mathbb{E}\lVert B_N\rVert_F^4\right)^{1/2}\left(\mathbb{E}\lVert Z\Gamma\rVert_{\mathrm{op}}^4\right)^{1/2}=O(T^{-2}). Var ( J 2 ) ≤ T 2 4 E ∥ B N Z Γ ∥ F 2 ≤ T 2 4 ( E ∥ B N ∥ F 4 ) 1/2 ( E ∥ Z Γ ∥ op 4 ) 1/2 = O ( T − 2 ) .
Thus ∥ Q 2 ∥ 2 = O ( T − 1 ) \lVert Q_2\rVert_2=O(T^{-1}) ∥ Q 2 ∥ 2 = O ( T − 1 ) . Finally, by Hölder and the uniform fixed-order quadratic-chaos moments,
∥ R 21 ε ∥ 2 ≤ C T − 3 / 2 ∑ i ≠ j ∣ E i j ∣ ≤ C n T − 3 / 2 ∥ E ∥ F = O ( T − 1 / 2 ) . \lVert R_{21}\varepsilon\rVert_2 \le CT^{-3/2}\sum_{i\ne j}|E_{ij}| \le CnT^{-3/2}\lVert E\rVert_F=O(T^{-1/2}). ∥ R 21 ε ∥ 2 ≤ C T − 3/2 i = j ∑ ∣ E ij ∣ ≤ C n T − 3/2 ∥ E ∥ F = O ( T − 1/2 ) .
Therefore ∥ T 2 ∥ 2 = O ( T − 1 / 2 ) \lVert T_2\rVert_2=O(T^{-1/2}) ∥ T 2 ∥ 2 = O ( T − 1/2 ) .
Term T 3 T_3 T 3 . Set A = R ∗ = I n + E = W 2 A=R_* = I_n+E=W^2 A = R ∗ = I n + E = W 2 and write T 3 = Q 3 + C 3 T_3=Q_3+C_3 T 3 = Q 3 + C 3 , where
Q 3 = ∑ i ≠ j { 2 N i j H i j + H i j 2 } = ∑ i ≠ j { ( W N W ) i j 2 − N i j 2 } . Q_3=\sum_{i\ne j}\{2N_{ij}H_{ij}+H_{ij}^2\}=\sum_{i\ne j}\{(WNW)_{ij}^2-N_{ij}^2\}. Q 3 = i = j ∑ { 2 N ij H ij + H ij 2 } = i = j ∑ {( W N W ) ij 2 − N ij 2 } .
is the unweighted sum and C 3 = ∑ i ≠ j ( 2 N i j H i j + H i j 2 ) { ( a i a j ) − 1 − 1 } C_3=\sum_{i\ne j}(2N_{ij}H_{ij}+H_{ij}^2)\{(a_i a_j)^{-1}-1\} C 3 = ∑ i = j ( 2 N ij H ij + H ij 2 ) {( a i a j ) − 1 − 1 } . We first show that
E Q 3 = τ 2 , T T ∥ E ∥ F 2 = O ( T − 1 ) , Var ( Q 3 ) = O ( T − 1 ) . \mathbb{E}Q_3=\frac{\tau_{2,T}}{T}\lVert E\rVert_F^2=O(T^{-1}),\qquad\operatorname{Var}(Q_3)=O(T^{-1}). E Q 3 = T τ 2 , T ∥ E ∥ F 2 = O ( T − 1 ) , Var ( Q 3 ) = O ( T − 1 ) .
To verify the mean, use Cov ( G i j , G i j ) = ( τ 2 , T / T ) ( A i i A j j + A i j 2 ) \operatorname{Cov}(G_{ij},G_{ij})=(\tau_{2,T}/T)(A_{ii}A_{jj}+A_{ij}^2) Cov ( G ij , G ij ) = ( τ 2 , T / T ) ( A ii A j j + A ij 2 ) , A i i = 1 A_{ii}=1 A ii = 1 and A i j = E i j A_{ij}=E_{ij} A ij = E ij for i ≠ j i\ne j i = j .
For the variance, first consider the full trace difference
V = Tr ( G 2 ) − Tr ( N 2 ) = Tr ( A N A N ) − Tr ( N 2 ) . V=\operatorname{Tr}(G^2)-\operatorname{Tr}(N^2)=\operatorname{Tr}(ANAN)-\operatorname{Tr}(N^2). V = Tr ( G 2 ) − Tr ( N 2 ) = Tr ( A N A N ) − Tr ( N 2 ) .
Its gradient is
∇ Z V = 4 T ( A N A − N ) Z Γ , A N A − N = E N + N E + E N E . \nabla_Z V=\frac{4}{T}(ANA-N)Z\Gamma,\qquad ANA-N=EN+NE+ENE. ∇ Z V = T 4 ( A N A − N ) Z Γ , A N A − N = E N + N E + E N E .
Gaussian Poincaré and Cauchy–Schwarz give
Var ( V ) ≤ E ∥ ∇ Z V ∥ F 2 ≤ 16 T 2 ( E ∥ A N A − N ∥ F 4 ) 1 / 2 ( E ∥ Z Γ ∥ o p 4 ) 1 / 2 . \operatorname{Var}(V)\le\mathbb{E}\lVert\nabla_ZV\rVert_F^2\le\frac{16}{T^2}\left(\mathbb{E}\lVert ANA-N\rVert_F^4\right)^{1/2}\left(\mathbb{E}\lVert Z\Gamma\rVert_{\mathrm{op}}^4\right)^{1/2}. Var ( V ) ≤ E ∥ ∇ Z V ∥ F 2 ≤ T 2 16 ( E ∥ A N A − N ∥ F 4 ) 1/2 ( E ∥ Z Γ ∥ op 4 ) 1/2 .
Put K E = sup n ∥ E ∥ F K_E = \sup_n \|E\|_F K E = sup n ∥ E ∥ F , which is finite in the local regime, and let b = 1 + K E b = 1+K_E b = 1 + K E and a = K E ( 2 + K E ) a = K_E(2+K_E) a = K E ( 2 + K E ) . Since ∥ W ∥ o p ≤ b \|W\|_{op} \le b ∥ W ∥ o p ≤ b and ∥ Θ ∥ F ≤ K E \|\Theta\|_F \le K_E ∥Θ ∥ F ≤ K E , the deterministic Schatten-norm inequalities give, pathwise,
∥ A N A − N ∥ F ≤ a ∥ N ∥ o p , ∥ H ∥ F ≤ a ∥ N ∥ o p , ∥ G ∥ o p ≤ b 2 ∥ N ∥ o p . \|ANA-N\|_F \le a\|N\|_{op}, \qquad\|H\|_F \le a\|N\|_{op}, \qquad\|G\|_{op} \le b^2\|N\|_{op}. ∥ A N A − N ∥ F ≤ a ∥ N ∥ o p , ∥ H ∥ F ≤ a ∥ N ∥ o p , ∥ G ∥ o p ≤ b 2 ∥ N ∥ o p .
Consequently, by (S.C.2),
E ∥ A N A − N ∥ F 4 ≤ a 4 m 4 4 , E ∥ Z Γ ∥ o p 4 ≤ K Γ 4 a 4 T 2 . (34) \mathbb{E}\|ANA-N\|_F^4 \le a^4m_4^4, \qquad\mathbb{E}\|Z\Gamma\|_{op}^4 \le K_\Gamma^4a^4T^2. \tag*{(34)} E ∥ A N A − N ∥ F 4 ≤ a 4 m 4 4 , E ∥ Z Γ ∥ o p 4 ≤ K Γ 4 a 4 T 2 . (34)
and Var ( V ) = O ( T − 1 ) \operatorname{Var}(V)=O(T^{-1}) Var ( V ) = O ( T − 1 ) ; the explicit constant is recorded in (S.C.6) below. The difference between V V V and Q 3 Q_3 Q 3 is the diagonal sum Δ 3 = ∑ i ( G i i 2 − N i i 2 ) = ∑ i ( 2 N i i H i i + H i i 2 ) \Delta_3=\sum_i(G_{ii}^2-N_{ii}^2)=\sum_i(2N_{ii}H_{ii}+H_{ii}^2) Δ 3 = ∑ i ( G ii 2 − N ii 2 ) = ∑ i ( 2 N ii H ii + H ii 2 ) . By the moment bounds stated at the outset,
∥ Δ 3 ∥ 2 ≤ C T ∑ i ( f i + f i 2 ) ≤ C T ( n ∥ Θ ∥ F + ∥ Θ ∥ F 2 ) = O ( T − 1 / 2 ) . \|\Delta_3\|_2 \le\frac{C}{T}\sum_i(f_i+f_i^2) \le\frac{C}{T}(\sqrt{n}\|\Theta\|_F+\|\Theta\|_F^2)=O(T^{-1/2}). ∥ Δ 3 ∥ 2 ≤ T C i ∑ ( f i + f i 2 ) ≤ T C ( n ∥Θ ∥ F + ∥Θ ∥ F 2 ) = O ( T − 1/2 ) .
The estimates for Q 3 = V − Δ 3 Q_3=V-\Delta_3 Q 3 = V − Δ 3 follow.
With D = diag ( G ) D=\operatorname{diag}(G) D = diag ( G ) as before, put P = G ∘ G − N ∘ N P=G\circ G-N\circ N P = G ∘ G − N ∘ N (Hadamard product), so that P i j = H i j ( G i j + N i j ) P_{ij}=H_{ij}(G_{ij}+N_{ij}) P ij = H ij ( G ij + N ij ) , and let P ˉ i = ∑ j ≠ i P i j \bar P_i=\sum_{j\ne i}P_{ij} P ˉ i = ∑ j = i P ij . On E \mathcal{E} E , the first-order Taylor formula for ( a i a j ) − 1 (a_i a_j)^{-1} ( a i a j ) − 1 with quadratic remainder and the symmetry of P P P give
C 3 = − 2 ∑ i G i i P ˉ i + R 3 , ∣ R 3 ∣ ≤ C ∑ i G i i 2 ∑ j ≠ i ∣ P i j ∣ . C_3=-2\sum_i G_{ii}\bar P_i+\mathcal{R}_3,\qquad|\mathcal{R}_3|\le C\sum_iG_{ii}^2\sum_{j\ne i}|P_{ij}|. C 3 = − 2 i ∑ G ii P ˉ i + R 3 , ∣ R 3 ∣ ≤ C i ∑ G ii 2 j = i ∑ ∣ P ij ∣.
The leading term has the exact trace representation
C 3 , 1 = − 2 Tr { D ( G 2 − N 2 ) } + 2 ∑ i G i i ( G i i 2 − N i i 2 ) . C_{3,1}=-2\operatorname{Tr}\{D(G^2-N^2)\}+2\sum_iG_{ii}(G_{ii}^2-N_{ii}^2). C 3 , 1 = − 2 Tr { D ( G 2 − N 2 )} + 2 i ∑ G ii ( G ii 2 − N ii 2 ) .
For the trace polynomial J = Tr { D ( G 2 − N 2 ) } J=\operatorname{Tr}\{D(G^2-N^2)\} J = Tr { D ( G 2 − N 2 )} , differentiation with respect to a symmetric perturbation d N dN d N yields
d J = Tr ( B J d N ) , B J = W diag ( diag ( G 2 − N 2 ) ) W + W ( G D + D G ) W − ( N D + D N ) . dJ = \operatorname{Tr}(B_J\,dN), \qquad B_J = W\operatorname{diag}\left(\operatorname{diag}(G^2-N^2)\right)W + W(GD+DG)W-(ND+DN). dJ = Tr ( B J d N ) , B J = W diag ( diag ( G 2 − N 2 ) ) W + W ( G D + D G ) W − ( N D + D N ) .
Consequently, since N = T − 1 Z Z ′ − I n N=T^{-1}ZZ^{\prime}-I_n N = T − 1 Z Z ′ − I n ,
∇ Z J = 2 T B J Z Γ , Var ( J ) ≤ 4 T 2 { E ∥ B J ∥ F 4 } 1 / 2 { E ∥ Z Γ ∥ o p 4 } 1 / 2 . \nabla_ZJ=\frac{2}{T}B_JZ\Gamma,\qquad\operatorname{Var}(J)\leq\frac{4}{T^2}\left\{\mathbb{E}\lVert B_J\rVert_F^4\right\}^{1/2}\left\{\mathbb{E}\lVert Z\Gamma\rVert_{\mathrm{op}}^4\right\}^{1/2}. ∇ Z J = T 2 B J Z Γ , Var ( J ) ≤ T 2 4 { E ∥ B J ∥ F 4 } 1/2 { E ∥ Z Γ ∥ op 4 } 1/2 .
To bound B J B_J B J , note that each G i i = T − 1 w i ′ Z Γ Z ′ w i − 1 G_{ii}=T^{-1}w_i^{\prime}Z\Gamma Z^{\prime}w_i-1 G ii = T − 1 w i ′ Z Γ Z ′ w i − 1 is a centered quadratic form with ∥ w i ∥ = 1 \lVert w_i\rVert=1 ∥ w i ∥ = 1 and variance 2 τ 2 , T / T ≤ 2 K Γ / T 2\tau_{2,T}/T\leq2K_\Gamma/T 2 τ 2 , T / T ≤ 2 K Γ / T ; degree-two Gaussian hypercontractivity gives ∥ G i i ∥ 8 ≤ 7 2 K Γ / T \lVert G_{ii}\rVert_8\leq7\sqrt{2K_\Gamma/T} ∥ G ii ∥ 8 ≤ 7 2 K Γ / T , and Minkowski’s inequality in L 4 L^4 L 4 applied to ∥ D ∥ F 2 \lVert D\rVert_F^2 ∥ D ∥ F 2 yields
∥ ∥ D ∥ F ∥ 8 ≤ 7 2 K Γ n / T ≤ d 8 , d 8 = 7 2 K Γ c + . \left\lVert\lVert D\rVert_F\right\rVert_8\leq7\sqrt{2K_\Gamma n/T}\leq d_8,\qquad d_8=7\sqrt{2K_\Gamma c_+}. ∥ ∥ D ∥ F ∥ 8 ≤ 7 2 K Γ n / T ≤ d 8 , d 8 = 7 2 K Γ c + .
The three terms of B J B_J B J satisfy the deterministic inequality
∥ B J ∥ F ≤ b 2 ( ∥ G ∥ o p + ∥ N ∥ o p ) ∥ H ∥ F + 2 ( b 2 ∥ G ∥ o p + ∥ N ∥ o p ) ∥ D ∥ F , \lVert B_J\rVert_F\leq b^2(\lVert G\rVert_{\mathrm{op}}+\lVert N\rVert_{\mathrm{op}})\lVert H\rVert_F+2(b^2\lVert G\rVert_{\mathrm{op}}+\lVert N\rVert_{\mathrm{op}})\lVert D\rVert_F, ∥ B J ∥ F ≤ b 2 (∥ G ∥ op + ∥ N ∥ op ) ∥ H ∥ F + 2 ( b 2 ∥ G ∥ op + ∥ N ∥ op ) ∥ D ∥ F ,
since ∥ diag ( diag M ) ∥ F ≤ ∥ M ∥ F \lVert\operatorname{diag}(\operatorname{diag}M)\rVert_F\leq\lVert M\rVert_F ∥ diag ( diag M ) ∥ F ≤ ∥ M ∥ F and G 2 − N 2 = G H + H N G^2-N^2=GH+HN G 2 − N 2 = G H + H N . Hölder’s inequality, with the eighth moments above, gives the dimension-free bound
( E ∥ B J ∥ F 4 ) 1 / 4 ≤ b 2 ( b 2 + 1 ) a m 8 2 + 2 ( b 4 + 1 ) m 8 d 8 = : C J < ∞ . (35) \left(\mathbb{E}\lVert B_J\rVert_F^4\right)^{1/4}\leq b^2(b^2+1)a\,m_8^2+2(b^4+1)m_8d_8=:C_J<\infty. \tag*{(35)} ( E ∥ B J ∥ F 4 ) 1/4 ≤ b 2 ( b 2 + 1 ) a m 8 2 + 2 ( b 4 + 1 ) m 8 d 8 =: C J < ∞. (35)
The two gradient bounds, together with (S.C.4) and (S.C.5) therefore give
Var ( V ) ≤ 16 a 2 m 4 2 K Γ 2 τ 4 2 / T , Var ( J ) ≤ 4 C J 2 K Γ 2 τ 4 2 / T . (36) \operatorname{Var}(V)\leq16a^2m_4^2K_\Gamma^2\tau_4^2/T,\qquad\operatorname{Var}(J)\leq4C_J^2K_\Gamma^2\tau_4^2/T. \tag*{(36)} Var ( V ) ≤ 16 a 2 m 4 2 K Γ 2 τ 4 2 / T , Var ( J ) ≤ 4 C J 2 K Γ 2 τ 4 2 / T . (36)
In particular Var ( J ) = O ( T − 1 ) \operatorname{Var}(J)=O(T^{-1}) Var ( J ) = O ( T − 1 ) . For the mean, recall τ 3 , T = T − 1 Tr ( Γ 3 ) \tau_{3,T}=T^{-1}\operatorname{Tr}(\Gamma^3) τ 3 , T = T − 1 Tr ( Γ 3 ) . For deterministic symmetric B 1 , B 2 , B 3 B_1,B_2,B_3 B 1 , B 2 , B 3 , diagonalize Γ \Gamma Γ and write N = T − 1 ∑ t = 1 T ν t , T ( z t z t ′ − I n ) N=T^{-1}\sum_{t=1}^{T}\nu_{t,T}(z_tz_t^{\prime}-I_n) N = T − 1 ∑ t = 1 T ν t , T ( z t z t ′ − I n ) , where the z t z_t z t are independent N ( 0 , I n ) N(0,I_n) N ( 0 , I n ) vectors. Independence and centering imply that only equal temporal indices survive in the third product. Wick’s formula (equivalently the third joint cumulant of Gaussian quadratic forms) therefore gives the exact identity
E ∏ r = 1 3 Tr ( B r N ) = 4 τ 3 , T T 2 Tr ( B 1 B 2 B 3 + B 1 B 3 B 2 ) . (37) \mathbb{E}\prod_{r=1}^{3}\operatorname{Tr}(B_rN)=\frac{4\tau_{3,T}}{T^2}\operatorname{Tr}(B_1B_2B_3+B_1B_3B_2). \tag*{(37)} E r = 1 ∏ 3 Tr ( B r N ) = T 2 4 τ 3 , T Tr ( B 1 B 2 B 3 + B 1 B 3 B 2 ) . (37)
For deterministic x , y x,y x , y and symmetric A A A , let B = x x ′ B=xx^{\prime} B = x x ′ and C k = ( y e k ′ + e k y ′ ) / 2 C_k=(ye_k^{\prime}+e_ky^{\prime})/2 C k = ( y e k ′ + e k y ′ ) /2 . Since y ′ N A N y = ∑ k , l A k l Tr ( C k N ) Tr ( C l N ) y^{\prime}NANy=\sum_{k,l}A_{kl}\operatorname{Tr}(C_kN)\operatorname{Tr}(C_lN) y ′ N A N y = ∑ k , l A k l Tr ( C k N ) Tr ( C l N ) , equation (S.C.7) yields
E { ( x ′ N x ) ( y ′ N A N y ) } = 8 τ 3 , T T 2 x ′ ( ∑ k , l A k l C k C l ) x = 2 τ 3 , T T 2 { Tr ( A ) ( x ′ y ) 2 + ∥ y ∥ 2 x ′ A x + 2 ( x ′ y ) ( x ′ A y ) } . (38) \begin{aligned}
\mathbb{E}\{(x^{\prime}Nx)(y^{\prime}NANy)\}
&=\frac{8\tau_{3,T}}{T^2}x^{\prime}\left(\sum_{k,l}A_{kl}C_kC_l\right)x\\
&=\frac{2\tau_{3,T}}{T^2}\{\operatorname{Tr}(A)(x^{\prime}y)^2+\lVert y\rVert^2x^{\prime}Ax+2(x^{\prime}y)(x^{\prime}Ay)\}. \tag*{(38)}
\end{aligned} E {( x ′ N x ) ( y ′ N A N y )} = T 2 8 τ 3 , T x ′ k , l ∑ A k l C k C l x = T 2 2 τ 3 , T { Tr ( A ) ( x ′ y ) 2 + ∥ y ∥ 2 x ′ A x + 2 ( x ′ y ) ( x ′ A y )} . (38)
Here the matrix product in the first line is exactly
∑ k , l A k l C k C l = 1 4 { y ( A y ) ′ + ( A y ) y ′ + Tr ( A ) y y ′ + ∥ y ∥ 2 A } . \sum_{k,l}A_{kl}C_kC_l=\frac{1}{4}\{y(Ay)^{\prime}+(Ay)y^{\prime}+\operatorname{Tr}(A)yy^{\prime}+\lVert y\rVert^2A\}. k , l ∑ A k l C k C l = 4 1 { y ( A y ) ′ + ( A y ) y ′ + Tr ( A ) y y ′ + ∥ y ∥ 2 A } .
Let w i = W e i w_i=We_i w i = W e i , A = W 2 = I n + E A=W^2=I_n+E A = W 2 = I n + E , and recall ∥ w i ∥ 2 = A i i = 1 \lVert w_i\rVert^2=A_{ii}=1 ∥ w i ∥ 2 = A ii = 1 . The exact trace identity is
J = ∑ i { ( w i ′ N w i ) ( w i ′ N A N w i ) − ( w i ′ N w i ) ( e i ′ N 2 e i ) } . J=\sum_i\{(w_i^{\prime}Nw_i)(w_i^{\prime}NANw_i)-(w_i^{\prime}Nw_i)(e_i^{\prime}N^2e_i)\}. J = i ∑ {( w i ′ N w i ) ( w i ′ N A N w i ) − ( w i ′ N w i ) ( e i ′ N 2 e i )} .
Apply (S.C.8) to the first summand with x = y = w i x=y=w_i x = y = w i and matrix A A A , and to the second with x = w i x=w_i x = w i , y = e i y=e_i y = e i , and matrix I n I_n I n . The respective expectations are
2 τ 3 , T T 2 { n + 3 ( A 2 ) i i } , 2 τ 3 , T T 2 { ( n + 2 ) W i i 2 + 1 } . \frac{2\tau_{3,T}}{T^2}\{n+3(A^2)_{ii}\},\qquad\frac{2\tau_{3,T}}{T^2}\{(n+2)W_{ii}^2+1\}. T 2 2 τ 3 , T { n + 3 ( A 2 ) ii } , T 2 2 τ 3 , T {( n + 2 ) W ii 2 + 1 } .
Thus
E J = 2 τ 3 , T T 2 { n 2 + 3 Tr ( A 2 ) − ( n + 2 ) ∑ i W i i 2 − n } . (39) \mathbb{E}J=\frac{2\tau_{3,T}}{T^2}\left\{n^2+3\operatorname{Tr}(A^2)-(n+2)\sum_iW_{ii}^2-n\right\}. \tag*{(39)} E J = T 2 2 τ 3 , T { n 2 + 3 Tr ( A 2 ) − ( n + 2 ) i ∑ W ii 2 − n } . (39)
Since Tr ( E ) = 0 \operatorname{Tr}(E)=0 Tr ( E ) = 0 , Tr ( A 2 ) = n + ∥ E ∥ F 2 \operatorname{Tr}(A^2)=n+\lVert E\rVert_F^2 Tr ( A 2 ) = n + ∥ E ∥ F 2 . Also W = I n + Θ W=I_n+\Theta W = I n + Θ and diag ( W 2 ) = 1 \operatorname{diag}(W^2)=1 diag ( W 2 ) = 1 imply 2 Θ i i + ( Θ 2 ) i i = 0 2\Theta_{ii}+(\Theta^2)_{ii}=0 2 Θ ii + ( Θ 2 ) ii = 0 , hence W i i 2 = 1 − ∑ j ≠ i Θ i j 2 W_{ii}^2=1-\sum_{j\ne i}\Theta_{ij}^2 W ii 2 = 1 − ∑ j = i Θ ij 2 . Substitution into (S.C.9) gives
E J = 2 τ 3 , T T 2 { 3 ∥ E ∥ F 2 + ( n + 2 ) ∑ i ≠ j Θ i j 2 } . (40) \mathbb{E}J=\frac{2\tau_{3,T}}{T^2}\left\{3\lVert E\rVert_F^2+(n+2)\sum_{i\ne j}\Theta_{ij}^2\right\}. \tag*{(40)} E J = T 2 2 τ 3 , T ⎩ ⎨ ⎧ 3 ∥ E ∥ F 2 + ( n + 2 ) i = j ∑ Θ ij 2 ⎭ ⎬ ⎫ . (40)
Because ∥ E ∥ F = O ( 1 ) \lVert E\rVert_F=O(1) ∥ E ∥ F = O ( 1 ) , ∥ Θ ∥ F ≤ ∥ E ∥ F = O ( 1 ) \lVert\Theta\rVert_F\le\lVert E\rVert_F=O(1) ∥ Θ ∥ F ≤ ∥ E ∥ F = O ( 1 ) , n ≍ T n\asymp T n ≍ T , and τ 3 , T ≤ ∥ Γ ∥ op 2 = O ( 1 ) \tau_{3,T}\le\lVert\Gamma\rVert_{\operatorname{op}}^2=O(1) τ 3 , T ≤ ∥ Γ ∥ op 2 = O ( 1 ) , we obtain the explicit bound
∣ E J ∣ ≤ C T − 2 { ∥ E ∥ F 2 + n ∥ Θ ∥ F 2 } = O ( T − 1 ) . |\mathbb{E}J|\le CT^{-2}\{\lVert E\rVert_F^2+n\lVert\Theta\rVert_F^2\}=O(T^{-1}). ∣ E J ∣ ≤ C T − 2 {∥ E ∥ F 2 + n ∥ Θ ∥ F 2 } = O ( T − 1 ) .
Combining this with the preceding Gaussian Poincaré estimate Var ( J ) = O ( T − 1 ) \operatorname{Var}(J)=O(T^{-1}) Var ( J ) = O ( T − 1 ) proves
∥ J ∥ 2 = O ( T − 1 / 2 ) . \lVert J\rVert_2=O(T^{-1/2}). ∥ J ∥ 2 = O ( T − 1/2 ) .
The diagonal correction ∑ i G i i ( G i i 2 − N i i 2 ) \sum_iG_{ii}(G_{ii}^2-N_{ii}^2) ∑ i G ii ( G ii 2 − N ii 2 ) has L 2 L^2 L 2 norm O ( T − 1 / 2 ) O(T^{-1/2}) O ( T − 1/2 ) by ∥ G i i ∥ 6 = O ( T − 1 / 2 ) \lVert G_{ii}\rVert_6=O(T^{-1/2}) ∥ G ii ∥ 6 = O ( T − 1/2 ) , ∥ N i i ∥ 6 = O ( T − 1 / 2 ) \lVert N_{ii}\rVert_6=O(T^{-1/2}) ∥ N ii ∥ 6 = O ( T − 1/2 ) , and Minkowski’s inequality. Therefore ∥ C 3 , 1 ∥ 2 = O ( T − 1 / 2 ) \lVert C_{3,1}\rVert_2=O(T^{-1/2}) ∥ C 3 , 1 ∥ 2 = O ( T − 1/2 ) .
For the Taylor remainder, write f = ∥ Θ ∥ F = O ( 1 ) f=\lVert\Theta\rVert_F=O(1) f = ∥ Θ ∥ F = O ( 1 ) and use the row estimates
∥ ∥ N i ⋅ ∥ 2 ∥ 8 + ∥ ∥ G i ⋅ ∥ 2 ∥ 8 ≤ C , ∥ ∥ H i ⋅ ∥ 2 ∥ 8 ≤ C ( f i + f T ) . (41) \lVert\lVert N_{i\cdot}\rVert_2\rVert_8+\lVert\lVert G_{i\cdot}\rVert_2\rVert_8\le C,\qquad\lVert\lVert H_{i\cdot}\rVert_2\rVert_8\le C\left(f_i+\frac{f}{\sqrt{T}}\right). \tag*{(41)} ∥∥ N i ⋅ ∥ 2 ∥ 8 + ∥∥ G i ⋅ ∥ 2 ∥ 8 ≤ C , ∥∥ H i ⋅ ∥ 2 ∥ 8 ≤ C ( f i + T f ) . (41)
To prove (S.C.11), use the exact covariance formula
E ( N a b N c d ) = τ 2 , T T ( δ a c δ b d + δ a d δ b c ) . \mathbb{E}(N_{ab}N_{cd})=\frac{\tau_{2,T}}{T}(\delta_{ac}\delta_{bd}+\delta_{ad}\delta_{bc}). E ( N ab N c d ) = T τ 2 , T ( δ a c δ b d + δ a d δ b c ) .
For a deterministic symmetric matrix Θ \Theta Θ this gives
E ∥ e i ′ N ∥ 2 2 = τ 2 , T T ( n + 1 ) , E ∥ e i ′ Θ N ∥ 2 2 = τ 2 , T T ( n + 1 ) f i 2 , E ∥ e i ′ N Θ ∥ 2 2 = τ 2 , T T ( f 2 + ∥ Θ e i ∥ 2 2 ) . \mathbb{E}\lVert e_i^{\prime}N\rVert_2^2=\frac{\tau_{2,T}}{T}(n+1),\qquad\mathbb{E}\lVert e_i^{\prime}\Theta N\rVert_2^2=\frac{\tau_{2,T}}{T}(n+1)f_i^2,\qquad\mathbb{E}\lVert e_i^{\prime}N\Theta\rVert_2^2=\frac{\tau_{2,T}}{T}(f^2+\lVert\Theta e_i\rVert_2^2). E ∥ e i ′ N ∥ 2 2 = T τ 2 , T ( n + 1 ) , E ∥ e i ′ Θ N ∥ 2 2 = T τ 2 , T ( n + 1 ) f i 2 , E ∥ e i ′ N Θ ∥ 2 2 = T τ 2 , T ( f 2 + ∥ Θ e i ∥ 2 2 ) .
Since Θ \Theta Θ is symmetric, ∥ Θ e i ∥ 2 = f i \lVert\Theta e_i\rVert_2=f_i ∥ Θ e i ∥ 2 = f i . Moreover, e i ′ Θ N Θ = ( e i ′ Θ N ) Θ e_i^{\prime}\Theta N\Theta=(e_i^{\prime}\Theta N)\Theta e i ′ Θ N Θ = ( e i ′ Θ N ) Θ , so its Euclidean norm is bounded by ∥ Θ ∥ op ∥ e i ′ Θ N ∥ 2 \lVert\Theta\rVert_{\operatorname{op}}\lVert e_i^{\prime}\Theta N\rVert_2 ∥ Θ ∥ op ∥ e i ′ Θ N ∥ 2 . Thus the L 2 L^2 L 2 row estimates follow from H = Θ N + N Θ + Θ N Θ H=\Theta N+N\Theta+\Theta N\Theta H = Θ N + N Θ + Θ N Θ and ∥ Θ ∥ op = O ( 1 ) \lVert\Theta\rVert_{\operatorname{op}}=O(1) ∥ Θ ∥ op = O ( 1 ) . Since the entries of N N N , H H H , and G G G are centered Gaussian quadratic forms, the dimension-free chaos bound in Section S.F, with p = 8 p=8 p = 8 , upgrades these L 2 L^2 L 2 row estimates to L 8 L^8 L 8 bounds with constants independent of n n n and T T T . This proves the N N N and H H H bounds; the G G G bound follows from G = N + H G=N+H G = N + H and sup i f i ≤ f = O ( 1 ) \sup_i f_i\le f=O(1) sup i f i ≤ f = O ( 1 ) . Also, by moment bounds stated at the outset,
sup i ∥ G i i ∥ 8 ≤ C T − 1 / 2 . (42) \sup_i \lVert G_{ii}\rVert_8 \leq C T^{-1/2}. \tag*{(42)} i sup ∥ G ii ∥ 8 ≤ C T − 1/2 . (42)
By Cauchy–Schwarz in the row index,
∑ j ≠ i ∣ P i j ∣ ≤ ∥ H i ⋅ ∥ 2 ( ∥ G i ⋅ ∥ 2 + ∥ N i ⋅ ∥ 2 ) . \sum_{j\ne i} |P_{ij}| \leq\lVert H_{i\cdot}\rVert_2(\lVert G_{i\cdot}\rVert_2+\lVert N_{i\cdot}\rVert_2). j = i ∑ ∣ P ij ∣ ≤ ∥ H i ⋅ ∥ 2 (∥ G i ⋅ ∥ 2 + ∥ N i ⋅ ∥ 2 ) .
Hölder’s inequality with exponents 4, 8, 8, followed by (S.C.11)–(S.C.12), yields
∥ R 31 E ∥ 2 ≤ C ∑ i ∥ G i i ∥ 8 2 ∥ H i ⋅ ∥ 8 ( ∥ ∥ G i ⋅ ∥ 2 ∥ 8 + ∥ ∥ N i ⋅ ∥ 2 ∥ 8 ) ≤ C T ∑ i ( f i + f T ) ≤ C T ( n f + n f T ) = O ( T − 1 / 2 ) . (43) \begin{aligned}
\lVert\mathcal{R}_{31E}\rVert_2
&\leq C\sum_i \lVert G_{ii}\rVert_8^2\lVert H_{i\cdot}\rVert_8(\lVert\lVert G_{i\cdot}\rVert_2\rVert_8+\lVert\lVert N_{i\cdot}\rVert_2\rVert_8)\\
&\leq\frac{C}{T}\sum_i\left(f_i+\frac{f}{\sqrt{T}}\right)
\leq\frac{C}{T}\left(\sqrt{n}f+\frac{nf}{\sqrt{T}}\right)
=O(T^{-1/2}). \tag*{(43)}
\end{aligned} ∥ R 31 E ∥ 2 ≤ C i ∑ ∥ G ii ∥ 8 2 ∥ H i ⋅ ∥ 8 (∥∥ G i ⋅ ∥ 2 ∥ 8 + ∥∥ N i ⋅ ∥ 2 ∥ 8 ) ≤ T C i ∑ ( f i + T f ) ≤ T C ( n f + T n f ) = O ( T − 1/2 ) . (43)
Combining (S.C.13) with ∥ C 3 , 1 ∥ 2 = O ( T − 1 / 2 ) \lVert C_{3,1}\rVert_2=O(T^{-1/2}) ∥ C 3 , 1 ∥ 2 = O ( T − 1/2 ) gives
∥ C 3 ∥ 2 = O ( T − 1 / 2 ) , \lVert C_3\rVert_2=O(T^{-1/2}), ∥ C 3 ∥ 2 = O ( T − 1/2 ) ,
and hence ∥ T 3 ∥ 2 ≤ ∥ Q 3 ∥ 2 + ∥ C 3 ∥ 2 = O ( T − 1 / 2 ) \lVert T_3\rVert_2\leq\lVert Q_3\rVert_2+\lVert C_3\rVert_2=O(T^{-1/2}) ∥ T 3 ∥ 2 ≤ ∥ Q 3 ∥ 2 + ∥ C 3 ∥ 2 = O ( T − 1/2 ) .
Term T 4 T_4 T 4 . On E \mathcal{E} E , Taylor’s formula gives
1 a i a j − 1 b i b j = − ( H i i + H j j ) + ω i j ′ , ∣ ω i j ′ ∣ ≤ C ( ∣ H i i ∣ + ∣ H j j ∣ ) ( ∣ G i i ∣ + ∣ G j j ∣ + ∣ N i i ∣ + ∣ N j j ∣ ) . \frac{1}{a_i a_j}-\frac{1}{b_i b_j}=-(H_{ii}+H_{jj})+\omega'_{ij},\qquad|\omega'_{ij}|\leq C(|H_{ii}|+|H_{jj}|)(|G_{ii}|+|G_{jj}|+|N_{ii}|+|N_{jj}|). a i a j 1 − b i b j 1 = − ( H ii + H j j ) + ω ij ′ , ∣ ω ij ′ ∣ ≤ C ( ∣ H ii ∣ + ∣ H j j ∣ ) ( ∣ G ii ∣ + ∣ G j j ∣ + ∣ N ii ∣ + ∣ N j j ∣ ) .
Put m i = ∑ j ≠ i N i j 2 m_i=\sum_{j\ne i}N_{ij}^2 m i = ∑ j = i N ij 2 and m ˉ = ( n − 1 ) τ 2 , T / T \bar{m}=(n-1)\tau_{2,T}/T m ˉ = ( n − 1 ) τ 2 , T / T . The first-order part is − 2 ∑ i H i i m i -2\sum_i H_{ii}m_i − 2 ∑ i H ii m i . Conditional on row Z i = z Z_i=z Z i = z , the variables N i j = T − 1 z ′ Γ Z j N_{ij}=T^{-1}z'\Gamma Z_j N ij = T − 1 z ′ Γ Z j for j ≠ i j\ne i j = i are independent centered normals with common variance ς i = T − 2 z ′ Γ 2 z \varsigma_i=T^{-2}z'\Gamma^2z ς i = T − 2 z ′ Γ 2 z . Hence
m i ∣ Z i = d ς i χ n − 1 2 , E ς i = τ 2 , T T , ∥ ς i − E ς i ∥ 4 = O ( T − 3 / 2 ) . m_i\mid Z_i\overset{d}{=}\varsigma_i\chi^2_{n-1},\qquad\mathbb{E}\varsigma_i=\frac{\tau_{2,T}}{T},\qquad\lVert\varsigma_i-\mathbb{E}\varsigma_i\rVert_4=O(T^{-3/2}). m i ∣ Z i = d ς i χ n − 1 2 , E ς i = T τ 2 , T , ∥ ς i − E ς i ∥ 4 = O ( T − 3/2 ) .
The last bound follows from the fourth moment of a centered Gaussian quadratic form and T r ( Γ 2 k ) = O ( T ) Tr(\Gamma^{2k})=O(T) T r ( Γ 2 k ) = O ( T ) for fixed k k k . Using conditional fourth moments of the chi-square variable and Hölder,
∥ m i − m ˉ ∥ 4 ≤ C ( n T + n T 3 / 2 ) = O ( T − 1 / 2 ) , ∥ m i ∥ 4 = O ( 1 ) . \lVert m_i-\bar{m}\rVert_4 \le C\left(\frac{\sqrt{n}}{T}+\frac{n}{T^{3/2}}\right)=O(T^{-1/2}),\qquad\lVert m_i\rVert_4=O(1). ∥ m i − m ˉ ∥ 4 ≤ C ( T n + T 3/2 n ) = O ( T − 1/2 ) , ∥ m i ∥ 4 = O ( 1 ) .
Because ∑ i H i i = Tr ( ( 2 Θ + Θ 2 ) N ) = Tr ( E N ) \sum_i H_{ii}=\operatorname{Tr}((2\Theta+\Theta^2)N)=\operatorname{Tr}(EN) ∑ i H ii = Tr (( 2Θ + Θ 2 ) N ) = Tr ( E N ) ,
∥ m ˉ ∑ i H i i ∥ 2 ≤ C τ 2 , T T ∥ E ∥ F = O ( T − 1 / 2 ) . \left\lVert\bar{m}\sum_i H_{ii}\right\rVert_2\le C\sqrt{\frac{\tau_{2,T}}{T}}\lVert E\rVert_F=O(T^{-1/2}). m ˉ i ∑ H ii 2 ≤ C T τ 2 , T ∥ E ∥ F = O ( T − 1/2 ) .
For the centered row sums, Minkowski and Hölder yield
∥ ∑ i H i i ( m i − m ˉ ) ∥ 2 ≤ C T − 1 ∑ i f i ≤ C n T ∥ Θ ∥ F = O ( T − 1 / 2 ) . \left\lVert\sum_i H_{ii}(m_i-\bar{m})\right\rVert_2\le CT^{-1}\sum_i f_i\le C\frac{\sqrt{n}}{T}\lVert\Theta\rVert_F=O(T^{-1/2}). i ∑ H ii ( m i − m ˉ ) 2 ≤ C T − 1 i ∑ f i ≤ C T n ∥ Θ ∥ F = O ( T − 1/2 ) .
For the Taylor remainder, set u i = ∣ G i i ∣ + ∣ N i i ∣ u_i=|G_{ii}|+|N_{ii}| u i = ∣ G ii ∣ + ∣ N ii ∣ and h i = ∣ H i i ∣ h_i=|H_{ii}| h i = ∣ H ii ∣ . On E \mathcal{E} E , the two-variable Taylor formula, applied to g ( x , y ) = [ ( 1 + x ) ( 1 + y ) ] − 1 g(x,y)=[(1+x)(1+y)]^{-1} g ( x , y ) = [( 1 + x ) ( 1 + y ) ] − 1 , gives
∣ ω i j ′ ∣ ≤ C ( h i + h j ) ( u i + u j ) . |\omega'_{ij}|\le C(h_i+h_j)(u_i+u_j). ∣ ω ij ′ ∣ ≤ C ( h i + h j ) ( u i + u j ) .
This bound holds because the gradient of g g g is Lipschitz on [ − 1 / 2 , 1 ] 2 [-1/2,1]^2 [ − 1/2 , 1 ] 2 , and G i i − N i i = H i i G_{ii}-N_{ii}=H_{ii} G ii − N ii = H ii ; this is a difference of Taylor expansions at ( G i i , G j j ) (G_{ii},G_{jj}) ( G ii , G j j ) and ( N i i , N j j ) (N_{ii},N_{jj}) ( N ii , N j j ) . By the moment bounds stated at the outset, ∥ u i ∥ p ≤ C p T − 1 / 2 \lVert u_i\rVert_p\le C_pT^{-1/2} ∥ u i ∥ p ≤ C p T − 1/2 and ∥ h i ∥ p ≤ C p T − 1 / 2 f i \lVert h_i\rVert_p\le C_pT^{-1/2}f_i ∥ h i ∥ p ≤ C p T − 1/2 f i for each fixed p p p , so Hölder with exponents ( 4 , 8 , 8 ) (4,8,8) ( 4 , 8 , 8 ) gives the per-pair bound
∥ N i j 2 h i u j ∥ 2 ≤ ∥ N i j 2 ∥ 4 ∥ h i ∥ 8 ∥ u j ∥ 8 ≤ C T − 2 f i . \lVert N_{ij}^2h_iu_j\rVert_2\le\lVert N_{ij}^2\rVert_4\lVert h_i\rVert_8\lVert u_j\rVert_8\le CT^{-2}f_i. ∥ N ij 2 h i u j ∥ 2 ≤ ∥ N ij 2 ∥ 4 ∥ h i ∥ 8 ∥ u j ∥ 8 ≤ C T − 2 f i .
The same estimate holds for N i j 2 h i u i N_{ij}^2h_iu_i N ij 2 h i u i , and, upon exchanging i , j i,j i , j , for N i j 2 h j u i N_{ij}^2h_ju_i N ij 2 h j u i and N i j 2 h j u j N_{ij}^2h_ju_j N ij 2 h j u j .
By Minkowski’s inequality, and since n / T n/T n / T is bounded,
∥ 1 E ∑ i ≠ j N i j 2 ω i j ′ ∥ 2 ≤ C T − 2 ∑ i ≠ j ( f i + f j ) ≤ C n T 2 ∑ i = 1 n f i ≤ C n 3 / 2 T 2 ∥ Θ ∥ F = O ( T − 1 / 2 ) . \begin{aligned}
\left\|1_{\mathcal{E}}\sum_{i\ne j}N_{ij}^{2}\omega'_{ij}\right\|_{2}
\le CT^{-2}\sum_{i\ne j}(f_i+f_j) \\
&\le C\frac{n}{T^{2}}\sum_{i=1}^{n}f_i
\le C\frac{n^{3/2}}{T^{2}}\|\Theta\|_{F}=O(T^{-1/2}).
\end{aligned} 1 E i = j ∑ N ij 2 ω ij ′ 2 ≤ C T − 2 i = j ∑ ( f i + f j ) ≤ C T 2 n i = 1 ∑ n f i ≤ C T 2 n 3/2 ∥Θ ∥ F = O ( T − 1/2 ) .
Consequently ∥ T 4 ∥ 2 = O ( T − 1 / 2 ) \|T_4\|_2=O(T^{-1/2}) ∥ T 4 ∥ 2 = O ( T − 1/2 ) .
The four contributions have the following L 2 L^2 L 2 bounds, all on E \mathcal{E} E :
∥ T 1 ∥ 2 ≤ C T − 1 / 2 ∥ E ∥ F 2 , ∥ T 2 ∥ 2 ≤ C { T − 1 / 2 + n T − 3 / 2 ∥ E ∥ F } , ∥ T 3 ∥ 2 ≤ C { T − 1 / 2 + T − 1 ∑ i f i + T − 1 n f / T } , ∥ T 4 ∥ 2 ≤ C { T − 1 / 2 + T − 1 ∑ i f i + n T − 2 ∑ i f i } . \begin{aligned}
\|T_1\|_2 &\le CT^{-1/2}\|E\|_F^2,\\
\|T_2\|_2 &\le C\{T^{-1/2}+nT^{-3/2}\|E\|_F\},\\
\|T_3\|_2 &\le C\{T^{-1/2}+T^{-1}\sum_i f_i+T^{-1}nf/\sqrt{T}\},\\
\|T_4\|_2 &\le C\{T^{-1/2}+T^{-1}\sum_i f_i+nT^{-2}\sum_i f_i\}.
\end{aligned} ∥ T 1 ∥ 2 ∥ T 2 ∥ 2 ∥ T 3 ∥ 2 ∥ T 4 ∥ 2 ≤ C T − 1/2 ∥ E ∥ F 2 , ≤ C { T − 1/2 + n T − 3/2 ∥ E ∥ F } , ≤ C { T − 1/2 + T − 1 i ∑ f i + T − 1 n f / T } , ≤ C { T − 1/2 + T − 1 i ∑ f i + n T − 2 i ∑ f i } .
Here f i = ∥ Θ i ⋅ ∥ 2 f_i=\|\Theta_{i\cdot}\|_2 f i = ∥ Θ i ⋅ ∥ 2 , f = ∥ Θ ∥ F f=\|\Theta\|_F f = ∥Θ ∥ F , and ∑ i f i ≤ n f \sum_i f_i\le\sqrt{n}f ∑ i f i ≤ n f . In the local regime ∥ E ∥ F = O ( 1 ) \|E\|_F=O(1) ∥ E ∥ F = O ( 1 ) and n / T n/T n / T bounded above and away from zero, every displayed term is O ( T − 1 / 2 ) O(T^{-1/2}) O ( T − 1/2 ) , so ∥ D n 1 E ∥ 2 ≤ ∑ j = 1 4 ∥ T j 1 E ∥ 2 = O ( T − 1 / 2 ) \|D_n1_{\mathcal{E}}\|_2\le\sum_{j=1}^{4}\|T_j1_{\mathcal{E}}\|_2=O(T^{-1/2}) ∥ D n 1 E ∥ 2 ≤ ∑ j = 1 4 ∥ T j 1 E ∥ 2 = O ( T − 1/2 ) . Together with the exceptional-event bound, E ( D n 2 ) = O ( T − 1 ) = O ( n − 1 ) \mathbb{E}(D_n^2)=O(T^{-1})=O(n^{-1}) E ( D n 2 ) = O ( T − 1 ) = O ( n − 1 ) . All constants depend only on the bounds for n / T n/T n / T , K Γ K_\Gamma K Γ , K E K_E K E , and fixed moment orders. ■
Remark S.1 The unit-diagonal restriction on W 2 = R ∗ W^2=R_* W 2 = R ∗ , namely 2 Θ i i + ( Θ 2 ) i i = 0 2\Theta_{ii}+(\Theta^2)_{ii}=0 2 Θ ii + ( Θ 2 ) ii = 0 for each i i i (whose trace is the identity in (S.C.1)), is used in the exact evaluation of E J \mathbb{E}J E J in (S.C.10); no assumption ∥ R ∗ − I ∥ o p = o ( 1 ) \|R_*-I\|_{\mathrm{op}}=o(1) ∥ R ∗ − I ∥ op = o ( 1 ) is needed. Section 5 of the main paper reports a Monte Carlo check of the size of D n D_n D n .
Auxiliary lemma (coefficientwise Gaussian cumulant bound). Let V V V be a fixed index set with ∣ V ∣ ≤ 4 |V|\le4 ∣ V ∣ ≤ 4 , let R V R_V R V be a correlation matrix, and let g ∼ N ( 0 , R V ⊗ I T ) g\sim N(0,R_V\otimes I_T) g ∼ N ( 0 , R V ⊗ I T ) . For j = 1 , … , r j=1,\ldots,r j = 1 , … , r , let
Q j = T − 1 { g ′ ( C j ⊗ Γ ) g − Tr [ ( C j R V ) ⊗ Γ ] } , Q_j=T^{-1}\{g'(C_j\otimes\Gamma)g-\operatorname{Tr}[(C_jR_V)\otimes\Gamma]\}, Q j = T − 1 { g ′ ( C j ⊗ Γ ) g − Tr [( C j R V ) ⊗ Γ ]} , :::`
where the C j C_j C j are fixed symmetric matrices of order ∣ V ∣ |V| ∣ V ∣ with uniformly bounded entries.
Assume Tr Γ = T \operatorname{Tr}\Gamma= T Tr Γ = T and ∥ Γ ∥ o p ≤ K \lVert\Gamma\rVert_{\mathrm{op}} \le K ∥ Γ ∥ op ≤ K . For every fixed r ≥ 2 r \ge2 r ≥ 2 , the joint cumulant κ r ( Q 1 , … , Q r ) \kappa_r(Q_1,\ldots,Q_r) κ r ( Q 1 , … , Q r ) is a polynomial in the off-diagonal entries of R V R_V R V of degree at most r r r , and every coefficient of this polynomial is O ( T 1 − r ) O(T^{1-r}) O ( T 1 − r ) , uniformly in R V R_V R V , T T T , and the choice of indices in V V V . More generally, each coefficient of a joint moment of d d d such centered forms is O ( T − ⌈ d / 2 ⌉ ) O(T^{-\lceil d/2\rceil}) O ( T − ⌈ d /2 ⌉ ) ; the same bound holds for a covariance of two nonempty products, with only partitions connecting the products retained.
Proof. For a real Gaussian vector, polarization of the quadratic-form cumulant identity yields
κ r ( Q 1 , … , Q r ) = 2 r − 1 r T 1 − r ∑ π ∈ S r Tr ( ( C π ( 1 ) R V ) ⋯ ( C π ( r ) R V ) ) Tr ( Γ r ) . \kappa_r(Q_1,\ldots,Q_r)=\frac{2^{r-1}}{r}T^{1-r}\sum_{\pi\in\mathfrak{S}_r}\operatorname{Tr}\left((C_{\pi(1)}R_V)\cdots(C_{\pi(r)}R_V)\right)\operatorname{Tr}(\Gamma^r). κ r ( Q 1 , … , Q r ) = r 2 r − 1 T 1 − r π ∈ S r ∑ Tr ( ( C π ( 1 ) R V ) ⋯ ( C π ( r ) R V ) ) Tr ( Γ r ) .
(The factor 1 / r 1/r 1/ r accounts for the r r r cyclic rotations; in particular, the formula gives 2 T − 2 Tr ( C 1 R V C 2 R V ) Tr ( Γ 2 ) 2T^{-2}\operatorname{Tr}(C_1R_VC_2R_V)\operatorname{Tr}(\Gamma^2) 2 T − 2 Tr ( C 1 R V C 2 R V ) Tr ( Γ 2 ) when r = 2 r=2 r = 2 .) Each trace in dimension ∣ V ∣ ≤ 4 |V|\le4 ∣ V ∣ ≤ 4 is a sum of at most a constant (depending on r r r ) number of products of r r r entries of R V R_V R V and entries of the C j C_j C j . On substituting R V = I V + E V R_V=I_V+E_V R V = I V + E V , this is a polynomial of degree at most r r r in the off-diagonal correlations, with each coefficient bounded independently of T T T and V V V . Since ∣ Tr ( Γ r ) ∣ ≤ T K r − 1 |\operatorname{Tr}(\Gamma^r)|\le TK^{r-1} ∣ Tr ( Γ r ) ∣ ≤ T K r − 1 , every coefficient has absolute value at most C r T 1 − r C_rT^{1-r} C r T 1 − r .
For the moment assertion, expand by the moment–cumulant formula. Singleton blocks vanish because the Q j Q_j Q j are centered. A partition of d d d factors into b b b blocks, each of size at least two, has coefficientwise order ∏ B T 1 − ∣ B ∣ = T b − d \prod_B T^{1-|B|}=T^{b-d} ∏ B T 1 − ∣ B ∣ = T b − d , and b ≤ ⌊ d / 2 ⌋ b\le\lfloor d/2\rfloor b ≤ ⌊ d /2 ⌋ . The number of partitions and the polynomial degrees depend only on d d d , so products and sums retain the stated coefficientwise order. For a covariance of two products, partitions with no block meeting both products cancel against the product of expectations. This leaves the same bound for the remaining connecting partitions. □
Auxiliary lemma (Eulerian correlation-monomial summation). Let E = E ′ E=E' E = E ′ have zero diagonal and ∥ E ∥ F ≤ M \lVert E\rVert_F\le M ∥ E ∥ F ≤ M . Let μ \mu μ be a monomial in off-diagonal entries E a b E_{ab} E ab of fixed positive degree m m m , whose associated multigraph has even degree at every incident vertex. If ℓ \ell ℓ additional labels do not occur in μ \mu μ , then the sum of ∣ μ ∣ |\mu| ∣ μ ∣ over all assignments of its incident vertices and those ℓ \ell ℓ labels is bounded by
n ℓ ∏ j = 1 q ∥ E ∥ F m j = n ℓ ∥ E ∥ F m , m 1 + ⋯ + m q = m , n^\ell\prod_{j=1}^{q} \|E\|_{F}^{m_j}=n^\ell\|E\|_{F}^{m},\qquad m_1+\cdots+m_q=m, n ℓ j = 1 ∏ q ∥ E ∥ F m j = n ℓ ∥ E ∥ F m , m 1 + ⋯ + m q = m ,
where m j m_j m j are the edge counts of its connected components. In particular, a coefficient bounded by C T − ℓ − 1 CT^{-\ell-1} C T − ℓ − 1 contributes at most C M m n ℓ T − ℓ − 1 = O ( n − 1 ) CM^{m}n^\ell T^{-\ell-1}=O(n^{-1}) C M m n ℓ T − ℓ − 1 = O ( n − 1 ) when n / T n/T n / T is bounded above and below.
Proof. Every connected component with nonzero degree has an Eulerian closed walk, since all degrees are even. Summing the absolute value of the product of edge weights over its vertex assignments is no larger than the sum over all closed walks of length m j m_j m j in the nonnegative symmetric matrix ∣ E ∣ = ( ∣ E a b ∣ ) |E|=(|E_{ab}|) ∣ E ∣ = ( ∣ E ab ∣ ) , namely Tr ( ∣ E ∣ m j ) \operatorname{Tr}(|E|^{m_j}) Tr ( ∣ E ∣ m j ) . As ∣ E ∣ |E| ∣ E ∣ is symmetric and m j ≥ 2 m_j\ge2 m j ≥ 2 ,
0 ≤ Tr ( ∣ E ∣ m j ) ≤ ∑ k ∣ λ k ( ∣ E ∣ ) ∣ m j ≤ ∥ E ∥ F m j . 0\le\operatorname{Tr}(|E|^{m_j})\le\sum_k|\lambda_k(|E|)|^{m_j}\le\|E\|_{F}^{m_j}. 0 ≤ Tr ( ∣ E ∣ m j ) ≤ k ∑ ∣ λ k ( ∣ E ∣ ) ∣ m j ≤ ∥ E ∥ F m j .
The label assignments of distinct components can be summed independently to obtain an upper bound, even if the original configuration requires distinct labels. Each absent vertex adds at most n n n choices. Multiplication gives the asserted inequality. The zero diagonal makes assignments collapsing an edge vanish. □
Lemma S.3 (Local covariance expansion). Under Assumptions 1–4, suppose n ( ϕ 2 − 1 ) = O ( 1 ) n(\phi_2-1)=O(1) n ( ϕ 2 − 1 ) = O ( 1 ) . Then Var A ( Tr ( R 2 ) ) − Var 0 ( Tr ( R 2 ) ) = O ( n − 1 ) \operatorname{Var}_A(\operatorname{Tr}(R^2))-\operatorname{Var}_0(\operatorname{Tr}(R^2))=O(n^{-1}) Var A ( Tr ( R 2 )) − Var 0 ( Tr ( R 2 )) = O ( n − 1 ) , where Var 0 \operatorname{Var}_0 Var 0 denotes the variance under H 0 H_0 H 0 .
Proof. With w a = W e a w_a=W e_a w a = W e a , write v a = Z ′ w a v_a=Z'w_a v a = Z ′ w a , so that v a ∼ N ( 0 , I T ) v_a\sim N(0,I_T) v a ∼ N ( 0 , I T ) , Cov ( v a , v b ) = ( R ∗ ) a b I T \operatorname{Cov}(v_a,v_b)=(R_*)_{ab}I_T Cov ( v a , v b ) = ( R ∗ ) ab I T , and r a b = v a ′ Γ v b / ( v a ′ Γ v a v b ′ Γ v b ) 1 / 2 r_{ab}=v_a'\Gamma v_b/(v_a'\Gamma v_a\,v_b'\Gamma v_b)^{1/2} r ab = v a ′ Γ v b / ( v a ′ Γ v a v b ′ Γ v b ) 1/2 . Put ρ a b = E a b \rho_{ab}=E_{ab} ρ ab = E ab for a ≠ b a\ne b a = b , so that ∑ a ≠ b ρ a b 2 = ∥ E ∥ F 2 = n ( ϕ 2 − 1 ) = O ( 1 ) \sum_{a\ne b}\rho_{ab}^2=\|E\|_F^2=n(\phi_2-1)=O(1) ∑ a = b ρ ab 2 = ∥ E ∥ F 2 = n ( ϕ 2 − 1 ) = O ( 1 ) . All constants below depend only on fixed moment orders, the uniform bound on ∥ Γ ∥ o p \|\Gamma\|_{\mathrm{op}} ∥Γ ∥ op , and upper and lower bounds for T / n T/n T / n ; they do not depend on the individual correlations or on n , T n,T n , T .
Step 1: Locality. Let I \mathcal{I} I be the set of ordered pairs e = ( a , b ) e=(a,b) e = ( a , b ) with a ≠ b a\ne b a = b . Then Tr ( R 2 ) − n = \operatorname{Tr}(R^2)-n= Tr ( R 2 ) − n =
∑ e ∈ I r e 2 and \sum_{e\in I} r_e^2 \text{ and} e ∈ I ∑ r e 2 and
Var A ( Tr ( R 2 ) ) = ∑ e , e ′ ∈ I c e e ′ ( R ⋆ ) , c e e ′ ( R ) = Cov R ( r e 2 , r e ′ 2 ) . \operatorname{Var}_{A}\left(\operatorname{Tr}(R^2)\right)=\sum_{e,e'\in I} c_{ee'}(R_\star),\qquad c_{ee'}(R)=\operatorname{Cov}_{R}\left(r_e^2,r_{e'}^2\right). Var A ( Tr ( R 2 ) ) = e , e ′ ∈ I ∑ c e e ′ ( R ⋆ ) , c e e ′ ( R ) = Cov R ( r e 2 , r e ′ 2 ) .
Since r e 2 r_e^2 r e 2 depends only on the two vectors v a v_a v a , a ∈ e a\in e a ∈ e , c e e ′ ( R ) c_{ee'}(R) c e e ′ ( R ) depends on R R R only through the principal submatrix R V R_V R V , V = e ∪ e ′ V=e\cup e' V = e ∪ e ′ , and ∣ V ∣ ∈ { 2 , 3 , 4 } |V|\in\{2,3,4\} ∣ V ∣ ∈ { 2 , 3 , 4 } . The same identity with R = I n R=I_n R = I n gives Var 0 ( Tr ( R 2 ) ) \operatorname{Var}_0(\operatorname{Tr}(R^2)) Var 0 ( Tr ( R 2 )) . Ordered pairs are retained throughout. Step 2: Polynomial truncation. For e = ( a , b ) e=(a,b) e = ( a , b ) put s e = T − 1 v a ′ Γ v b s_e=T^{-1}v_a'\Gamma v_b s e = T − 1 v a ′ Γ v b , δ a = T − 1 v a ′ Γ v a − 1 \delta_a=T^{-1}v_a'\Gamma v_a-1 δ a = T − 1 v a ′ Γ v a − 1 , define the degree-five Taylor polynomial by
Π 5 ( x , y ) = ∑ k = 0 5 ( − 1 ) k ∑ j = 0 k x j y k − j , 1 ( 1 + x ) ( 1 + y ) = Π 5 ( x , y ) + Q 6 ( x , y ) . \Pi_5(x,y)=\sum_{k=0}^{5}(-1)^k\sum_{j=0}^{k}x^j y^{k-j},\qquad\frac{1}{(1+x)(1+y)}=\Pi_5(x,y)+Q_6(x,y). Π 5 ( x , y ) = k = 0 ∑ 5 ( − 1 ) k j = 0 ∑ k x j y k − j , ( 1 + x ) ( 1 + y ) 1 = Π 5 ( x , y ) + Q 6 ( x , y ) .
where, for ∣ x ∣ ∨ ∣ y ∣ ≤ 1 / 2 |x|\vee|y|\le1/2 ∣ x ∣ ∨ ∣ y ∣ ≤ 1/2 ,
∣ Q 6 ( x , y ) ∣ ≤ C ( ∣ x ∣ + ∣ y ∣ ) 6 . |Q_6(x,y)|\le C(|x|+|y|)^6. ∣ Q 6 ( x , y ) ∣ ≤ C ( ∣ x ∣ + ∣ y ∣ ) 6 .
Put Π e = Π 5 ( δ a , δ b ) \Pi_e=\Pi_5(\delta_a,\delta_b) Π e = Π 5 ( δ a , δ b ) and set P e = s e 2 Π e P_e=s_e^2\Pi_e P e = s e 2 Π e and R e = r e 2 − P e R_e=r_e^2-P_e R e = r e 2 − P e . On the event ∣ δ a ∣ ∨ ∣ δ b ∣ ≤ 1 / 2 |\delta_a|\vee|\delta_b|\le1/2 ∣ δ a ∣ ∨ ∣ δ b ∣ ≤ 1/2 the multivariate Taylor formula gives ∣ R e ∣ ≤ C ∣ s e ∣ 2 ( ∣ δ a ∣ + ∣ δ b ∣ ) 6 |R_e|\le C|s_e|^2(|\delta_a|+|\delta_b|)^6 ∣ R e ∣ ≤ C ∣ s e ∣ 2 ( ∣ δ a ∣ + ∣ δ b ∣ ) 6 . For every fixed q q q , Gaussian quadratic-form moment bounds imply ∥ δ a ∥ q ≤ C q T − 1 / 2 \|\delta_a\|_q\le C_qT^{-1/2} ∥ δ a ∥ q ≤ C q T − 1/2 and ∥ s e ∥ q ≤ C q ( ∣ ρ e ∣ + T − 1 / 2 ) \|s_e\|_q\le C_q(|\rho_e|+T^{-1/2}) ∥ s e ∥ q ≤ C q ( ∣ ρ e ∣ + T − 1/2 ) . Hölder’s inequality therefore yields the first L 2 L^2 L 2 bound displayed below. The complement has probability at most C e − c T Ce^{-cT} C e − c T uniformly in the alternative, since each v a v_a v a is marginally standard Gaussian; as 0 ≤ r e 2 ≤ 1 0\le r_e^2\le1 0 ≤ r e 2 ≤ 1 and P e P_e P e has uniformly bounded moments, its contribution is exponentially small. Thus
∥ R e ∥ 2 ≤ C ( ρ e 2 + T − 1 ) T − 3 , ∥ r e 2 − E r e 2 ∥ 2 + ∥ P e − E P e ∥ 2 ≤ C ( ∣ ρ e ∣ T − 1 / 2 + T − 1 ) . \|R_e\|_2\le C(\rho_e^2+T^{-1})T^{-3},\qquad\|r_e^2-\mathbb{E}r_e^2\|_2+\|P_e-\mathbb{E}P_e\|_2\le C(|\rho_e|T^{-1/2}+T^{-1}). ∥ R e ∥ 2 ≤ C ( ρ e 2 + T − 1 ) T − 3 , ∥ r e 2 − E r e 2 ∥ 2 + ∥ P e − E P e ∥ 2 ≤ C ( ∣ ρ e ∣ T − 1/2 + T − 1 ) .
where ρ e = ρ a b \rho_e=\rho_{ab} ρ e = ρ ab . Writing r e 2 = P e + R e r_e^2=P_e+R_e r e 2 = P e + R e , Cauchy–Schwarz and symmetry in ( e , e ′ ) (e,e') ( e , e ′ ) give
∑ e , e ′ ∣ c e e ′ ( R ⋆ ) − Cov ( P e , P e ′ ) ∣ ≤ C T − 3 ∑ e ( ρ e 2 + T − 1 ) ∑ e ′ ( ∣ ρ e ′ ∣ T − 1 / 2 + T − 1 ) . \sum_{e,e'}|c_{ee'}(R_\star)-\operatorname{Cov}(P_e,P_{e'})|\le CT^{-3}\sum_e(\rho_e^2+T^{-1})\sum_{e'}(|\rho_{e'}|T^{-1/2}+T^{-1}). e , e ′ ∑ ∣ c e e ′ ( R ⋆ ) − Cov ( P e , P e ′ ) ∣ ≤ C T − 3 e ∑ ( ρ e 2 + T − 1 ) e ′ ∑ ( ∣ ρ e ′ ∣ T − 1/2 + T − 1 ) .
For completeness, the preceding covariance approximation is uniform in all pair configurations, including e = e ′ e=e' e = e ′ , pairs sharing one index, and disjoint pairs. Indeed, with E e = { max ( ∣ δ a ∣ , ∣ δ b ∣ ) ≤ 1 / 2 } \mathcal{E}_e=\{\max(|\delta_a|,|\delta_b|)\le1/2\} E e = { max ( ∣ δ a ∣ , ∣ δ b ∣ ) ≤ 1/2 } , the fixed-order Gaussian quadratic-form moment bounds imply, uniformly in a , b a,b a , b and the admissible R ∗ R_* R ∗ ,
∥ s e ∥ 8 ≤ C ( ∣ ρ e ∣ + T − 1 / 2 ) , ∥ δ a ∥ 24 + ∥ δ b ∥ 24 ≤ C T − 1 / 2 . \|s_e\|_8\le C(|\rho_e|+T^{-1/2}),\qquad\|\delta_a\|_{24}+\|\delta_b\|_{24}\le CT^{-1/2}. ∥ s e ∥ 8 ≤ C ( ∣ ρ e ∣ + T − 1/2 ) , ∥ δ a ∥ 24 + ∥ δ b ∥ 24 ≤ C T − 1/2 .
Consequently, Hölder’s inequality on E e \mathcal{E}_e E e yields
∥ R e 1 E e ∥ 2 ≤ C ∥ s e ∥ 8 2 ( ∥ δ a ∥ 24 + ∥ δ b ∥ 24 ) 6 ≤ C ( ρ e 2 + T − 1 ) T − 3 . \|\mathcal{R}_e\mathbf{1}_{\mathcal{E}_e}\|_2\le C\|s_e\|_8^2(\|\delta_a\|_{24}+\|\delta_b\|_{24})^6\le C(\rho_e^2+T^{-1})T^{-3}. ∥ R e 1 E e ∥ 2 ≤ C ∥ s e ∥ 8 2 ( ∥ δ a ∥ 24 + ∥ δ b ∥ 24 ) 6 ≤ C ( ρ e 2 + T − 1 ) T − 3 .
On E e c \mathcal{E}_e^c E e c , Gaussian concentration for the normalized quadratic form T − 1 v a ′ Γ v a T^{-1}v_a' \Gamma v_a T − 1 v a ′ Γ v a gives P ( E e c ) ≤ C e − c T \mathbb{P}(\mathcal{E}_e^c)\le Ce^{-cT} P ( E e c ) ≤ C e − c T , the marginal law of each v a v_a v a is standard Gaussian, so this estimate is independent of the cross-sectional correlation. The inequality 0 ≤ r e 2 ≤ 1 0\le r_e^2\le1 0 ≤ r e 2 ≤ 1 and uniform fixed-order moments of the polynomial P e \mathcal{P}_e P e give an O ( e − c ′ T ) O(e^{-c'T}) O ( e − c ′ T ) contribution to the L 2 L^2 L 2 norm by Hölder’s inequality (with possibly smaller c ′ > 0 c'>0 c ′ > 0 ).
To check the centered-factor estimate, write s e = ρ e + s ~ e s_e=\rho_e+\widetilde{s}_e s e = ρ e + s e , where ∥ s ~ e ∥ q ≤ C q T − 1 / 2 \|\widetilde{s}_e\|_q\le C_qT^{-1/2} ∥ s e ∥ q ≤ C q T − 1/2 , and note that ∥ Π e − 1 ∥ q ≤ C q T − 1 / 2 \|\Pi_e-1\|_q\le C_qT^{-1/2} ∥ Π e − 1 ∥ q ≤ C q T − 1/2 for fixed q q q . Then
P e − ρ e 2 = 2 ρ e s ~ e + s ~ e 2 + s e 2 ( Π e − 1 ) , \mathcal{P}_e-\rho_e^2=2\rho_e\widetilde{s}_e+\widetilde{s}_e^2+s_e^2(\Pi_e-1), P e − ρ e 2 = 2 ρ e s e + s e 2 + s e 2 ( Π e − 1 ) ,
so centering and Hölder’s inequality yield ∥ P e − E P e ∥ 2 ≤ C ( ∣ ρ e ∣ T − 1 / 2 + T − 1 ) \|\mathcal{P}_e-\mathbb{E}\mathcal{P}_e\|_2\le C(|\rho_e|T^{-1/2}+T^{-1}) ∥ P e − E P e ∥ 2 ≤ C ( ∣ ρ e ∣ T − 1/2 + T − 1 ) . The same estimate for r e 2 r_e^2 r e 2 follows by adding R e \mathcal{R}_e R e ; ∣ ρ e ∣ ≤ 1 |\rho_e|\le1 ∣ ρ e ∣ ≤ 1 ensures that its remainder is no larger in order. Finally, Cauchy–Schwarz applied to Cov ( R e , P e ′ ) \operatorname{Cov}(\mathcal{R}_e,\mathcal{P}_{e'}) Cov ( R e , P e ′ ) , Cov ( P e , R e ′ ) \operatorname{Cov}(\mathcal{P}_e,\mathcal{R}_{e'}) Cov ( P e , R e ′ ) and Cov ( R e , R e ′ ) \operatorname{Cov}(\mathcal{R}_e,\mathcal{R}_{e'}) Cov ( R e , R e ′ ) produces the displayed double-sum bound. In particular, this argument does not assume independence between e e e and e ′ e' e ′ .
The first sum is ∥ E ∥ F 2 + n ( n − 1 ) / T = O ( n ) \|\mathsf{E}\|_F^2+n(n-1)/T=O(n) ∥ E ∥ F 2 + n ( n − 1 ) / T = O ( n ) , and the second is O ( n ) O(n) O ( n ) because ∑ e ∣ ρ e ∣ ≤ n ∥ E ∥ F \sum_e|\rho_e|\le n\|\mathsf{E}\|_F ∑ e ∣ ρ e ∣ ≤ n ∥ E ∥ F . The right-hand side is therefore O ( n 2 T − 3 ) = O ( n − 1 ) O(n^2T^{-3})=O(n^{-1}) O ( n 2 T − 3 ) = O ( n − 1 ) . The same bound holds under H 0 H_0 H 0 . Hence
Var A ( Tr ( R 2 ) ) − Var 0 ( Tr ( R 2 ) ) = ∑ e , e ′ { ϑ e e ′ ( R ∗ ) − ϑ e e ′ ( I n ) } + O ( n − 1 ) , (44) \operatorname{Var}_A(\operatorname{Tr}(R^2))-\operatorname{Var}_0(\operatorname{Tr}(R^2))=\sum_{e,e'}\{\vartheta_{ee'}(R_*)-\vartheta_{ee'}(I_n)\}+O(n^{-1}), \tag*{(44)} Var A ( Tr ( R 2 )) − Var 0 ( Tr ( R 2 )) = e , e ′ ∑ { ϑ e e ′ ( R ∗ ) − ϑ e e ′ ( I n )} + O ( n − 1 ) , (44)
where ϑ e e ′ ( R ) = Cov R ( P e , P e ′ ) \vartheta_{ee'}(R)=\operatorname{Cov}_R(P_e,P_{e'}) ϑ e e ′ ( R ) = Cov R ( P e , P e ′ ) .
Step 3: Polynomial coefficients and the free-vertex bound. For V = e ∪ e ′ V=e\cup e' V = e ∪ e ′ write R V = I V + ( ρ a b ) a , b ∈ V R_V=I_V+(\rho_{ab})_{a,b\in V} R V = I V + ( ρ ab ) a , b ∈ V , with zero diagonal in the perturbation. Each δ a \delta_a δ a and s ~ a b = s a b − ρ a b \widetilde{s}_{ab}=s_{ab}-\rho_{ab} s ab = s ab − ρ ab is a centered Gaussian quadratic form. For fixed r ≥ 2 r\ge2 r ≥ 2 , the auxiliary lemma above applies to their joint cumulants and to the moments and covariances of products of these centered forms. In particular, each coefficient in the finite-dimensional correlation polynomial of an r r r th cumulant is O ( T 1 − r ) O(T^{1-r}) O ( T 1 − r ) , and a connected partition of d d d centered factors into b b b blocks contributes coefficientwise O ( T b − d ) O(T^{b-d}) O ( T b − d ) .
To account explicitly for powers of ρ e \rho_e ρ e , write
P e = ∑ α + β ≤ 5 c α β ( ρ e 2 + 2 ρ e s ~ e + s ~ e 2 ) δ a α δ b β . P_e=\sum_{\alpha+\beta\le5}c_{\alpha\beta}(\rho_e^2+2\rho_e\widetilde{s}_e+\widetilde{s}_e^2)\delta_a^\alpha\delta_b^\beta. P e = α + β ≤ 5 ∑ c α β ( ρ e 2 + 2 ρ e s e + s e 2 ) δ a α δ b β .
Every summand is a deterministic correlation monomial multiplied by a product of centered quadratic forms. Terms with no centered factors are deterministic and disappear from covariances. More generally, a nonzero covariance contribution requires at least one centered factor from each of the two pairwise polynomials. The explicit factors ρ e \rho_e ρ e and ρ e ′ \rho_{e'} ρ e ′ are counted as edges of the correlation monomial, not as centered factors. Consequently, their presence never improves the T T T -order of a coefficient; it can only reduce the number of absent vertices. The preceding connected-partition bound applies term by term to all three choices ρ e 2 \rho_e^2 ρ e 2 , 2 ρ e s ~ e 2\rho_e\widetilde{s}_e 2 ρ e s e , and s ~ e 2 \widetilde{s}_e^2 s e 2 and likewise for e ′ e' e ′ .
It follows that ϑ e e ′ ( R V ) − ϑ e e ′ ( I V ) \vartheta_{ee'}(R_V)-\vartheta_{ee'}(I_V) ϑ e e ′ ( R V ) − ϑ e e ′ ( I V ) is a polynomial ∑ μ ≠ 1 a μ μ \sum_{\mu\ne1}a_\mu\mu ∑ μ = 1 a μ μ in the off-diagonal correlations, of degree bounded independently of n , T n,T n , T . Let ℓ ( μ ) \ell(\mu) ℓ ( μ ) count the vertices of V V V absent from μ \mu μ . We claim
deg a ( μ ) is even for every vertex a , ∣ a μ ∣ ≤ C T − ℓ ( μ ) − 1 . (45) \deg_a(\mu)\text{ is even for every vertex }a,\qquad|a_\mu|\le CT^{-\ell(\mu)-1}. \tag*{(45)} deg a ( μ ) is even for every vertex a , ∣ a μ ∣ ≤ C T − ℓ ( μ ) − 1 . (45)
The parity follows from simultaneous sign changes of v a v_a v a and the correlations incident to a a a : each P e \mathcal{P}_e P e is invariant. For ℓ = 0 \ell= 0 ℓ = 0 , a nonzero covariance term contains at least two centered factors, and the connected-partition bound is O ( T − 1 ) O(T^{-1}) O ( T − 1 ) . For ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 , choose an absent vertex a a a and an edge e e e containing it. Neither an explicit factor ρ e \rho_e ρ e nor a correlation factor incident to a a a can occur in the coefficient of μ \mu μ . Therefore the contribution from P e \mathcal{P}_e P e necessarily uses s ~ e 2 \tilde{s}_e^2 s ~ e 2 and at least two centered factors. The other pairwise polynomial must contribute at least one centered factor for a nonzero covariance. Thus d ≥ 3 d \ge3 d ≥ 3 and the connected-partition bound is O ( T − 2 ) O(T^{-2}) O ( T − 2 ) , proving the claim for ℓ = 1 \ell= 1 ℓ = 1 .
For ℓ = 2 \ell= 2 ℓ = 2 , a nonconstant monomial with even degree at each vertex has at least two vertices and hence ∣ V ∣ = 4 |V| = 4 ∣ V ∣ = 4 . The pairs are disjoint. If the two absent vertices form one of the pairs, say e e e , then μ \mu μ involves only ρ e ′ \rho_{e'} ρ e ′ ; setting every other correlation to zero makes ( v a ) a ∈ e (v_a)_{a\in e} ( v a ) a ∈ e independent of ( v a ) a ∈ e ′ (v_a)_{a\in e'} ( v a ) a ∈ e ′ , so ϑ e e ′ \vartheta_{ee'} ϑ e e ′ vanishes identically and all such coefficients are zero. Otherwise, after relabeling, the only possible monomial uses correlations between one endpoint i i i of e = ( i , j ) e = (i,j) e = ( i , j ) and one endpoint k k k of e ′ = ( k , l ) e' = (k,l) e ′ = ( k , l ) , with j , l j,l j , l absent. To bound all its coefficients simultaneously, set ρ i k = t \rho_{ik} = t ρ ik = t and every other off-diagonal correlation to zero. This is a valid correlation matrix for ∣ t ∣ ≤ 1 / 2 |t| \le1/2 ∣ t ∣ ≤ 1/2 . Conditional on ( v i , v k ) (v_i,v_k) ( v i , v k ) , v j , v l v_j,v_l v j , v l are independent, so
ϑ e e ′ ( t ) = Cov { h ( v i ) , h ′ ( v k ) } , h ( v i ) = E ( P i j ∣ v i ) . \vartheta_{ee'}(t) = \operatorname{Cov}\{h(v_i),h'(v_k)\}, \qquad h(v_i) = \mathbb{E}(P_{ij}\mid v_i). ϑ e e ′ ( t ) = Cov { h ( v i ) , h ′ ( v k )} , h ( v i ) = E ( P ij ∣ v i ) .
Since ρ i j = 0 \rho_{ij} = 0 ρ ij = 0 , expand Π i j = ∑ α , β c α β δ i α δ j β \Pi_{ij} = \sum_{\alpha,\beta} c_{\alpha\beta}\delta_i^\alpha\delta_j^\beta Π ij = ∑ α , β c α β δ i α δ j β and write
h ( v i ) = ∑ α , β c α β δ i α v i ′ A β v i T 2 , A β = E ( v v ′ δ β ) Γ , h(v_i) = \sum_{\alpha,\beta} c_{\alpha\beta}\delta_i^\alpha\frac{v_i' A_\beta v_i}{T^2}, \qquad A_\beta= \mathbb{E}(vv'\delta^\beta)\Gamma, h ( v i ) = α , β ∑ c α β δ i α T 2 v i ′ A β v i , A β = E ( v v ′ δ β ) Γ ,
where v ∼ N ( 0 , I T ) v \sim N(0,I_T) v ∼ N ( 0 , I T ) and δ = T − 1 v ′ Γ v − 1 \delta= T^{-1}v'\Gamma v - 1 δ = T − 1 v ′ Γ v − 1 . For every fixed β \beta β , ∥ A β ∥ o p ≤ C β \|A_\beta\|_{\mathrm{op}} \le C_\beta ∥ A β ∥ op ≤ C β uniformly: for a unit vector u u u , Cauchy–Schwarz bounds E { ( u ′ v ) 2 ∣ δ ∣ β } \mathbb{E}\{(u'v)^2|\delta|^\beta\} E {( u ′ v ) 2 ∣ δ ∣ β } by a constant depending only on β \beta β and K Γ K_\Gamma K Γ , so ∥ E ( v v ′ δ β ) ∥ o p ≤ C β \|\mathbb{E}(vv'\delta^\beta)\|_{\mathrm{op}} \le C_\beta ∥ E ( v v ′ δ β ) ∥ op ≤ C β , and ∥ A β ∥ o p ≤ ∥ Γ ∥ o p ∥ E ( v v ′ δ β ) ∥ o p \|A_\beta\|_{\mathrm{op}} \le\|\Gamma\|_{\mathrm{op}}\|\mathbb{E}(vv'\delta^\beta)\|_{\mathrm{op}} ∥ A β ∥ op ≤ ∥Γ ∥ op ∥ E ( v v ′ δ β ) ∥ op . Set ϖ β = T − 1 Tr ( A β ) \varpi_\beta= T^{-1}\operatorname{Tr}(A_\beta) ϖ β = T − 1 Tr ( A β ) and B β = A β − ϖ β Γ B_\beta= A_\beta- \varpi_\beta\Gamma B β = A β − ϖ β Γ . Then Tr ( B β ) = 0 \operatorname{Tr}(B_\beta) = 0 Tr ( B β ) = 0 , ∥ B β ∥ F ≤ C T \|B_\beta\|_F \le C\sqrt{T} ∥ B β ∥ F ≤ C T , and
v i ′ A β v i T 2 = ϖ β T ( 1 + δ i ) + v i ′ B β v i T 2 , ∥ v i ′ B β v i T 2 ∥ q ≤ C q T − 3 / 2 . \frac{v_i'A_\beta v_i}{T^2} = \frac{\varpi_\beta}{T}(1+\delta_i) + \frac{v_i'B_\beta v_i}{T^2}, \qquad\left\|\frac{v_i'B_\beta v_i}{T^2}\right\|_q \le C_qT^{-3/2}. T 2 v i ′ A β v i = T ϖ β ( 1 + δ i ) + T 2 v i ′ B β v i , T 2 v i ′ B β v i q ≤ C q T − 3/2 .
Here the last estimate follows from Gaussian quadratic-form moments, and ∥ δ i ∥ q ≤ C q T − 1 / 2 \|\delta_i\|_q \le C_qT^{-1/2} ∥ δ i ∥ q ≤ C q T − 1/2 for each fixed q q q . To make the centering argument explicit, set U β , i = T − 2 v i ′ B β v i U_{\beta,i}=T^{-2}v_i^\prime B_\beta v_i U β , i = T − 2 v i ′ B β v i and H α β , i = δ i α { ϖ β T − 1 ( 1 + δ i ) + U β , i } H_{\alpha\beta,i}=\delta_i^\alpha\{\varpi_\beta T^{-1}(1+\delta_i)+U_{\beta,i}\} H α β , i = δ i α { ϖ β T − 1 ( 1 + δ i ) + U β , i } . The deterministic coefficients satisfy ∣ ϖ β ∣ ≤ C β |\varpi_\beta|\le C_\beta ∣ ϖ β ∣ ≤ C β . For α = 0 \alpha=0 α = 0 , subtracting the expectation gives
∥ H 0 β , i − E H 0 β , i ∥ 2 ≤ C T − 1 ∥ δ i ∥ 2 + 2 ∥ U β , i ∥ 2 ≤ C T − 3 / 2 . \|H_{0\beta,i}-\mathbb{E}H_{0\beta,i}\|_2\le CT^{-1}\|\delta_i\|_2+2\|U_{\beta,i}\|_2\le CT^{-3/2}. ∥ H 0 β , i − E H 0 β , i ∥ 2 ≤ C T − 1 ∥ δ i ∥ 2 + 2∥ U β , i ∥ 2 ≤ C T − 3/2 .
For every fixed α ≥ 1 \alpha\ge1 α ≥ 1 , Hölder’s inequality and the fixed-order Gaussian moment estimates give
∥ H α β , i − E H α β , i ∥ 2 ≤ 2 ∥ H α β , i ∥ 2 ≤ C T − 1 ∥ δ i ∥ 4 α α ( 1 + ∥ δ i ∥ 4 ) + C ∥ δ i ∥ 4 α α ∥ U β , i ∥ 4 ≤ C T − 1 − α / 2 + C T − ( 3 + α ) / 2 ≤ C T − 3 / 2 . \|H_{\alpha\beta,i}-\mathbb{E}H_{\alpha\beta,i}\|_2\le2\|H_{\alpha\beta,i}\|_2
\le CT^{-1}\|\delta_i\|_{4\alpha}^{\alpha}(1+\|\delta_i\|_4)+C\|\delta_i\|_{4\alpha}^{\alpha}\|U_{\beta,i}\|_4
\le CT^{-1-\alpha/2}+CT^{-(3+\alpha)/2}\le CT^{-3/2}. ∥ H α β , i − E H α β , i ∥ 2 ≤ 2∥ H α β , i ∥ 2 ≤ C T − 1 ∥ δ i ∥ 4 α α ( 1 + ∥ δ i ∥ 4 ) + C ∥ δ i ∥ 4 α α ∥ U β , i ∥ 4 ≤ C T − 1 − α /2 + C T − ( 3 + α ) /2 ≤ C T − 3/2 .
Here the case α = 0 \alpha=0 α = 0 is treated separately so that the nonvanishing deterministic part of T − 1 ( 1 + δ i ) T^{-1}(1+\delta_i) T − 1 ( 1 + δ i ) is not mistaken for a centered fluctuation. Since the Taylor polynomial contains only finitely many ( α , β ) (\alpha,\beta) ( α , β ) , summing these bounds shows ∥ h − E h ∥ 2 ≤ C T − 3 / 2 \|h-\mathbb{E}h\|_2\le CT^{-3/2} ∥ h − E h ∥ 2 ≤ C T − 3/2 ; the same argument applies to h ′ h' h ′ . The marginal bounds are uniform in t t t , because each v i v_i v i has the same N ( 0 , I T ) N(0,I_T) N ( 0 , I T ) law for ∣ t ∣ ≤ 1 / 2 |t|\le1/2 ∣ t ∣ ≤ 1/2 . Cauchy–Schwarz therefore gives ∣ ϑ e e ′ ( t ) ∣ ≤ C T − 3 |\vartheta_{ee'}(t)|\le CT^{-3} ∣ ϑ e e ′ ( t ) ∣ ≤ C T − 3 uniformly on [ − 1 / 2 , 1 / 2 ] [-1/2,1/2] [ − 1/2 , 1/2 ] . Because this is a polynomial in t t t of fixed degree, its individual coefficients are O ( T − 3 ) O(T^{-3}) O ( T − 3 ) by finite-dimensional polynomial norm equivalence. To make the coefficient identification explicit, fix the ordered-pair configuration e = ( i , j ) e=(i,j) e = ( i , j ) , e ′ = ( k , l ) e'=(k,l) e ′ = ( k , l ) with four distinct indices. A monomial in which j j j and l l l are absent can involve only the single off-diagonal variable ρ i k \rho_{ik} ρ ik : all other correlations have an endpoint in { j , l } \{j,l\} { j , l } . Thus the specialization ρ i k = t \rho_{ik}=t ρ ik = t , with all other off-diagonal entries zero, preserves every coefficient in this class and eliminates every monomial outside it. In particular, no cancellation with monomials involving an absent vertex is possible. For the fixed Taylor degree, the univariate polynomial has uniformly bounded degree; evaluating it at a fixed set of distinct points in [ − 1 / 2 , 1 / 2 ] [-1/2,1/2] [ − 1/2 , 1/2 ] and inverting the corresponding Vandermonde matrix bounds each coefficient by C T − 3 CT^{-3} C T − 3 . This establishes the claimed bound for the two-absent-vertex class. No nonconstant even-degree monomial has ℓ > 2 \ell> 2 ℓ > 2 when ∣ V ∣ ≤ 4 |V| \le4 ∣ V ∣ ≤ 4 . This completes the coefficient proof.
Step 4: Summation. We apply the Eulerian correlation-monomial summation lemma above. Fix one of the finitely many types of pairs ( e , e ′ ) (e,e') ( e , e ′ ) , a nonconstant monomial μ \mu μ , and its factor multigraph G μ \mathcal{G}_{\mu} G μ on the vertices of μ \mu μ . By the parity property in (S.C.15), every connected component of G μ \mathcal{G}_{\mu} G μ is Eulerian. The auxiliary summation lemma bounds the sum of ∣ μ ∣ |\mu| ∣ μ ∣ over all assignments of its incident vertices and ℓ ( μ ) \ell(\mu) ℓ ( μ ) absent vertices by n ℓ ( μ ) ∥ E ∥ F deg ( μ ) n^{\ell(\mu)}\lVert E\rVert_{F}^{\deg(\mu)} n ℓ ( μ ) ∥ E ∥ F d e g ( μ ) . Combining this with (S.C.15), each fixed monomial type contributes at most
C T − ℓ ( μ ) − 1 n ℓ ( μ ) ∥ E ∥ F deg ( μ ) = O ( n − 1 ) . CT^{-\ell(\mu)-1}n^{\ell(\mu)}\lVert E\rVert_{F}^{\deg(\mu)} = O(n^{-1}). C T − ℓ ( μ ) − 1 n ℓ ( μ ) ∥ E ∥ F d e g ( μ ) = O ( n − 1 ) .
This estimate allows disconnected factor graphs and repeated labels; no distinctness assumption is needed for the upper bound. The number of abstract graph types and monomials is bounded by an absolute constant because ∣ V ∣ ≤ 4 |V| \le4 ∣ V ∣ ≤ 4 and the Taylor degree is fixed; thus this summation does not introduce a dimension-dependent factor. Summing over these types and monomials proves the lemma. ■
Completion of the proof of Lemma 6. Lemma S.1 establishes the asserted uniform O ( 1 ) O(1) O ( 1 ) variance bound. If n ( ϕ 2 − 1 ) = O ( 1 ) n(\phi_2-1)=O(1) n ( ϕ 2 − 1 ) = O ( 1 ) , the covariance comparison in Lemma S.3 gives
Var A ( Tr ( R 2 ) ) = Var 0 ( Tr ( R 0 2 ) ) + O ( n − 1 ) . \operatorname{Var}_{A}(\operatorname{Tr}(R^2))=\operatorname{Var}_{0}(\operatorname{Tr}(R_0^2))+O(n^{-1}). Var A ( Tr ( R 2 )) = Var 0 ( Tr ( R 0 2 )) + O ( n − 1 ) .
Applying the null expansion of Lemma 3 proves (4).
Proof of Theorem 2
Put F n = Tr ( R 2 ) F_n=\operatorname{Tr}(R^2) F n = Tr ( R 2 ) , F 0 n = Tr ( R 0 2 ) F_{0n}=\operatorname{Tr}(R_0^2) F 0 n = Tr ( R 0 2 ) , and s n = n ( ϕ 2 − 1 ) s_n=n(\phi_2-1) s n = n ( ϕ 2 − 1 ) . Under the local-alternative condition s n = O ( 1 ) s_n=O(1) s n = O ( 1 ) , Lemma S.2 constructs F n F_n F n and F 0 n F_{0n} F 0 n from the same Gaussian innovations and gives
D n = F n − F 0 n − s n , E A ( D n 2 ) = O ( n − 1 ) . D_n=F_n-F_{0n}-s_n,\qquad\mathbb{E}_{A}(D_n^2)=O(n^{-1}). D n = F n − F 0 n − s n , E A ( D n 2 ) = O ( n − 1 ) .
Thus D n = o P A ( 1 ) D_n = o_{P_A}(1) D n = o P A ( 1 ) , while F 0 n F_{0n} F 0 n has exactly the null distribution. Since τ 2 , T ≥ 1 \tau_{2,T} \ge1 τ 2 , T ≥ 1 , K T ≥ 0 K_T \ge0 K T ≥ 0 (Remark 2), τ 2 , T , K T = O ( 1 ) \tau_{2,T}, K_T = O(1) τ 2 , T , K T = O ( 1 ) , and λ → λ 0 ∈ ( 0 , ∞ ) \lambda\to\lambda_0 \in(0,\infty) λ → λ 0 ∈ ( 0 , ∞ ) , there are constants 0 < c ≤ C 0 < c \le C 0 < c ≤ C with c ≤ σ n 2 ≤ C c \le\sigma_n^2 \le C c ≤ σ n 2 ≤ C . With Δ n = s n / σ n \Delta_n = s_n/\sigma_n Δ n = s n / σ n , the definition of L M LM L M yields the exact identity
L M − Δ n = F 0 n − μ n σ n + D n σ n . LM-\Delta_n=\frac{F_{0n}-\mu_n}{\sigma_n}+\frac{D_n}{\sigma_n}. L M − Δ n = σ n F 0 n − μ n + σ n D n .
Since F 0 n F_{0n} F 0 n has the null distribution, Theorem 1 applies to the first term on the right, and the second is o P A ( 1 ) o_{P_A}(1) o P A ( 1 ) . Slutsky’s theorem gives L M − Δ n → d N ( 0 , 1 ) LM-\Delta_n \xrightarrow{d} N(0,1) L M − Δ n d N ( 0 , 1 ) ; no convergence of the bounded sequence Δ n \Delta_n Δ n is required.
Proof of Corollary 1 (Detection Boundary)
Write s n = n ( ϕ 2 − 1 ) = Tr ( R ∗ 2 ) − n = ∥ R ∗ − I n ∥ F 2 ≥ 0 s_n=n(\phi_2-1)=\operatorname{Tr}(R_*^2)-n=\lVert R_*-I_n\rVert_F^2\ge0 s n = n ( ϕ 2 − 1 ) = Tr ( R ∗ 2 ) − n = ∥ R ∗ − I n ∥ F 2 ≥ 0 and Δ n = s n / σ n \Delta_n=s_n/\sigma_n Δ n = s n / σ n . As in the proof of Theorem 2, there exist constants 0 < c < C < ∞ 0<c<C<\infty 0 < c < C < ∞ such that c ≤ σ n ≤ C c\le\sigma_n\le C c ≤ σ n ≤ C for all sufficiently large n n n . Thus s n → 0 s_n\to0 s n → 0 (respectively, s n → ∞ s_n\to\infty s n → ∞ ) if and only if Δ n → 0 \Delta_n\to0 Δ n → 0 (respectively, Δ n → ∞ \Delta_n\to\infty Δ n → ∞ ).
Set
b α ( d ) = Φ ( d − z α / 2 ) + Φ ( − d − z α / 2 ) , d ≥ 0. b_\alpha(d)=\Phi(d-z_{\alpha/2})+\Phi(-d-z_{\alpha/2}),\qquad d\ge0. b α ( d ) = Φ ( d − z α /2 ) + Φ ( − d − z α /2 ) , d ≥ 0.
For 0 < α < 1 0<\alpha<1 0 < α < 1 , b α ( 0 ) = α b_\alpha(0)=\alpha b α ( 0 ) = α , b α ( d ) < 1 b_\alpha(d)<1 b α ( d ) < 1 for every finite d d d , and b α ( d ) → 1 b_\alpha(d)\to1 b α ( d ) → 1 as d → ∞ d\to\infty d → ∞ . Moreover,
b α ′ ( d ) = φ ( d − z α / 2 ) − φ ( d + z α / 2 ) > 0 ( d > 0 ) , b_\alpha'(d)=\varphi(d-z_{\alpha/2})-\varphi(d+z_{\alpha/2})>0\qquad(d>0), b α ′ ( d ) = φ ( d − z α /2 ) − φ ( d + z α /2 ) > 0 ( d > 0 ) ,
where φ \varphi φ denotes the standard normal density. Hence b α b_\alpha b α is strictly increasing on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) .
Local power. Suppose s n = O ( 1 ) s_n=O(1) s n = O ( 1 ) . Let W n = L M − Δ n W_n=LM-\Delta_n W n = L M − Δ n and z = z α / 2 z=z_{\alpha/2} z = z α /2 . By Theorem 2, W n → d N ( 0 , 1 ) W_n\xrightarrow{d}N(0,1) W n d N ( 0 , 1 ) . Since Φ \Phi Φ is continuous, Pólya’s theorem gives sup x ∣ P A ( W n ≤ x ) − Φ ( x ) ∣ → 0 \sup_x\lvert P_A(W_n\le x)-\Phi(x)\rvert\to0 sup x ∣ P A ( W n ≤ x ) − Φ ( x )∣ → 0 , and likewise for P A ( W n < x ) P_A(W_n<x) P A ( W n < x ) ; this covers the moving thresholds z − Δ n z-\Delta_n z − Δ n and − z − Δ n -z-\Delta_n − z − Δ n . Since the two rejection events are disjoint,
P A ( ∣ L M ∣ > z ) = P A ( W n > z − Δ n ) + P A ( W n < − z − Δ n ) , P_A(\lvert LM\rvert>z)=P_A(W_n>z-\Delta_n)+P_A(W_n<-z-\Delta_n), P A (∣ L M ∣ > z ) = P A ( W n > z − Δ n ) + P A ( W n < − z − Δ n ) ,
and consequently
P A ( ∣ L M ∣ > z α / 2 ) = b α ( Δ n ) + o ( 1 ) . (46) P_A(|LM| > z_{\alpha/2}) = b_\alpha(\Delta_n) + o(1). \tag*{(46)} P A ( ∣ L M ∣ > z α /2 ) = b α ( Δ n ) + o ( 1 ) . (46)
No convergence of the bounded sequence Δ n \Delta_n Δ n is required.
(a) Sufficiency. Suppose s n → ∞ s_n \to\infty s n → ∞ . The mean expansion of Lemma 5, together with the boundedness of ϕ 2 − 1 \phi_2 - 1 ϕ 2 − 1 under Assumption 4, implies
m n : = E A [ Tr ( R 2 ) ] − μ n = s n + O ( 1 ) + O ( n − 1 ) = s n { 1 + o ( 1 ) } . m_n := E_A[\operatorname{Tr}(R^2)] - \mu_n = s_n + O(1) + O(n^{-1}) = s_n\{1 + o(1)\}. m n := E A [ Tr ( R 2 )] − μ n = s n + O ( 1 ) + O ( n − 1 ) = s n { 1 + o ( 1 )} .
Lemma 6 supplies Var A { Tr ( R 2 ) } = O ( 1 ) \operatorname{Var}_A\{\operatorname{Tr}(R^2)\} = O(1) Var A { Tr ( R 2 )} = O ( 1 ) . Consequently m n − z α / 2 σ n → ∞ m_n - z_{\alpha/2}\sigma_n \to\infty m n − z α /2 σ n → ∞ , and Chebyshev’s inequality yields
P A { ∣ L M ∣ ≤ z α / 2 } ≤ P A { Tr ( R 2 ) − E A Tr ( R 2 ) ≤ − ( m n − z α / 2 σ n ) } ≤ Var A { Tr ( R 2 ) } ( m n − z α / 2 σ n ) 2 ⟶ 0. \begin{aligned}
P_A\{|LM| \le z_{\alpha/2}\} \le P_A\{\operatorname{Tr}(R^2) - E_A\operatorname{Tr}(R^2) \le-(m_n - z_{\alpha/2}\sigma_n)\} \\
&\le\frac{\operatorname{Var}_A\{\operatorname{Tr}(R^2)\}}{(m_n - z_{\alpha/2}\sigma_n)^2} \longrightarrow0.
\end{aligned} P A { ∣ L M ∣ ≤ z α /2 } ≤ P A { Tr ( R 2 ) − E A Tr ( R 2 ) ≤ − ( m n − z α /2 σ n )} ≤ ( m n − z α /2 σ n ) 2 Var A { Tr ( R 2 )} ⟶ 0.
Thus the rejection probability tends to one. This covers every diverging signal, including s n = o ( n ) s_n = o(n) s n = o ( n ) .
(a) Necessity. If s n s_n s n does not tend to infinity, there is a subsequence on which s n s_n s n is bounded. Passing to a further subsequence, s n → s ∈ [ 0 , ∞ ) s_n \to s \in[0,\infty) s n → s ∈ [ 0 , ∞ ) and σ n → σ ∈ [ c , C ] \sigma_n \to\sigma\in[c,C] σ n → σ ∈ [ c , C ] . This is a local-alternative subsequence, so (S.E.1) gives rejection probability tending to b α ( s / σ ) < 1 b_\alpha(s/\sigma) < 1 b α ( s / σ ) < 1 . Hence power cannot tend to one along the full sequence.
(b) Sufficiency. If s n → 0 s_n \to0 s n → 0 , then Δ n → 0 \Delta_n \to0 Δ n → 0 ; (S.E.1) and continuity of b α b_\alpha b α give power tending to b α ( 0 ) = α b_\alpha(0) = \alpha b α ( 0 ) = α .
(b) Necessity. Suppose instead that s n s_n s n does not tend to zero. There are ε > 0 \varepsilon> 0 ε > 0 and a subsequence with s n ≥ ε s_n \ge\varepsilon s n ≥ ε . If this subsequence has an unbounded further subsequence, extract one with s n → ∞ s_n \to\infty s n → ∞ ; part (a) then gives power tending to one, contrary to convergence to α \alpha α . Otherwise s n s_n s n is bounded on a further subsequence; compactness permits s n → s ≥ ε s_n \to s \ge\varepsilon s n → s ≥ ε and σ n → σ ∈ [ c , C ] \sigma_n \to\sigma\in[c,C] σ n → σ ∈ [ c , C ] . Then (S.E.1) gives limiting power b α ( s / σ ) > α b_\alpha(s/\sigma) > \alpha b α ( s / σ ) > α , again a contradiction. Thus power tending to α \alpha α forces s n → 0 s_n \to0 s n → 0 .
(c) If Δ n → Δ ∈ ( 0 , ∞ ) \Delta_n \to\Delta\in(0,\infty) Δ n → Δ ∈ ( 0 , ∞ ) , then s n = O ( 1 ) s_n = O(1) s n = O ( 1 ) because σ n \sigma_n σ n is bounded, and (S.E.1) yields power tending to b α ( Δ ) ∈ ( α , 1 ) b_\alpha(\Delta) \in(\alpha,1) b α ( Δ ) ∈ ( α , 1 ) .
Remark S.1 The equivalences above concern the rejection probability of the stated two-sided L M LM L M test at a fixed nominal level α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . They rely on the boundedness and non-degeneracy of σ n \sigma_n σ n , the uniform O ( 1 ) O(1) O ( 1 ) alternative variance bound in Lemma 6, and the local Gaussian shift approximation of Theorem 2. Theorem 2 alone does not justify the s n → ∞ s_n \to\infty s n → ∞ case; this is why part (a) uses Chebyshev’s inequality.
Auxiliary Gaussian and operator inequalities
This section proves the analytic inputs used in the null martingale argument and in the local-alternative coupling. Constants below depend only on K Γ K_\Gamma K Γ and on fixed moment orders.
Quadratic-form cumulants and moments. Let z ∼ N ( 0 , I d ) z \sim N(0,I_d) z ∼ N ( 0 , I d ) and B = B ′ B = B^\prime B = B ′ be deterministic. Diagonalization and the moment generating function of a squared standard normal yield, for ∣ t ∣ |t| ∣ t ∣ sufficiently small,
log E exp { t ( z ′ B z − Tr B ) } = ∑ r = 2 ∞ 2 r − 1 r t r Tr ( B r ) . \log\mathbb{E}\exp\{t(z^\prime Bz-\operatorname{Tr}B)\}
=\sum_{r=2}^{\infty}\frac{2^{r-1}}{r}t^r\operatorname{Tr}(B^r). log E exp { t ( z ′ B z − Tr B )} = r = 2 ∑ ∞ r 2 r − 1 t r Tr ( B r ) .
Thus κ r = 2 r − 1 ( r − 1 ) ! Tr ( B r ) \kappa_r = 2^{r-1}(r-1)!\operatorname{Tr}(B^r) κ r = 2 r − 1 ( r − 1 )! Tr ( B r ) for r ≥ 2 r \ge2 r ≥ 2 . For each fixed integer m m m , the moment–cumulant formula, together with ∣ Tr ( B r ) ∣ ≤ ∥ B ∥ o p 2 ∥ B ∥ F r − 2 |\operatorname{Tr}(B^r)| \le\|B\|_{\mathrm{op}}^2\|B\|_{\mathrm{F}}^{r-2} ∣ Tr ( B r ) ∣ ≤ ∥ B ∥ op 2 ∥ B ∥ F r − 2 , implies
E ∣ z ′ B z − Tr B ∣ 2 m ≤ C m ∥ B ∥ F 2 m . \mathbb{E}|z^\prime Bz-\operatorname{Tr}B|^{2m} \le C_m\|B\|_{\mathrm{F}}^{2m}. E ∣ z ′ B z − Tr B ∣ 2 m ≤ C m ∥ B ∥ F 2 m .
For symmetric B 1 , B 2 , B 3 B_1,B_2,B_3 B 1 , B 2 , B 3 , polarization of the third cumulant gives
E ∏ j = 1 3 ( z ′ B j z − Tr B j ) = 4 Tr ( B 1 B 2 B 3 + B 1 B 3 B 2 ) . \mathbb{E}\prod_{j=1}^{3}(z^\prime B_jz-\operatorname{Tr}B_j)
=4\operatorname{Tr}(B_1B_2B_3+B_1B_3B_2). E j = 1 ∏ 3 ( z ′ B j z − Tr B j ) = 4 Tr ( B 1 B 2 B 3 + B 1 B 3 B 2 ) .
These are exact identities, including when the matrices are singular.
Inverse quadratic forms without a spectral gap. Suppose Γ ⪰ 0 \Gamma\succeq0 Γ ⪰ 0 , Tr Γ = T \operatorname{Tr}\Gamma= T Tr Γ = T , and ∥ Γ ∥ o p ≤ K Γ \lVert\Gamma\rVert_{\mathrm{op}} \le K_{\Gamma} ∥ Γ ∥ op ≤ K Γ . If exactly k k k eigenvalues are at least 1 / 2 1/2 1/2 , then T ≤ k K Γ + ( T − k ) / 2 T \le kK_{\Gamma} + (T-k)/2 T ≤ k K Γ + ( T − k ) /2 , so k ≥ T / ( 2 K Γ − 1 ) k \ge T/(2K_{\Gamma}-1) k ≥ T / ( 2 K Γ − 1 ) . Consequently z ′ Γ z ≥ 1 2 χ k 2 z^{\prime}\Gamma z \ge\frac{1}{2}\chi_k^2 z ′ Γ z ≥ 2 1 χ k 2 under an eigenbasis coupling, and for every fixed positive integer m m m and all sufficiently large T T T ,
E ( z ′ Γ z ) − m ≤ E ( 1 2 χ k 2 ) − m = Γ ( k / 2 − m ) Γ ( k / 2 ) = O ( T − m ) . \mathbb{E}(z^{\prime}\Gamma z)^{-m} \le\mathbb{E}\left(\frac{1}{2}\chi_k^2\right)^{-m} = \frac{\Gamma(k/2-m)}{\Gamma(k/2)} = O(T^{-m}). E ( z ′ Γ z ) − m ≤ E ( 2 1 χ k 2 ) − m = Γ ( k /2 ) Γ ( k /2 − m ) = O ( T − m ) .
This is the bound (B.2); in particular, no bound on the smallest eigenvalue of Γ \Gamma Γ is needed.
Hilbert-valued Gaussian-chaos moment bound. Let Y = ( Y 1 , … , Y d ) Y = (Y_1,\ldots,Y_d) Y = ( Y 1 , … , Y d ) be a vector of polynomials of total Gaussian degree at most two, including deterministic terms. Scalar Gaussian hypercontractivity gives ∥ Y j ∥ L p ≤ ( p − 1 ) ∥ Y j ∥ L 2 \lVert Y_j\rVert_{L^p} \le(p-1)\lVert Y_j\rVert_{L^2} ∥ Y j ∥ L p ≤ ( p − 1 ) ∥ Y j ∥ L 2 for p ≥ 2 p \ge2 p ≥ 2 . Minkowski’s inequality in L p / 2 L^{p/2} L p /2 then yields
∥ ∥ Y ∥ 2 2 ∥ L p = ∥ ∑ j Y j 2 ∥ L p / 2 ≤ ∑ j ∥ Y j ∥ L p 2 ≤ ( p − 1 ) 2 E ∥ Y ∥ 2 2 . \lVert\lVert Y\rVert_2^2\rVert_{L^p} = \left\lVert\sum_j Y_j^2\right\rVert_{L^{p/2}} \le\sum_j \lVert Y_j\rVert_{L^p}^2 \le(p-1)^2\mathbb{E}\lVert Y\rVert_2^2. ∥∥ Y ∥ 2 2 ∥ L p = j ∑ Y j 2 L p /2 ≤ j ∑ ∥ Y j ∥ L p 2 ≤ ( p − 1 ) 2 E ∥ Y ∥ 2 2 .
Thus the constant is independent of the dimension d d d . This yields the rowwise L 8 L^8 L 8 bounds in the coupling proof once their second moments have been computed from the exact covariance of N N N .
References