Asymptotic Theory for Sample Canonical Direction Alignment in High-Dimensional Spiked CCA
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Mathematics. The day in mathematics: new theorems, new proofs, and what they open.
Ada: We're looking at the paper: "Asymptotic Theory for Sample Canonical Direction Alignment in High-Dimensional Spiked CCA".
Emmy: Here's the gist of the paper. This paper studies the asymptotic behavior of sample canonical directions in a Gaussian spiked high-dimensional CCA model and quantifies their directional recovery quality through squared alignment measures.
Ada: Core Problem and Model Setup The main objective of this paper is to quantify, in a finite-rank spiked high-dimensional CCA model, how well a sample canonical direction aligns with its population counterpart Under the asymptotic regime in which the dimensions of the two data blocks grow proportionally with the sample size, sample canonical directions are generally not consistent estimators of their population counterparts, even when the corresponding sample canonical correlations separate from the bulk spectrum The study works under Assumption 1.1 (On the dimensions), which assumes that p:= p(n), q:= q(n), and, as n → ∞, p/n → c1 ∈ (0, 1), q/n → c2 ∈ (0, 1), s.t.
Emmy: c1 + c2 ∈ (0, 1) Assumption 1.2 (On the population spiked model) assumes that rank(Σxy) = k for some fixed positive integer k, and that the associated population spikes are above a phase-transition threshold rc.
Ada: Asymptotic Results for Alignment The paper develops a first- and second-order asymptotic theory for the squared alignment hui, νii squared, which measures the amount of population-level directional information retained by the sample canonical direction The first contribution is a deterministic first-order characterization of this squared alignment, proving that hui, νii squared p → 1 + d(ri), where d(ri) is an explicit function of the population spike ri and c1, c2 The main contribution is a CLT for the fluctuation of the squared alignment, showing that √n (hui, νii squared − 1 − d(ri)) converges in distribution to a centered normal distribution with variance σ 2(ri)/(1 + d(ri)) 4 Estimation of Asymptotic Quantities To make the theoretical quantities computable from data, plug-in estimators are constructed for both the limiting mean and the asymptotic variance based on the sample spiked eigenvalue li The estimator for the limiting mean is µˆi = 1 / (1 + d(ˆri)), where rˆi is estimated by inverting the deterministic outlier eigenvalue map rˆi = γ-1(li) Similarly, the asymptotic variance estimator is τ 2 i = σ 2(ˆri) / (1 + d(ˆri)) 4 These plug-in estimators are proven to consistently estimate the limiting mean and asymptotic variance of the squared alignment Real-Data Illustration and Simulation A real-data illustration using limestone grassland community data shows that for the leading sample spiked eigenvalue l1 = 0.8293, the plug-in estimator µˆ1 = 0.8155 provides a quantitative assessment of the recovery quality Numerical simulations support these results, illustrating the convergence of hui, νii squared to its deterministic limit and demonstrating that the standardized statistics follow an asymptotic normal distribution under Theorem 2.5 Technical Proof Strategy The proofs hinge on reducing the eigenvector projection problem to a problem concerning quadratic forms of resolvent matrices using the characteristic equation The analysis involves decomposing the matrix Mi into four components, Ik1, Ik2, Ik3, and Ik4 Lemma 5.1 establishes that the diagonal entries of Mi satisfy
Mi: ii = 1 + d(ri) + Op(n − 1/2), while Lemma 5.2 characterizes the limiting variances of the fluctuation terms X1n, X2n, and X3n These results are used to prove Theorem 2.4 by showing that hui, νii squared = 1 + Op(n−1) and Theorem 2.5 by showing the standardized statistic converges to N(0, σ 2(ri))/(1 + d(ri)) 4 The final proof of Lemma 6.1 combines these estimates to show that the variance of the normalized trace functionals is bounded by O(n−2).
Paper summary: Ada: We've just covered the core findings about sample canonical direction alignment in the paper "Asymptotic Theory for Sample Canonical Direction Alignment in High-Dimensional Spiked CCA". To recap, this work deals with quantifying the directional recovery quality of sample directions when p and q grow with n.
Emmy: The thesis is that while these sample directions aren't always consistent estimators of their population counterparts, the paper establishes a deterministic first-order limit for the squared alignment and a central limit theorem for its fluctuations.
Kolya: It essentially provides a way to measure how well we can recover population information from our noisy, high-dimensional sample data by looking at this squared alignment measure.
Ada: Exactly. The context is that we are in a regime where the dimensions of the two data blocks scale with the sample size, which is very different from classical low-dimensional asymptotics where p and q are fixed while n goes to infinity.
Emmy: They define their model with specific assumptions about dimensionality, assuming p and q grow proportionally to n by constants c1 and c2 that sum up to less than one. They also assume a fixed rank k for the population cross-covariance matrix Σxy.
Kolya: And they set a threshold rc based on those constants, which dictates when the sample spikes are large enough to be considered significant and separate from the noise bulk of the limiting spectral distribution.
Ada: If your population spike ri is larger than that threshold rc, then you can analyze each sample eigenvector individually because they behave nicely under these assumptions.
Emmy: The theoretical machinery involves reducing the eigenvector projection problem to analyzing quadratic forms of resolvent matrices using a characteristic equation. They decompose the matrix Mi into four components, Ik1, Ik2, Ik3 and Ik4.
Kolya: So, what's the practical implication of that decomposition for understanding the alignment? Does it simplify things enough?
Ada: Lemma five point one establishes something about those diagonal entries of Mi: they satisfy Miii equals one plus d of ri plus some small error term that goes with n minus half.
Emmy: And lemma five point two characterizes the limiting variances of the fluctuation terms X1n, X2n, and X3n. These characterizations are what allow them to prove theorem two point four and theorem two point five.
Kolya: So those lemmas are the tools that bridge the gap between abstract matrix decomposition and a concrete probabilistic statement about convergence?
Ada: Precisely. Theorem two point four shows that hui p equals one plus some small error term that goes with n minus one, which means the alignment is almost always close to its deterministic limit.
Emmy: And theorem two point five shows that the standardized statistic converges to a normal distribution with a variance related to sigma squared ri divided by one plus d of ri raised to the fourth power.
Kolya: So, in short, they've given us both a deterministic limit and the statistical behavior around that limit. That's pretty comprehensive for this type of analysis.
Ada: And the paper does this using plug-in estimators for the limiting mean and variance based on observed sample spiked eigenvalue li to make it computable from data.
Conclusion: Ada: So wrapping up, the paper "Asymptotic Theory for Sample Canonical Direction Alignment in High-Dimensional Spiked CCA" by Zhangni Pu, Zhangxiao Zhuo, and Jiang Hu addresses the behavior of sample canonical directions under high-dimensional settings.
Emmy: The authors establish that even though these sample directions aren't consistent estimators of their population counterparts in these high-dimensional regimes, they can still quantify the recovery quality.
Kolya: They prove that they can establish a deterministic first-order limit for the squared alignment and provide a statistical bound on how much it fluctuates around that limit using a normal distribution.
Ada: The title itself highlights the focus on asymptotic theory, which means it's about understanding what happens when the dimensions scale with n.
Emmy: The implication is that instead of just saying sample directions are wrong, we get a concrete mathematical description of *how* they are wrong and how much error there is.
Kolya: So for someone who just listens, it means we move from a vague statement about inconsistency to a specific measure of recovery quality based on the limiting mean and variance that you can estimate from your data.
Ada: It turns the non-consistency of sample canonical directions into a data-based measure of recovery quality. That's what they are really doing here.
Key Laboratory of Big Data Analysis of Jilin Province · School of Mathematics & Statistics, Northeast Normal University
math.ST, stat.ME, stat.TH
Submitted: 2026-06-08
Updated: 2026-11-10
Comments: Added the subcritical alignment result, refined the central limit theorem and its asymptotic variance, and improved the proofs and presentation
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 83/100
The gist: This paper studies the asymptotic behavior of sample canonical directions in a Gaussian spiked high-dimensional CCA model and quantifies their directional recovery quality through squared alignment…
Key concepts
- Sample Canonical Directions
- These are the principal directions derived from the sample covariance matrix in a high-dimensional setting. The paper examines how closely these sample directions resemble the true underlying population structure when dimensions are large.
- Squared Alignment Measure ($ u_{ii}^2$)
- This measure quantifies the amount of population-level directional information retained by a specific sample canonical direction. It is used to assess the quality of the alignment between the sample and population directions.
- Asymptotic Regime (Assumption 1.1 & 1.2)
- This describes a high-dimensional scenario where both data block dimensions grow proportionally with the sample size ($n$). The model assumes a fixed, finite rank for the true population structure, which is crucial for analyzing the asymptotic behavior.
Terminology
Summary
This paper studies the asymptotic behavior of sample canonical directions in a Gaussian spiked high-dimensional CCA model and quantifies their directional recovery quality through squared alignment measures. p. 0
Core Problem and Model Setup
The main objective of this paper is to quantify, in a finite-rank spiked high-dimensional CCA model, how well a sample canonical direction aligns with its population counterpart Under the asymptotic regime in which the dimensions of the two data blocks grow proportionally with the sample size, sample canonical directions are generally not consistent estimators of their population counterparts, even when the corresponding sample canonical correlations separate from the bulk spectrum The study works under Assumption 1.1 (On the dimensions), which assumes that p:= p(n), q:= q(n), and, as n → ∞, p/n → c1 ∈ (0, 1), q/n → c2 ∈ (0, 1), s. p. 0t. c1 + c2 ∈ (0, 1) Assumption 1. p. 02 (On the population spiked model) assumes that rank(Σxy) = k for some fixed positive integer k, and that the associated population spikes are above a phase-transition threshold rc. p. 2
Asymptotic Results for Alignment
The paper develops a first- and second-order asymptotic theory for the squared alignment hui, νii squared, which measures the amount of population-level directional information retained by the sample canonical direction The first contribution is a deterministic first-order characterization of this squared alignment, proving that hui, νii squared p → 1 + d(ri), where d(ri) is an explicit function of the population spike ri and c1, c2 The main contribution is a CLT for the fluctuation of the squared alignment, showing that √n (hui, νii squared − 1 − d(ri)) converges in distribution to a centered normal distribution with variance σ 2(ri)/(1 + d(ri)) 4
Estimation of Asymptotic Quantities
To make the theoretical quantities computable from data, plug-in estimators are constructed for both the limiting mean and the asymptotic variance based on the sample spiked eigenvalue li The estimator for the limiting mean is µˆi = 1 / (1 + d(ˆri)), where rˆi is estimated by inverting the deterministic outlier eigenvalue map rˆi = γ-1(li) Similarly, the asymptotic variance estimator is τ 2 i = σ 2(ˆri) / (1 + d(ˆri)) 4 These plug-in estimators are proven to consistently estimate the limiting mean and asymptotic variance of the squared alignment
Real-Data Illustration and Simulation
A real-data illustration using limestone grassland community data shows that for the leading sample spiked eigenvalue l1 = 0.8293, the plug-in estimator µˆ1 = 0. p. 48155 provides a quantitative assessment of the recovery quality Numerical simulations support these results, illustrating the convergence of hui, νii squared to its deterministic limit and demonstrating that the standardized statistics follow an asymptotic normal distribution under Theorem 2. p. 05
Technical Proof Strategy
The proofs hinge on reducing the eigenvector projection problem to a problem concerning quadratic forms of resolvent matrices using the characteristic equation The analysis involves decomposing the matrix Mi into four components, Ik1, Ik2, Ik3, and Ik4 Lemma 5.1 establishes that the diagonal entries of Mi satisfy [Mi]ii = 1 + d(ri) + Op(n − 1/2), while Lemma 5. p. 12 characterizes the limiting variances of the fluctuation terms X1n, X2n, and X3n These results are used to prove Theorem 2. p. 74 by showing that hui, νii squared = 1 + Op(n−1) and Theorem 2. p. 75 by showing the standardized statistic converges to N(0, σ 2(ri))/(1 + d(ri)) 4 The final proof of Lemma 6. p. 41 combines these estimates to show that the variance of the normalized trace functionals is bounded by O(n−2). p. 4
Key Findings on Recovery Quality
The paper concludes that since sample canonical directions are generally not consistent estimators of their population counterparts, the relevant question is how large the limiting squared alignment is and how it fluctuates around its limit Theorem 2.4 identifies the asymptotic center of this alignment as µi:= 1 / (1 + d(ri)) Proposition 3. p. 71 shows that this limiting mean can be consistently estimated from the observed sample spiked eigenvalue li This turns the non-consistency of sample canonical directions into a data-based measure of recovery quality The example shows that µˆ1 = 0. p. 88155 suggests that the leading sample canonical direction retains a substantial amount of population-level directional information. p. 3
References
[1] Anderson, T. W. (2003). An introduction to multivariate statistical analysis, volume 3. p. 0 p. 0 Wiley New York
[2] Bai, Z. and feng Yao, J. (2008). Central limit theorems for eigenvalues in a spiked population model. p. 2 p. 2 Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, 44(3):447 – 474
[3] Bai, Z., Hou, Z., Hu, J., Jiang, D. p. 0 p. 0, and Zhang, X. (2022). Limiting canonical distribution of two large-dimensional random vectors. p. 0 p. 0 In Methodology and Applications of Statistics: A Volume in Honor of CR Rao on the Occasion of his 100th Birthday, pages 213–238. Springer
[4] Bai, Z., Miao, B., and Pan, G. (2007). On asymptotics of eigenvectors of large sample covariance matrix. p. 1 The Annals of Probability, 35(4):1532–1572
[5] Bai, Z. D. and Silverstein, J. W. (2004). CLT for linear spectral statistics of large-dimensional sample covariance matrices. p. 2 The Annals of Probability, 32(1A):553 – 605
[6] Bao, Z., Ding, X., Wang, J., and Wang, K. (2022). Statistical inference for principal components of spiked covariance matrices. p. 3 p. 3 The Annals of Statistics, 50(2):1144–1169
[7] Bao, Z., Ding, X., and Wang, K. (2021). Singular vector and singular subspace distribution for the matrix denoising model. p. 2 p. 2 Ann. Statist., 49(1):370–392
[8] Bao, Z., Hu, J., Pan, G., and Zhou, W. (2019). Canonical correlation coefficients of high-dimensional gaussian vectors: Finite rank case. p. 1 p. 1 The Annals of Statistics, 47(1):612–640
[9] Bao, Z., Wang, D., and Zhu, Y. (2026). Eigenvector distribution of random matrices under critical finite-rank deformations
[10] Bykhovskaya, A. and Gorin, V. p. 3 p. 3 (2023). High-dimensional canonical correlation analysis
[11] Fan, J., Fan, Y., Han, X., and Lv, J. (2022). Asymptotic theory of eigenvectors for random matrices with diverging spikes. p. 2 p. 2 Journal of the American Statistical Association, 117(538):996– 1009
[12] Gittins, R. (1985). Canonical Analysis: A Review with Applications in Ecology, volume 12 of Biomathematics. p. 1 p. 1 Springer, Berlin, Heidelberg
[13] Harold, H. (1936). Relations between two sets of variables. p. 3 p. 3 Biometrika, 28(3):321–377
[14] Hotelling, H. (1935). The most predictable criterion. Journal of educational Psychology, 26(2):139
[15] Hou, Z., Zhang, X., Bai, Z., and Hu, J. (2023). Spiked eigenvalues of noncentral fisher matrix with applications. p. 2 p. 2 Bernoulli, 29(4):3171–3197
[16] Lei, J. (2016). A goodness-of-fit test for stochastic block models. p. 3 p. 3 The Annals of Statistics, 44(1):401–424
[17] Li, Y., Zhou, H., and Hu, J. (2023). The eigenvector lsd of information plus noise matrices and its application to linear regression model. p. 6 p. 6 Statistics & Probability Letters, 197:109811
[18] Liu, X., Liu, Y., Pan, G., Zhang, L., and Zhang, Z. (2023). Asymptotic properties of spiked eigenvalues and eigenvectors of signal-plus-noise matrices with their applications. p. 6 p. 6 arXiv preprint arXiv:2310. p. 0 p. 013939
[19] Lytova, A. and Pastur, L. (2009). Central limit theorem for linear eigenvalue statistics of random matrices with independent entries. p. 0 p. 0 The Annals of Probability, 37(5):1778 – 1840
[20] Ma, Z. and Yang, F. (2023). Sample canonical correlation coefficients of high-dimensional random vectors with finite rank correlations. p. 0 p. 0 Bernoulli, 29(3):1905–1932
[21] Muirhead, R. J. (1982). Aspects of multivariate statistical theory. p. 1 p. 1 John Wiley & Sons
[22] Paul, D. (2007). Asymptotics of sample eigenstructure for a large dimensional spiked covariance model. p. 1 p. 1 Statist. Sinica, 17(4):1617–1642
[23] Pu, Z., Zhang, X., Hu, J., and Bai, Z. (2024). The asymptotic properties of the extreme eigenvectors of high-dimensional generalized spiked covariance model. p. 1 arXiv preprint arXiv:2405. p. 008524
[24] Wachter, K. W. (1980). The Limiting Empirical Measure of Multiple Discriminant Ratios. p. 7 The Annals of Statistics, 8(5):937 – 957
[25] Xi, H., Yang, F., and Yin, J. (2020). Convergence of eigenvector empirical spectral distribution of sample covariance matrices. p. 2 The Annals of Statistics, 48(2):953–982
[26] Xia, N., Qin, Y., and Bai, Z. (2013). Convergence rates of eigenvector empirical spectral distribution of large dimensional sample covariance matrix. p. 2 The Annals of Statistics, 41(5):2572–2607
[27] Yang, F. (2022). Limiting distribution of the sample canonical correlation coefficients of high-dimensional random vectors. p. 0 Electronic Journal of Probability, 27:1–71
[28] Yang, Y. and Pan, G. (2012). The convergence of the empirical distribution of canonical correlation coefficients. p. 2 Electronic Journal of Probability, 17:1–13
[29] Yang, Y. and Pan, G. (2015). Independence test for high dimensional data based on regularized canonical correlation coefficients. p. 0 The Annals of Statistics, 43(2):467–500
[30] Zhang, X. (2023). The limiting spectral distribution of the sample canonical correlation matrix. p. 2 arXiv preprint arXiv:2309. p. 013369
[44] Acknowledgments The authors would like to thank Zhigang Bao for many illuminating discussions in an early stage of this research
[45] References [1] Anderson, T. W. (2003). An introduction to multivariate statistical analysis, volume 3. Wiley New York
[45] [2] Bai, Z. and feng Yao, J. (2008). Central limit theorems for eigenvalues in a spiked population model. Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, 44(3):447 – 474
[5] Bai, Z., Hou, Z., Hu, J., Jiang, D., and Zhang, X. (2022). Limiting canonical distribution of two large-dimensional random vectors. In Methodology and Applications of Statistics: A Volume in Honor of CR Rao on the Occasion of his 100th Birthday, pages 213–238. Springer
[6] Bao, Z., Ding, X., Wang, J., and Wang, K. (2022). Statistical inference for principal components of spiked covariance matrices. The Annals of Statistics, 50(2):1144–1169
[7] Bao, Z., Ding, X., and Wang, K. (2021). Singular vector and singular subspace distribution for the matrix denoising model. Ann. Statist., 49(1):370–392
[8] Bao, Z., Hu, J., Pan, G., and Zhou, W. (2019). Canonical correlation coefficients of high-dimensional gaussian vectors: Finite rank case. The Annals of Statistics, 47(1):612–640
[9] Bao, Z., Wang, D., and Zhu, Y. (2026). Eigenvector distribution of random matrices under critical finite-rank deformations
[10] Bykhovskaya, A. and Gorin, V. (2023). High-dimensional canonical correlation analysis arXiv preprint arXiv:2306. p. 016393
[11] Fan, J., Fan, Y., Han, X., and Lv, J. (2022). Asymptotic theory of eigenvectors for random matrices with diverging spikes. Journal of the American Statistical Association, 117(538):996– 1009
[12] Gittins, R. (1985). Canonical Analysis: A Review with Applications in Ecology, volume 12 of Biomathematics. Springer, Berlin, Heidelberg
[13] Harold, H. (1936). Relations between two sets of variables. Biometrika, 28(3):321–377
[14] Hotelling, H. (1935). The most predictable criterion. Journal of educational Psychology, 26(2):139
[15] Hou, Z., Zhang, X., Bai, Z., and Hu, J. (2023). Spiked eigenvalues of noncentral fisher matrix with applications. Bernoulli, 29(4):3171–3197
[16] Lei, J. (2016). A goodness-of-fit test for stochastic block models. The Annals of Statistics, 44(1):401–424
[17] Li, Y., Zhou, H., and Hu, J. (2023). The eigenvector lsd of information plus noise matrices and its application to linear regression model. Statistics & Probability Letters, 197:109811
[18] Liu, X., Liu, Y., Pan, G., Zhang, L., and Zhang, Z. (2023). Asymptotic properties of spiked eigenvalues and eigenvectors of signal-plus-noise matrices with their applications. arXiv preprint arXiv:2310.13939
[19] Lytova, A. and Pastur, L. (2009). Central limit theorem for linear eigenvalue statistics of random matrices with independent entries. The Annals of Probability, 37(5):1778 – 1840
[20] Ma, Z. and Yang, F. (2023). Sample canonical correlation coefficients of high-dimensional random vectors with finite rank correlations. Bernoulli, 29(3):1905–1932
[21] Muirhead, R. J. (1982). Aspects of multivariate statistical theory. John Wiley & Sons
[22] Paul, D. (2007). Asymptotics of sample eigenstructure for a large dimensional spiked covariance model. Statist. Sinica, 17(4):1617–1642
[23] Pu, Z., Zhang, X., Hu, J., and Bai, Z.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Consistent estimation of recovery quality: The improved system can estimate not only whether a sample canonical direction is detectable from the spectrum but also its
recovery quality
by providing a data-based measure of how accurately it represents its population counterpart, quantified by the plug-in estimator mu hat i = 1 over 1 plus d hat ri. -
Quantification of uncertainty: The system can estimate not only the limiting mean of the squared alignment but also the scale of its random fluctuation by computing tau hat i = sigma square (hat ri) over 1 plus d hat ri, providing a sample-based measure of both strength and uncertainty.
-
Robust asymptotic behavior analysis: The improved system can analyze the convergence in distribution for standardized squared alignments, showing that sqrt n [hu i, nu ii 2 - 1 over 1 plus d(ri)] converges to a normal distribution with variance sigma square (ri) over (1 plus d(ri)) to support Gaussian approximation.
-
Real-world assessment of direction quality: The system can assess the reliability of leading sample canonical directions in practice by using the plug-in estimator mu hat i as an empirical indicator of high population-level directional information retention, distinguishing it from simple eigenvalue detection.
Sources
- High-Dimensional Canonical Correlation Analysis
- Asymptotic limits of spiked eigenvalues and eigenvectors of signal-plus-noise matrices with weak signals and heteroskedastic noise
- The Asymptotic Properties of the Extreme Eigenvectors of High-dimensional Generalized Spiked Covariance Model
- The limiting spectral distribution of the sample canonical correlation matrix
Related papers
- Sigmoid-FTRL: Design-Based Adaptive Neyman Allocation for AIPW Estimators
- Conformal Prediction for Dyadic Regression Under Complex Missingness
- Sharp Asymptotic Theory of Maximum Likelihood Estimation for Gaussian Processes with an RBF Kernel
- Bentkus-type asymptotic e-values
- Canonical Layer Inference for Deep Gaussian Processes via Symmetry-Breaking Augmentation
- High-Dimensional Asymptotics of Differentially Private PCA