Introduction
Many real-world systems exhibit abrupt structural shifts. Financial crises reshape cross-sectional correlations of asset returns; central-bank policy regime changes alter macroeconomic dynamics; climate regimes reorganize correlated environmental variables; experimental conditions in neuroscience modify neural population dynamics. Knowing the moment of the regime change is essential for forecasting, risk management, model adaptation, and interpretation of post-shift behavior.
Mathematically, a sudden structural shift means that the underlying data time series X = ( x t ) 1 ≤ t ≤ n ∈ R p × n X = (x_t)_{1 \le t \le n} \in\mathbb{R}^{p \times n} X = ( x t ) 1 ≤ t ≤ n ∈ R p × n is generated by two different models: one before the breakpoint τ = ⌊ θ n ⌋ \tau= \lfloor\theta n \rfloor τ = ⌊ θ n ⌋ , and one after τ \tau τ (see Figure 1 ). The central inferential task is then to estimate τ \tau τ , or at least the relative break time θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) , and to prove consistency of the estimator as the sample size n n n grows. The quasi-maximum-likelihood (QML) breakpoint estimator θ ^ n \hat{\theta}_n θ ^ n minimizes the regularized log-likelihood objective
Figure 1. Example of a time series with a single breakpoint (top): regime 1 (blue) gives way to regime 2 (red) at the breakpoint τ = ⌊ θ n ⌋ \tau= \lfloor\theta n \rfloor τ = ⌊ θ n ⌋ for θ = 0.45 \theta= 0.45 θ = 0.45 . The regularized QML objective J n ( ν ) J_n(\nu) J n ( ν ) it induces (bottom) attains its global minimum at the grid point corresponding to the breakpoint in this example.
J n ( ν ) = ν log det ( Σ ^ 0 : ν + ε I q ) + ( 1 − ν ) log det ( Σ ^ ν : 1 + ε I q ) J_n(\nu) = \nu\log\det\left(\widehat{\Sigma}_{0:\nu} + \varepsilon I_q\right) + (1-\nu) \log\det\left(\widehat{\Sigma}_{\nu:1} + \varepsilon I_q\right) J n ( ν ) = ν log det ( Σ 0 : ν + ε I q ) + ( 1 − ν ) log det ( Σ ν : 1 + ε I q )
over candidate fractions ν ∈ ( 0 , 1 ) \nu\in(0,1) ν ∈ ( 0 , 1 ) . Here, Σ ^ a : b \widehat{\Sigma}_{a:b} Σ a : b is the empirical second moment of a q q q -dimensional signal series y t y_t y t on the macroscopic block { ⌊ a n ⌋ + 1 , … , ⌊ b n ⌋ } \{\lfloor an \rfloor+ 1,\ldots,\lfloor bn \rfloor\} {⌊ an ⌋ + 1 , … , ⌊ bn ⌋} . The constant ε ≥ 0 \varepsilon\ge0 ε ≥ 0 is a regularization parameter. Figure 1 illustrates the idea on a simple series with a single breakpoint: J n ( ν ) J_n(\nu) J n ( ν ) attains its global minimum at the grid point corresponding to the break in this example.
Classical change-point theory imposes distributional or independence structure on the data — from CUSUM/MOSUM [25 , 10 , 13 ] to nonparametric feature-map / kernel-discrepancy methods [20 , 24 ] : consistency is typically proved under structural assumptions — independence or weak dependence, moment conditions, often homogeneity — strong enough to force a (quantitative) law of large numbers on the relevant empirical statistics. In many real-world settings these assumptions are unrealistic. For example, considering financial time series: returns are not independent and are heavy-tailed, and insider trading makes the idiosyncratic component correlated with the history and non-Markov. In general physical systems, the measurement itself can introduce systemic and persistent artifacts that are hard to detect and to remove. Overall, demanding that statistics satisfy a SLLN is too strong. For breakpoint detection it should only matter whether there is enough evidence of a regime change, regardless of the microstructure of the signal, such as whether the noise is independent, idiosyncratic, or even random. Estimating the breakpoint should be possible as soon as the signal overcomes the noise.
This intuition has been formalized in a substantial line of work on model-free and nonparametric change-point detection. The monographs of Brodsky and Darkhovsky [12 ] and Csörgő and Horváth [14 ] treat consistency under jump-size or moment conditions rather than full distributional models; in the econometric literature, Bai and Perron [8 ] and Lavielle and Moulines [22 ] prove consistency of least-squares multi-break estimators under jump-size hypotheses and weak-dependence assumptions on the noise. More recently, consistency and optimal detection have been established under explicit lower bounds on the size and spacing of the changes, in works such as Frick–Munk–Sieling [18 ] (multiscale, SMUCE), the kernel-based methods of Garreau–Arlot [19 ] and Arlot–Celisse–Harchaoui [2 ] , the sparse-projection approach of Wang and Samworth [29 ] , and the minimax-optimality theory of Verzelen et al. [28 ] ; see also the survey [27 ] . Figure 2 shows a data set of the kind we have in mind: block averages converge locally, but no regimewise homogeneous law of large numbers holds.
Figure 2. A time series with a single breakpoint at θ = 0.4 \theta= 0.4 θ = 0.4 (base variance 1 → 2 1 \to2 1 → 2 ; red: ± 2 σ \pm2\sigma ± 2 σ envelope) under a slowly varying modulation: every macroscopic block average converges, but to limits that vary along the sample — no regimewise homogeneous law of large numbers holds. The pathwise framework of Section 2 covers such data.
We bring this viewpoint to the QML criterion — in a model-free setting — identifying the sharp signal-to-noise threshold separating consistency from inconsistency. We measure the strength of the signal via
Δ b e t w e e n = inf I 1 ⊂ [ 0 , θ ] , I 2 ⊂ [ θ , 1 ] lim inf n → ∞ ∥ Σ ^ I 1 − Σ ^ I 2 ∥ F . \Delta_{\mathrm{between}} = \inf_{\substack{I_1 \subset[0,\theta], I_2 \subset[\theta,1]}} \liminf_{n \to\infty} \left\lVert\widehat{\Sigma}_{I_1} - \widehat{\Sigma}_{I_2} \right\rVert_F. Δ between = I 1 ⊂ [ 0 , θ ] , I 2 ⊂ [ θ , 1 ] inf n → ∞ lim inf Σ I 1 − Σ I 2 F .
It measures the minimal between-regime contrast. The strength of the noise is measured via
Δ w i t h i n = sup same regime lim sup n → ∞ ∥ Σ ^ I 1 − Σ ^ I 2 ∥ F . \Delta_{\mathrm{within}} = \sup_{\text{same regime}} \limsup_{n \to\infty} \left\lVert\widehat{\Sigma}_{I_1} - \widehat{\Sigma}_{I_2} \right\rVert_F. Δ within = same regime sup n → ∞ lim sup Σ I 1 − Σ I 2 F .
It measures the maximal within-regime fluctuation. In the main result of this paper, see Theorem 23 in Section 2 below, we show that the QML estimator θ ^ n \widehat{\theta}_n θ n is consistent (θ ^ n → θ \widehat{\theta}_n \to\theta θ n → θ as n → ∞ n \to\infty n → ∞ ) whenever the signal-to-noise ratio (SNR) exceeds
Δ b e t w e e n Δ w i t h i n > 1 2 θ ( 1 − θ ) ⋅ M + ε m + ε . (1) \frac{\Delta_{\mathrm{between}}}{\Delta_{\mathrm{within}}} > \frac{1}{2\sqrt{\theta(1-\theta)}} \cdot\frac{M+\varepsilon}{m+\varepsilon}. \tag*{(1)} Δ within Δ between > 2 θ ( 1 − θ ) 1 ⋅ m + ε M + ε . (1)
where M M M and m m m are asymptotic upper and lower spectral bounds for the empirical block second moments involved; see (2) below for the precise definition. No SLLN, no independence, no idiosyncratic structure, no stochastic model whatsoever is required of the data. The classical SLLN-based consistency result (see Theorem 21 below) is recovered as the SLLN forces Δ w i t h i n = 0 \Delta_{\mathrm{within}} = 0 Δ within = 0 corresponding to an infinite SNR. We also show that weakening the SNR by removing the first factor on the right-hand side of (1) still forces a local minimum of the QML objective function at the breakpoint. The proof of Theorem 21 rests on a quantitative version of the strict concavity of log det \log\det log det — a Jensen-gap inequality with explicit constants (see Lemma 44 below).
An additional structural novelty is the additive ridge regularization ε I q \varepsilon I_q ε I q inside each log det. It entails three important consequences. First, the ridge keeps the objective well defined when a regime second moment is singular, e.g. when a factor disappears or emerges at the break (see Remark 16 below); without regularization, positive definiteness of both regime second moments is needed, see Assumption (PD).
Second, the ridge controls the threshold (1) itself. The threshold is the product of two factors: the geometric factor { 2 θ ( 1 − θ ) } − 1 \{2\sqrt{\theta(1-\theta)}\}^{-1} { 2 θ ( 1 − θ ) } − 1 , which only reflects the position θ \theta θ of the break, and the conditioning factor ( M + ε ) / ( m + ε ) (M+\varepsilon)/(m+\varepsilon) ( M + ε ) / ( m + ε ) , which increasing ε \varepsilon ε drives to 1 (see Remark 17 below for the explicit choice). The weakest condition obtainable by tuning the ridge is therefore that the signal-to-noise ratio exceeds the geometric factor alone, and Proposition 24 shows that this residual threshold is sharp: below it, global recovery can fail. Third, and specific to the breakpoint problem, the ridge removes the endpoint boundary layers (see Figure 3 ). For ε > 0 \varepsilon> 0 ε > 0 the QML criterion can be minimized over the full grid D n \mathcal{D}_n D n rather than over a fixed trimmed subset (see Remark 18 and Theorems 21 and 23 below).
Figure 3. A break inside the endpoint boundary layer (θ = 0.05 \theta= 0.05 θ = 0.05 , dashed): without regularization the QML objective dives at the endpoints and its global minimizer is a boundary artifact; with the ridge it attains its minimum at the break.
In modern applications, data are often high-dimensional (large cross-sections, panels, portfolios), where classical low-dimensional change-point methods either break down or require additional structure. The main reason is that generically eigenvalues and eigenvectors are not consistent in the regime p / n → γ > 0 p/n \to\gamma> 0 p / n → γ > 0 . For such data it is common that the strongly correlated observed variables have much of their variation driven by a small number of latent factors, which is captured by factor models . Without claiming completeness, let us list several examples. In finance, factor models are standard tools for portfolio construction, risk management, and performance attribution (see e.g. [][1 ] ). In macroeconomics, dynamic factor models are widely used for nowcasting and forecasting from large panels of economic indicators, including real-time monitoring of GDP and business-cycle conditions by central banks (see e.g. [26 ] and also [9 ] ). In neuroscience, factor models extract low-dimensional latent structure from large neural population recordings (cf. e.g. [15 ] ), and in climate science they summarize and forecast large-scale correlated environmental variability (cf. e.g. [23 ] ).
Breakpoint estimation in high-dimensional factor models was studied by Bai, Han, and Shi [6 ] using least-squares/PCA-based methods; the closest work to ours is the QML/log-determinant approach of Duan, Bai, and Han [17 ] , without ridge regularization, building on the factor-model/PCA theory of Bai [5 ] and Bai–Ng [7 ] . Under structural assumptions on the noise — weak cross-sectional and temporal dependence, moment conditions, often homogeneity — strong enough to force a quantitative law of large numbers on the empirical second moments, [17 ] locate the break up to an error of size O ( 1 ) O(1) O ( 1 ) . We therefore apply in Section 3 the abstract framework of Section 2 to a pervasive factor model x t = Λ ∙ f t + e t x_t = \Lambda_\bullet f_t + e_t x t = Λ ∙ f t + e t — the subscript ∙ ∈ { ≤ , > } \bullet\in\{\le, >\} ∙ ∈ { ≤ , > } is a placeholder for the regime, with loadings Λ ≤ \Lambda_{\le} Λ ≤ before and Λ > \Lambda_{>} Λ > after the break — in the high-dimensional high-sample-size (HDHSS) regime, which means that the dimension p = p n → ∞ p = p_n \to\infty p = p n → ∞ diverges with the sample size n n n . Pervasive means that the leading eigenvalues of Λ ∙ Λ ∙ T \Lambda_\bullet\Lambda_\bullet^\mathsf{T} Λ ∙ Λ ∙ T scale like the underlying dimension p p p ; this is a natural assumption, as factors like the market factor or industrial sectors usually affect a macroscopic proportion of the assets (see Remark 30 below). We recover and extend the QML breakpoint result of Duan, Bai, and Han [17 ] by allowing general error terms e t e_t e t whose blockwise spectral contribution is negligible on the pervasive scale, i.e. of order o ( p ) o(p) o ( p ) . The error term is not necessarily independent, uncorrelated, idiosyncratic, or even random — and may carry any non-pervasive, lower-order structure (weak factors, idiosyncratic noise, dependence). In Corollary 32 below, we then show consistency of the QML estimator for this factor model via a combination of Theorem 23 and the Davis–Kahan sin θ \theta θ theorem. The factor-model approach carries a second, practical advantage: projecting onto the leading left singular basis of the data — the leading principal components — averages out idiosyncratic, name-specific fluctuation and thereby smooths the QML objective. Figure 4 displays this on daily U.S. equity returns spanning 2005–2010: the projected objective is visibly smoother than the SPY-only objective. Breakpoint estimation in the high-dimensional low-sample-size (HDLSS) regime, where p → ∞ p \to\infty p → ∞ while n n n remains fixed, remains an open direction; the inconsistency of sample PCA eigenvectors is a central obstacle to extending the present QML theory.
Figure 4. Smoothing by projection on daily U.S. equity returns, 2005–2010. Top: SPY price. Middle and bottom: normalized QML objectives J n ( ν ) J_n(\nu) J n ( ν ) on the same sample, for SPY alone (p = 1 p = 1 p = 1 ) and for the large-cap universe (p = 200 p = 200 p = 200 names) projected onto its leading q = 5 q = 5 q = 5 principal components (ε = 10 − 8 \varepsilon= 10^{-8} ε = 1 0 − 8 ). The projected objective is visibly smoother.
The theory of this paper applies to data carrying one dominant breakpoint: a location at which the between-regime contrast exceeds the within-regime fluctuation (Definition 8); our theorems require the excess to reach the threshold factor of (1). A dominant breakpoint is unique whenever it exists (Lemma 9), and every minimizer of J n J_n J n converges to it (Theorem 23). In this sense, being a global minimizer of J n J_n J n is a necessary condition for the dominant breakpoint, but not a sufficient one.
If no location meets the threshold, the theory makes no statement, and a global minimizer of J n J_n J n may then be produced by a transient anomaly, a corrupted stretch of data, or several breaks of comparable size; and even a genuine breakpoint whose contrast falls short of the threshold can be missed (Proposition 24). A minimizer of J n J_n J n should therefore be read as a candidate breakpoint whose signal-to-noise ratio remains to be checked.
The paper is organized as follows. Section 2 develops the abstract, model-free framework and establishes the sharp signal-to-noise consistency threshold for the regularized QML estimator, together with its ridge consequences and the sharpness construction. Section 3 specializes the framework to high-dimensional pervasive factor models. Section 4 contains the proofs, and Appendix A documents the code generating the figures.
Conventions and notation
For a matrix M M M , we denote with ∥ M ∥ \lVert M\rVert ∥ M ∥ the operator norm of M = ( M i j ) M = (M_{ij}) M = ( M ij ) with respect to the standard Euclidean norm and with ∥ M ∥ F \lVert M\rVert_F ∥ M ∥ F its Frobenius norm, i.e.
∥ M ∥ : = sup ∣ x ∣ = 1 ∣ M x ∣ and ∥ M ∥ F : = ∑ i , j ∣ M i j ∣ 2 \lVert M\rVert:= \sup_{|x|=1} |Mx| \quad\text{and} \quad\lVert M\rVert_F := \sqrt{\sum_{i,j} |M_{ij}|^2} ∥ M ∥ := ∣ x ∣ = 1 sup ∣ M x ∣ and ∥ M ∥ F := i , j ∑ ∣ M ij ∣ 2
A hat denotes an empirical (sample) quantity or an estimator, e.g. Σ ^ a : b \widehat{\Sigma}_{a:b} Σ a : b , θ ^ n \widehat{\theta}_n θ n , and U ^ \widehat{U} U .
For a location ξ ∈ ( 0 , 1 ) \xi\in(0,1) ξ ∈ ( 0 , 1 ) we write k ξ : = ⌊ ξ n ⌋ k_\xi:= \lfloor\xi n\rfloor k ξ := ⌊ ξ n ⌋ for its discretization, so that the breakpoint is τ = k θ \tau= k_\theta τ = k θ ; for a sequence ( ν n ) (\nu_n) ( ν n ) we abbreviate k n : = ⌊ ν n n ⌋ k_n := \lfloor\nu_n n\rfloor k n := ⌊ ν n n ⌋ .
A subscript or superscript ∙ \bullet ∙ is a placeholder for a regime index, ∙ ∈ { ≤ , > } \bullet\in\{\leq, >\} ∙ ∈ { ≤ , > } , where ≤ \leq ≤ refers to the regime before the breakpoint τ \tau τ and > > > to the regime after it; e.g. Λ ∙ \Lambda_\bullet Λ ∙ , Σ F ∙ \Sigma_{F\bullet} Σ F ∙ , and Σ ∙ \Sigma_\bullet Σ ∙ denote the corresponding pre-/post-break objects.
For a finite set, e.g. a macroscopic index block I I I , we denote with ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ its cardinality; whether ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ means the Euclidean norm of a vector or the cardinality of a set will always be clear from the context.
Abstract framework for QML-Breakpoint detection
The data is given by a p p p -dimensional time series X = ( x t ) 1 ≤ t ≤ n ∈ R p × n X = (x_t)_{1 \leq t \leq n} \in\mathbb{R}^{p \times n} X = ( x t ) 1 ≤ t ≤ n ∈ R p × n . We work in a high-sample-size asymptotic regime in which the number of observations n n n increases. The analysis of this section is carried out entirely on a fixed, q q q -dimensional signal series (see Definition 2 below) and imposes no restriction on the ambient dimension p p p , which may be held fixed or diverge with n n n ; only the signal dimension q q q is fixed. This generality is what lets us instantiate the same framework in the high-dimensional regime p = p n → ∞ p = p_n \to\infty p = p n → ∞ in Section 3. We assume that the data suddenly switches regimes at time τ : = ⌊ θ n ⌋ ∈ N \tau:= \lfloor\theta n\rfloor\in\mathbb{N} τ := ⌊ θ n ⌋ ∈ N for some fixed θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) . The time τ \tau τ is called the breakpoint of the data. A model-free definition of a dominant breakpoint, formulated in terms of the signal series alone, is given in Definition 8 below; the data may contain several breakpoints, but by Lemma 9 at most one of them is dominant. The main purpose of this section is to show that a quasi-maximum-likelihood (QML) estimator is able to detect the dominant breakpoint provided that the data contains a signal that is sufficiently distinct between the regimes.
As we work with asymptotics, i.e. sending the number of observations n n n to infinity, let us be more precise about how the data is generated.
Assumption 1. Let θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) . We assume that there is a latent two-sided p p p -dimensional sequence Z = ( z s ) s ∈ Z \mathbf{Z} = (z_s)_{s\in\mathbb{Z}} Z = ( z s ) s ∈ Z , fixed and independent of n n n , which is in the first regime for s ≤ 0 s \le0 s ≤ 0 and in the second regime for s ≥ 1 s \ge1 s ≥ 1 . For a sample of size n n n , the observed data X = ( x t ) 1 ≤ t ≤ n ∈ R p × n \mathbf{X} = (x_t)_{1\le t\le n} \in\mathbb{R}^{p\times n} X = ( x t ) 1 ≤ t ≤ n ∈ R p × n is the length-n n n window of Z \mathbf{Z} Z that straddles the origin at the fraction θ \theta θ ,
x t : = z t − ⌊ θ n ⌋ , 1 ≤ t ≤ n , x_t := z_{t-\lfloor\theta n \rfloor}, \qquad1 \le t \le n, x t := z t − ⌊ θ n ⌋ , 1 ≤ t ≤ n ,
i.e. the window − ⌊ θ n ⌋ + 1 ≤ s ≤ ⌈ ( 1 − θ ) n ⌉ -\lfloor\theta n \rfloor+ 1 \le s \le\lceil(1-\theta)n \rceil − ⌊ θ n ⌋ + 1 ≤ s ≤ ⌈( 1 − θ ) n ⌉ re-indexed to start at t = 1 t=1 t = 1 .
In practical terms, Assumption 1 means that as the sample size n n n grows, the observation window of the fixed latent sequence Z \mathbf{Z} Z expands on both sides of the break at the origin, while the break itself is held at the fixed sample fraction θ \theta θ . This construction has several advantages:
The breakpoint τ n = ⌊ θ n ⌋ \tau_n = \lfloor\theta n \rfloor τ n = ⌊ θ n ⌋ is kept at a fixed fraction θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) of the sample.
It guarantees that both regimes have a sample size growing proportional to n n n .
It provides a clean coupling across different n n n , simplifying the asymptotic analysis of estimators.
This coupling is only a convenient canonical construction: all results below use only the stated blockwise empirical limits and therefore apply unchanged to triangular arrays satisfying the assumptions of the result in question.
Instead of considering the full p p p -dimensional time series ( x t ) (x_t) ( x t ) , we only consider a q q q -dimensional signal time series ( y t ) (y_t) ( y t ) , carefully chosen to separate the regimes. Choosing the right signal is important as it allows for filtering noise and amplifying the regime separation. The signal time series ( y t ) (y_t) ( y t ) is the only object on which the QML estimator below acts.
Definition 2 (Signal time series and signal separation space). The signal time series ( y t ) (y_t) ( y t ) is a q q q -dimensional sequence y t ∈ R q y_t \in\mathbb{R}^q y t ∈ R q , 1 ≤ t ≤ n 1 \le t \le n 1 ≤ t ≤ n , obtained from the data by a map x t ↦ y t x_t \mapsto y_t x t ↦ y t . This map can be nonlinear but the canonical construction is linear. Assuming that the separating signal lives in a q q q -dimensional subspace of R p \mathbb{R}^p R p , we fix an orthonormal basis u i ∈ R p u_i \in\mathbb{R}^p u i ∈ R p , 1 ≤ i ≤ q 1 \le i \le q 1 ≤ i ≤ q , of this space, and let U q = ( u 1 , … , u q ) ∈ R p × q U_q = (u_1,\ldots,u_q) \in\mathbb{R}^{p\times q} U q = ( u 1 , … , u q ) ∈ R p × q denote the associated matrix. This subspace span ( U q ) \operatorname{span}(U_q) span ( U q ) is called the signal separation space, and the signal time series records the coordinates of the orthogonal projection of x t x_t x t onto it in the basis U q U_q U q , i.e.
y t : = U q T x t ∈ R q , 1 ≤ t ≤ n . y_t := U_q^{\mathsf{T}}x_t \in\mathbb{R}^q, \qquad1 \le t \le n. y t := U q T x t ∈ R q , 1 ≤ t ≤ n .
Working with the coordinate vector y t ∈ R q y_t \in\mathbb{R}^q y t ∈ R q rather than the projection in R p \mathbb{R}^p R p is what reduces the ambient dimension to the fixed number q q q . The signal-separation space might depend on the data X \mathbf{X} X and therefore also on n n n or $p.
The choice of the signal time series — equivalently, of the map x t ↦ y t x_t \mapsto y_t x t ↦ y t that produces it — is crucial for the success of breakpoint detection, and a priori knowledge is valuable. Many different constructions of the signal have been proposed, ranging from contrast-based methods like CUSUM (e.g. [25 , 10 ] ) and MOSUM (e.g. [13 , 14 ] ).
to global dimension reduction via PCA/SVD factors (e.g. [6 , 17 ] ). Other popular alternatives include data-driven projection directions [29 ] and nonparametric feature-map discrepancies (kernel/energy-type signals) [20 , 24 ] . The framework also accommodates feature engineering by absorbing it into the construction of x t x_t x t . The raw inputs can be augmented with engineered features (e.g. quadratic interactions, lagged variables, kernel evaluations) designed to amplify the between-regime contrast. We assume that the data time series x t x_t x t already contains those augmented features. We want to point out one trivial case: choosing y t = x t y_t = x_t y t = x t , i.e. the identity map, is a valid choice for defining the signal time series.
Remark 3 (PCA/SVD signals and their limitations). If a priori knowledge is not available, one data-driven choice of the signal separation space is as follows (see for example [5 , 6 , 17 ] ): The signal separation space is spanned by the left leading singular basis vectors obtained by the singular value decomposition of the data matrix X X X . Equivalently this corresponds to choosing the leading eigenspace of the empirical covariance matrix 1 n X X T \frac{1}{n}XX^{\mathsf{T}} n 1 X X T . We use this approach when studying breakpoint detection in factor models (see Section 3 below for more details). A possible problem is that PCA/SVD signal spaces can fail to separate regimes when the directions that change across regimes are not aligned with the leading principal components [3 ] . Concretely, in the retained signal coordinates y t = U q T x t y_t = U_q^{\mathsf{T}}x_t y t = U q T x t the detector observes only the compressed regime moment U q T Σ U q U_q^{\mathsf{T}}\Sigma U_q U q T Σ U q of each regime second moment Σ \Sigma Σ . Writing D = Σ 1 − Σ 2 D = \Sigma_1-\Sigma_2 D = Σ 1 − Σ 2 for the difference between the two regimes, the break is visible in the signal time series if and only if U q T D U q ≠ 0 U_q^{\mathsf{T}}DU_q \ne0 U q T D U q = 0 . Choosing the leading-q q q eigenspace as signal separation space therefore hides the break exactly when D D D is supported on directions orthogonal to that eigenspace — typically trailing, low-variance directions that PCA discards — whereas a change in the variance along a retained leading direction leaves U q T D U q ≠ 0 U_q^{\mathsf{T}}DU_q \ne0 U q T D U q = 0 and remains visible.
As we send n → ∞ n \to\infty n → ∞ , let us introduce the notion of macroscopic intervals.
Definition 4 (Macroscopic intervals). For 0 ≤ a < b ≤ 1 0 \le a < b \le1 0 ≤ a < b ≤ 1 let I ( a , b ) I(a,b) I ( a , b ) denote the macroscopic interval
I ( a , b ) = { ⌊ a n ⌋ + 1 , … , ⌊ b n ⌋ } . I(a,b) = \{\lfloor an\rfloor+ 1,\ldots,\lfloor bn\rfloor\}. I ( a , b ) = {⌊ an ⌋ + 1 , … , ⌊ bn ⌋} .
We define the family I \mathcal{I} I of all macroscopic intervals by
I : = { I ( a , b ) ∣ 0 ≤ a < b ≤ 1 } . \mathcal{I} := \{I(a,b) \mid0 \le a < b \le1\}. I := { I ( a , b ) ∣ 0 ≤ a < b ≤ 1 } .
Let θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) denote the (unknown) breakpoint. We define the sets of macroscopic intervals that lie entirely in the left- and right-hand regimes by
I ≤ : = { I ∈ I : I ⊂ { 1 , … , ⌊ θ n ⌋ } } , I > : = { I ∈ I : I ⊂ { ⌊ θ n ⌋ + 1 , … , n } } . \mathcal{I}_{\le} := \{I \in\mathcal{I}: I \subset\{1,\ldots,\lfloor\theta n\rfloor\}\}, \qquad\mathcal{I}_{>} := \{I \in\mathcal{I}: I \subset\{\lfloor\theta n\rfloor+1,\ldots,n\}\}. I ≤ := { I ∈ I : I ⊂ { 1 , … , ⌊ θ n ⌋}} , I > := { I ∈ I : I ⊂ {⌊ θ n ⌋ + 1 , … , n }} .
Definition 5. For a nonempty macroscopic interval I = { ⌊ a n ⌋ + 1 , … , ⌊ b n ⌋ } ∈ I I = \{\lfloor an\rfloor+1,\ldots,\lfloor bn\rfloor\} \in\mathcal{I} I = {⌊ an ⌋ + 1 , … , ⌊ bn ⌋} ∈ I we define the empirical signal second moment matrix as
Σ ^ a : b : = 1 ∣ I ∣ ∑ t ∈ I y t y t T ∈ R q × q . \widehat{\Sigma}_{a:b} := \frac{1}{|I|}\sum_{t\in I} y_t y_t^{\mathsf{T}} \in\mathbb{R}^{q\times q}. Σ a : b := ∣ I ∣ 1 t ∈ I ∑ y t y t T ∈ R q × q .
We distinguish two families of macroscopic blocks. The break-anchored blocks are
I θ b r : = { I ( s , θ ) : 0 ≤ s < θ } ∪ { I ( θ , s ) : θ < s ≤ 1 } , \mathcal{I}_{\theta}^{\mathrm{br}} := \{I(s,\theta): 0 \le s < \theta\} \cup\{I(\theta,s): \theta< s \le1\}, I θ br := { I ( s , θ ) : 0 ≤ s < θ } ∪ { I ( θ , s ) : θ < s ≤ 1 } ,
and the regime-cut blocks are all pieces obtained by cutting either regime once,
I θ c u t : = I θ b r ∪ { I ( 0 , s ) : 0 < s < θ } ∪ { I ( s , 1 ) : θ < s < 1 } . \mathcal{I}_{\theta}^{\mathrm{cut}} := \mathcal{I}_{\theta}^{\mathrm{br}} \cup\{I(0,s):0<s<\theta\} \cup\{I(s,1):\theta<s<1\}. I θ cut := I θ br ∪ { I ( 0 , s ) : 0 < s < θ } ∪ { I ( s , 1 ) : θ < s < 1 } .
We define the spectral ceiling and floor by
M : = sup I ( a , b ) ∈ I θ b r lim sup n → ∞ ∥ Σ ^ a : b ∥ , m : = inf I ( a , b ) ∈ I θ c u t lim inf n → ∞ λ min ( Σ ^ a : b ) . (2) \begin{aligned}
M &:= \sup_{I(a,b)\in\mathcal{I}_{\theta}^{\mathrm{br}}} \limsup_{n\to\infty} \lVert\widehat{\Sigma}_{a:b}\rVert,\\
m &:= \inf_{I(a,b)\in\mathcal{I}_{\theta}^{\mathrm{cut}}} \liminf_{n\to\infty} \lambda_{\min}(\widehat{\Sigma}_{a:b}). \tag*{(2)}
\end{aligned} M m := I ( a , b ) ∈ I θ br sup n → ∞ lim sup ∥ Σ a : b ∥ , := I ( a , b ) ∈ I θ cut inf n → ∞ lim inf λ m i n ( Σ a : b ) . (2)
We additionally define
Δ w i t h i n : = sup 0 ≤ a i < b i ≤ θ , i = 1 , 2 or θ ≤ a i < b i ≤ 1 , i = 1 , 2 lim sup n → ∞ ∥ Σ ^ a 1 : b 1 − Σ ^ a 2 : b 2 ∥ F , (3) \Delta_{\mathrm{within}} := \sup_{\substack{0\leq a_i<b_i\leq\theta,\ i=1,2\ \text{or}\\ \theta\leq a_i<b_i\leq1,\ i=1,2}} \limsup_{n\to\infty} \lVert\widehat{\Sigma}_{a_1:b_1}-\widehat{\Sigma}_{a_2:b_2}\rVert_F, \tag*{(3)} Δ within := 0 ≤ a i < b i ≤ θ , i = 1 , 2 or θ ≤ a i < b i ≤ 1 , i = 1 , 2 sup n → ∞ lim sup ∥ Σ a 1 : b 1 − Σ a 2 : b 2 ∥ F , (3)
Δ b e t w e e n : = inf 0 ≤ a 1 < b 1 ≤ θ , θ ≤ a 2 < b 2 ≤ 1 lim inf n → ∞ ∥ Σ ^ a 1 : b 1 − Σ ^ a 2 : b 2 ∥ F . (4) \Delta_{\mathrm{between}} := \inf_{\substack{0\leq a_1<b_1\leq\theta,\ \theta\leq a_2<b_2\leq1}} \liminf_{n\to\infty} \lVert\widehat{\Sigma}_{a_1:b_1}-\widehat{\Sigma}_{a_2:b_2}\rVert_F. \tag*{(4)} Δ between := 0 ≤ a 1 < b 1 ≤ θ , θ ≤ a 2 < b 2 ≤ 1 inf n → ∞ lim inf ∥ Σ a 1 : b 1 − Σ a 2 : b 2 ∥ F . (4)
For later use, if ξ ∈ ( 0 , 1 ) \xi\in(0,1) ξ ∈ ( 0 , 1 ) is treated as a candidate split, we denote by Δ w i t h i n ( ξ ) \Delta_{\mathrm{within}}(\xi) Δ within ( ξ ) and Δ b e t w e e n ( ξ ) \Delta_{\mathrm{between}}(\xi) Δ between ( ξ ) the quantities defined in (3) and (4), respectively, with the regime boundary θ \theta θ replaced by ξ \xi ξ . Thus, Δ w i t h i n = Δ w i t h i n ( θ ) \Delta_{\mathrm{within}}=\Delta_{\mathrm{within}}(\theta) Δ within = Δ within ( θ ) and Δ b e t w e e n = Δ b e t w e e n ( θ ) \Delta_{\mathrm{between}}=\Delta_{\mathrm{between}}(\theta) Δ between = Δ between ( θ ) .
Remark 6. We want to point out that we do not make structural assumptions on how the data matrix is generated (e.g. via a factor model), or require that the data follows a strong law of large numbers.
Remark 7. The empirical signal second moment matrices will be the pivotal tool to detect the breakpoint. The quantity Δ w i t h i n \Delta_{\mathrm{within}} Δ within measures the worst-case fluctuation of the empirical signal second moment matrix within a regime; it therefore measures the noise. The quantity Δ b e t w e e n \Delta_{\mathrm{between}} Δ between measures the minimal difference between signal second moment matrices of different regimes. It therefore measures the strength of the regime separation signal.
Definition 8 (Dominant breakpoint). A candidate split ξ ∈ ( 0 , 1 ) \xi\in(0,1) ξ ∈ ( 0 , 1 ) is called a dominant breakpoint of the signal time series if
Δ b e t w e e n ( ξ ) > Δ w i t h i n ( ξ ) , (5) \Delta_{\mathrm{between}}(\xi)>\Delta_{\mathrm{within}}(\xi), \tag*{(5)} Δ between ( ξ ) > Δ within ( ξ ) , (5)
i.e. if blocks on opposite sides of ξ \xi ξ differ asymptotically more than any two blocks on the same side of ξ \xi ξ .
The data may contain several breakpoints, but the dominant breakpoint is unique.
Lemma 9 (At most one dominant breakpoint). There is at most one dominant breakpoint, i.e. at most one ξ ∈ ( 0 , 1 ) \xi\in(0,1) ξ ∈ ( 0 , 1 ) satisfying (5). Therefore, we reserve the symbol θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) to denote the dominant breakpoint in the signal time series.
Proof. We argue by contradiction and let 0 < ν 1 < ν 2 < 1 0<\nu_1<\nu_2<1 0 < ν 1 < ν 2 < 1 be two dominant breakpoints. The macroscopic blocks I ( 0 , ν 1 ) I(0,\nu_1) I ( 0 , ν 1 ) and I ( ν 1 , ν 2 ) I(\nu_1,\nu_2) I ( ν 1 , ν 2 ) lie on opposite sides of ν 1 \nu_1 ν 1 but on the same side of ν 2 \nu_2 ν 2 , so the definitions (3) and (4) give
Δ b e t w e e n ( ν 1 ) ≤ lim inf n → ∞ ∥ Σ ^ 0 : ν 1 − Σ ^ ν 1 : ν 2 ∥ F ≤ lim sup n → ∞ ∥ Σ ^ 0 : ν 1 − Σ ^ ν 1 : ν 2 ∥ F ≤ Δ w i t h i n ( ν 2 ) . \Delta_{\mathrm{between}}(\nu_1)\leq\liminf_{n\to\infty}\lVert\widehat{\Sigma}_{0:\nu_1}-\widehat{\Sigma}_{\nu_1:\nu_2}\rVert_F\leq\limsup_{n\to\infty}\lVert\widehat{\Sigma}_{0:\nu_1}-\widehat{\Sigma}_{\nu_1:\nu_2}\rVert_F\leq\Delta_{\mathrm{within}}(\nu_2). Δ between ( ν 1 ) ≤ n → ∞ lim inf ∥ Σ 0 : ν 1 − Σ ν 1 : ν 2 ∥ F ≤ n → ∞ lim sup ∥ Σ 0 : ν 1 − Σ ν 1 : ν 2 ∥ F ≤ Δ within ( ν 2 ) .
Exchanging the roles of the two candidates and using the blocks I ( ν 1 , ν 2 ) I(\nu_1,\nu_2) I ( ν 1 , ν 2 ) and I ( ν 2 , 1 ) I(\nu_2,1) I ( ν 2 , 1 ) gives Δ b e t w e e n ( ν 2 ) ≤ Δ w i t h i n ( ν 1 ) \Delta_{\mathrm{between}}(\nu_2)\leq\Delta_{\mathrm{within}}(\nu_1) Δ between ( ν 2 ) ≤ Δ within ( ν 1 ) . If both ν 1 \nu_1 ν 1 and ν 2 \nu_2 ν 2 satisfied (5), chaining these inequalities would yield
Δ w i t h i n ( ν 1 ) < Δ b e t w e e n ( ν 1 ) ≤ Δ w i t h i n ( ν 2 ) < Δ b e t w e e n ( ν 2 ) ≤ Δ w i t h i n ( ν 1 ) . \Delta_{\mathrm{within}}(\nu_1)<\Delta_{\mathrm{between}}(\nu_1)\leq\Delta_{\mathrm{within}}(\nu_2)<\Delta_{\mathrm{between}}(\nu_2)\leq\Delta_{\mathrm{within}}(\nu_1). Δ within ( ν 1 ) < Δ between ( ν 1 ) ≤ Δ within ( ν 2 ) < Δ between ( ν 2 ) ≤ Δ within ( ν 1 ) .
a contradiction. □
Remark 10. A dominant breakpoint need not exist: for scalar data taking the second-moment levels a → b → a \mathrm{a} \to\mathrm{b} \to\mathrm{a} a → b → a on three equal segments, every ξ ∈ ( 0 , 1 ) \xi\in(0,1) ξ ∈ ( 0 , 1 ) has level-a blocks on both sides, so Δ b e t w e e n ( ξ ) = 0 \Delta_{\mathrm{between}}(\xi)=0 Δ between ( ξ ) = 0 for all ξ \xi ξ , although the data has two genuine breaks, neither of which is dominant.
In this section we work with the following assumptions:
Assumption 11. (A1) Bounded operator norm. We assume that M < ∞ M < \infty M < ∞ .
(A2) SLLN in each regime. We assume that there are positive semidefinite matrices Σ ≤ ≠ Σ ≥ ∈ R q × q \Sigma_{\leq} \ne\Sigma_{\geq} \in\mathbb{R}^{q \times q} Σ ≤ = Σ ≥ ∈ R q × q such that for all 0 ≤ a 1 < b 1 ≤ θ 0 \leq a_1 < b_1 \leq\theta 0 ≤ a 1 < b 1 ≤ θ and θ ≤ a 2 < b 2 ≤ 1 \theta\leq a_2 < b_2 \leq1 θ ≤ a 2 < b 2 ≤ 1 it holds
lim n → ∞ Σ ^ a 1 : b 1 = Σ ≤ and lim n → ∞ Σ ^ a 2 : b 2 = Σ ≥ . \lim_{n \to\infty} \widehat{\Sigma}_{a_1:b_1} = \Sigma_{\leq}
\quad\text{and} \quad
\lim_{n \to\infty} \widehat{\Sigma}_{a_2:b_2} = \Sigma_{\geq}. n → ∞ lim Σ a 1 : b 1 = Σ ≤ and n → ∞ lim Σ a 2 : b 2 = Σ ≥ .
(A3) Weak signal-separation condition. For ε ≥ 0 \varepsilon\geq0 ε ≥ 0
Δ b e t w e e n > M + ε m + ε Δ w i t h i n . \Delta_{\mathrm{between}} >
\frac{M+\varepsilon}{m+\varepsilon}\Delta_{\mathrm{within}}. Δ between > m + ε M + ε Δ within .
(A4) A posteriori strong signal-separation condition. For ε ≥ 0 \varepsilon\geq0 ε ≥ 0
Δ b e t w e e n > 1 2 θ ( 1 − θ ) M + ε m + ε Δ w i t h i n . \Delta_{\mathrm{between}} >
\frac{1}{2\sqrt{\theta(1-\theta)}}\frac{M+\varepsilon}{m+\varepsilon}\Delta_{\mathrm{within}}. Δ between > 2 θ ( 1 − θ ) 1 m + ε M + ε Δ within .
(A5) A priori strong signal-separation condition. For 0 ≤ δ ≤ θ ≤ 1 − δ 0 \leq\delta\leq\theta\leq1-\delta 0 ≤ δ ≤ θ ≤ 1 − δ and ε ≥ 0 \varepsilon\geq0 ε ≥ 0
Δ b e t w e e n > 1 2 δ ( 1 − δ ) M + ε m + ε Δ w i t h i n . \Delta_{\mathrm{between}} >
\frac{1}{2\sqrt{\delta(1-\delta)}}\frac{M+\varepsilon}{m+\varepsilon}\Delta_{\mathrm{within}}. Δ between > 2 δ ( 1 − δ ) 1 m + ε M + ε Δ within .
(PD) Positive definiteness. We assume that m > 0 m > 0 m > 0 .
Remark 12. The second moment bound of Assumption (A1) is standard and does not rule out the analysis of heavy-tailed data ( x t ) (x_t) ( x t ) as it applies to the signal time series ( y t ) (y_t) ( y t ) . A nonlinear mapping x t ↦ y t x_t \mapsto y_t x t ↦ y t can ensure Assumption (A1), for example by enforcing a uniform bound on sup t ∣ y t ∣ \sup_t |y_t| sup t ∣ y t ∣ via truncation.
Remark 13. (Second moments versus covariances). Throughout the manuscript we work with the empirical second moment matrices
Σ ^ a : b : = 1 ∣ I ∣ ∑ t ∈ I y t y t T , \widehat{\Sigma}_{a:b} := \frac{1}{|I|}\sum_{t \in I} y_t y_t^{\mathsf{T}}, Σ a : b := ∣ I ∣ 1 t ∈ I ∑ y t y t T ,
i.e. uncentered averages of y t y t T y_t y_t^{\mathsf{T}} y t y t T . The framework extends naturally to the centered covariance matrices
Σ ^ a : b c o v : = Σ ^ a : b − y ˉ a : b y ˉ a : b T , y ˉ a : b : = 1 ∣ I ∣ ∑ t ∈ I y t , \widehat{\Sigma}^{\mathrm{cov}}_{a:b} := \widehat{\Sigma}_{a:b} - \bar{y}_{a:b}\bar{y}_{a:b}^{\mathsf{T}},
\qquad
\bar{y}_{a:b} := \frac{1}{|I|}\sum_{t \in I} y_t, Σ a : b cov := Σ a : b − y ˉ a : b y ˉ a : b T , y ˉ a : b := ∣ I ∣ 1 t ∈ I ∑ y t ,
which differ from the second moments by a rank-one correction. Using the same argument, i.e. quantifying the Jensen gap, naturally gives the correct definition of signal Δ b e t w e e n \Delta_{\mathrm{between}} Δ between and noise Δ w i t h i n \Delta_{\mathrm{within}} Δ within , leading to variants of the main results Theorem 21, Theorem 23, and Corollary 32. We chose to work with the second moments throughout because they have clean additivity on macroscopic intervals (cf. (40) in the proof of Theorem 23), which simplifies the calculations.
Remark 14. Building upon Remark 7, the ratio Δ b e t w e e n Δ w i t h i n \frac{\Delta_{\mathrm{between}}}{\Delta_{\mathrm{within}}} Δ within Δ between can be interpreted as the signal-to-noise ratio, and therefore the Assumptions (A2)–(A5) quantify different signal-to-noise ratios. We observe that Assumption (A2) implies
Δ b e t w e e n < ∞ and Δ w i t h i n = 0 , \Delta_{\mathrm{between}} < \infty\quad\text{and} \quad\Delta_{\mathrm{within}} = 0, Δ between < ∞ and Δ within = 0 ,
having the interpretation that the signal-to-noise ratio is infinite. We immediately see that (A2) implies (A3), (A4), and (A5) for all choices of δ > 0 \delta> 0 δ > 0 and ε > 0 \varepsilon> 0 ε > 0 . Moreover, under (A2) every block in G θ b r \mathcal{G}_{\theta}^{\mathrm{br}} G θ br and G θ c u t \mathcal{G}_{\theta}^{\mathrm{cut}} G θ cut converges to the corresponding regime matrix, and therefore
M = max { ∥ Σ ≤ ∥ , ∥ Σ > ∥ } , m = min { λ min ( Σ ≤ ) , λ min ( Σ > ) } . M = \max\{\lVert\Sigma_{\le}\rVert,\lVert\Sigma_{>}\rVert\}, \qquad m = \min\{\lambda_{\min}(\Sigma_{\le}),\lambda_{\min}(\Sigma_{>})\}. M = max {∥ Σ ≤ ∥ , ∥ Σ > ∥} , m = min { λ m i n ( Σ ≤ ) , λ m i n ( Σ > )} .
In particular, (A2) implies (A1). Additionally, we see that if θ ∈ [ δ , 1 − δ ] \theta\in[\delta,1-\delta] θ ∈ [ δ , 1 − δ ] then δ ( 1 − δ ) ≤ θ ( 1 − θ ) \delta(1-\delta) \le\theta(1-\theta) δ ( 1 − δ ) ≤ θ ( 1 − θ ) , and hence the condition (A5) implies (A4). Finally, since 2 θ ( 1 − θ ) ≤ 1 2\sqrt{\theta(1-\theta)} \le1 2 θ ( 1 − θ ) ≤ 1 , the prefactor in (A4) is at least one, so (A4) implies (A3). Altogether we obtain the implication chain
(A5) ⟹ (A4) ⟹ (A3) , \text{(A5)} \Longrightarrow\text{(A4)} \Longrightarrow\text{(A3)}, (A5) ⟹ (A4) ⟹ (A3) ,
where the first implication holds for θ ∈ [ δ , 1 − δ ] \theta\in[\delta,1-\delta] θ ∈ [ δ , 1 − δ ] .
Definition 15 (QML objective function and QML breakpoint estimator θ ^ n \widehat{\theta}_n θ n ). For fixed ε ≥ 0 \varepsilon\ge0 ε ≥ 0 and every ν ∈ ( 0 , 1 ) \nu\in(0,1) ν ∈ ( 0 , 1 ) such that both empirical blocks are nonempty, equivalently 1 ≤ ⌊ ν n ⌋ ≤ n − 1 1 \le\lfloor\nu n \rfloor\le n-1 1 ≤ ⌊ ν n ⌋ ≤ n − 1 , we define the QML objective by
J n ( ν ) : = ν log det ( Σ ^ 0 : ν + ε I q ) + ( 1 − ν ) log det ( Σ ^ ν : 1 + ε I q ) . J_n(\nu) := \nu\log\det\left(\widehat{\Sigma}_{0:\nu}+\varepsilon I_q\right) + (1-\nu)\log\det\left(\widehat{\Sigma}_{\nu:1}+\varepsilon I_q\right). J n ( ν ) := ν log det ( Σ 0 : ν + ε I q ) + ( 1 − ν ) log det ( Σ ν : 1 + ε I q ) .
Here the empirical second moments are understood according to Definition 5, in particular with the floor convention in their block endpoints. Thus the formula also defines J n J_n J n at off-grid fractions whenever both blocks are nonempty. For breakpoint estimation, the data are discrete and it suffices to evaluate J n J_n J n at the grids
D n : = { k n : 1 ≤ k ≤ n − 1 } , D n ( δ ) : = D n ∩ [ δ , 1 − δ ] , 0 < δ < 1 / 2. \mathcal{D}_n := \left\{\frac{k}{n}:1\le k\le n-1\right\}, \qquad\mathcal{D}_n(\delta):=\mathcal{D}_n\cap[\delta,1-\delta], \qquad0<\delta<1/2. D n := { n k : 1 ≤ k ≤ n − 1 } , D n ( δ ) := D n ∩ [ δ , 1 − δ ] , 0 < δ < 1/2.
The untrimmed and trimmed QML estimators of the breakpoint θ \theta θ are obtained by minimizing J n J_n J n and formally defined via
θ ^ n ∈ arg min ν ∈ D n J n ( ν ) , θ ^ n , δ ∈ arg min ν ∈ D n ( δ ) J n ( ν ) . \widehat{\theta}_n \in\operatorname*{arg\,min}_{\nu\in\mathcal{D}_n} J_n(\nu), \qquad\widehat{\theta}_{n,\delta} \in\operatorname*{arg\,min}_{\nu\in\mathcal{D}_n(\delta)} J_n(\nu). θ n ∈ ν ∈ D n arg min J n ( ν ) , θ n , δ ∈ ν ∈ D n ( δ ) arg min J n ( ν ) .
There are three important reasons to include the ε \varepsilon ε -regularization + ε I q +\varepsilon I_q + ε I q in the definition of the QML-objective function J n J_n J n :
Remark 16. The ε \varepsilon ε -regularization + ε I q +\varepsilon I_q + ε I q in the definition of the QML-objective function J n J_n J n allows eigenvalues of the signal second moment matrix to take on the value 0 0 0 . For example, this allows one to model the situation where a signal completely vanishes in the other regime. The main financial application is the rank collapse in a crisis regime, where at the breakpoint all assets are strongly correlated and the cross-sectional correlations are negligible in comparison.
Remark 17. Adding the ε \varepsilon ε -regularization + ε I q +\varepsilon I_q + ε I q in the definition of the QML-objective function J n J_n J n gives one additional degree of freedom to choose from. By choosing ε \varepsilon ε appropriately, one can weaken the signal-separation conditions. In particular (A3) can be weakened to
(A3*) We assume that
Δ b e t w e e n > Δ w i t h i n . \Delta_{\mathrm{between}} > \Delta_{\mathrm{within}}. Δ between > Δ within .
Then ( A 3 ∗ ) (A3^*) ( A 3 ∗ ) implies ( A 3 ) (A3) ( A 3 ) for
ε ≥ max { 0 , M Δ w i t h i n − m Δ b e t w e e n Δ b e t w e e n − Δ w i t h i n } . \varepsilon\ge\max\left\{0,\frac{M\Delta_{\mathrm{within}}-m\Delta_{\mathrm{between}}}{\Delta_{\mathrm{between}}-\Delta_{\mathrm{within}}}\right\}. ε ≥ max { 0 , Δ between − Δ within M Δ within − m Δ between } .
Thus ( A 3 ∗ ) (A3^*) ( A 3 ∗ ) is the best condition obtained by optimizing the sufficient local condition (A3) over the ridge parameter.
The assumption ( A 4 ) (A4) ( A 4 ) can be weakened to:
( A 4 ∗ ) (A4^*) ( A 4 ∗ ) We assume that
Δ b e t w e e n > 1 2 θ ( 1 − θ ) Δ w i t h i n . \Delta_{\mathrm{between}}>\frac{1}{2\sqrt{\theta(1-\theta)}}\Delta_{\mathrm{within}}. Δ between > 2 θ ( 1 − θ ) 1 Δ within .
Then ( A 4 ∗ ) (A4^*) ( A 4 ∗ ) implies ( A 4 ) (A4) ( A 4 ) for
ε ≥ max { 0 , M Δ w i t h i n − 2 θ ( 1 − θ ) m Δ b e t w e e n 2 θ ( 1 − θ ) Δ b e t w e e n − Δ w i t h i n } . \varepsilon\ge\max\left\{0,\frac{M\Delta_{\mathrm{within}}-2\sqrt{\theta(1-\theta)}m\Delta_{\mathrm{between}}}{2\sqrt{\theta(1-\theta)}\Delta_{\mathrm{between}}-\Delta_{\mathrm{within}}}\right\}. ε ≥ max { 0 , 2 θ ( 1 − θ ) Δ between − Δ within M Δ within − 2 θ ( 1 − θ ) m Δ between } .
In contrast to the preceding local condition, the ridge-optimized global threshold in ( A 4 ∗ ) (A4^*) ( A 4 ∗ ) is shown to be sharp in Proposition 24 and Remark 25. We want to point out that in the special case θ = 0.5 \theta=0.5 θ = 0.5 the right-hand side of Assumption ( A 4 ∗ ) (A4^*) ( A 4 ∗ ) equals 1. The assumption ( A 5 ) (A5) ( A 5 ) can be weakened similarly.
Remark 18 (The ridge removes the endpoint boundary layers). The third reason is specific to the breakpoint problem. For ε = 0 \varepsilon=0 ε = 0 the minimization of J n J_n J n has to be restricted to a trimmed grid D n ( δ ) \mathcal{D}_n(\delta) D n ( δ ) : a block second moment built from fewer than q q q samples is rank-deficient, so J n = − ∞ J_n=-\infty J n = − ∞ on the short endpoint windows, and, more generally, the empirical block second moments near the endpoints need not remain uniformly positive definite, so that no uniform control of J n J_n J n is available there. For ε > 0 \varepsilon>0 ε > 0 the ridge bounds the spectrum of the regularized block second moments from below by ε \varepsilon ε , uniformly over all windows, which removes these endpoint boundary layers (see Lemma 38 and Lemma 39 below). As a consequence, the QML criterion can be minimized over the full grid D n \mathcal{D}_n D n (see Theorem 21, Theorem 23, and Corollary 32 below).
Remark 19 (Interpretation of QML objective function). The name QML estimator comes from the fact that J n J_n J n is the profile log-likelihood of the following Gaussian block model of a regime change (for simplicity, ε = 0 \varepsilon=0 ε = 0 ). For a candidate breakpoint 1 ≤ t < n 1\le t<n 1 ≤ t < n , assume the data Y : = ( y ( s ) ) 1 ≤ s ≤ n Y:=(y(s))_{1\le s\le n} Y := ( y ( s ) ) 1 ≤ s ≤ n , written in signal coordinates, is generated with unknown covariance matrices Σ ≤ t , Σ > t \Sigma_{\le t},\Sigma_{>t} Σ ≤ t , Σ > t as
{ y ( s ) } 1 ≤ s ≤ t ∼ i . i . d . N ( 0 , Σ ≤ t ) , { y ( s ) } t < s ≤ n ∼ i . i . d . N ( 0 , Σ > t ) . \{y(s)\}_{1\le s\le t}\mathrel{\mathop{\sim}^{\mathrm{i.i.d.}}\mathcal{N}(0,\Sigma_{\le t}),\qquad\{y(s)\}_{t<s\le n}\mathrel{\mathop{\sim}^{\mathrm{i.i.d.}}\mathcal{N}(0,\Sigma_{>t}).}} { y ( s ) } 1 ≤ s ≤ t ∼ i.i.d. N ( 0 , Σ ≤ t ) , { y ( s ) } t < s ≤ n ∼ i.i.d. N ( 0 , Σ > t ) .
The Gaussian log-likelihood ℓ ( Y ∣ t , Σ ≤ t , Σ > t ) \ell(Y\mid t,\Sigma_{\le t},\Sigma_{>t}) ℓ ( Y ∣ t , Σ ≤ t , Σ > t ) is maximized uniquely at Σ ≤ t = Σ ^ 0 : t \Sigma_{\le t}=\widehat{\Sigma}_{0:t} Σ ≤ t = Σ 0 : t and Σ > t = Σ ^ t : n \Sigma_{>t}=\widehat{\Sigma}_{t:n} Σ > t = Σ t : n , and a standard trace computation (the quadratic terms evaluate to t q tq tq and ( n − t ) q (n-t)q ( n − t ) q at the maximizer) gives
max Σ ≤ t , Σ > t ℓ ( Y ∣ t , Σ ≤ t , Σ > t ) = − n 2 J n ( t n ) − n q 2 ( 1 + log ( 2 π ) ) . \max_{\Sigma_{\le t},\Sigma_{>t}}\ell(Y\mid t,\Sigma_{\le t},\Sigma_{>t})=-\frac{n}{2}J_n\left(\frac{t}{n}\right)-\frac{nq}{2}(1+\log(2\pi)). Σ ≤ t , Σ > t max ℓ ( Y ∣ t , Σ ≤ t , Σ > t ) = − 2 n J n ( n t ) − 2 n q ( 1 + log ( 2 π )) .
Minimizing the QML objective over t t t is therefore maximizing the profile likelihood over the candidate breakpoint.
First we consider the classical case, which means that the empirical second moment matrices satisfy the SLLN of Assumption (A2).
Definition 20 (Population targets for left/right segments). For ν ∈ ( 0 , 1 ) \nu\in(0,1) ν ∈ ( 0 , 1 ) we define the population targets for the left and the right segment by
Σ 0 : ν : = { Σ ≤ , ν ≤ θ , θ ν Σ ≤ + ν − θ ν Σ > , ν > θ , Σ ν : 1 : = { θ − ν 1 − ν Σ ≤ + 1 − θ 1 − ν Σ > , ν < θ , Σ > , ν ≥ θ . \Sigma_{0:\nu}:=
\begin{cases}
\Sigma_{\leq}, & \nu\leq\theta,\\
\frac{\theta}{\nu}\Sigma_{\leq}+\frac{\nu-\theta}{\nu}\Sigma_{>}, & \nu>\theta,
\end{cases}
\qquad
\Sigma_{\nu:1}:=
\begin{cases}
\frac{\theta-\nu}{1-\nu}\Sigma_{\leq}+\frac{1-\theta}{1-\nu}\Sigma_{>}, & \nu<\theta,\\
\Sigma_{>}, & \nu\geq\theta.
\end{cases} Σ 0 : ν := { Σ ≤ , ν θ Σ ≤ + ν ν − θ Σ > , ν ≤ θ , ν > θ , Σ ν : 1 := { 1 − ν θ − ν Σ ≤ + 1 − ν 1 − θ Σ > , Σ > , ν < θ , ν ≥ θ .
For ε ≥ 0 \varepsilon\geq0 ε ≥ 0 , we define the limiting QML-objective function J ∞ ( ν ) J_{\infty}(\nu) J ∞ ( ν ) for ν ∈ ( 0 , 1 ) \nu\in(0,1) ν ∈ ( 0 , 1 ) via
J ∞ ( ν ) : = ν log det ( Σ 0 : ν + ε I q ) + ( 1 − ν ) log det ( Σ ν : 1 + ε I q ) . J_{\infty}(\nu):=\nu\log\det\left(\Sigma_{0:\nu}+\varepsilon I_q\right)+(1-\nu)\log\det\left(\Sigma_{\nu:1}+\varepsilon I_q\right). J ∞ ( ν ) := ν log det ( Σ 0 : ν + ε I q ) + ( 1 − ν ) log det ( Σ ν : 1 + ε I q ) .
We define its values at ν = 0 \nu=0 ν = 0 and ν = 1 \nu=1 ν = 1 by continuous extension.
Theorem 21. We assume (A2) (which also immediately implies (A1)). If ε = 0 \varepsilon=0 ε = 0 , we additionally assume (PD). Then:
(i) The limiting QML objective function J ∞ J_{\infty} J ∞ is strictly concave on ( 0 , θ ) (0,\theta) ( 0 , θ ) and on ( θ , 1 ) (\theta,1) ( θ , 1 ) and has a cusp at its unique minimizer θ \theta θ . More precisely, for all ν ∈ [ 0 , 1 ] \nu\in[0,1] ν ∈ [ 0 , 1 ]
J ∞ ( ν ) − J ∞ ( θ ) ≥ min { θ , 1 − θ } 2 ( M + ε ) 2 ∥ Σ > − Σ ≤ ∥ F 2 ∣ ν − θ ∣ . (6) J_{\infty}(\nu)-J_{\infty}(\theta)\geq
\frac{\min\{\theta,1-\theta\}}{2(M+\varepsilon)^2}
\left\lVert\Sigma_{>}-\Sigma_{\leq}\right\rVert_F^2
\lvert\nu-\theta\rvert. \tag*{(6)} J ∞ ( ν ) − J ∞ ( θ ) ≥ 2 ( M + ε ) 2 min { θ , 1 − θ } ∥ Σ > − Σ ≤ ∥ F 2 ∣ ν − θ ∣ . (6)
(ii) If ε > 0 \varepsilon>0 ε > 0 , then
lim n → ∞ sup ν ∈ D n ∣ J n ( ν ) − J ∞ ( ν ) ∣ = 0. (7) \lim_{n\to\infty}\sup_{\nu\in\mathcal{D}_n}
\left\lvert J_n(\nu)-J_{\infty}(\nu)\right\rvert=0. \tag*{(7)} n → ∞ lim ν ∈ D n sup ∣ J n ( ν ) − J ∞ ( ν ) ∣ = 0. (7)
If ε = 0 \varepsilon=0 ε = 0 , then, for every 0 < δ < 1 2 0<\delta<\frac{1}{2} 0 < δ < 2 1 ,
lim n → ∞ sup ν ∈ D n ( δ ) ∣ J n ( ν ) − J ∞ ( ν ) ∣ = 0. (8) \lim_{n\to\infty}\sup_{\nu\in\mathcal{D}_n(\delta)}
\left\lvert J_n(\nu)-J_{\infty}(\nu)\right\rvert=0. \tag*{(8)} n → ∞ lim ν ∈ D n ( δ ) sup ∣ J n ( ν ) − J ∞ ( ν ) ∣ = 0. (8)
(iii) If ε > 0 \varepsilon>0 ε > 0 , every untrimmed QML breakpoint estimator θ ^ n ∈ arg min ν ∈ D n J n ( ν ) \hat{\theta}_n\in\arg\min_{\nu\in\mathcal{D}_n}J_n(\nu) θ ^ n ∈ arg min ν ∈ D n J n ( ν ) is consistent, i.e. lim n → ∞ θ ^ n = θ \lim_{n\to\infty}\hat{\theta}_n=\theta lim n → ∞ θ ^ n = θ .
If ε = 0 \varepsilon=0 ε = 0 and δ ∈ ( 0 , min { θ , 1 − θ } ) \delta\in(0,\min\{\theta,1-\theta\}) δ ∈ ( 0 , min { θ , 1 − θ }) , then every trimmed QML breakpoint estimator θ ^ n , δ ∈ arg min ν ∈ D n ( δ ) J n ( ν ) \hat{\theta}_{n,\delta}\in\arg\min_{\nu\in\mathcal{D}_n(\delta)}J_n(\nu) θ ^ n , δ ∈ arg min ν ∈ D n ( δ ) J n ( ν ) is consistent, i.e. lim n → ∞ θ ^ n , δ = θ \lim_{n\to\infty}\hat{\theta}_{n,\delta}=\theta lim n → ∞ θ ^ n , δ = θ .
Though the case ε = 0 \varepsilon=0 ε = 0 of Theorem 21 is classical, its part for ε > 0 \varepsilon>0 ε > 0 is new: the ridge removes the endpoint boundary layers, so the QML objective can be minimized over the full grid and no trimming of the candidate set is needed (see Remark 18). In Section 2.1 we give a heuristic argument for Theorem 21, built from the two ingredients that also drive the proof of its robust counterpart (Theorem 23 below): the concavity of log det \log\det log det and the explicit two-sided Jensen-gap estimate of Lemma 44. The rigorous proof is given in Section 4.2.1.
Remark 22 (Why consistency without a rate). Under stronger, quantitative assumptions the breakpoint can be located much more precisely: for pervasive factor models with i.i.d. factors and bounded higher moments, [17 ] estimate the break index up to a bounded error, i.e. ∣ θ ^ n − θ ∣ = O P ( 1 / n ) \lvert\hat{\theta}_n-\theta\rvert=O_{\mathrm{P}}(1/n) ∣ θ ^ n − θ ∣ = O P ( 1/ n ) ; see also the classical change-point rates in [21 , 4 ] . Theorem 21 only concludes θ ^ n → θ \hat{\theta}_n\to\theta θ ^ n → θ . This is not a shortcoming of the method but the exact price of the qualitative Assumption (A2), which demands convergence of the block second moments without quantifying its speed. The following example shows that under our assumptions no rate of convergence can hold.
Let q = 1 q = 1 q = 1 and let the data take just two values: fix 0 < a ≠ b 0 < a \ne b 0 < a = b , a null sequence δ n ↓ 0 \delta_n \downarrow0 δ n ↓ 0 , and set
y t ( n ) : = a for t ≤ ⌊ ( θ + δ n ) n ⌋ , y t ( n ) : = b for t > ⌊ ( θ + δ n ) n ⌋ , y_t^{(n)} := \sqrt{a}\quad\text{for }t \le\lfloor(\theta+\delta_n)n\rfloor,\qquad y_t^{(n)} := \sqrt{b}\quad\text{for }t > \lfloor(\theta+\delta_n)n\rfloor, y t ( n ) := a for t ≤ ⌊( θ + δ n ) n ⌋ , y t ( n ) := b for t > ⌊( θ + δ n ) n ⌋ ,
so the jump sits at θ + δ n \theta+\delta_n θ + δ n , slightly to the right of θ \theta θ . Every fixed macroscopic interval contained in ( 0 , θ ) (0,\theta) ( 0 , θ ) or in ( θ , 1 ) (\theta,1) ( θ , 1 ) eventually contains no jump, so Assumption (A2) holds with breakpoint θ \theta θ , Σ ≤ = a \Sigma_{\le} = a Σ ≤ = a , Σ > = b \Sigma_{>} = b Σ > = b , and moreover Δ w i t h i n = 0 \Delta_{\mathrm{within}} = 0 Δ within = 0 and Δ b e t w e e n = ∣ a − b ∣ > 0 \Delta_{\mathrm{between}} = |a-b| > 0 Δ between = ∣ a − b ∣ > 0 : all assumptions hold with constants independent of the sequence δ n \delta_n δ n . On the other hand, at every finite n n n the data are exactly two-level with jump at θ + δ n \theta+\delta_n θ + δ n , and an elementary computation shows that J n J_n J n is minimized at this jump up to O ( 1 / n ) O(1/n) O ( 1/ n ) , i.e. θ ^ n = θ + δ n + O ( 1 / n ) \hat{\theta}_n = \theta+\delta_n+O(1/n) θ ^ n = θ + δ n + O ( 1/ n ) . Since δ n \delta_n δ n may tend to zero arbitrarily slowly, no rate for ∣ θ ^ n − θ ∣ |\hat{\theta}_n-\theta| ∣ θ ^ n − θ ∣ can be deduced: consistency is the exact conclusion, and Theorem 21 is sharp in this sense. As the example satisfies even Assumption (A2), the same applies a fortiori to Theorem 23.
The example also shows where a rate must come from: the estimation error δ n \delta_n δ n is exactly the speed at which the block second moments in Assumption (A2) converge. Quantifying that speed — which is what the stochastic assumptions of [17 ] do — is precisely what recovers a rate. In short: rates are the price of unquantified hypotheses, and the example shows the price is unavoidable.
Using Remark 14, the interpretation of Theorem 21 is that the QML breakpoint estimator is consistent under infinite signal-to-noise ratio. However, data is usually corrupted and the corruption can be systemic over the whole dataset. There is no reason for the noise to be independent from the data making it hard or impossible to filter it out. One needs to model the in-regime empirical second moments as Σ ^ a : b = Σ ≤ + E a : b \widehat{\Sigma}_{a:b} = \Sigma_{\le} + E_{a:b} Σ a : b = Σ ≤ + E a : b , where E a : b E_{a:b} E a : b denotes the measurement error, and there is no reason to believe that lim n → ∞ E a : b = 0 \lim_{n\to\infty} E_{a:b} = 0 lim n → ∞ E a : b = 0 , implying that Assumption (A2) is too strong; there might be no underlying SLLN at all. We henceforth take on a signal-theoretic point of view and show that a high enough signal-to-noise ratio should be sufficient to detect the breakpoint, leading us to the main result of this article.
Theorem 23 (Consistency of the QML breakpoint estimator). We consider ε ≥ 0 \varepsilon\ge0 ε ≥ 0 , and if ε = 0 \varepsilon= 0 ε = 0 we assume (PD). It holds:
(i) Under the Assumption (A1) and (A3), there are δ ˉ , c > 0 \bar{\delta}, c > 0 δ ˉ , c > 0 such that
lim inf n → ∞ J n ( ν ) − J n ( θ ) ≥ c ∣ ν − θ ∣ , for all ν ∈ ( θ − δ ˉ , θ + δ ˉ ) ∩ ( 0 , 1 ) . \liminf_{n\to\infty} J_n(\nu) - J_n(\theta) \ge c|\nu-\theta|,\qquad\text{for all }\nu\in(\theta-\bar{\delta},\theta+\bar{\delta})\cap(0,1). n → ∞ lim inf J n ( ν ) − J n ( θ ) ≥ c ∣ ν − θ ∣ , for all ν ∈ ( θ − δ ˉ , θ + δ ˉ ) ∩ ( 0 , 1 ) .
This means that asymptotically the breakpoint θ \theta θ is a local minimum of the limiting QML objective function lim inf n J n \liminf_n J_n lim inf n J n .
(ii) Under the Assumption (A1) and (A4), there is c > 0 c > 0 c > 0 such that
lim inf n → ∞ J n ( ν ) − J n ( θ ) ≥ c ∣ ν − θ ∣ , for all ν ∈ ( 0 , 1 ) . (9) \liminf_{n\to\infty} J_n(\nu) - J_n(\theta) \ge c|\nu-\theta|,\qquad\text{for all }\nu\in(0,1). \tag*{(9)} n → ∞ lim inf J n ( ν ) − J n ( θ ) ≥ c ∣ ν − θ ∣ , for all ν ∈ ( 0 , 1 ) . (9)
In particular, if ε > 0 \varepsilon> 0 ε > 0 , every untrimmed QML breakpoint estimator θ ^ n ∈ arg min ν ∈ D n J n ( ν ) \hat{\theta}_n \in\operatorname*{arg\,min}_{\nu\in\mathcal{D}_n} J_n(\nu) θ ^ n ∈ arg min ν ∈ D n J n ( ν ) is consistent, i.e. lim n → ∞ θ ^ n = θ \lim_{n\to\infty}\hat{\theta}_n = \theta lim n → ∞ θ ^ n = θ . If ε = 0 \varepsilon= 0 ε = 0 , then, for every δ ∈ ( 0 , min { θ , 1 − θ } ) \delta\in(0,\min\{\theta,1-\theta\}) δ ∈ ( 0 , min { θ , 1 − θ }) , every trimmed QML breakpoint estimator θ ^ n , δ ∈ arg min ν ∈ D n ( δ ) J n ( ν ) \hat{\theta}_{n,\delta} \in\operatorname*{arg\,min}_{\nu\in\mathcal{D}_n(\delta)} J_n(\nu) θ ^ n , δ ∈ arg min ν ∈ D n ( δ ) J n ( ν ) is consistent, i.e. lim n → ∞ θ ^ n , δ = θ \lim_{n\to\infty}\hat{\theta}_{n,\delta} = \theta lim n → ∞ θ ^ n , δ = θ .
The proof of Theorem 23 is carried out in Section 4.2.2, building on the structure of the proof of Theorem 21 and on the same two ingredients (see Section 2.1). constants c c c are explicit. Writing S : = Δ b e t w e e n 2 ( M + ε ) 2 S := \frac{\Delta_{\mathrm{between}}^{2}}{(M+\varepsilon)^{2}} S := ( M + ε ) 2 Δ between 2 and W : = Δ w i t h i n 2 ( m + ε ) 2 W := \frac{\Delta_{\mathrm{within}}^{2}}{(m+\varepsilon)^{2}} W := ( m + ε ) 2 Δ within 2 for the normalized signal and noise levels, one may take c = ( S − W ) / 4 c = (S-W)/4 c = ( S − W ) /4 in part (i), see (47), and c c c as in (49) in part (ii), namely c = 1 2 min { θ , 1 − θ } S c = \frac{1}{2}\min\{\theta,1-\theta\}S c = 2 1 min { θ , 1 − θ } S if W = 0 W=0 W = 0 and c = 2 θ ( 1 − θ ) S W − W 2 max { θ , 1 − θ } c = \frac{2\sqrt{\theta(1-\theta)SW}-W}{2\max\{\theta,1-\theta\}} c = 2 m a x { θ , 1 − θ } 2 θ ( 1 − θ ) S W − W if W > 0 W>0 W > 0 . Both constants depend only on the gap in Assumptions (A3) and (A4), i.e. on the amount by which the signal-to-noise ratio Δ b e t w e e n / Δ w i t h i n \Delta_{\mathrm{between}}/\Delta_{\mathrm{within}} Δ between / Δ within exceeds the threshold on the respective right-hand side.
Proposition 24 below shows that the ridge-optimized signal-to-noise threshold in (A4*) is necessary for a uniform global-recovery guarantee. In this sense Theorem 23 is optimal. This does not assert that the additional conditioning factor in the fixed-ε \varepsilon ε condition (A4) is itself necessary.
Proposition 24. For θ ∈ ( 0 , 1 2 ) \theta\in(0,\frac{1}{2}) θ ∈ ( 0 , 2 1 ) and 0 < K < r < ∞ 0<K<r<\infty 0 < K < r < ∞ we define the deterministic signal time series ( y t ) 1 ≤ t ≤ n (y_t)_{1\le t\le n} ( y t ) 1 ≤ t ≤ n via
y t = { 1 + K + r , 1 ≤ t ≤ ⌊ θ n ⌋ , 1 + K , ⌊ θ n ⌋ < t ≤ ⌊ 0.5 n ⌋ , 1 , ⌊ 0.5 n ⌋ < t ≤ n . (10) y_t =
\begin{cases}
\sqrt{1+K+r}, & 1 \le t \le\lfloor\theta n\rfloor,\\
\sqrt{1+K}, & \lfloor\theta n\rfloor< t \le\lfloor0.5n\rfloor,\\
\sqrt{1}, & \lfloor0.5n\rfloor< t \le n.
\end{cases} \tag*{(10)} y t = ⎩ ⎨ ⎧ 1 + K + r , 1 + K , 1 , 1 ≤ t ≤ ⌊ θ n ⌋ , ⌊ θ n ⌋ < t ≤ ⌊ 0.5 n ⌋ , ⌊ 0.5 n ⌋ < t ≤ n . (10)
Then at θ \theta θ and 0.5 0.5 0.5 it holds
Δ w i t h i n ( θ ) = K , Δ b e t w e e n ( θ ) = r , Δ b e t w e e n ( θ ) Δ w i t h i n ( θ ) = r K , (11) \Delta_{\mathrm{within}}(\theta)=K,\qquad\Delta_{\mathrm{between}}(\theta)=r,\qquad\frac{\Delta_{\mathrm{between}}(\theta)}{\Delta_{\mathrm{within}}(\theta)}=\frac{r}{K}, \tag*{(11)} Δ within ( θ ) = K , Δ between ( θ ) = r , Δ within ( θ ) Δ between ( θ ) = K r , (11)
and
Δ w i t h i n ( 0.5 ) = r , Δ b e t w e e n ( 0.5 ) = K , Δ b e t w e e n ( 0.5 ) Δ w i t h i n ( 0.5 ) = K r . (12) \Delta_{\mathrm{within}}(0.5)=r,\qquad\Delta_{\mathrm{between}}(0.5)=K,\qquad\frac{\Delta_{\mathrm{between}}(0.5)}{\Delta_{\mathrm{within}}(0.5)}=\frac{K}{r}. \tag*{(12)} Δ within ( 0.5 ) = r , Δ between ( 0.5 ) = K , Δ within ( 0.5 ) Δ between ( 0.5 ) = r K . (12)
Moreover, if
r K = Δ b e t w e e n ( θ ) Δ w i t h i n ( θ ) < 1 2 θ ( 1 − θ ) , (13) \frac{r}{K}=\frac{\Delta_{\mathrm{between}}(\theta)}{\Delta_{\mathrm{within}}(\theta)}<\frac{1}{2\sqrt{\theta(1-\theta)}}, \tag*{(13)} K r = Δ within ( θ ) Δ between ( θ ) < 2 θ ( 1 − θ ) 1 , (13)
then, for every regularization parameter ε > 0 \varepsilon>0 ε > 0 , the pointwise limit J ∞ ( ν ) : = lim n J n ( ν ) J_{\infty}(\nu):=\lim_{n}J_n(\nu) J ∞ ( ν ) := lim n J n ( ν ) exists for every ν ∈ ( 0 , 1 ) \nu\in(0,1) ν ∈ ( 0 , 1 ) and has a unique minimizer at 0.5 0.5 0.5 . In particular, it satisfies
J ∞ ( 0.5 ) < J ∞ ( θ ) . (14) J_{\infty}(0.5)<J_{\infty}(\theta). \tag*{(14)} J ∞ ( 0.5 ) < J ∞ ( θ ) . (14)
For an illustration of Proposition 24 we refer to Figure 5 . The proof is given in Section 4.2.3.
Figure 5. Illustration of the limiting objective J ∞ ( ν ) J_{\infty}(\nu) J ∞ ( ν ) in the deterministic three-level example from Proposition 24 (with θ = 0.30 \theta= 0.30 θ = 0.30 , ν ⋆ = 0.5 \nu_{\star} = 0.5 ν ⋆ = 0.5 , K = 1 K = 1 K = 1 , r = 1.05 r = 1.05 r = 1.05 , ε = 0.1 \varepsilon= 0.1 ε = 0.1 ). Dashed lines indicate θ \theta θ and ν ⋆ \nu_{\star} ν ⋆ .
Remark 25 (Optimality of (A4*) for QML-breakpoint detection). In the situation of Proposition 24 there are two structural changes, at θ \theta θ and at ν ⋆ = 1 2 \nu_\star= \frac{1}{2} ν ⋆ = 2 1 . The designation of θ \theta θ as the genuine break is not a convention: according to the signal-to-noise criterion Δ b e t w e e n ( ν ) / Δ w i t h i n ( ν ) > 1 \Delta_{\mathrm{between}}(\nu)/\Delta_{\mathrm{within}}(\nu) > 1 Δ between ( ν ) / Δ within ( ν ) > 1 at a candidate ν \nu ν (the candidate-level form of (A3 ) in Remark 17; the display (3) –(4) defines the ν \nu ν -dependent quantities), equations (11) and (12) give*
Δ b e t w e e n ( θ ) Δ w i t h i n ( θ ) > 1 and Δ b e t w e e n ( ν ⋆ ) Δ w i t h i n ( ν ⋆ ) < 1. \frac{\Delta_{\mathrm{between}}(\theta)}{\Delta_{\mathrm{within}}(\theta)} > 1 \quad\text{and} \quad\frac{\Delta_{\mathrm{between}}(\nu_\star)}{\Delta_{\mathrm{within}}(\nu_\star)} < 1. Δ within ( θ ) Δ between ( θ ) > 1 and Δ within ( ν ⋆ ) Δ between ( ν ⋆ ) < 1.
Nevertheless, under the insufficient global SNR in (13) , the QML objective has its unique global minimum at the non-dominant change ν ⋆ \nu_\star ν ⋆ for every ε > 0 \varepsilon> 0 ε > 0 . Although (A3 ) holds and a sufficiently large ridge therefore makes θ \theta θ a local minimum, that local condition cannot force global recovery.*
More precisely, fix θ ∈ ( 0 , 1 2 ) \theta\in(0,\frac{1}{2}) θ ∈ ( 0 , 2 1 ) and write c θ : = { 2 θ ( 1 − θ ) } − 1 c_\theta:= \{2\sqrt{\theta(1-\theta)}\}^{-1} c θ := { 2 θ ( 1 − θ ) } − 1 . For every s ∈ ( 1 , c θ ) s \in(1,c_\theta) s ∈ ( 1 , c θ ) , choosing r = s K r=sK r = sK in Proposition 24 produces a scalar deterministic array with Δ b e t w e e n ( θ ) / Δ w i t h i n ( θ ) = s \Delta_{\mathrm{between}}(\theta)/\Delta_{\mathrm{within}}(\theta)=s Δ between ( θ ) / Δ within ( θ ) = s whose QML objective is minimized at ν ⋆ \nu_\star ν ⋆ for every ridge parameter. Conversely, (A4 ) and Remark 17 provide a ridge for which Theorem 23 guarantees global recovery whenever the ratio exceeds c θ c_\theta c θ . Hence the threshold*
Δ b e t w e e n Δ w i t h i n > 1 2 θ ( 1 − θ ) \frac{\Delta_{\mathrm{between}}}{\Delta_{\mathrm{within}}} > \frac{1}{2\sqrt{\theta(1-\theta)}} Δ within Δ between > 2 θ ( 1 − θ ) 1
is exact for a uniform ridge-optimized global-recovery guarantee.
Proof heuristic and the role of the regularization parameter ε \varepsilon ε
This subsection explains the heuristic behind the proof of Theorem 21 and Theorem 23. For that purpose let us assume that
Σ ^ a : b ≈ Σ a : b for all 0 ≤ a < b ≤ 1. \widehat{\Sigma}_{a:b} \approx\Sigma_{a:b} \quad\text{for all } 0 \leq a < b \leq1. Σ a : b ≈ Σ a : b for all 0 ≤ a < b ≤ 1.
This heuristic operates within the SLLN picture of Theorem 21, in which the population targets Σ a : b \Sigma_{a:b} Σ a : b are defined; the proof of the robust Theorem 23 does not rely on these limiting objects and instead works directly with the empirical block moments Σ ^ a : b \widehat{\Sigma}_{a:b} Σ a : b (see Section 4.2.2). Strict concavity of the map A ↦ log det ( A ) A \mapsto\log\det(A) A ↦ log det ( A ) then yields that for ν ∈ ( θ , 1 ) \nu\in(\theta,1) ν ∈ ( θ , 1 ) (the case ν ∈ ( 0 , θ ) \nu\in(0,\theta) ν ∈ ( 0 , θ ) is symmetric)
J n ( ν ) ≈ J ∞ ( ν ) = ν log det ( Σ 0 : ν ) + ( 1 − ν ) log det ( Σ ν : 1 ) = ν log det ( θ ν Σ ≤ + ν − θ ν Σ > ) + ( 1 − ν ) log det ( Σ > ) > θ log det ( Σ ≤ ) + ( 1 − θ ) log det ( Σ > ) = J ∞ ( θ ) ≈ J n ( θ ) . \begin{aligned}
J_n(\nu) &\approx J_\infty(\nu) = \nu\log\det(\Sigma_{0:\nu}) + (1-\nu)\log\det(\Sigma_{\nu:1}) \\
&= \nu\log\det\left(\frac{\theta}{\nu}\Sigma_{\leq} + \frac{\nu-\theta}{\nu}\Sigma_{>}\right) + (1-\nu)\log\det(\Sigma_{>}) \\
&> \theta\log\det(\Sigma_{\leq}) + (1-\theta)\log\det(\Sigma_{>}) = J_\infty(\theta) \approx J_n(\theta).
\end{aligned} J n ( ν ) ≈ J ∞ ( ν ) = ν log det ( Σ 0 : ν ) + ( 1 − ν ) log det ( Σ ν : 1 ) = ν log det ( ν θ Σ ≤ + ν ν − θ Σ > ) + ( 1 − ν ) log det ( Σ > ) > θ log det ( Σ ≤ ) + ( 1 − θ ) log det ( Σ > ) = J ∞ ( θ ) ≈ J n ( θ ) .
Therefore, the breakpoint θ \theta θ is a global minimizer of J ∞ J_\infty J ∞ , and as J n J_n J n converges to J ∞ J_\infty J ∞ , the QML estimator should be consistent.
The proof of Theorem 21 and Theorem 23 then consists of quantifying the concavity estimate, more precisely by calculating and comparing two Jensen gaps. The main ingredient in Theorem 23 is then the estimate (cf. (45) below):
lim inf n → ∞ ( J n ( ν ; ε ) − J n ( θ ; ε ) ) ≥ 1 2 T + ( ν ) ( ν − θ ) , \liminf_{n\to\infty}\bigl(J_n(\nu;\varepsilon)-J_n(\theta;\varepsilon)\bigr) \geq\frac{1}{2}T_+(\nu)(\nu-\theta), n → ∞ lim inf ( J n ( ν ; ε ) − J n ( θ ; ε ) ) ≥ 2 1 T + ( ν ) ( ν − θ ) ,
where
T + ( ν ) = θ ν ( M + ε ) 2 Δ b e t w e e n 2 − 1 − ν ( 1 − θ ) ( m + ε ) 2 Δ w i t h i n 2 . (15) T_+(\nu)=\frac{\theta}{\nu(M+\varepsilon)^2}\Delta_{\mathrm{between}}^2-\frac{1-\nu}{(1-\theta)(m+\varepsilon)^2}\Delta_{\mathrm{within}}^2. \tag*{(15)} T + ( ν ) = ν ( M + ε ) 2 θ Δ between 2 − ( 1 − θ ) ( m + ε ) 2 1 − ν Δ within 2 . (15)
Thus the “cusp sharpness” at the true breakpoint ν = θ \nu= \theta ν = θ is governed by the contrast between two quadratic contributions: a positive signal term of size Δ b e t w e e n 2 / ( M + ε ) 2 \Delta_{\mathrm{between}}^{2}/(M+\varepsilon)^{2} Δ between 2 / ( M + ε ) 2 and a negative noise term of size Δ w i t h i n 2 / ( m + ε ) 2 \Delta_{\mathrm{within}}^{2}/(m+\varepsilon)^{2} Δ within 2 / ( m + ε ) 2 . The sufficient signal-to-noise conditions (A3)–(A5) are precisely those that force T + ( ν ) T_{+}(\nu) T + ( ν ) , together with its left-hand analogue, to stay uniformly positive, either in a neighborhood of θ \theta θ for a local minimum, or on the whole open interval ( 0 , 1 ) (0,1) ( 0 , 1 ) for the global energy gap under (A4).
Remark 26 (The QML objective is biased toward a central split). The estimate (15) also reveals a structural bias of the QML criterion toward a balanced split at the center ν = 1 2 \nu= \frac{1}{2} ν = 2 1 . Indeed, the geometric factor 1 2 θ ( 1 − θ ) \frac{1}{2\sqrt{\theta(1-\theta)}} 2 θ ( 1 − θ ) 1 entering (A4) and (A4 ), viewed as a function of the break fraction θ \theta θ , is minimized at θ = 1 2 \theta= \frac{1}{2} θ = 2 1 , where it equals 1, and grows without bound as θ → 0 \theta\to0 θ → 0 or θ → 1 \theta\to1 θ → 1 . Hence a central breakpoint is the easiest to recover: a signal-to-noise ratio Δ b e t w e e n / Δ w i t h i n \Delta_{\mathrm{between}}/\Delta_{\mathrm{within}} Δ between / Δ within only slightly above 1 already certifies it as the global minimizer. An off-center breakpoint, by contrast, must clear the higher bar 1 2 θ ( 1 − θ ) > 1 \frac{1}{2\sqrt{\theta(1-\theta)}} > 1 2 θ ( 1 − θ ) 1 > 1 . Proposition 24 shows that this geometric bar is sharp for ridge-optimized global recovery at every off-center break fraction (see also Remark 25).*
Remark 27 (Correct choice of the regularization parameter ε > 0 \varepsilon> 0 ε > 0 ). For a candidate ν > θ \nu> \theta ν > θ , the best choice of ε > 0 \varepsilon> 0 ε > 0 is the one that maximizes T + ( ν ) T_{+}(\nu) T + ( ν ) in equation (15); for ν < θ \nu< \theta ν < θ , the left-hand analogue applies. The parameter ε \varepsilon ε influences the size of ( M + ε ) − 2 (M+\varepsilon)^{-2} ( M + ε ) − 2 and ( m + ε ) − 2 (m+\varepsilon)^{-2} ( m + ε ) − 2 . Let us first consider the two extremes. Choosing ε ≪ 1 \varepsilon\ll1 ε ≪ 1 very small: If m m m is small or zero, then ( m + ε ) − 2 (m+\varepsilon)^{-2} ( m + ε ) − 2 is very large and the negative noise contribution gets amplified, potentially flattening or even moving the minimum away from θ \theta θ . Choosing ε ≫ 1 \varepsilon\gg1 ε ≫ 1 very large: If ε \varepsilon ε dominates M M M and m m m then for a minimum of lim inf n J n \liminf_{n} J_{n} lim inf n J n one only needs a stronger signal than noise, i.e. Δ b e t w e e n > Δ w i t h i n \Delta_{\mathrm{between}} > \Delta_{\mathrm{within}} Δ between > Δ within . The problem is that the contrast of the cusp scales like 1 ε \frac{1}{\varepsilon} ε 1 , making the objective function nearly flat around θ \theta θ and therefore the QML estimator sensitive to noise. The best performance is expected at an intermediate “sweet spot” where ε \varepsilon ε is large enough to remove the small-eigenvalue bottleneck, but still small enough that between-regime separation produces a visibly stronger contrast than within-regime heterogeneity.
QML-breakpoint detection in factor models
In this section we apply the abstract framework of Section 2 to high-dimensional data. The data is generated by a factor model whose loadings switch at the breakpoint (Definition 28 below). The detection strategy consists of a single preprocessing step: we take the leading q q q PCA scores of the p p p -dimensional observations and rescale them by p − 1 / 2 p^{-1/2} p − 1/2 , obtaining a q q q -dimensional signal series to which the framework of Section 2 applies.
The key structural hypothesis is that the factors are pervasive — each moves a macroscopic proportion of the coordinates, so the signal eigenvalues grow like the dimension p p p (see Remark 30 below) — while the error term only needs to be of lower order o ( p ) o(p) o ( p ) and may otherwise be completely general; the order-p p p spectral gap then lets the Davis–Kahan inequality align the estimated projection with the pervasive directions. The main results are Proposition 31, which verifies the assumptions of the abstract framework for the projected series — vanishing within-regime noise, positive between-regime contrast — and Corollary 32, which concludes consistency of the untrimmed QML breakpoint estimator for every ridge ε > 0 \varepsilon> 0 ε > 0 . Throughout we work in the high-dimensional high-sample-size (HDHSS) regime: p = p n → ∞ p = p_n \to\infty p = p n → ∞ jointly with n n n , while the numbers of factors stay fixed.
Let us start with the definition of a factor model and what it means to have a regime change at the breakpoint τ \tau τ .
Definition 28 (Regime-switching pervasive factor model). Let p = p n p = p_n p = p n denote the dimension of the observations, allowed to depend on the number n n n of observations. Let X = ( x t ) 1 ≤ t ≤ n ∈ R p × n X = (x_t)_{1 \le t \le n} \in\mathbb{R}^{p \times n} X = ( x t ) 1 ≤ t ≤ n ∈ R p × n denote the observed data. We say that X X X follows a regime-switching pervasive factor model with breakpoint τ = ⌊ θ n ⌋ \tau= \lfloor\theta n \rfloor τ = ⌊ θ n ⌋ , for some fixed θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) , if
x t = { Λ ≤ f t + e t , 1 ≤ t ≤ τ , Λ > f t + e t , τ < t ≤ n , (16) x_t =
\begin{cases}
\Lambda_{\leq} f_t + e_t, & 1 \leq t \leq\tau, \\
\Lambda_{>} f_t + e_t, & \tau< t \leq n,
\end{cases} \tag*{(16)} x t = { Λ ≤ f t + e t , Λ > f t + e t , 1 ≤ t ≤ τ , τ < t ≤ n , (16)
where:
Λ ≤ = Λ ≤ , n ∈ R p × q ≤ \Lambda_{\leq} = \Lambda_{\leq,n} \in\mathbb{R}^{p \times q_{\leq}} Λ ≤ = Λ ≤ , n ∈ R p × q ≤ and Λ > = Λ > , n ∈ R p × q > \Lambda_{>} = \Lambda_{>,n} \in\mathbb{R}^{p \times q_{>}} Λ > = Λ > , n ∈ R p × q > are deterministic loading matrices, with q ≤ , q > q_{\leq}, q_{>} q ≤ , q > fixed, not depending on n n n ;
f t f_t f t are the factors, with f t ∈ R q ≤ f_t \in\mathbb{R}^{q_{\leq}} f t ∈ R q ≤ for t ≤ τ t \leq\tau t ≤ τ and f t ∈ R q > f_t \in\mathbb{R}^{q_{>}} f t ∈ R q > for t > τ t > \tau t > τ ;
e t ∈ R p e_t \in\mathbb{R}^{p} e t ∈ R p is a general error vector absorbing all non-pervasive, lower-order structure, e.g. weak factors, idiosyncratic noise, dependence, or correlation with the factors.
We suppress the dependence of p , Λ ∙ , f t , e t p,\Lambda_{\bullet}, f_t, e_t p , Λ ∙ , f t , e t on n n n whenever harmless. For convenience we write s t : = Λ ≤ f t s_t := \Lambda_{\leq}f_t s t := Λ ≤ f t for t ≤ τ t \leq\tau t ≤ τ and s t : = Λ > f t s_t := \Lambda_{>}f_t s t := Λ > f t for t > τ t > \tau t > τ , so that x t = s t + e t x_t = s_t + e_t x t = s t + e t . For a macroscopic interval I = I ( a , b ) I = I(a,b) I = I ( a , b ) we set
S a : b : = 1 ∣ I ∣ ∑ t ∈ I s t s t T , E a : b : = 1 ∣ I ∣ ∑ t ∈ I e t e t T . S_{a:b} := \frac{1}{|I|}\sum_{t \in I}s_ts_t^{\mathsf{T}}, \qquad E_{a:b} := \frac{1}{|I|}\sum_{t \in I}e_te_t^{\mathsf{T}}. S a : b := ∣ I ∣ 1 t ∈ I ∑ s t s t T , E a : b := ∣ I ∣ 1 t ∈ I ∑ e t e t T .
Classical results on QML-based breakpoint estimation in factor models — most prominently Duan, Bai, and Han [17 ] — concern the unregularized objective (ε = 0 \varepsilon= 0 ε = 0 ) under the structural assumptions on the idiosyncratic error e t e_t e t recalled in the introduction, which force a quantitative law of large numbers on the block second moments. In contrast (see Corollary 32 below), we establish consistency allowing e t e_t e t to be completely general , requiring only that its blockwise energy is negligible at the pervasive scale (see (F4) below).
We now describe the signal time series ( y t ) 1 ≤ t ≤ n (y_t)_{1 \leq t \leq n} ( y t ) 1 ≤ t ≤ n used for the detection of the breakpoint. As we cannot rely on domain knowledge for generic data, we define the signal space in a data-driven way (see also Remark 3). We fix a working rank q q q , theoretically the number of pervasive directions (see Assumption (F1) below), chosen in practice by thresholding the singular values (cf. Remark 34). Let U ^ ∈ R p × q \widehat{U} \in\mathbb{R}^{p \times q} U ∈ R p × q denote the matrix whose columns are the leading q q q eigenvectors of the empirical second moment matrix of the full dataset
1 n ∑ t = 1 n x t x t T ∈ R p × p . \frac{1}{n}\sum_{t=1}^{n}x_tx_t^{\mathsf{T}} \in\mathbb{R}^{p \times p}. n 1 t = 1 ∑ n x t x t T ∈ R p × p .
We define the p − 1 / 2 p^{-1/2} p − 1/2 -rescaled PCA-score time series by
y t : = 1 p U ^ T x t ∈ R q , 1 ≤ t ≤ n . (17) y_t := \frac{1}{\sqrt{p}}\widehat{U}^{\mathsf{T}}x_t \in\mathbb{R}^{q}, \qquad1 \le t \le n. \tag*{(17)} y t := p 1 U T x t ∈ R q , 1 ≤ t ≤ n . (17)
The entries of y t y_t y t are the coordinates of the orthogonal projection U ^ U ^ T x t \widehat{U}\widehat{U}^{\mathsf{T}}x_t U U T x t in the basis U ^ \widehat{U} U .
For a macroscopic interval I = I ( a , b ) I=I(a,b) I = I ( a , b ) the empirical second moment of the rescaled PCA scores y t y_t y t on I I I is then defined as in Definition 5 as
Σ ^ a : b : = 1 ∣ I ∣ ∑ t ∈ I y t y t T ∈ R q × q . (18) \widehat{\Sigma}_{a:b}:=\frac{1}{|I|}\sum_{t\in I}y_ty_t^{\mathsf{T}}\in\mathbb{R}^{q\times q}. \tag*{(18)} Σ a : b := ∣ I ∣ 1 t ∈ I ∑ y t y t T ∈ R q × q . (18)
The normalization by 1 / p 1/p 1/ p makes Σ ^ a : b \widehat{\Sigma}_{a:b} Σ a : b an order-one q × q q\times q q × q covariance (Remark 30). The PCA step serves two purposes:
Dimension reduction: going from p p p dimensions to the fixed dimension q q q makes the asymptotic analysis feasible even as p = p n → ∞ p=p_n\to\infty p = p n → ∞ ;
Denoising: replacing x t x_t x t by its orthogonal projection U ^ U ^ T x t \widehat{U}\widehat{U}^{\mathsf{T}}x_t U U T x t onto the leading left-singular eigenspace of X X X is a standard denoising device; the scores U ^ T x t \widehat{U}^{\mathsf{T}}x_t U T x t describe the data in its natural feature coordinates; this is the smoothing displayed in Figure 4 .
We now collect the asymptotic conditions needed to detect the breakpoint: pervasiveness, a factor SLLN, the regime change, and an asymptotic control on the error. Stacking the two regimes, the data of a regime-switching model is itself a factor model with loadings [ Λ ≤ , Λ > ] [\Lambda_{\leq},\Lambda_{>}] [ Λ ≤ , Λ > ] ,
X = [ Λ ≤ , Λ > ] [ F ≤ 0 0 F > ] + [ E ≤ , E > ] , X=[\Lambda_{\leq},\Lambda_{>}]\begin{bmatrix}F_{\leq}&0\\0&F_{>}\end{bmatrix}+[E_{\leq},E_{>}], X = [ Λ ≤ , Λ > ] [ F ≤ 0 0 F > ] + [ E ≤ , E > ] ,
which is why the pervasiveness condition (F1) below is formulated for the combined loadings.
Assumption 29. In addition to the model of Definition 28, we assume:
(F1) Pervasive loadings. The combined normalized loading Gram matrix converges to a positive-semidefinite limit Σ Λ ⪰ 0 \Sigma_{\Lambda}\succeq0 Σ Λ ⪰ 0 , i.e.
lim p → ∞ 1 p [ Λ ≤ , Λ > ] T [ Λ ≤ , Λ > ] = Σ Λ ∈ R ( q ≤ + q > ) × ( q ≤ + q > ) . \lim_{p\to\infty}\frac{1}{p}[\Lambda_{\leq},\Lambda_{>}]^{\mathsf{T}}[\Lambda_{\leq},\Lambda_{>}]=\Sigma_{\Lambda}\in\mathbb{R}^{(q_{\leq}+q_{>})\times(q_{\leq}+q_{>})}. p → ∞ lim p 1 [ Λ ≤ , Λ > ] T [ Λ ≤ , Λ > ] = Σ Λ ∈ R ( q ≤ + q > ) × ( q ≤ + q > ) .
We define q : = rank ( Σ Λ ) ≤ q ≤ + q > q:=\operatorname{rank}(\Sigma_{\Lambda})\leq q_{\leq}+q_{>} q := rank ( Σ Λ ) ≤ q ≤ + q > as the pervasive rank. We have a strict inequality exactly when the two regimes share at least one pervasive direction, i.e. a factor unchanged across the breakpoint.
(F2) Factor SLLN. There exist positive definite matrices Σ F ≤ ≻ 0 \Sigma_{F_{\leq}}\succ0 Σ F ≤ ≻ 0 and Σ F > ≻ 0 \Sigma_{F_{>}}\succ0 Σ F > ≻ 0 such that for every macroscopic interval I ( a , b ) I(a,b) I ( a , b ) contained in the first regime and I ( c , d ) I(c,d) I ( c , d ) contained in the second regime
lim n → ∞ 1 ∣ I ( a , b ) ∣ ∑ t ∈ I ( a , b ) f t f t T = Σ F ≤ and lim n → ∞ 1 ∣ I ( c , d ) ∣ ∑ t ∈ I ( c , d ) f t f t T = Σ F > . \lim_{n\to\infty}\frac{1}{|I(a,b)|}\sum_{t\in I(a,b)}f_tf_t^{\mathsf{T}}=\Sigma_{F_{\leq}}\qquad\text{and}\qquad\lim_{n\to\infty}\frac{1}{|I(c,d)|}\sum_{t\in I(c,d)}f_tf_t^{\mathsf{T}}=\Sigma_{F_{>}}. n → ∞ lim ∣ I ( a , b ) ∣ 1 t ∈ I ( a , b ) ∑ f t f t T = Σ F ≤ and n → ∞ lim ∣ I ( c , d ) ∣ 1 t ∈ I ( c , d ) ∑ f t f t T = Σ F > .
(F3) Regime change in the leading part. For ∙ ∈ { ≤ , > } \bullet\in\{\leq,>\} ∙ ∈ { ≤ , > } write Σ ∙ s i g : = Λ ∙ Σ F ∙ Λ ∙ T \Sigma^{\mathrm{sig}}_{\bullet}:=\Lambda_{\bullet}\Sigma_{F_{\bullet}}\Lambda_{\bullet}^{\mathsf{T}} Σ ∙ sig := Λ ∙ Σ F ∙ Λ ∙ T . We assume that the two regimes stay asymptotically separated on the pervasive scale, i.e.
lim inf n → ∞ 1 p n ∥ Σ ≤ s i g − Σ > s i g ∥ F > 0. \liminf_{n\to\infty}\frac{1}{p_n}\left\lVert\Sigma^{\mathrm{sig}}_{\leq}-\Sigma^{\mathrm{sig}}_{>}\right\rVert_{F}>0. n → ∞ lim inf p n 1 Σ ≤ sig − Σ > sig F > 0.
(F4) Negligible error on the pervasive scale. For every fixed macroscopic interval I ( a , b ) I(a,b) I ( a , b ) , 0 ≤ a < b ≤ 1 0 \le a < b \le1 0 ≤ a < b ≤ 1 ,
lim n → ∞ 1 p n ∥ E a : b ∥ = 0. \lim_{n \to\infty} \frac{1}{p_n} \lVert E_{a:b} \rVert= 0. n → ∞ lim p n 1 ∥ E a : b ∥ = 0.
Let us explain the intuition of the pervasive order-p p p scaling of the eigenvalues in Assumption (F1).
Remark 30 (Pervasive scaling in financial factor models). To illustrate pervasive scaling in financial factor models, let us consider the simplest model, where there is only a single market factor with loading vector Λ = ( 1 , … , 1 ) T ∈ R p × 1 \Lambda= (1,\ldots,1)^{\mathsf{T}} \in\mathbb{R}^{p \times1} Λ = ( 1 , … , 1 ) T ∈ R p × 1 . This means that assets move as the market factor plus some idiosyncratic noise. Then
∥ Λ ∥ = sup ∣ f ∣ = 1 ∣ Λ f ∣ = p , ∥ Λ ∥ 2 = p , \lVert\Lambda\rVert= \sup_{\lvert f \rvert= 1} \lvert\Lambda f \rvert= \sqrt{p}, \qquad\lVert\Lambda\rVert^2 = p, ∥ Λ ∥ = ∣ f ∣ = 1 sup ∣ Λ f ∣ = p , ∥ Λ ∥ 2 = p ,
and the covariance contribution Λ Λ T \Lambda\Lambda^{\mathsf{T}} Λ Λ T is the all-ones matrix, which has a single non-zero eigenvalue equal to p p p . Thus a pervasive factor contributes an eigenvalue of order p p p . It is important that all pervasive factors are caught when choosing the dimension q q q , as we need in the proof of Proposition 31 that the associated q q q -eigengap is also of order p p p .
Let us now turn to the main result of this section.
Proposition 31. Consider a regime-switching pervasive factor model (16) satisfying (F1)–(F4), and let ( y t ) (y_t) ( y t ) be the p − 1 / 2 p^{-1/2} p − 1/2 -rescaled PCA-score series (17), with block second moments Σ ^ a : b \widehat{\Sigma}_{a:b} Σ a : b given by (18). Then:
(i) Assumption (A1) holds; that is, the break-anchored spectral ceiling in (2) satisfies M < ∞ M < \infty M < ∞ ;
(ii) the within- and between-regime separations of Section 2 satisfy
Δ w i t h i n = 0 and Δ b e t w e e n = lim inf n → ∞ 1 p n ∥ Σ ^ ≤ s i g − Σ ^ > s i g ∥ F , \Delta_{\mathrm{within}} = 0 \qquad\text{and} \qquad\Delta_{\mathrm{between}} = \liminf_{n \to\infty} \frac{1}{p_n} \left\lVert\widehat{\Sigma}^{\mathrm{sig}}_{\leq} - \widehat{\Sigma}^{\mathrm{sig}}_{>} \right\rVert_F, Δ within = 0 and Δ between = n → ∞ lim inf p n 1 Σ ≤ sig − Σ > sig F ,
and therefore, together with (F3), Assumption (A4) holds for every ε > 0 \varepsilon> 0 ε > 0 .
The proof of Proposition 31 is given in Section 4.2.4. A direct combination with Theorem 23 yields consistency of the QML-breakpoint estimator:
Corollary 32 (HDHSS consistency of QML-breakpoint estimation in pervasive factor models). We consider a regime-switching pervasive factor model (16) satisfying (F1)–(F4). Then, for every ε > 0 \varepsilon> 0 ε > 0 , every untrimmed regularized QML breakpoint estimator θ ^ n ∈ arg min ν ∈ D n J n ( ν ) \widehat{\theta}_n \in\arg\min_{\nu\in\mathcal{D}_n} J_n(\nu) θ n ∈ arg min ν ∈ D n J n ( ν ) satisfies θ ^ n → θ \widehat{\theta}_n \to\theta θ n → θ .
Remark 33 (Beyond infinite SNR). Assumption (F4) makes the error budget vanish at the pervasive scale, and Proposition 31 accordingly lands in the infinite signal-to-noise regime Δ w i t h i n = 0 \Delta_{\mathrm{within}} = 0 Δ within = 0 of the abstract framework. Nothing in the proof is tied to this extreme case: all estimates are quantitative, and a sufficiently small non-vanishing error budget yields a positive Δ w i t h i n \Delta_{\mathrm{within}} Δ within and a reduced Δ b e t w e e n \Delta_{\mathrm{between}} Δ between that still satisfy the finite signal-to-noise condition (A4), so Theorem 23 continues to give consistency; we do not track the explicit constants.
Remark 34 (Estimating the working rank). In practice the pervasive rank q q q *is unknown and is estimated from the data, for example by thresholding the singular values of the data matrix. The message is: overestimation is harmless, underestimation is not. If the estimated rank eventually stabilizes at a value ≥ q \ge q ≥ q , the projected series still contains the consistent q q q -dimensional pervasive component, while the extra directions contribute only o ( 1 ) o(1) o ( 1 ) to the normalized block second moments, by (F4) and the Davis–Kahan estimate; the proof of Proposition 31 goes through unchanged and the regularized QML breakpoint estimator (any ϵ > 0 \epsilon> 0 ϵ > 0 ) remains consistent. If the rank is underestimated, the projection can collapse the two regimes onto each other, and the break may become undetectable.
Proofs
The conceptual motivation for the arguments, including the two competing Jensen gaps and the role of the ridge parameter, was given in Section 2.1. Here we supply the technical details in two parts. Subsection 4.1 collects the auxiliary tools: Lemmas 36–43 provide the boundedness, continuity, and endpoint control needed to pass from pointwise estimates to minimizers of J n J_n J n , while Lemma 44 quantifies the curvature of log det. Subsection 4.2 proves the main results: Theorem 21 gives classical SLLN-based consistency, Theorem 23 gives finite-signal-to-noise consistency, Proposition 24 establishes sharpness of the global threshold, and Proposition 31 verifies the abstract assumptions for the pervasive factor model. Combining the last proposition with Theorem 23 yields the factor-model consistency result, Corollary 32.
Auxiliary results.
Lemma 35 (Block additivity of empirical second moments). Let ( z t ) 1 ≤ t ≤ n (z_t)_{1 \le t \le n} ( z t ) 1 ≤ t ≤ n be vectors in R d \mathbb{R}^d R d and, for a nonempty index set I ⊂ { 1 , … , n } I \subset\{1,\ldots,n\} I ⊂ { 1 , … , n } , set Σ ^ I : = 1 ∣ I ∣ ∑ t ∈ I z t z t T \widehat{\Sigma}_I := \frac{1}{|I|}\sum_{t \in I} z_t z_t^{\mathsf T} Σ I := ∣ I ∣ 1 ∑ t ∈ I z t z t T . If I , J I,J I , J are nonempty and disjoint, then
Σ ^ I ∪ J = ∣ I ∣ ∣ I ∣ + ∣ J ∣ Σ ^ I + ∣ J ∣ ∣ I ∣ + ∣ J ∣ Σ ^ J . (19) \widehat{\Sigma}_{I \cup J} =
\frac{|I|}{|I|+|J|}\widehat{\Sigma}_I+
\frac{|J|}{|I|+|J|}\widehat{\Sigma}_J. \tag*{(19)} Σ I ∪ J = ∣ I ∣ + ∣ J ∣ ∣ I ∣ Σ I + ∣ I ∣ + ∣ J ∣ ∣ J ∣ Σ J . (19)
and consequently
Σ ^ J = ∣ I ∣ + ∣ J ∣ ∣ J ∣ Σ ^ I ∪ J − ∣ I ∣ ∣ J ∣ Σ ^ I , ∥ Σ ^ J − Σ ^ I ∪ J ∥ ≤ ∣ I ∣ ∣ J ∣ ( ∥ Σ ^ I ∪ J ∥ + ∥ Σ ^ I ∥ ) . (20) \widehat{\Sigma}_J =
\frac{|I|+|J|}{|J|}\widehat{\Sigma}_{I \cup J}-
\frac{|I|}{|J|}\widehat{\Sigma}_I,
\qquad
\left\|\widehat{\Sigma}_J-\widehat{\Sigma}_{I \cup J}\right\|
\le
\frac{|I|}{|J|}
\left(\left\|\widehat{\Sigma}_{I \cup J}\right\|+
\left\|\widehat{\Sigma}_I\right\|\right). \tag*{(20)} Σ J = ∣ J ∣ ∣ I ∣ + ∣ J ∣ Σ I ∪ J − ∣ J ∣ ∣ I ∣ Σ I , Σ J − Σ I ∪ J ≤ ∣ J ∣ ∣ I ∣ ( Σ I ∪ J + Σ I ) . (20)
Proof. Multiplying out, ( ∣ I ∣ + ∣ J ∣ ) Σ ^ I ∪ J = ∑ t ∈ I ∪ J z t z t T = ∣ I ∣ Σ ^ I + ∣ J ∣ Σ ^ J (|I|+|J|)\widehat{\Sigma}_{I \cup J}=\sum_{t \in I \cup J}z_tz_t^{\mathsf T}=|I|\widehat{\Sigma}_I+|J|\widehat{\Sigma}_J ( ∣ I ∣ + ∣ J ∣ ) Σ I ∪ J = ∑ t ∈ I ∪ J z t z t T = ∣ I ∣ Σ I + ∣ J ∣ Σ J , which is (19). Solving for Σ ^ J \widehat{\Sigma}_J Σ J gives the identity in (20); subtracting Σ ^ I ∪ J \widehat{\Sigma}_{I \cup J} Σ I ∪ J from it yields Σ ^ J − Σ ^ I ∪ J = ∣ I ∣ ∣ J ∣ ( Σ ^ I ∪ J − Σ ^ I ) \widehat{\Sigma}_J-\widehat{\Sigma}_{I \cup J}=\frac{|I|}{|J|}(\widehat{\Sigma}_{I \cup J}-\widehat{\Sigma}_I) Σ J − Σ I ∪ J = ∣ J ∣ ∣ I ∣ ( Σ I ∪ J − Σ I ) , and the triangle inequality gives the norm bound. □
Lemma 36 (Bounds on empirical second moments). If 0 ≤ a < b ≤ 1 0 \le a < b \le1 0 ≤ a < b ≤ 1 , then
lim sup n → ∞ ∥ Σ ^ a : b ∥ ≤ Δ w i t h i n + M . (21) \limsup_{n \to\infty}\left\|\widehat{\Sigma}_{a:b}\right\|\le\Delta_{\mathrm{within}}+M. \tag*{(21)} n → ∞ lim sup Σ a : b ≤ Δ within + M . (21)
Proof of Lemma 36. We first prove (21) for 0 ≤ a < b ≤ θ 0 \le a < b \le\theta 0 ≤ a < b ≤ θ and for θ ≤ a < b ≤ 1 \theta\le a < b \le1 θ ≤ a < b ≤ 1 . The general statement for 0 ≤ a < θ < b ≤ 1 0 \le a < \theta< b \le1 0 ≤ a < θ < b ≤ 1 follows then from the triangle inequality and Lemma 35, applied to I ( a , θ ) I(a,\theta) I ( a , θ ) and I ( θ , b ) I(\theta,b) I ( θ , b ) :
Σ ^ a : b = ⌊ θ n ⌋ − ⌊ a n ⌋ ⌊ b n ⌋ − ⌊ a n ⌋ Σ ^ a : θ + ⌊ b n ⌋ − ⌊ θ n ⌋ ⌊ b n ⌋ − ⌊ a n ⌋ Σ ^ θ : b . \widehat{\Sigma}_{a:b}
=
\frac{\lfloor\theta n\rfloor-\lfloor an\rfloor}{\lfloor bn\rfloor-\lfloor an\rfloor}\widehat{\Sigma}_{a:\theta}
+
\frac{\lfloor bn\rfloor-\lfloor\theta n\rfloor}{\lfloor bn\rfloor-\lfloor an\rfloor}\widehat{\Sigma}_{\theta:b}. Σ a : b = ⌊ bn ⌋ − ⌊ an ⌋ ⌊ θ n ⌋ − ⌊ an ⌋ Σ a : θ + ⌊ bn ⌋ − ⌊ an ⌋ ⌊ bn ⌋ − ⌊ θ n ⌋ Σ θ : b .
Hence, let us only consider the case 0 ≤ a < b ≤ θ 0 \le a < b \le\theta 0 ≤ a < b ≤ θ . We observe that
∥ Σ ^ a : b ∥ ≤ ∥ Σ ^ a : b − Σ ^ 0 : θ ∥ + ∥ Σ ^ 0 : θ ∥ . \left\|\widehat{\Sigma}_{a:b}\right\|
\le
\left\|\widehat{\Sigma}_{a:b}-\widehat{\Sigma}_{0:\theta}\right\|
+
\left\|\widehat{\Sigma}_{0:\theta}\right\|. Σ a : b ≤ Σ a : b − Σ 0 : θ + Σ 0 : θ .
from which the desired inequality (21) follows by unwinding the involved definitions.
Lemma 37 (Modulus of continuity for empirical second moments). For all 0 < ν 1 < ν 2 < 1 0 < \nu_1 < \nu_2 < 1 0 < ν 1 < ν 2 < 1 ,
lim sup n → ∞ ∥ Σ ^ 0 : ν 2 − Σ ^ 0 : ν 1 ∥ ≤ 2 ν 2 ( Δ w i t h i n + M ) ∣ ν 2 − ν 1 ∣ , \limsup_{n\to\infty}\left\|\widehat{\Sigma}_{0:\nu_2}-\widehat{\Sigma}_{0:\nu_1}\right\| \le\frac{2}{\nu_2}\left(\Delta_{\mathrm{within}}+M\right)|\nu_2-\nu_1|, n → ∞ lim sup Σ 0 : ν 2 − Σ 0 : ν 1 ≤ ν 2 2 ( Δ within + M ) ∣ ν 2 − ν 1 ∣ ,
and
lim sup n → ∞ ∥ Σ ^ ν 2 : 1 − Σ ^ ν 1 : 1 ∥ ≤ 2 1 − ν 1 ( Δ w i t h i n + M ) ∣ ν 2 − ν 1 ∣ . \limsup_{n\to\infty}\left\|\widehat{\Sigma}_{\nu_2:1}-\widehat{\Sigma}_{\nu_1:1}\right\| \le\frac{2}{1-\nu_1}\left(\Delta_{\mathrm{within}}+M\right)|\nu_2-\nu_1|. n → ∞ lim sup Σ ν 2 : 1 − Σ ν 1 : 1 ≤ 1 − ν 1 2 ( Δ within + M ) ∣ ν 2 − ν 1 ∣.
Proof of Lemma 37. We only consider ∥ Σ ^ 0 : ν 2 − Σ ^ 0 : ν 1 ∥ \left\|\widehat{\Sigma}_{0:\nu_2}-\widehat{\Sigma}_{0:\nu_1}\right\| Σ 0 : ν 2 − Σ 0 : ν 1 as the term ∥ Σ ^ ν 2 : 1 − Σ ^ ν 1 : 1 ∥ \left\|\widehat{\Sigma}_{\nu_2:1}-\widehat{\Sigma}_{\nu_1:1}\right\| Σ ν 2 : 1 − Σ ν 1 : 1 can be estimated in the same way. Set k i : = ⌊ ν i n ⌋ k_i:=\lfloor\nu_i n\rfloor k i := ⌊ ν i n ⌋ , i = 1 , 2 i=1,2 i = 1 , 2 . For all large enough n n n , 1 ≤ k 1 < k 2 ≤ n − 1 1\le k_1<k_2\le n-1 1 ≤ k 1 < k 2 ≤ n − 1 , and (19) gives
Σ ^ 0 : ν 2 = k 1 k 2 Σ ^ 0 : ν 1 + k 2 − k 1 k 2 Σ ^ ν 1 : ν 2 . \widehat{\Sigma}_{0:\nu_2}=\frac{k_1}{k_2}\widehat{\Sigma}_{0:\nu_1}+\frac{k_2-k_1}{k_2}\widehat{\Sigma}_{\nu_1:\nu_2}. Σ 0 : ν 2 = k 2 k 1 Σ 0 : ν 1 + k 2 k 2 − k 1 Σ ν 1 : ν 2 .
Hence
Σ ^ 0 : ν 2 − Σ ^ 0 : ν 1 = k 2 − k 1 k 2 ( Σ ^ ν 1 : ν 2 − Σ ^ 0 : ν 1 ) . \widehat{\Sigma}_{0:\nu_2}-\widehat{\Sigma}_{0:\nu_1}=\frac{k_2-k_1}{k_2}\left(\widehat{\Sigma}_{\nu_1:\nu_2}-\widehat{\Sigma}_{0:\nu_1}\right). Σ 0 : ν 2 − Σ 0 : ν 1 = k 2 k 2 − k 1 ( Σ ν 1 : ν 2 − Σ 0 : ν 1 ) .
Taking operator norms and using the triangle inequality gives
∥ Σ ^ 0 : ν 2 − Σ ^ 0 : ν 1 ∥ ≤ k 2 − k 1 k 2 ( ∥ Σ ^ 0 : ν 1 ∥ + ∥ Σ ^ ν 1 : ν 2 ∥ ) . \left\|\widehat{\Sigma}_{0:\nu_2}-\widehat{\Sigma}_{0:\nu_1}\right\|\le\frac{k_2-k_1}{k_2}\left(\left\|\widehat{\Sigma}_{0:\nu_1}\right\|+\left\|\widehat{\Sigma}_{\nu_1:\nu_2}\right\|\right). Σ 0 : ν 2 − Σ 0 : ν 1 ≤ k 2 k 2 − k 1 ( Σ 0 : ν 1 + Σ ν 1 : ν 2 ) .
Since ( k 2 − k 1 ) / k 2 → ( ν 2 − ν 1 ) / ν 2 (k_2-k_1)/k_2\to(\nu_2-\nu_1)/\nu_2 ( k 2 − k 1 ) / k 2 → ( ν 2 − ν 1 ) / ν 2 , applying Lemma 36 yields the desired estimate.
Lemma 38 (Vanishing energy in short windows). *Assume (A1) and Δ w i t h i n < ∞ \Delta_{\mathrm{within}}<\infty Δ within < ∞ . Then, for every fixed ν ˉ ∈ [ 0 , 1 ] \bar{\nu}\in[0,1] ν ˉ ∈ [ 0 , 1 ] and ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) , setting a ρ : = ( ν ˉ − ρ ) ∨ 0 a_\rho:=(\bar{\nu}-\rho)\vee0 a ρ := ( ν ˉ − ρ ) ∨ 0 and b ρ : = ( ν ˉ + ρ ) ∧ 1 b_\rho:=(\bar{\nu}+\rho)\wedge1 b ρ := ( ν ˉ + ρ ) ∧ 1 ,
lim sup n → ∞ 1 n ∑ t = ⌊ a ρ n ⌋ + 1 ⌊ b ρ n ⌋ ∥ y t ∥ 2 ≤ 2 ρ q ( M + Δ w i t h i n ) . (22) \limsup_{n\to\infty}\frac{1}{n}\sum_{t=\lfloor a_\rho n\rfloor+1}^{\lfloor b_\rho n\rfloor}\|y_t\|^2\le2\rho q\left(M+\Delta_{\mathrm{within}}\right). \tag*{(22)} n → ∞ lim sup n 1 t = ⌊ a ρ n ⌋ + 1 ∑ ⌊ b ρ n ⌋ ∥ y t ∥ 2 ≤ 2 ρq ( M + Δ within ) . (22)
In particular, combining the windows at ν ˉ = 0 \bar{\nu}=0 ν ˉ = 0 and ν ˉ = 1 \bar{\nu}=1 ν ˉ = 1 and sending ρ ↓ 0 \rho\downarrow0 ρ ↓ 0 ,
lim ρ ↓ 0 lim sup n → ∞ 1 n ( ∑ t = 1 ⌊ ρ n ⌋ ∥ y t ∥ 2 + ∑ t = n − ⌊ ρ n ⌋ + 1 n ∥ y t ∥ 2 ) = 0. (23) \lim_{\rho\downarrow0}\limsup_{n\to\infty}\frac{1}{n}\left(\sum_{t=1}^{\lfloor\rho n\rfloor}\|y_t\|^2+\sum_{t=n-\lfloor\rho n\rfloor+1}^{n}\|y_t\|^2\right)=0. \tag*{(23)} ρ ↓ 0 lim n → ∞ lim sup n 1 t = 1 ∑ ⌊ ρ n ⌋ ∥ y t ∥ 2 + t = n − ⌊ ρ n ⌋ + 1 ∑ n ∥ y t ∥ 2 = 0. (23)
Consequently, if k n , ℓ n ∈ { 0 , … , n } k_n,\ell_n\in\{0,\ldots,n\} k n , ℓ n ∈ { 0 , … , n } satisfy k n / n → 0 k_n/n\to0 k n / n → 0 and ℓ n / n → 0 \ell_n/n\to0 ℓ n / n → 0 , then
1 n ( ∑ t = 1 k n ∥ y t ∥ 2 + ∑ t = n − ℓ n + 1 n ∥ y t ∥ 2 ) ⟶ 0. (24) \frac{1}{n}\left(\sum_{t=1}^{k_n}\|y_t\|^2+\sum_{t=n-\ell_n+1}^{n}\|y_t\|^2\right)\longrightarrow0. \tag*{(24)} n 1 ( t = 1 ∑ k n ∥ y t ∥ 2 + t = n − ℓ n + 1 ∑ n ∥ y t ∥ 2 ) ⟶ 0. (24)
Proof of Lemma 38. Fix ν ˉ ∈ [ 0 , 1 ] \bar{\nu}\in[0,1] ν ˉ ∈ [ 0 , 1 ] and ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) . By construction 0 ≤ a ρ < b ρ ≤ 1 0\le a_\rho<b_\rho\le1 0 ≤ a ρ < b ρ ≤ 1 and b ρ − a ρ ≤ 2 ρ b_\rho-a_\rho\le2\rho b ρ − a ρ ≤ 2 ρ . Hence
1 n ∑ t = ⌊ a ρ n ⌋ + 1 ⌊ b ρ n ⌋ ∥ y t ∥ 2 = ⌊ b ρ n ⌋ − ⌊ a ρ n ⌋ n Tr Σ ^ a ρ : b ρ ≤ ( 2 ρ + 1 n ) q ∥ Σ ^ a ρ : b ρ ∥ . \frac{1}{n}\sum_{t=\lfloor a_\rho n\rfloor+1}^{\lfloor b_\rho n\rfloor}\|y_t\|^2=\frac{\lfloor b_\rho n\rfloor-\lfloor a_\rho n\rfloor}{n}\operatorname{Tr}\widehat{\Sigma}_{a_\rho:b_\rho}\le\left(2\rho+\frac{1}{n}\right)q\left\|\widehat{\Sigma}_{a_\rho:b_\rho}\right\|. n 1 t = ⌊ a ρ n ⌋ + 1 ∑ ⌊ b ρ n ⌋ ∥ y t ∥ 2 = n ⌊ b ρ n ⌋ − ⌊ a ρ n ⌋ Tr Σ a ρ : b ρ ≤ ( 2 ρ + n 1 ) q Σ a ρ : b ρ .
Lemma 36 yields
lim sup n → ∞ 1 n ∑ t = ⌊ a ρ n ⌋ + 1 ⌊ b ρ n ⌋ ∥ y t ∥ 2 ≤ 2 ρ q ( M + Δ w i t h i n ) . \limsup_{n\to\infty}\frac{1}{n}\sum_{t=\lfloor a\rho n\rfloor+1}^{\lfloor b\rho n\rfloor}\lVert y_t\rVert^2 \le2\rho q\left(M+\Delta_{\mathrm{within}}\right). n → ∞ lim sup n 1 t = ⌊ a ρ n ⌋ + 1 ∑ ⌊ b ρ n ⌋ ∥ y t ∥ 2 ≤ 2 ρq ( M + Δ within ) .
which is (22) . For (23) it suffices to note that the two boundary index sets are contained in the windows of (22) with ν ˉ = 0 \bar{\nu}=0 ν ˉ = 0 and ν ˉ = 1 \bar{\nu}=1 ν ˉ = 1 , respectively, and to send ρ ↓ 0 \rho\downarrow0 ρ ↓ 0 . Finally, for every fixed ρ > 0 \rho>0 ρ > 0 , the moving boundary sets in the sequential conclusion above are contained in those same fixed ρ \rho ρ -windows for all large enough n n n . Taking lim sup n \limsup_n lim sup n and then sending ρ ↓ 0 \rho\downarrow0 ρ ↓ 0 proves that conclusion.
Lemma 39 (Short endpoint windows under ridge regularization). Assume (A1), Δ w i t h i n < ∞ \Delta_{\mathrm{within}}<\infty Δ within < ∞ , and ε > 0 \varepsilon>0 ε > 0 . If ν n ∈ D n \nu_n\in\mathcal{D}_n ν n ∈ D n and ν n → 0 \nu_n\to0 ν n → 0 , then
ν n ∣ log det ( Σ ^ 0 : ν n + ε I q ) ∣ ⟶ 0 ; \nu_n\left|\log\det\left(\widehat{\Sigma}_{0:\nu_n}+\varepsilon I_q\right)\right|\longrightarrow0; ν n log det ( Σ 0 : ν n + ε I q ) ⟶ 0 ;
if ν n ∈ D n \nu_n\in\mathcal{D}_n ν n ∈ D n and ν n → 1 \nu_n\to1 ν n → 1 , then
( 1 − ν n ) ∣ log det ( Σ ^ ν n : 1 + ε I q ) ∣ ⟶ 0. (1-\nu_n)\left|\log\det\left(\widehat{\Sigma}_{\nu_n:1}+\varepsilon I_q\right)\right|\longrightarrow0. ( 1 − ν n ) log det ( Σ ν n : 1 + ε I q ) ⟶ 0.
Proof of Lemma 39. We prove the left endpoint statement; the right endpoint is identical. Write ν n = k n / n \nu_n=k_n/n ν n = k n / n with 1 ≤ k n ≤ n − 1 1\le k_n\le n-1 1 ≤ k n ≤ n − 1 . The lower spectral bound Σ ^ 0 : ν n + ε I q ⪰ ε I q \widehat{\Sigma}_{0:\nu_n}+\varepsilon I_q\succeq\varepsilon I_q Σ 0 : ν n + ε I q ⪰ ε I q gives
ν n log det ( Σ ^ 0 : ν n + ε I q ) ≥ q ν n log ε ⟶ 0. \nu_n\log\det\left(\widehat{\Sigma}_{0:\nu_n}+\varepsilon I_q\right)\ge q\nu_n\log\varepsilon\longrightarrow0. ν n log det ( Σ 0 : ν n + ε I q ) ≥ q ν n log ε ⟶ 0.
For the upper bound, equation (24) of Lemma 38 gives
e n : = 1 n ∑ t = 1 k n ∥ y t ∥ 2 = ν n Tr Σ ^ 0 : ν n ⟶ 0. e_n:=\frac{1}{n}\sum_{t=1}^{k_n}\lVert y_t\rVert^2=\nu_n\operatorname{Tr}\widehat{\Sigma}_{0:\nu_n}\longrightarrow0. e n := n 1 t = 1 ∑ k n ∥ y t ∥ 2 = ν n Tr Σ 0 : ν n ⟶ 0.
Overestimating the eigenvalues by the trace, we get
log det ( Σ ^ 0 : ν n + ε I q ) ≤ q log ( ε + Tr Σ ^ 0 : ν n ) = q log ( ε + e n ν n ) . \log\det\left(\widehat{\Sigma}_{0:\nu_n}+\varepsilon I_q\right)\le q\log\left(\varepsilon+\operatorname{Tr}\widehat{\Sigma}_{0:\nu_n}\right)=q\log\left(\varepsilon+\frac{e_n}{\nu_n}\right). log det ( Σ 0 : ν n + ε I q ) ≤ q log ( ε + Tr Σ 0 : ν n ) = q log ( ε + ν n e n ) .
Overestimating log ( ε + e n ν n ) \log\left(\varepsilon+\frac{e_n}{\nu_n}\right) log ( ε + ν n e n ) by linearization at ε \varepsilon ε shows that the product ν n log ( ε + e n / ν n ) \nu_n\log(\varepsilon+e_n/\nu_n) ν n log ( ε + e n / ν n ) tends to zero since e n → 0 e_n\to0 e n → 0 . This proves the claim. □
Lemma 40 (Lipschitz continuity of log det \log\det log det ). Let us consider symmetric positive definite matrices A , B ∈ R q × q A,B\in\mathbb{R}^{q\times q} A , B ∈ R q × q and define
C ( t ) : = ( 1 − t ) A + t B , for t ∈ [ 0 , 1 ] . C(t):=(1-t)A+tB,\qquad\text{for }t\in[0,1]. C ( t ) := ( 1 − t ) A + tB , for t ∈ [ 0 , 1 ] .
Assume that sup t ∈ [ 0 , 1 ] ∥ C ( t ) − 1 ∥ < ∞ \sup_{t\in[0,1]}\lVert C(t)^{-1}\rVert<\infty sup t ∈ [ 0 , 1 ] ∥ C ( t ) − 1 ∥ < ∞ . Then
∣ log det ( A ) − log det ( B ) ∣ ≤ q sup t ∈ [ 0 , 1 ] ∥ C ( t ) − 1 ∥ ∥ A − B ∥ F ≤ q sup t ∈ [ 0 , 1 ] ∥ C ( t ) − 1 ∥ ∥ A − B ∥ . (25) \left|\log\det(A)-\log\det(B)\right|
\le\sqrt{q}\sup_{t\in[0,1]}\lVert C(t)^{-1}\rVert\lVert A-B\rVert_F
\le q\sup_{t\in[0,1]}\lVert C(t)^{-1}\rVert\lVert A-B\rVert. \tag*{(25)} ∣ log det ( A ) − log det ( B ) ∣ ≤ q t ∈ [ 0 , 1 ] sup ∥ C ( t ) − 1 ∥ ∥ A − B ∥ F ≤ q t ∈ [ 0 , 1 ] sup ∥ C ( t ) − 1 ∥ ∥ A − B ∥ . (25)
Proof of Lemma 40. Set ϕ ( t ) : = log det C ( t ) \phi(t):=\log\det C(t) ϕ ( t ) := log det C ( t ) . Jacobi’s differential identity d ( log det X ) = Tr ( X − 1 d X ) \mathrm{d}(\log\det X)=\operatorname{Tr}(X^{-1}\mathrm{d}X) d ( log det X ) = Tr ( X − 1 d X ) (see e.g. [11, Section A.4.1]) gives ϕ ′ ( t ) = Tr ( C ( t ) − 1 ( B − A ) ) \phi'(t)=\operatorname{Tr}(C(t)^{-1}(B-A)) ϕ ′ ( t ) = Tr ( C ( t ) − 1 ( B − A )) , so by the fundamental theorem of calculus
log det ( A ) − log det ( B ) = ϕ ( 0 ) − ϕ ( 1 ) = ∫ 0 1 Tr ( C ( t ) − 1 ( A − B ) ) d t . \log\det(A)-\log\det(B)=\phi(0)-\phi(1)=\int_0^1\operatorname{Tr}(C(t)^{-1}(A-B))\,\mathrm{d}t. log det ( A ) − log det ( B ) = ϕ ( 0 ) − ϕ ( 1 ) = ∫ 0 1 Tr ( C ( t ) − 1 ( A − B )) d t .
Bounding the integrand by Cauchy–Schwarz for the Frobenius inner product, ∣ Tr ( C ( t ) − 1 ( A − B ) ) ∣ ≤ ∥ C ( t ) − 1 ∥ F ∥ A − B ∥ F ≤ q sup t ∈ [ 0 , 1 ] ∥ C ( t ) − 1 ∥ ∥ A − B ∥ F \left|\operatorname{Tr}(C(t)^{-1}(A-B))\right|\le\lVert C(t)^{-1}\rVert_F\lVert A-B\rVert_F\le\sqrt{q}\sup_{t\in[0,1]}\lVert C(t)^{-1}\rVert\lVert A-B\rVert_F Tr ( C ( t ) − 1 ( A − B )) ≤ ∥ C ( t ) − 1 ∥ F ∥ A − B ∥ F ≤ q sup t ∈ [ 0 , 1 ] ∥ C ( t ) − 1 ∥ ∥ A − B ∥ F , proves the first inequality of (25) ; the second follows from ∥ A − B ∥ F ≤ q ∥ A − B ∥ \lVert A-B\rVert_F\le\sqrt{q}\lVert A-B\rVert ∥ A − B ∥ F ≤ q ∥ A − B ∥ . □
Lemma 41 (Sequential stability of the QML objective). Assume (A1), Δ w i t h i n < ∞ \Delta_{\mathrm{within}} < \infty Δ within < ∞ , and that either ε > 0 \varepsilon> 0 ε > 0 or (PD) holds. If ν n → ν ˉ ∈ ( 0 , 1 ) \nu_n \to\bar{\nu} \in(0,1) ν n → ν ˉ ∈ ( 0 , 1 ) , then J n ( ν n ) − J n ( ν ˉ ) ⟶ 0 J_n(\nu_n)-J_n(\bar{\nu}) \longrightarrow0 J n ( ν n ) − J n ( ν ˉ ) ⟶ 0 .
Proof of Lemma 41. Set
k n : = ⌊ ν n n ⌋ , k ˉ n : = ⌊ ν ˉ n ⌋ , r n : = k n ∧ k ˉ n , s n : = k n ∨ k ˉ n . k_n := \lfloor\nu_n n \rfloor,\qquad\bar{k}_n := \lfloor\bar{\nu}n \rfloor,\qquad r_n := k_n \wedge\bar{k}_n,\qquad s_n := k_n \vee\bar{k}_n. k n := ⌊ ν n n ⌋ , k ˉ n := ⌊ ν ˉ n ⌋ , r n := k n ∧ k ˉ n , s n := k n ∨ k ˉ n .
For all large enough n n n , 1 ≤ r n ≤ s n ≤ n − 1 1 \le r_n \le s_n \le n-1 1 ≤ r n ≤ s n ≤ n − 1 . We first compare the two prefix blocks. Lemma 35, applied to the shorter prefix and the intervening window { r n + 1 , … , s n } \{r_n+1,\ldots,s_n\} { r n + 1 , … , s n } (identity (19) if k n ≥ k ˉ n k_n \ge\bar{k}_n k n ≥ k ˉ n , identity (20) otherwise), gives the common bound
∥ Σ ^ 0 : ν n − Σ ^ 0 : ν ˉ ∥ ≤ ∣ k n − k ˉ n ∣ r n ∥ Σ ^ 0 : ν ˉ ∥ + 1 r n ∑ t = r n + 1 s n ∥ y t ∥ 2 . \left\|\widehat{\Sigma}_{0:\nu_n}-\widehat{\Sigma}_{0:\bar{\nu}}\right\| \le\frac{|k_n-\bar{k}_n|}{r_n}\left\|\widehat{\Sigma}_{0:\bar{\nu}}\right\|+\frac{1}{r_n}\sum_{t=r_n+1}^{s_n}\|y_t\|^2. Σ 0 : ν n − Σ 0 : ν ˉ ≤ r n ∣ k n − k ˉ n ∣ Σ 0 : ν ˉ + r n 1 t = r n + 1 ∑ s n ∥ y t ∥ 2 .
Here r n / n → ν ˉ r_n/n \to\bar{\nu} r n / n → ν ˉ and ∣ k n − k ˉ n ∣ / n → 0 |k_n-\bar{k}_n|/n \to0 ∣ k n − k ˉ n ∣/ n → 0 , while Lemma 36 bounds the fixed-block norm. Moreover, for every fixed ρ > 0 \rho> 0 ρ > 0 , the window { r n + 1 , … , s n } \{r_n+1,\ldots,s_n\} { r n + 1 , … , s n } is contained, for all large n n n , in the ρ \rho ρ -window around ν ˉ \bar{\nu} ν ˉ . Hence (22) and then ρ ↓ 0 \rho\downarrow0 ρ ↓ 0 show that the second term also tends to zero. Therefore
∥ Σ ^ 0 : ν n − Σ ^ 0 : ν ˉ ∥ ⟶ 0. \left\|\widehat{\Sigma}_{0:\nu_n}-\widehat{\Sigma}_{0:\bar{\nu}}\right\| \longrightarrow0. Σ 0 : ν n − Σ 0 : ν ˉ ⟶ 0.
The same argument applied to suffix blocks gives
∥ Σ ^ ν n : 1 − Σ ^ ν ˉ : 1 ∥ ⟶ 0. \left\|\widehat{\Sigma}_{\nu_n:1}-\widehat{\Sigma}_{\bar{\nu}:1}\right\| \longrightarrow0. Σ ν n : 1 − Σ ν ˉ : 1 ⟶ 0.
By Definition 5, each of Σ ^ 0 : ν ˉ \widehat{\Sigma}_{0:\bar{\nu}} Σ 0 : ν ˉ and Σ ^ ν ˉ : 1 \widehat{\Sigma}_{\bar{\nu}:1} Σ ν ˉ : 1 is either a member of I θ c u t \mathcal{I}^{\mathrm{cut}}_{\theta} I θ cut or a convex combination of two members of that family (by (19)). Consequently,
lim inf n → ∞ λ min ( Σ ^ 0 : ν ˉ ) ≥ m , lim inf n → ∞ λ min ( Σ ^ ν ˉ : 1 ) ≥ m . \liminf_{n\to\infty}\lambda_{\min}\left(\widehat{\Sigma}_{0:\bar{\nu}}\right) \ge m,\qquad
\liminf_{n\to\infty}\lambda_{\min}\left(\widehat{\Sigma}_{\bar{\nu}:1}\right) \ge m. n → ∞ lim inf λ m i n ( Σ 0 : ν ˉ ) ≥ m , n → ∞ lim inf λ m i n ( Σ ν ˉ : 1 ) ≥ m .
We now turn these facts into a statement about the QML objectives. Abbreviate
A n : = Σ ^ 0 : ν n + ε I q , A ˉ n : = Σ ^ 0 : ν ˉ + ε I q , B n : = Σ ^ ν n : 1 + ε I q , B ˉ n : = Σ ^ ν ˉ : 1 + ε I q , A_n := \widehat{\Sigma}_{0:\nu_n}+\varepsilon I_q,\qquad\bar{A}_n := \widehat{\Sigma}_{0:\bar{\nu}}+\varepsilon I_q,\qquad B_n := \widehat{\Sigma}_{\nu_n:1}+\varepsilon I_q,\qquad\bar{B}_n := \widehat{\Sigma}_{\bar{\nu}:1}+\varepsilon I_q, A n := Σ 0 : ν n + ε I q , A ˉ n := Σ 0 : ν ˉ + ε I q , B n := Σ ν n : 1 + ε I q , B ˉ n := Σ ν ˉ : 1 + ε I q ,
so that J n ( ν n ) = ν n log det A n + ( 1 − ν n ) log det B n J_n(\nu_n)=\nu_n\log\det A_n+(1-\nu_n)\log\det B_n J n ( ν n ) = ν n log det A n + ( 1 − ν n ) log det B n and J n ( ν ˉ ) = ν ˉ log det A ˉ n + ( 1 − ν ˉ ) log det B ˉ n J_n(\bar{\nu})=\bar{\nu}\log\det\bar{A}_n+(1-\bar{\nu})\log\det\bar{B}_n J n ( ν ˉ ) = ν ˉ log det A ˉ n + ( 1 − ν ˉ ) log det B ˉ n .
Common spectral bounds. By the two displays above, lim inf n λ min ( A ˉ n ) ≥ m + ε \liminf_n\lambda_{\min}(\bar{A}_n)\ge m+\varepsilon lim inf n λ m i n ( A ˉ n ) ≥ m + ε and lim inf n λ min ( B ˉ n ) ≥ m + ε \liminf_n\lambda_{\min}(\bar{B}_n)\ge m+\varepsilon lim inf n λ m i n ( B ˉ n ) ≥ m + ε , where m + ε > 0 m+\varepsilon>0 m + ε > 0 by (PD) if ε = 0 \varepsilon=0 ε = 0 and trivially if ε > 0 \varepsilon>0 ε > 0 . Since ∥ A n − A ˉ n ∥ → 0 \|A_n-\bar{A}_n\|\to0 ∥ A n − A ˉ n ∥ → 0 and ∥ B n − B ˉ n ∥ → 0 \|B_n-\bar{B}_n\|\to0 ∥ B n − B ˉ n ∥ → 0 , the same lower bounds hold for A n A_n A n and B n B_n B n . Lemma 36 gives lim sup n ∥ A ˉ n ∥ ≤ M + Δ w i t h i n + ε \limsup_n\|\bar{A}_n\|\le M+\Delta_{\mathrm{within}}+\varepsilon lim sup n ∥ A ˉ n ∥ ≤ M + Δ within + ε , and the same bound holds for B ˉ n \bar{B}_n B ˉ n , and hence also for A n A_n A n and B n B_n B n . Consequently, for all large enough n n n , every convex combination C ( t ) = ( 1 − t ) A n + t A ˉ n C(t)=(1-t)A_n+t\bar{A}_n C ( t ) = ( 1 − t ) A n + t A ˉ n , t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , satisfies
1 2 ( m + ε ) I q ⪯ C ( t ) ⪯ ( M + Δ w i t h i n + ε + 1 ) I q , \frac{1}{2}(m+\varepsilon)I_q \preceq C(t) \preceq(M+\Delta_{\mathrm{within}}+\varepsilon+1)I_q, 2 1 ( m + ε ) I q ⪯ C ( t ) ⪯ ( M + Δ within + ε + 1 ) I q ,
and the same holds for the pair B n , B ˉ n B_n,\bar{B}_n B n , B ˉ n .
Control of the log-determinants. Lemma 40 with sup t ∈ [ 0 , 1 ] ∥ C ( t ) − 1 ∥ ≤ 2 / ( m + ε ) \sup_{t\in[0,1]}\|C(t)^{-1}\|\le2/(m+\varepsilon) sup t ∈ [ 0 , 1 ] ∥ C ( t ) − 1 ∥ ≤ 2/ ( m + ε ) yields
∣ log det A n − log det A ˉ n ∣ ≤ 2 q m + ε ∥ A n − A ˉ n ∥ ⟶ 0 , |\log\det A_n-\log\det\bar{A}_n|\le\frac{2q}{m+\varepsilon}\|A_n-\bar{A}_n\|\longrightarrow0, ∣ log det A n − log det A ˉ n ∣ ≤ m + ε 2 q ∥ A n − A ˉ n ∥ ⟶ 0 ,
∣ log det B n − log det B ˉ n ∣ ≤ 2 q m + ε ∥ B n − B ˉ n ∥ ⟶ 0 , |\log\det B_n-\log\det\bar{B}_n|\le\frac{2q}{m+\varepsilon}\|B_n-\bar{B}_n\|\longrightarrow0, ∣ log det B n − log det B ˉ n ∣ ≤ m + ε 2 q ∥ B n − B ˉ n ∥ ⟶ 0 ,
while the two-sided spectral bounds give, for all large n n n ,
∣ log det A ˉ n ∣ + ∣ log det B ˉ n ∣ ≤ 2 q max { ∣ log m + ε 2 ∣ , ∣ log ( M + Δ w i t h i n + ε + 1 ) ∣ } = : K < ∞ . \left\lvert\log\det\bar{A}_n \right\rvert+\left\lvert\log\det\bar{B}_n \right\rvert\le2q\max\left\{\left\lvert\log\frac{m+\varepsilon}{2}\right\rvert,\left\lvert\log\left(M+\Delta_{\mathrm{within}}+\varepsilon+1\right)\right\rvert\right\}=:K<\infty. log det A ˉ n + log det B ˉ n ≤ 2 q max { log 2 m + ε , ∣ log ( M + Δ within + ε + 1 ) ∣ } =: K < ∞.
Convergence of the objectives. Writing
J n ( ν n ) − J n ( ν ˉ ) = ν n ( log det A n − log det A ˉ n ) + ( 1 − ν n ) ( log det B n − log det B ˉ n ) J_n(\nu_n)-J_n(\bar{\nu})=\nu_n\left(\log\det A_n-\log\det\bar{A}_n\right)+(1-\nu_n)\left(\log\det B_n-\log\det\bar{B}_n\right) J n ( ν n ) − J n ( ν ˉ ) = ν n ( log det A n − log det A ˉ n ) + ( 1 − ν n ) ( log det B n − log det B ˉ n )
+ ( ν n − ν ˉ ) ( log det A ˉ n − log det B ˉ n ) , \qquad+(\nu_n-\bar{\nu})\left(\log\det\bar{A}_n-\log\det\bar{B}_n\right), + ( ν n − ν ˉ ) ( log det A ˉ n − log det B ˉ n ) ,
the first two terms tend to zero by the previous step, and the last one is bounded by ∣ ν n − ν ˉ ∣ K → 0 \lvert\nu_n-\bar{\nu}\rvert K\to0 ∣ ν n − ν ˉ ∣ K → 0 . This proves the claim.
Lemma 42 (Stability of macroscopic blocks near the endpoints). Assume (A1) and Δ w i t h i n < ∞ \Delta_{\mathrm{within}}<\infty Δ within < ∞ . It holds
lim D n ∋ ν n → 1 ∥ Σ ^ θ : ν n − Σ ^ θ : 1 ∥ = 0 and lim D n ∋ ν n → 0 ∥ Σ ^ ν n : θ − Σ ^ 0 : θ ∥ = 0. \lim_{\mathcal{D}_n\ni\nu_n\to1}\left\lVert\widehat{\Sigma}_{\theta:\nu_n}-\widehat{\Sigma}_{\theta:1}\right\rVert=0
\quad\text{and}\quad
\lim_{\mathcal{D}_n\ni\nu_n\to0}\left\lVert\widehat{\Sigma}_{\nu_n:\theta}-\widehat{\Sigma}_{0:\theta}\right\rVert=0. D n ∋ ν n → 1 lim Σ θ : ν n − Σ θ : 1 = 0 and D n ∋ ν n → 0 lim Σ ν n : θ − Σ 0 : θ = 0.
Consequently, if ε > 0 \varepsilon>0 ε > 0 and ν n → 1 \nu_n\to1 ν n → 1 , then a direct application of Lemma 40 yields
∣ log det ( Σ ^ θ : ν n + ε I q ) − log det ( Σ ^ θ : 1 + ε I q ) ∣ ≤ q ε ∥ Σ ^ θ : ν n − Σ ^ θ : 1 ∥ ⟶ 0 , \left\lvert\log\det\left(\widehat{\Sigma}_{\theta:\nu_n}+\varepsilon I_q\right)-\log\det\left(\widehat{\Sigma}_{\theta:1}+\varepsilon I_q\right)\right\rvert\le\frac{q}{\varepsilon}\left\lVert\widehat{\Sigma}_{\theta:\nu_n}-\widehat{\Sigma}_{\theta:1}\right\rVert\longrightarrow0, log det ( Σ θ : ν n + ε I q ) − log det ( Σ θ : 1 + ε I q ) ≤ ε q Σ θ : ν n − Σ θ : 1 ⟶ 0 ,
and symmetrically, if ε > 0 \varepsilon>0 ε > 0 and ν n → 0 \nu_n\to0 ν n → 0 , the same bound holds for the pair Σ ^ ν n : θ \widehat{\Sigma}_{\nu_n:\theta} Σ ν n : θ , Σ ^ 0 : θ \widehat{\Sigma}_{0:\theta} Σ 0 : θ .
Proof of Lemma 42. We only prove the case ν n → 1 \nu_n\to1 ν n → 1 . For all large enough n n n we have ν n > θ \nu_n>\theta ν n > θ . Let
N n : = n − ⌊ θ n ⌋ , ℓ n : = n − ⌊ ν n n ⌋ , B n : = ∑ t = n − ℓ n + 1 n y t y t T . N_n:=n-\lfloor\theta n\rfloor,\qquad\ell_n:=n-\lfloor\nu_n n\rfloor,\qquad B_n:=\sum_{t=n-\ell_n+1}^{n}y_t y_t^{\mathsf{T}}. N n := n − ⌊ θ n ⌋ , ℓ n := n − ⌊ ν n n ⌋ , B n := t = n − ℓ n + 1 ∑ n y t y t T .
Then ℓ n / N n → 0 \ell_n/N_n\to0 ℓ n / N n → 0 , and by (20) with I = { n − ℓ n + 1 , … , n } I=\{n-\ell_n+1,\ldots,n\} I = { n − ℓ n + 1 , … , n } and J = I ( θ , ν n ) J=I(\theta,\nu_n) J = I ( θ , ν n ) ,
Σ ^ θ : ν n = N n N n − ℓ n Σ ^ θ : 1 − 1 N n − ℓ n B n . \widehat{\Sigma}_{\theta:\nu_n}=\frac{N_n}{N_n-\ell_n}\widehat{\Sigma}_{\theta:1}-\frac{1}{N_n-\ell_n}B_n. Σ θ : ν n = N n − ℓ n N n Σ θ : 1 − N n − ℓ n 1 B n .
Therefore,
∥ Σ ^ θ : ν n − Σ ^ θ : 1 ∥ ≤ ℓ n N n − ℓ n ∥ Σ ^ θ : 1 ∥ + Tr B n N n − ℓ n . \left\lVert\widehat{\Sigma}_{\theta:\nu_n}-\widehat{\Sigma}_{\theta:1}\right\rVert\le\frac{\ell_n}{N_n-\ell_n}\left\lVert\widehat{\Sigma}_{\theta:1}\right\rVert+\frac{\operatorname{Tr}B_n}{N_n-\ell_n}. Σ θ : ν n − Σ θ : 1 ≤ N n − ℓ n ℓ n Σ θ : 1 + N n − ℓ n Tr B n .
The first term tends to zero by Assumption (A1), and the second term tends to zero by (24), since ℓ n / n → 0 \ell_n/n\to0 ℓ n / n → 0 and ( N n − ℓ n ) / n → 1 − θ > 0 (N_n-\ell_n)/n\to1-\theta>0 ( N n − ℓ n ) / n → 1 − θ > 0 . The left endpoint is symmetric. The log-determinant estimates are a direct application of Lemma 40: for the regularized blocks, every interpolant ( 1 − t ) ( Σ ^ θ : ν n + ε I q ) + t ( Σ ^ θ : 1 + ε I q ) (1-t)(\widehat{\Sigma}_{\theta:\nu_n}+\varepsilon I_q)+t(\widehat{\Sigma}_{\theta:1}+\varepsilon I_q) ( 1 − t ) ( Σ θ : ν n + ε I q ) + t ( Σ θ : 1 + ε I q ) dominates ε I q \varepsilon I_q ε I q , so the supremum in (25) is at most ε − 1 \varepsilon^{-1} ε − 1 , and the claimed bounds follow; the case ν n → 0 \nu_n\to0 ν n → 0 is identical. □ \square □
Lemma 43 (Lipschitz continuity of the QML-objective functions). Assume (A1), Δ w i t h i n < ∞ \Delta_{\mathrm{within}}<\infty Δ within < ∞ , and that either ε > 0 \varepsilon>0 ε > 0 or (PD) holds. There exists a constant C < ∞ C<\infty C < ∞ only depending on q q q , ε \varepsilon ε , m m m , M M M and Δ w i t h i n \Delta_{\mathrm{within}} Δ within such that for all 0 < ν 1 < ν 2 < 1 0<\nu_1<\nu_2<1 0 < ν 1 < ν 2 < 1
lim sup n → ∞ ∣ J n ( ν 1 ) − J n ( ν 2 ) ∣ ≤ C ∣ ν 2 − ν 1 ∣ (26) \limsup_{n\to\infty}\left\lvert J_n(\nu_1)-J_n(\nu_2)\right\rvert\le C\lvert\nu_2-\nu_1\rvert \tag*{(26)} n → ∞ lim sup ∣ J n ( ν 1 ) − J n ( ν 2 ) ∣ ≤ C ∣ ν 2 − ν 1 ∣ (26)
If, in addition, Assumption (A2) holds, then
∣ J ∞ ( ν 1 ) − J ∞ ( ν 2 ) ∣ ≤ C ∣ ν 2 − ν 1 ∣ . (27) \lvert J_{\infty}(\nu_1)-J_{\infty}(\nu_2)\rvert\le C\lvert\nu_2-\nu_1\rvert. \tag*{(27)} ∣ J ∞ ( ν 1 ) − J ∞ ( ν 2 )∣ ≤ C ∣ ν 2 − ν 1 ∣ . (27)
For every η ∈ ( 0 , 1 / 2 ) \eta\in(0,1/2) η ∈ ( 0 , 1/2 ) , the estimate (26) holds uniformly on [ η , 1 − η ] [\eta,1-\eta] [ η , 1 − η ] in the sense that
lim sup n → ∞ sup ν 1 , ν 2 ∈ [ η , 1 − η ] ( ∣ J n ( ν 1 ) − J n ( ν 2 ) ∣ − C ∣ ν 2 − ν 1 ∣ ) ≤ 0. (28) \limsup_{n\to\infty}\ \sup_{\nu_1,\nu_2\in[\eta,1-\eta]}\left(\lvert J_n(\nu_1)-J_n(\nu_2)\rvert-C\lvert\nu_2-\nu_1\rvert\right)\le0. \tag*{(28)} n → ∞ lim sup ν 1 , ν 2 ∈ [ η , 1 − η ] sup ( ∣ J n ( ν 1 ) − J n ( ν 2 )∣ − C ∣ ν 2 − ν 1 ∣ ) ≤ 0. (28)
In contrast to (26), the uniform version (28) may be applied to points ν i = ν i ( n ) \nu_i=\nu_i(n) ν i = ν i ( n ) that vary with n n n , as they do in the proof of Theorem 21 below.
Proof of Lemma 43. Argument for (26): For every fixed ν ∈ ( 0 , 1 ) \nu\in(0,1) ν ∈ ( 0 , 1 ) , Definition 5 shows that each of Σ ^ 0 : ν \widehat{\Sigma}_{0:\nu} Σ 0 : ν and Σ ^ ν : 1 \widehat{\Sigma}_{\nu:1} Σ ν : 1 is either the empirical second moment of an interval in I θ c u t \mathcal{I}_{\theta}^{\mathrm{cut}} I θ cut or a convex combination of empirical second moments of two intervals in I θ c u t \mathcal{I}_{\theta}^{\mathrm{cut}} I θ cut . Hence their lower spectral limits are at least m m m . Lemma 36 bounds their upper spectral limits by M + Δ w i t h i n M+\Delta_{\mathrm{within}} M + Δ within . Consequently, for all sufficiently large n n n , their regularized eigenvalues lie in
[ m + ε 2 , 2 ( M + Δ w i t h i n + ε ) ] . \left[\frac{m+\varepsilon}{2},2\left(M+\Delta_{\mathrm{within}}+\varepsilon\right)\right]. [ 2 m + ε , 2 ( M + Δ within + ε ) ] .
To simplify calculations we introduce the notation
L n ( ν ) : = log det ( Σ ^ 0 : ν + ε I q ) , R n ( ν ) : = log det ( Σ ^ ν : 1 + ε I q ) , L_n(\nu):=\log\det\left(\widehat{\Sigma}_{0:\nu}+\varepsilon I_q\right),\qquad R_n(\nu):=\log\det\left(\widehat{\Sigma}_{\nu:1}+\varepsilon I_q\right), L n ( ν ) := log det ( Σ 0 : ν + ε I q ) , R n ( ν ) := log det ( Σ ν : 1 + ε I q ) ,
which allows us to write
J n ( ν ) = ν L n ( ν ) + ( 1 − ν ) R n ( ν ) . J_n(\nu)=\nu L_n(\nu)+(1-\nu)R_n(\nu). J n ( ν ) = ν L n ( ν ) + ( 1 − ν ) R n ( ν ) .
Adding and subtracting terms gives
J n ( ν 2 ) − J n ( ν 1 ) = ν 2 L n ( ν 2 ) − ν 1 L n ( ν 1 ) + ( 1 − ν 2 ) R n ( ν 2 ) − ( 1 − ν 1 ) R n ( ν 1 ) = ( ν 2 − ν 1 ) L n ( ν 2 ) + ν 1 ( L n ( ν 2 ) − L n ( ν 1 ) ) − ( ν 2 − ν 1 ) R n ( ν 2 ) + ( 1 − ν 1 ) ( R n ( ν 2 ) − R n ( ν 1 ) ) . \begin{aligned}
J_n(\nu_2)-J_n(\nu_1)
&=\nu_2L_n(\nu_2)-\nu_1L_n(\nu_1)+(1-\nu_2)R_n(\nu_2)-(1-\nu_1)R_n(\nu_1)\\
&=(\nu_2-\nu_1)L_n(\nu_2)+\nu_1\left(L_n(\nu_2)-L_n(\nu_1)\right)\\
&\quad-(\nu_2-\nu_1)R_n(\nu_2)+(1-\nu_1)\left(R_n(\nu_2)-R_n(\nu_1)\right).
\end{aligned} J n ( ν 2 ) − J n ( ν 1 ) = ν 2 L n ( ν 2 ) − ν 1 L n ( ν 1 ) + ( 1 − ν 2 ) R n ( ν 2 ) − ( 1 − ν 1 ) R n ( ν 1 ) = ( ν 2 − ν 1 ) L n ( ν 2 ) + ν 1 ( L n ( ν 2 ) − L n ( ν 1 ) ) − ( ν 2 − ν 1 ) R n ( ν 2 ) + ( 1 − ν 1 ) ( R n ( ν 2 ) − R n ( ν 1 ) ) .
Hence, by the triangle inequality, we obtain
∣ J n ( ν 2 ) − J n ( ν 1 ) ∣ ≤ ∣ ν 2 − ν 1 ∣ ( ∣ L n ( ν 2 ) ∣ + ∣ R n ( ν 2 ) ∣ ) + ν 1 ∣ L n ( ν 2 ) − L n ( ν 1 ) ∣ + ( 1 − ν 1 ) ∣ R n ( ν 2 ) − R n ( ν 1 ) ∣ . (29) \lvert J_n(\nu_2)-J_n(\nu_1)\rvert\le\lvert\nu_2-\nu_1\rvert\left(\lvert L_n(\nu_2)\rvert+\lvert R_n(\nu_2)\rvert\right)+\nu_1\lvert L_n(\nu_2)-L_n(\nu_1)\rvert+(1-\nu_1)\lvert R_n(\nu_2)-R_n(\nu_1)\rvert. \tag*{(29)} ∣ J n ( ν 2 ) − J n ( ν 1 )∣ ≤ ∣ ν 2 − ν 1 ∣ ( ∣ L n ( ν 2 )∣ + ∣ R n ( ν 2 )∣ ) + ν 1 ∣ L n ( ν 2 ) − L n ( ν 1 )∣ + ( 1 − ν 1 ) ∣ R n ( ν 2 ) − R n ( ν 1 )∣ . (29)
As the determinant equals the product of eigenvalues, set
C log : = 2 q max { ∣ log ( 2 ( M + Δ w i t h i n + ε ) ) ∣ , ∣ log ( m + ε 2 ) ∣ } . C_{\log}:=2q\max\left\{\left\lvert\log\left(2(M+\Delta_{\mathrm{within}}+\varepsilon)\right)\right\rvert,\left\lvert\log\left(\frac{m+\varepsilon}{2}\right)\right\rvert\right\}. C l o g := 2 q max { ∣ log ( 2 ( M + Δ within + ε ) ) ∣ , log ( 2 m + ε ) } .
The preceding two-sided spectral bound gives
∣ L n ( ν 2 ) ∣ + ∣ R n ( ν 2 ) ∣ ≤ C log . (30) \lvert L_n(\nu_2)\rvert+\lvert R_n(\nu_2)\rvert\le C_{\log}. \tag*{(30)} ∣ L n ( ν 2 )∣ + ∣ R n ( ν 2 )∣ ≤ C l o g . (30)
Setting A : = Σ ^ 0 : ν 2 A:=\widehat{\Sigma}_{0:\nu_2} A := Σ 0 : ν 2 and B : = Σ ^ 0 : ν 1 B:=\widehat{\Sigma}_{0:\nu_1} B := Σ 0 : ν 1 and using Lemma 40, we observe that for large enough n n n
∣ L n ( ν 2 ) − L n ( ν 1 ) ∣ ≤ q sup t ∈ [ 0 , 1 ] ∥ ( A t + ( 1 − t ) B + ε I q ) − 1 ∥ ∥ A − B ∥ ≤ 4 q M + Δ w i t h i n ν 2 ( m + ε ) ∣ ν 2 − ν 1 ∣ . (31) \begin{aligned}
\lvert L_n(\nu_2)-L_n(\nu_1)\rvert
&\le q\sup_{t\in[0,1]}\left\|\left(At+(1-t)B+\varepsilon I_q\right)^{-1}\right\|\lVert A-B\rVert\\
&\le4q\frac{M+\Delta_{\mathrm{within}}}{\nu_2(m+\varepsilon)}\lvert\nu_2-\nu_1\rvert. \tag*{(31)}
\end{aligned} ∣ L n ( ν 2 ) − L n ( ν 1 )∣ ≤ q t ∈ [ 0 , 1 ] sup ( A t + ( 1 − t ) B + ε I q ) − 1 ∥ A − B ∥ ≤ 4 q ν 2 ( m + ε ) M + Δ within ∣ ν 2 − ν 1 ∣ . (31)
where we used the lower bound on the eigenvalues in combination with Lemma 37. Similarly, we get
∣ R n ( ν 2 ) − R n ( ν 1 ) ∣ ≤ 4 q M + Δ w i t h i n ( 1 − ν 1 ) ( m + ε ) ∣ ν 2 − ν 1 ∣ , (32) \lvert R_n(\nu_2)-R_n(\nu_1)\rvert\leq4q\frac{M+\Delta_{\mathrm{within}}}{(1-\nu_1)(m+\varepsilon)}\lvert\nu_2-\nu_1\rvert, \tag*{(32)} ∣ R n ( ν 2 ) − R n ( ν 1 )∣ ≤ 4 q ( 1 − ν 1 ) ( m + ε ) M + Δ within ∣ ν 2 − ν 1 ∣ , (32)
When we insert the bounds (30), (31), and (32) into (29), we get
∣ J n ( ν 2 ) − J n ( ν 1 ) ∣ ≤ ∣ ν 2 − ν 1 ∣ ( 4 q M + Δ w i t h i n m + ε ( ν 1 ν 2 + 1 ) + C log ) . \lvert J_n(\nu_2)-J_n(\nu_1)\rvert\leq\lvert\nu_2-\nu_1\rvert\left(4q\frac{M+\Delta_{\mathrm{within}}}{m+\varepsilon}\left(\frac{\nu_1}{\nu_2}+1\right)+C_{\log}\right). ∣ J n ( ν 2 ) − J n ( ν 1 )∣ ≤ ∣ ν 2 − ν 1 ∣ ( 4 q m + ε M + Δ within ( ν 2 ν 1 + 1 ) + C l o g ) .
Incorporating the fact that ν 1 / ν 2 ≤ 1 \nu_1/\nu_2\leq1 ν 1 / ν 2 ≤ 1 , we get the desired estimate (26), i.e.
∣ J n ( ν 2 ) − J n ( ν 1 ) ∣ ≤ ∣ ν 2 − ν 1 ∣ C , \lvert J_n(\nu_2)-J_n(\nu_1)\rvert\leq\lvert\nu_2-\nu_1\rvert C, ∣ J n ( ν 2 ) − J n ( ν 1 )∣ ≤ ∣ ν 2 − ν 1 ∣ C ,
where C C C has no dependence on ν 2 , ν 1 \nu_2,\nu_1 ν 2 , ν 1 .
The argument for (27) is similar after observing under Assumption (A2) that the eigenvalues of Σ 0 : ν + ε I q \Sigma_{0:\nu}+\varepsilon I_q Σ 0 : ν + ε I q and Σ ν : 1 + ε I q \Sigma_{\nu:1}+\varepsilon I_q Σ ν : 1 + ε I q belong to [ m + ε , M + ε ] [m+\varepsilon,M+\varepsilon] [ m + ε , M + ε ] .
Argument for (28): Suppose the claim fails for some η ∈ ( 0 , 1 / 2 ) \eta\in(0,1/2) η ∈ ( 0 , 1/2 ) . Then there exist α > 0 \alpha>0 α > 0 , a subsequence (not relabelled), and points ν 1 ( n ) , ν 2 ( n ) ∈ [ η , 1 − η ] \nu_1(n),\nu_2(n)\in[\eta,1-\eta] ν 1 ( n ) , ν 2 ( n ) ∈ [ η , 1 − η ] such that
∣ J n ( ν 1 ( n ) ) − J n ( ν 2 ( n ) ) ∣ ≥ C ∣ ν 2 ( n ) − ν 1 ( n ) ∣ + α . \lvert J_n(\nu_1(n))-J_n(\nu_2(n))\rvert\geq C\lvert\nu_2(n)-\nu_1(n)\rvert+\alpha. ∣ J n ( ν 1 ( n )) − J n ( ν 2 ( n ))∣ ≥ C ∣ ν 2 ( n ) − ν 1 ( n )∣ + α .
By compactness of [ η , 1 − η ] [\eta,1-\eta] [ η , 1 − η ] , after passing to a further subsequence, ν i ( n ) → ν ˉ i ∈ [ η , 1 − η ] \nu_i(n)\to\bar{\nu}_i\in[\eta,1-\eta] ν i ( n ) → ν ˉ i ∈ [ η , 1 − η ] for i = 1 , 2 i=1,2 i = 1 , 2 . Lemma 41 gives
J n ( ν i ( n ) ) − J n ( ν ˉ i ) ⟶ 0 , i = 1 , 2. J_n(\nu_i(n))-J_n(\bar{\nu}_i)\longrightarrow0,\qquad i=1,2. J n ( ν i ( n )) − J n ( ν ˉ i ) ⟶ 0 , i = 1 , 2.
It follows that
lim sup n → ∞ ∣ J n ( ν 1 ( n ) ) − J n ( ν 2 ( n ) ) ∣ ≤ lim sup n → ∞ ∣ J n ( ν ˉ 1 ) − J n ( ν ˉ 2 ) ∣ ≤ C ∣ ν ˉ 2 − ν ˉ 1 ∣ , \limsup_{n\to\infty}\lvert J_n(\nu_1(n))-J_n(\nu_2(n))\rvert\leq\limsup_{n\to\infty}\lvert J_n(\bar{\nu}_1)-J_n(\bar{\nu}_2)\rvert\leq C\lvert\bar{\nu}_2-\bar{\nu}_1\rvert, n → ∞ lim sup ∣ J n ( ν 1 ( n )) − J n ( ν 2 ( n ))∣ ≤ n → ∞ lim sup ∣ J n ( ν ˉ 1 ) − J n ( ν ˉ 2 )∣ ≤ C ∣ ν ˉ 2 − ν ˉ 1 ∣ ,
where the last inequality is (26). This contradicts the preceding display because ∣ ν 2 ( n ) − ν 1 ( n ) ∣ → ∣ ν ˉ 2 − ν ˉ 1 ∣ \lvert\nu_2(n)-\nu_1(n)\rvert\to\lvert\bar{\nu}_2-\bar{\nu}_1\rvert ∣ ν 2 ( n ) − ν 1 ( n )∣ → ∣ ν ˉ 2 − ν ˉ 1 ∣ . □
Lemma 44 (Jensen gap for log det). Let A , B ∈ R q × q A,B\in\mathbb{R}^{q\times q} A , B ∈ R q × q be symmetric positive definite matrices, λ ∈ [ 0 , 1 ] \lambda\in[0,1] λ ∈ [ 0 , 1 ] , and C ( t ) : = ( 1 − t ) A + t B C(t):=(1-t)A+tB C ( t ) := ( 1 − t ) A + tB for t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . Define the Jensen gap as
G ( A , B ; λ ) : = log det ( C ( λ ) ) − ( ( 1 − λ ) log det A + λ log det B ) . G(A,B;\lambda):=\log\det(C(\lambda))-\left((1-\lambda)\log\det A+\lambda\log\det B\right). G ( A , B ; λ ) := log det ( C ( λ )) − ( ( 1 − λ ) log det A + λ log det B ) .
If there are μ min , μ max ∈ R + \mu_{\min},\mu_{\max}\in\mathbb{R}^{+} μ m i n , μ m a x ∈ R + such that for all t ∈ ( 0 , 1 ) t\in(0,1) t ∈ ( 0 , 1 ) and z ∈ R q z\in\mathbb{R}^{q} z ∈ R q
μ min ∥ z ∥ 2 ≤ z T C ( t ) z ≤ μ max ∥ z ∥ 2 (33) \mu_{\min}\lVert z\rVert^2\leq z^\mathsf{T}C(t)z\leq\mu_{\max}\lVert z\rVert^2 \tag*{(33)} μ m i n ∥ z ∥ 2 ≤ z T C ( t ) z ≤ μ m a x ∥ z ∥ 2 (33)
then
λ ( 1 − λ ) 2 μ max 2 ∥ A − B ∥ F 2 ≤ G ( A , B ; λ ) ≤ λ ( 1 − λ ) 2 μ min 2 ∥ A − B ∥ F 2 . (34) \frac{\lambda(1-\lambda)}{2\mu_{\max}^2}\lVert A-B\rVert_F^2\leq G(A,B;\lambda)\leq\frac{\lambda(1-\lambda)}{2\mu_{\min}^2}\lVert A-B\rVert_F^2. \tag*{(34)} 2 μ m a x 2 λ ( 1 − λ ) ∥ A − B ∥ F 2 ≤ G ( A , B ; λ ) ≤ 2 μ m i n 2 λ ( 1 − λ ) ∥ A − B ∥ F 2 . (34)
Moreover, setting D : = B − A D:=B-A D := B − A and g ( t ) : = log det C ( t ) g(t):=\log\det C(t) g ( t ) := log det C ( t ) , one has
g ′ ′ ( t ) = − ∥ C ( t ) − 1 / 2 D C ( t ) − 1 / 2 ∥ F 2 , t ∈ [ 0 , 1 ] . (35) g''(t)=-\left\lVert C(t)^{-1/2}DC(t)^{-1/2}\right\rVert_F^2,\qquad t\in[0,1]. \tag*{(35)} g ′′ ( t ) = − C ( t ) − 1/2 D C ( t ) − 1/2 F 2 , t ∈ [ 0 , 1 ] . (35)
Proof of Lemma 44. Set D : = B − A D := B - A D := B − A . The function g : [ 0 , 1 ] → R g : [0,1] \to\mathbb{R} g : [ 0 , 1 ] → R , g ( t ) : = log det C ( t ) g(t) := \log\det C(t) g ( t ) := log det C ( t ) , is twice continuously differentiable. Jacobi’s formula gives
d d t det C ( t ) = det C ( t ) Tr ( C ( t ) − 1 C ′ ( t ) ) . \frac{d}{dt}\det C(t) = \det C(t)\operatorname{Tr}\left(C(t)^{-1}C'(t)\right). d t d det C ( t ) = det C ( t ) Tr ( C ( t ) − 1 C ′ ( t ) ) .
Since C ′ ( t ) = D C'(t) = D C ′ ( t ) = D , it follows that
g ′ ( t ) = Tr ( C ( t ) − 1 D ) . g'(t) = \operatorname{Tr}\left(C(t)^{-1}D\right). g ′ ( t ) = Tr ( C ( t ) − 1 D ) .
Differentiating once more and using
d d t C ( t ) − 1 = − C ( t ) − 1 D C ( t ) − 1 , \frac{d}{dt}C(t)^{-1} = -C(t)^{-1}DC(t)^{-1}, d t d C ( t ) − 1 = − C ( t ) − 1 D C ( t ) − 1 ,
we obtain
g ′ ′ ( t ) = − Tr ( C ( t ) − 1 D C ( t ) − 1 D ) = − Tr ( ( C ( t ) − 1 / 2 D C ( t ) − 1 / 2 ) 2 ) = − ∥ C ( t ) − 1 / 2 D C ( t ) − 1 / 2 ∥ F 2 . \begin{aligned}
g''(t) &= -\operatorname{Tr}\left(C(t)^{-1}DC(t)^{-1}D\right) \\
&= -\operatorname{Tr}\left(\left(C(t)^{-1/2}DC(t)^{-1/2}\right)^2\right) \\
&= -\left\|C(t)^{-1/2}DC(t)^{-1/2}\right\|_F^2.
\end{aligned} g ′′ ( t ) = − Tr ( C ( t ) − 1 D C ( t ) − 1 D ) = − Tr ( ( C ( t ) − 1/2 D C ( t ) − 1/2 ) 2 ) = − C ( t ) − 1/2 D C ( t ) − 1/2 F 2 .
This proves (35).
Here we used the cyclicity of the trace and the fact that C ( t ) − 1 / 2 D C ( t ) − 1 / 2 C(t)^{-1/2}DC(t)^{-1/2} C ( t ) − 1/2 D C ( t ) − 1/2 is symmetric.
Applying the fundamental theorem of calculus twice gives the interpolation-error identity
g ( λ ) − ( 1 − λ ) g ( 0 ) − λ g ( 1 ) = − ( 1 − λ ) ∫ 0 λ t g ′ ′ ( t ) d t − λ ∫ λ 1 ( 1 − t ) g ′ ′ ( t ) d t . \begin{aligned}
g(\lambda) - (1-\lambda)g(0) - \lambda g(1)
= {}& -(1-\lambda)\int_0^\lambda t g''(t)\,dt \\
&-\lambda\int_\lambda^1(1-t)g''(t)\,dt.
\end{aligned} g ( λ ) − ( 1 − λ ) g ( 0 ) − λ g ( 1 ) = − ( 1 − λ ) ∫ 0 λ t g ′′ ( t ) d t − λ ∫ λ 1 ( 1 − t ) g ′′ ( t ) d t .
Substituting (35), we obtain
G ( A , B ; λ ) = ( 1 − λ ) ∫ 0 λ t ∥ C ( t ) − 1 / 2 D C ( t ) − 1 / 2 ∥ F 2 d t + λ ∫ λ 1 ( 1 − t ) ∥ C ( t ) − 1 / 2 D C ( t ) − 1 / 2 ∥ F 2 d t . (36) \begin{aligned}
G(A,B;\lambda) ={}& (1-\lambda)\int_0^\lambda t\left\|C(t)^{-1/2}DC(t)^{-1/2}\right\|_F^2\,dt \\
&+\lambda\int_\lambda^1(1-t)\left\|C(t)^{-1/2}DC(t)^{-1/2}\right\|_F^2\,dt. \tag*{(36)}
\end{aligned} G ( A , B ; λ ) = ( 1 − λ ) ∫ 0 λ t C ( t ) − 1/2 D C ( t ) − 1/2 F 2 d t + λ ∫ λ 1 ( 1 − t ) C ( t ) − 1/2 D C ( t ) − 1/2 F 2 d t . (36)
For every symmetric positive-definite matrix C C C and every matrix D D D , we have
1 λ max ( C ) ∥ D ∥ F ≤ ∥ C − 1 / 2 D C − 1 / 2 ∥ F ≤ 1 λ min ( C ) ∥ D ∥ F . \frac{1}{\lambda_{\max}(C)}\|D\|_F \le\left\|C^{-1/2}DC^{-1/2}\right\|_F \le\frac{1}{\lambda_{\min}(C)}\|D\|_F. λ m a x ( C ) 1 ∥ D ∥ F ≤ C − 1/2 D C − 1/2 F ≤ λ m i n ( C ) 1 ∥ D ∥ F .
Consequently, Assumption (33) implies
1 μ max 2 ∥ D ∥ F 2 ≤ ∥ C ( t ) − 1 / 2 D C ( t ) − 1 / 2 ∥ F 2 ≤ 1 μ min 2 ∥ D ∥ F 2 . \frac{1}{\mu_{\max}^2}\|D\|_F^2 \le\left\|C(t)^{-1/2}DC(t)^{-1/2}\right\|_F^2 \le\frac{1}{\mu_{\min}^2}\|D\|_F^2. μ m a x 2 1 ∥ D ∥ F 2 ≤ C ( t ) − 1/2 D C ( t ) − 1/2 F 2 ≤ μ m i n 2 1 ∥ D ∥ F 2 .
Finally,
( 1 − λ ) ∫ 0 λ t d t + λ ∫ λ 1 ( 1 − t ) d t = λ ( 1 − λ ) 2 . (1-\lambda)\int_0^\lambda t\,dt+\lambda\int_\lambda^1(1-t)\,dt=\frac{\lambda(1-\lambda)}{2}. ( 1 − λ ) ∫ 0 λ t d t + λ ∫ λ 1 ( 1 − t ) d t = 2 λ ( 1 − λ ) .
Inserting these estimates into (36) yields
λ ( 1 − λ ) 2 μ max 2 ∥ A − B ∥ F 2 ≤ G ( A , B ; λ ) ≤ λ ( 1 − λ ) 2 μ min 2 ∥ A − B ∥ F 2 , \frac{\lambda(1-\lambda)}{2\mu_{\max}^2}\|A-B\|_F^2 \le G(A,B;\lambda) \le\frac{\lambda(1-\lambda)}{2\mu_{\min}^2}\|A-B\|_F^2, 2 μ m a x 2 λ ( 1 − λ ) ∥ A − B ∥ F 2 ≤ G ( A , B ; λ ) ≤ 2 μ m i n 2 λ ( 1 − λ ) ∥ A − B ∥ F 2 ,
which proves (34). □
Proof of the main results.
Proof of Theorem 21.
Argument for (i): For easier notation we write A : = Σ ≤ + ε I q A := \Sigma_{\leq} + \varepsilon I_q A := Σ ≤ + ε I q , B : = Σ > + ε I q B := \Sigma_{>} + \varepsilon I_q B := Σ > + ε I q and D : = A − B ≠ 0 D := A - B \neq0 D := A − B = 0 by assumption (A2). We deduce the concavity of J ∞ J_{\infty} J ∞ on ( θ , 1 ) (\theta,1) ( θ , 1 ) ; the interval ( 0 , θ ) (0,\theta) ( 0 , θ ) uses the same argument. Setting
X ( ν ) : = θ ν A + ν − θ ν B = B + θ ν D , X(\nu) := \frac{\theta}{\nu} A + \frac{\nu-\theta}{\nu} B = B + \frac{\theta}{\nu}D, X ( ν ) := ν θ A + ν ν − θ B = B + ν θ D ,
we rewrite J ∞ ( ν ) J_{\infty}(\nu) J ∞ ( ν ) for ν > θ \nu> \theta ν > θ as
J ∞ ( ν ) = ν log det X ( ν ) + ( 1 − ν ) log det B = ν g ( θ ν ) + ( 1 − ν ) log det B , J_{\infty}(\nu) = \nu\log\det X(\nu) + (1-\nu)\log\det B = \nu g\left(\frac{\theta}{\nu}\right) + (1-\nu)\log\det B, J ∞ ( ν ) = ν log det X ( ν ) + ( 1 − ν ) log det B = ν g ( ν θ ) + ( 1 − ν ) log det B ,
where g ( s ) : = log det ( ( 1 − s ) B + s A ) g(s) := \log\det((1-s)B+sA) g ( s ) := log det (( 1 − s ) B + s A ) : the matrix X ( ν ) X(\nu) X ( ν ) runs along the affine interpolation of the Jensen-gap Lemma 44 (with the roles of A A A and B B B interchanged) at parameter s = θ / ν s = \theta/\nu s = θ / ν . Differentiating twice, the g ′ g' g ′ -terms cancel, and the second-derivative identity (35) in Lemma 44 yields
J ∞ ′ ′ ( ν ) = θ 2 ν 3 g ′ ′ ( θ ν ) = − θ 2 ν 3 ∥ X ( ν ) − 1 / 2 D X ( ν ) − 1 / 2 ∥ F 2 . J_{\infty}''(\nu) = \frac{\theta^2}{\nu^3}g''\left(\frac{\theta}{\nu}\right) = -\frac{\theta^2}{\nu^3}\left\|X(\nu)^{-1/2}D X(\nu)^{-1/2}\right\|_F^2. J ∞ ′′ ( ν ) = ν 3 θ 2 g ′′ ( ν θ ) = − ν 3 θ 2 X ( ν ) − 1/2 D X ( ν ) − 1/2 F 2 .
The Frobenius norm term is strictly positive since D ≠ 0 D \neq0 D = 0 and X ( ν ) X(\nu) X ( ν ) is positive definite, hence proving the strict concavity of J ∞ J_{\infty} J ∞ on ( θ , 1 ) (\theta,1) ( θ , 1 ) .
Argument for (6): We observe that by Assumption (A1) and (A2) it holds that for all z ∈ R q z \in\mathbb{R}^q z ∈ R q
z T A z ≤ ( M + ε ) ∣ z ∣ 2 and z T B z ≤ ( M + ε ) ∣ z ∣ 2 , z^{\mathsf{T}}Az \leq(M+\varepsilon)|z|^2 \quad\text{and} \quad z^{\mathsf{T}}Bz \leq(M+\varepsilon)|z|^2, z T A z ≤ ( M + ε ) ∣ z ∣ 2 and z T B z ≤ ( M + ε ) ∣ z ∣ 2 ,
and therefore also
z T ( ( 1 − t ) A + t B ) z ≤ ( M + ε ) ∣ z ∣ 2 for all t ∈ [ 0 , 1 ] . z^{\mathsf{T}}\bigl((1-t)A+tB\bigr)z \leq(M+\varepsilon)|z|^2 \quad\text{for all } t \in[0,1]. z T ( ( 1 − t ) A + tB ) z ≤ ( M + ε ) ∣ z ∣ 2 for all t ∈ [ 0 , 1 ] .
The complementary lower spectral hypothesis of Lemma 44 also holds: the ridge supplies it when ε > 0 \varepsilon> 0 ε > 0 , while (PD) supplies it when ε = 0 \varepsilon= 0 ε = 0 . This allows us to apply the lower bound of Lemma 44, which yields, for the Jensen gap G ( A , B ; λ ) G(A,B;\lambda) G ( A , B ; λ ) defined there and λ ∈ [ 0 , 1 ] \lambda\in[0,1] λ ∈ [ 0 , 1 ] , the estimate
G ( A , B ; λ ) ≥ λ ( 1 − λ ) 2 ( M + ε ) 2 ∥ A − B ∥ F 2 . (37) G(A,B;\lambda) \geq\frac{\lambda(1-\lambda)}{2(M+\varepsilon)^2}\lVert A-B\rVert_F^2. \tag*{(37)} G ( A , B ; λ ) ≥ 2 ( M + ε ) 2 λ ( 1 − λ ) ∥ A − B ∥ F 2 . (37)
For ν > θ \nu> \theta ν > θ , set λ = ( ν − θ ) / ν ∈ ( 0 , 1 ) \lambda= (\nu-\theta)/\nu\in(0,1) λ = ( ν − θ ) / ν ∈ ( 0 , 1 ) so that
θ ν A + ν − θ ν B = ( 1 − λ ) A + λ B . \frac{\theta}{\nu}A+\frac{\nu-\theta}{\nu}B=(1-\lambda)A+\lambda B. ν θ A + ν ν − θ B = ( 1 − λ ) A + λ B .
Using the definition of J ∞ J_{\infty} J ∞ (cf. Theorem 21), direct regrouping yields
J ∞ ( ν ) − J ∞ ( θ ) = ν G ( A , B ; λ ) . J_{\infty}(\nu)-J_{\infty}(\theta)=\nu G(A,B;\lambda). J ∞ ( ν ) − J ∞ ( θ ) = ν G ( A , B ; λ ) .
Combining this with (37) gives
J ∞ ( ν ) − J ∞ ( θ ) ≥ ν ⋅ λ ( 1 − λ ) 2 ( M + ε ) 2 ∥ A − B ∥ F 2 . J_{\infty}(\nu)-J_{\infty}(\theta) \geq\nu\cdot\frac{\lambda(1-\lambda)}{2(M+\varepsilon)^2}\lVert A-B\rVert_F^2. J ∞ ( ν ) − J ∞ ( θ ) ≥ ν ⋅ 2 ( M + ε ) 2 λ ( 1 − λ ) ∥ A − B ∥ F 2 .
Now λ = ( ν − θ ) / ν \lambda=(\nu-\theta)/\nu λ = ( ν − θ ) / ν and 1 − λ = θ / ν 1-\lambda=\theta/\nu 1 − λ = θ / ν , hence
ν λ ( 1 − λ ) = ν ⋅ ν − θ ν ⋅ θ ν = θ ( ν − θ ) ν ≥ θ ( ν − θ ) , \nu\lambda(1-\lambda)=\nu\cdot\frac{\nu-\theta}{\nu}\cdot\frac{\theta}{\nu}=\frac{\theta(\nu-\theta)}{\nu}\geq\theta(\nu-\theta), ν λ ( 1 − λ ) = ν ⋅ ν ν − θ ⋅ ν θ = ν θ ( ν − θ ) ≥ θ ( ν − θ ) ,
since ν ≤ 1 \nu\leq1 ν ≤ 1 . Therefore,
J ∞ ( ν ) − J ∞ ( θ ) ≥ θ 2 ( M + ε ) 2 ∥ A − B ∥ F 2 ( ν − θ ) , ν ∈ ( θ , 1 ] . J_{\infty}(\nu)-J_{\infty}(\theta)\geq\frac{\theta}{2(M+\varepsilon)^2}\lVert A-B\rVert_F^2(\nu-\theta),\qquad\nu\in(\theta,1]. J ∞ ( ν ) − J ∞ ( θ ) ≥ 2 ( M + ε ) 2 θ ∥ A − B ∥ F 2 ( ν − θ ) , ν ∈ ( θ , 1 ] .
For ν < θ \nu< \theta ν < θ , a similar argument yields the estimate
J ∞ ( ν ) − J ∞ ( θ ) ≥ 1 − θ 2 ( M + ε ) 2 ∥ A − B ∥ F 2 ( θ − ν ) , ν ∈ [ 0 , θ ) . J_{\infty}(\nu) - J_{\infty}(\theta) \ge\frac{1-\theta}{2(M+\varepsilon)^2}\lVert A-B\rVert_{F}^{2}(\theta-\nu), \qquad\nu\in[0,\theta). J ∞ ( ν ) − J ∞ ( θ ) ≥ 2 ( M + ε ) 2 1 − θ ∥ A − B ∥ F 2 ( θ − ν ) , ν ∈ [ 0 , θ ) .
Since the ridge cancels in the difference, A − B = Σ ≤ − Σ > A-B=\Sigma_{\leq}-\Sigma_{>} A − B = Σ ≤ − Σ > and hence ∥ A − B ∥ F = ∥ Σ > − Σ ≤ ∥ F \lVert A-B\rVert_{F}=\lVert\Sigma_{>}-\Sigma_{\leq}\rVert_{F} ∥ A − B ∥ F = ∥ Σ > − Σ ≤ ∥ F . Combining the two cases yields the desired inequality (6).
Argument for (ii): We first prove the uniform convergence on trimmed grids. Fix 0 < δ < 1 / 2 0<\delta<1/2 0 < δ < 1/2 and η ∈ ( 0 , 1 ) \eta\in(0,1) η ∈ ( 0 , 1 ) . Let G η ⊂ [ δ , 1 − δ ] \mathcal{G}_{\eta}\subset[\delta,1-\delta] G η ⊂ [ δ , 1 − δ ] be a finite grid with mesh at most η \eta η . Assumption (A2) and the definitions of J n J_n J n and J ∞ J_{\infty} J ∞ imply J n ( ν ♯ ) → J ∞ ( ν ♯ ) J_n(\nu^\sharp)\to J_{\infty}(\nu^\sharp) J n ( ν ♯ ) → J ∞ ( ν ♯ ) for every ν ♯ ∈ G η \nu^\sharp\in\mathcal{G}_{\eta} ν ♯ ∈ G η ; since the grid is finite,
max ν ♯ ∈ G η ∣ J n ( ν ♯ ) − J ∞ ( ν ♯ ) ∣ ⟶ 0. \max_{\nu^\sharp\in\mathcal{G}_{\eta}}\left|J_n(\nu^\sharp)-J_{\infty}(\nu^\sharp)\right|\longrightarrow0. ν ♯ ∈ G η max J n ( ν ♯ ) − J ∞ ( ν ♯ ) ⟶ 0.
For any ν ∈ D n ( δ ) \nu\in\mathcal{D}_n(\delta) ν ∈ D n ( δ ) choose ν ♯ ∈ G η \nu^\sharp\in\mathcal{G}_{\eta} ν ♯ ∈ G η with ∣ ν − ν ♯ ∣ ≤ η |\nu-\nu^\sharp|\leq\eta ∣ ν − ν ♯ ∣ ≤ η . We abbreviate the uniform Lipschitz defect
R n : = sup ν 1 , ν 2 ∈ [ δ , 1 − δ ] ( ∣ J n ( ν 1 ) − J n ( ν 2 ) ∣ − C ∣ ν 2 − ν 1 ∣ ) , lim sup n → ∞ R n ≤ 0 , R_n:=\sup_{\nu_1,\nu_2\in[\delta,1-\delta]}\left(\left|J_n(\nu_1)-J_n(\nu_2)\right|-C|\nu_2-\nu_1|\right),\qquad\limsup_{n\to\infty}R_n\leq0, R n := ν 1 , ν 2 ∈ [ δ , 1 − δ ] sup ( ∣ J n ( ν 1 ) − J n ( ν 2 ) ∣ − C ∣ ν 2 − ν 1 ∣ ) , n → ∞ lim sup R n ≤ 0 ,
where C C C is the Lipschitz constant of Lemma 43 and the estimate holds by (28) of Lemma 43; note that R n R_n R n controls the pair ( ν , ν ♯ ) ∈ [ δ , 1 − δ ] 2 (\nu,\nu^\sharp)\in[\delta,1-\delta]^2 ( ν , ν ♯ ) ∈ [ δ , 1 − δ ] 2 even though ν \nu ν varies with n n n . By the triangle inequality and Lemma 43,
∣ J n ( ν ) − J ∞ ( ν ) ∣ ≤ ∣ J n ( ν ) − J n ( ν ♯ ) ∣ + ∣ J n ( ν ♯ ) − J ∞ ( ν ♯ ) ∣ + ∣ J ∞ ( ν ♯ ) − J ∞ ( ν ) ∣ ≤ 2 C η + R n + max μ ∈ G η ∣ J n ( μ ) − J ∞ ( μ ) ∣ . \begin{aligned}
|J_n(\nu)-J_{\infty}(\nu)| &\leq|J_n(\nu)-J_n(\nu^\sharp)|+|J_n(\nu^\sharp)-J_{\infty}(\nu^\sharp)|\\
&\quad+|J_{\infty}(\nu^\sharp)-J_{\infty}(\nu)|\\
&\leq2C\eta+R_n+\max_{\mu\in\mathcal{G}_{\eta}}|J_n(\mu)-J_{\infty}(\mu)|.
\end{aligned} ∣ J n ( ν ) − J ∞ ( ν ) ∣ ≤ ∣ J n ( ν ) − J n ( ν ♯ ) ∣ + ∣ J n ( ν ♯ ) − J ∞ ( ν ♯ ) ∣ + ∣ J ∞ ( ν ♯ ) − J ∞ ( ν ) ∣ ≤ 2 C η + R n + μ ∈ G η max ∣ J n ( μ ) − J ∞ ( μ ) ∣.
Sending first n → ∞ n\to\infty n → ∞ and then η ↓ 0 \eta\downarrow0 η ↓ 0 proves (8). This proves the whole assertion in the case ε = 0 \varepsilon=0 ε = 0 . Since δ \delta δ was arbitrary, the same argument also gives, for ε > 0 \varepsilon>0 ε > 0 and every fixed s ∈ ( 0 , 1 / 2 ) s\in(0,1/2) s ∈ ( 0 , 1/2 ) ,
lim n → ∞ sup ν ∈ D n ∩ [ s , 1 − s ] ∣ J n ( ν ) − J ∞ ( ν ) ∣ = 0. (38) \lim_{n\to\infty}\sup_{\nu\in\mathcal{D}_n\cap[s,1-s]}|J_n(\nu)-J_{\infty}(\nu)|=0. \tag*{(38)} n → ∞ lim ν ∈ D n ∩ [ s , 1 − s ] sup ∣ J n ( ν ) − J ∞ ( ν ) ∣ = 0. (38)
It remains to remove the middle-region cutoff when ε > 0 \varepsilon>0 ε > 0 . We do this by a sequential endpoint estimate. For s > 0 s>0 s > 0 , set
E n L ( s ) : = sup ν ∈ D n ∩ ( 0 , s ] ∣ J n ( ν ) − J ∞ ( ν ) ∣ , E_n^L(s):=\sup_{\nu\in\mathcal{D}_n\cap(0,s]}|J_n(\nu)-J_{\infty}(\nu)|, E n L ( s ) := ν ∈ D n ∩ ( 0 , s ] sup ∣ J n ( ν ) − J ∞ ( ν ) ∣ ,
E n R ( s ) : = sup ν ∈ D n ∩ [ 1 − s , 1 ) ∣ J n ( ν ) − J ∞ ( ν ) ∣ . E_n^R(s):=\sup_{\nu\in\mathcal{D}_n\cap[1-s,1)}|J_n(\nu)-J_{\infty}(\nu)|. E n R ( s ) := ν ∈ D n ∩ [ 1 − s , 1 ) sup ∣ J n ( ν ) − J ∞ ( ν ) ∣.
We claim
lim s ↓ 0 lim sup n → ∞ E n L ( s ) = 0 , lim s ↓ 0 lim sup n → ∞ E n R ( s ) = 0. (39) \lim_{s\downarrow0}\limsup_{n\to\infty}E_n^L(s)=0,\qquad\lim_{s\downarrow0}\limsup_{n\to\infty}E_n^R(s)=0. \tag*{(39)} s ↓ 0 lim n → ∞ lim sup E n L ( s ) = 0 , s ↓ 0 lim n → ∞ lim sup E n R ( s ) = 0. (39)
By symmetry, it suffices to consider the left endpoint. We use the sequential criterion for the first limit in (39): it is enough to prove that
∣ J n ( ν n ) − J ∞ ( ν n ) ∣ ⟶ 0 |J_n(\nu_n)-J_{\infty}(\nu_n)|\longrightarrow0 ∣ J n ( ν n ) − J ∞ ( ν n ) ∣ ⟶ 0
for every sequence ν n ∈ D n \nu_n\in\mathcal{D}_n ν n ∈ D n with ν n → 0 \nu_n\to0 ν n → 0 . Indeed, failure of the first limit would yield, by a diagonal choice, a sequence of this form along which the difference stays bounded away from zero (completed by ν n : = 1 / n \nu_n:=1/n ν n := 1/ n at any unused indices). Fix such a sequence. Lemma 39 gives
ν n ∣ log det ( Σ ^ 0 : ν n + ε I q ) ∣ ⟶ 0. \nu_n \left\lvert\log\det\left(\widehat{\Sigma}_{0:\nu_n}+\varepsilon I_q\right)\right\rvert\longrightarrow0. ν n log det ( Σ 0 : ν n + ε I q ) ⟶ 0.
Moreover, eventually ν n < θ \nu_n<\theta ν n < θ , so Σ 0 : ν n = Σ ≤ \Sigma_{0:\nu_n}=\Sigma_{\leq} Σ 0 : ν n = Σ ≤ and
ν n ∣ log det ( Σ 0 : ν n + ε I q ) ∣ = ν n ∣ log det ( Σ ≤ + ε I q ) ∣ ⟶ 0. \nu_n\left\lvert\log\det\left(\Sigma_{0:\nu_n}+\varepsilon I_q\right)\right\rvert
=\nu_n\left\lvert\log\det\left(\Sigma_{\leq}+\varepsilon I_q\right)\right\rvert
\longrightarrow0. ν n ∣ log det ( Σ 0 : ν n + ε I q ) ∣ = ν n ∣ log det ( Σ ≤ + ε I q ) ∣ ⟶ 0.
It remains to compare the long right blocks. Set
Σ ∗ : = θ Σ ≤ + ( 1 − θ ) Σ > . \Sigma_*:=\theta\Sigma_{\leq}+(1-\theta)\Sigma_{>}. Σ ∗ := θ Σ ≤ + ( 1 − θ ) Σ > .
By (19) and Assumption (A2),
Σ ^ 0 : 1 = ⌊ θ n ⌋ n Σ ^ 0 : θ + n − ⌊ θ n ⌋ n Σ ^ θ : 1 ⟶ Σ ∗ . \widehat{\Sigma}_{0:1}
=\frac{\lfloor\theta n\rfloor}{n}\widehat{\Sigma}_{0:\theta}
+\frac{n-\lfloor\theta n\rfloor}{n}\widehat{\Sigma}_{\theta:1}
\longrightarrow\Sigma_*. Σ 0 : 1 = n ⌊ θ n ⌋ Σ 0 : θ + n n − ⌊ θ n ⌋ Σ θ : 1 ⟶ Σ ∗ .
Writing k n : = n ν n ∈ { 1 , … , n − 1 } k_n:=n\nu_n\in\{1,\ldots,n-1\} k n := n ν n ∈ { 1 , … , n − 1 } , identity (20) also gives
Σ ^ ν n : 1 − Σ ^ 0 : 1 = k n n − k n Σ ^ 0 : 1 − 1 n − k n ∑ t = 1 k n y t y t T . \widehat{\Sigma}_{\nu_n:1}-\widehat{\Sigma}_{0:1}
=\frac{k_n}{n-k_n}\widehat{\Sigma}_{0:1}
-\frac{1}{n-k_n}\sum_{t=1}^{k_n}y_ty_t^{\mathsf T}. Σ ν n : 1 − Σ 0 : 1 = n − k n k n Σ 0 : 1 − n − k n 1 t = 1 ∑ k n y t y t T .
Consequently,
∥ Σ ^ ν n : 1 − Σ ^ 0 : 1 ∥ ≤ k n n − k n ∥ Σ ^ 0 : 1 ∥ + 1 n − k n ∑ t = 1 k n ∥ y t ∥ 2 ⟶ 0 , \left\lVert\widehat{\Sigma}_{\nu_n:1}-\widehat{\Sigma}_{0:1}\right\rVert
\leq\frac{k_n}{n-k_n}\left\lVert\widehat{\Sigma}_{0:1}\right\rVert
+\frac{1}{n-k_n}\sum_{t=1}^{k_n}\left\lVert y_t\right\rVert^2
\longrightarrow0, Σ ν n : 1 − Σ 0 : 1 ≤ n − k n k n Σ 0 : 1 + n − k n 1 t = 1 ∑ k n ∥ y t ∥ 2 ⟶ 0 ,
where the first term vanishes because k n / n = ν n → 0 k_n/n=\nu_n\to0 k n / n = ν n → 0 and Σ ^ 0 : 1 \widehat{\Sigma}_{0:1} Σ 0 : 1 converges, while the second vanishes by (24) of Lemma 38. Thus Σ ^ ν n : 1 → Σ ∗ \widehat{\Sigma}_{\nu_n:1}\to\Sigma_* Σ ν n : 1 → Σ ∗ . On the other hand, the definition of the population target gives
Σ ν n : 1 = θ − ν n 1 − ν n Σ ≤ + 1 − θ 1 − ν n Σ > ⟶ Σ ∗ . \Sigma_{\nu_n:1}
=\frac{\theta-\nu_n}{1-\nu_n}\Sigma_{\leq}
+\frac{1-\theta}{1-\nu_n}\Sigma_{>}
\longrightarrow\Sigma_*. Σ ν n : 1 = 1 − ν n θ − ν n Σ ≤ + 1 − ν n 1 − θ Σ > ⟶ Σ ∗ .
Hence
∥ Σ ^ ν n : 1 − Σ ν n : 1 ∥ ⟶ 0. \left\lVert\widehat{\Sigma}_{\nu_n:1}-\Sigma_{\nu_n:1}\right\rVert\longrightarrow0. Σ ν n : 1 − Σ ν n : 1 ⟶ 0.
Since the two regularized matrices dominate ε I q \varepsilon I_q ε I q , Lemma 40 now yields
∣ log det ( Σ ^ ν n : 1 + ε I q ) − log det ( Σ ν n : 1 + ε I q ) ∣ ⟶ 0. \left\lvert\log\det\left(\widehat{\Sigma}_{\nu_n:1}+\varepsilon I_q\right)
-\log\det\left(\Sigma_{\nu_n:1}+\varepsilon I_q\right)\right\rvert
\longrightarrow0. log det ( Σ ν n : 1 + ε I q ) − log det ( Σ ν n : 1 + ε I q ) ⟶ 0.
Consequently,
∣ J n ( ν n ) − J ∞ ( ν n ) ∣ ≤ ν n ∣ log det ( Σ ^ 0 : ν n + ε I q ) ∣ + ν n ∣ log det ( Σ 0 : ν n + ε I q ) ∣ + ( 1 − ν n ) ∣ log det ( Σ ^ ν n : 1 + ε I q ) − log det ( Σ ν n : 1 + ε I q ) ∣ ⟶ 0. \begin{aligned}
\left\lvert J_n(\nu_n)-J_\infty(\nu_n)\right\rvert
&\leq\nu_n\left\lvert\log\det\left(\widehat{\Sigma}_{0:\nu_n}+\varepsilon I_q\right)\right\rvert\\
&\quad+\nu_n\left\lvert\log\det\left(\Sigma_{0:\nu_n}+\varepsilon I_q\right)\right\rvert\\
&\quad+(1-\nu_n)\left\lvert\log\det\left(\widehat{\Sigma}_{\nu_n:1}+\varepsilon I_q\right)
-\log\det\left(\Sigma_{\nu_n:1}+\varepsilon I_q\right)\right\rvert\\
&\longrightarrow0.
\end{aligned} ∣ J n ( ν n ) − J ∞ ( ν n ) ∣ ≤ ν n log det ( Σ 0 : ν n + ε I q ) + ν n ∣ log det ( Σ 0 : ν n + ε I q ) ∣ + ( 1 − ν n ) log det ( Σ ν n : 1 + ε I q ) − log det ( Σ ν n : 1 + ε I q ) ⟶ 0.
This proves the sequential claim and therefore the left endpoint limit in (39). The right endpoint follows symmetrically.
For fixed s ∈ ( 0 , 1 2 min { θ , 1 − θ } ) s\in\left(0,\frac{1}{2}\min\{\theta,1-\theta\}\right) s ∈ ( 0 , 2 1 min { θ , 1 − θ } ) , write
E n M ( s ) : = sup ν ∈ D n ∩ [ s , 1 − s ] ∣ J n ( ν ) − J ∞ ( ν ) ∣ . E_n^M(s):=\sup_{\nu\in\mathcal{D}_n\cap[s,1-s]}\left\lvert J_n(\nu)-J_\infty(\nu)\right\rvert. E n M ( s ) := ν ∈ D n ∩ [ s , 1 − s ] sup ∣ J n ( ν ) − J ∞ ( ν ) ∣ .
By (38), E n M ( s ) → 0 E_n^M(s)\to0 E n M ( s ) → 0 . Since
sup ν ∈ D n ∣ J n ( ν ) − J ∞ ( ν ) ∣ ≤ max { E n L ( s ) , E n M ( s ) , E n R ( s ) } , \sup_{\nu\in\mathcal{D}_n}\left\lvert J_n(\nu)-J_\infty(\nu)\right\rvert
\leq\max\left\{E_n^L(s),E_n^M(s),E_n^R(s)\right\}, ν ∈ D n sup ∣ J n ( ν ) − J ∞ ( ν ) ∣ ≤ max { E n L ( s ) , E n M ( s ) , E n R ( s ) } ,
taking first lim sup n → ∞ \limsup_{n\to\infty} lim sup n → ∞ and then s ↓ 0 s \downarrow0 s ↓ 0 in (39) proves (7).
Argument for (iii): Let θ n : = ⌊ θ n ⌋ / n \theta_n := \lfloor\theta n\rfloor/n θ n := ⌊ θ n ⌋ / n . Then θ n ∈ D n \theta_n \in\mathcal{D}_n θ n ∈ D n for all large enough n n n and θ n → θ \theta_n \to\theta θ n → θ . We first consider ε > 0 \varepsilon> 0 ε > 0 and let θ ^ n ∈ arg min ν ∈ D n J n ( ν ) \hat{\theta}_n \in\arg\min_{\nu\in\mathcal{D}_n} J_n(\nu) θ ^ n ∈ arg min ν ∈ D n J n ( ν ) . Every subsequence has a further subsequence, again denoted by θ ^ n \hat{\theta}_n θ ^ n , converging to some ν ∈ [ 0 , 1 ] \nu\in[0,1] ν ∈ [ 0 , 1 ] . We prove that ν = θ \nu= \theta ν = θ .
Assume ν ≠ θ \nu\ne\theta ν = θ and set γ : = ∣ ν − θ ∣ > 0 \gamma:= |\nu-\theta| > 0 γ := ∣ ν − θ ∣ > 0 . Let
c ∞ : = min { θ , 1 − θ } 2 ( M + ε ) 2 ∥ Σ > − Σ ≤ ∥ F 2 > 0 c_\infty:= \frac{\min\{\theta,1-\theta\}}{2(M+\varepsilon)^2}\lVert\Sigma_{>} - \Sigma_{\leq}\rVert_F^2 > 0 c ∞ := 2 ( M + ε ) 2 min { θ , 1 − θ } ∥ Σ > − Σ ≤ ∥ F 2 > 0
be the cusp constant in (6). For all large enough n n n , ∣ θ ^ n − θ ∣ ≥ γ / 2 |\hat{\theta}_n-\theta| \ge\gamma/2 ∣ θ ^ n − θ ∣ ≥ γ /2 . By part (ii), for all large enough n n n ,
sup μ ∈ D n ∣ J n ( μ ) − J ∞ ( μ ) ∣ ≤ c ∞ γ 16 . \sup_{\mu\in\mathcal{D}_n}|J_n(\mu)-J_\infty(\mu)| \le\frac{c_\infty\gamma}{16}. μ ∈ D n sup ∣ J n ( μ ) − J ∞ ( μ ) ∣ ≤ 16 c ∞ γ .
Let C ∞ C_\infty C ∞ be the Lipschitz constant in (27). Since θ n → θ \theta_n \to\theta θ n → θ , for all large enough n n n ,
∣ J ∞ ( θ n ) − J ∞ ( θ ) ∣ ≤ C ∞ ∣ θ n − θ ∣ ≤ c ∞ γ 16 . |J_\infty(\theta_n)-J_\infty(\theta)| \le C_\infty|\theta_n-\theta| \le\frac{c_\infty\gamma}{16}. ∣ J ∞ ( θ n ) − J ∞ ( θ ) ∣ ≤ C ∞ ∣ θ n − θ ∣ ≤ 16 c ∞ γ .
By (6) and ∣ θ ^ n − θ ∣ ≥ γ / 2 |\hat{\theta}_n-\theta| \ge\gamma/2 ∣ θ ^ n − θ ∣ ≥ γ /2 , we have J ∞ ( θ ^ n ) − J ∞ ( θ ) ≥ c ∞ ∣ θ ^ n − θ ∣ ≥ c ∞ γ / 2 J_\infty(\hat{\theta}_n)-J_\infty(\theta) \ge c_\infty|\hat{\theta}_n-\theta| \ge c_\infty\gamma/2 J ∞ ( θ ^ n ) − J ∞ ( θ ) ≥ c ∞ ∣ θ ^ n − θ ∣ ≥ c ∞ γ /2 . Together with the two estimates above and the comparison point θ n \theta_n θ n , we get
J n ( θ ^ n ) − J n ( θ n ) ≥ J ∞ ( θ ^ n ) − J ∞ ( θ ) − 2 sup μ ∈ D n ∣ J n ( μ ) − J ∞ ( μ ) ∣ − ∣ J ∞ ( θ n ) − J ∞ ( θ ) ∣ ≥ c ∞ ∣ θ ^ n − θ ∣ − c ∞ γ 8 − c ∞ γ 16 ≥ 5 16 c ∞ γ > 0 , \begin{aligned}
J_n(\hat{\theta}_n)-J_n(\theta_n)
&\ge J_\infty(\hat{\theta}_n)-J_\infty(\theta)-2\sup_{\mu\in\mathcal{D}_n}|J_n(\mu)-J_\infty(\mu)|-|J_\infty(\theta_n)-J_\infty(\theta)| \\
&\ge c_\infty|\hat{\theta}_n-\theta|-\frac{c_\infty\gamma}{8}-\frac{c_\infty\gamma}{16}
\ge\frac{5}{16}c_\infty\gamma> 0,
\end{aligned} J n ( θ ^ n ) − J n ( θ n ) ≥ J ∞ ( θ ^ n ) − J ∞ ( θ ) − 2 μ ∈ D n sup ∣ J n ( μ ) − J ∞ ( μ ) ∣ − ∣ J ∞ ( θ n ) − J ∞ ( θ ) ∣ ≥ c ∞ ∣ θ ^ n − θ ∣ − 8 c ∞ γ − 16 c ∞ γ ≥ 16 5 c ∞ γ > 0 ,
contradicting J n ( θ ^ n ) ≤ J n ( θ n ) J_n(\hat{\theta}_n) \le J_n(\theta_n) J n ( θ ^ n ) ≤ J n ( θ n ) . Hence every subsequential limit equals θ \theta θ , and θ ^ n → θ \hat{\theta}_n \to\theta θ ^ n → θ .
If ε = 0 \varepsilon= 0 ε = 0 , fix 0 < δ < min { θ , 1 − θ } 0 < \delta< \min\{\theta,1-\theta\} 0 < δ < min { θ , 1 − θ } and let θ ^ n , δ ∈ arg min ν ∈ D n ( δ ) J n ( ν ) \hat{\theta}_{n,\delta} \in\arg\min_{\nu\in\mathcal{D}_n(\delta)}J_n(\nu) θ ^ n , δ ∈ arg min ν ∈ D n ( δ ) J n ( ν ) . The same argument applies with D n ( δ ) \mathcal{D}_n(\delta) D n ( δ ) in place of D n \mathcal{D}_n D n , because θ n ∈ D n ( δ ) \theta_n \in\mathcal{D}_n(\delta) θ n ∈ D n ( δ ) for all large enough n n n and the trimmed uniform convergence (8) holds on that grid. Therefore θ ^ n , δ → θ \hat{\theta}_{n,\delta} \to\theta θ ^ n , δ → θ . □
Proof of Theorem 23
The proof implements the two-gap mechanism described in Section 2.1: the objective difference is decomposed into a positive Jensen gap measuring separation between the regimes and a negative gap measuring variation within a regime. The two-sided estimate of Lemma 44 reduces positivity to (A3) locally and (A4) globally, after which the continuity and endpoint lemmas transfer the fixed-candidate bounds to minimizers on the grid.
We first note that the hypotheses imply Δ w i t h i n < ∞ \Delta_{\mathrm{within}} < \infty Δ within < ∞ . Indeed, choosing the fixed blocks I ( 0 , θ ) I(0,\theta) I ( 0 , θ ) and I ( θ , 1 ) I(\theta,1) I ( θ , 1 ) in the definition of Δ b e t w e e n \Delta_{\mathrm{between}} Δ between , Assumption (A1), the triangle inequality, and ∥ ⋅ ∥ F ≤ q ∥ ⋅ ∥ \lVert\cdot\rVert_F \le\sqrt{q}\lVert\cdot\rVert ∥ ⋅ ∥ F ≤ q ∥ ⋅ ∥ give Δ b e t w e e n ≤ 2 q M < ∞ \Delta_{\mathrm{between}} \le2\sqrt{q}M < \infty Δ between ≤ 2 q M < ∞ . Under either (A3) or (A4), the corresponding strict inequality then forces Δ w i t h i n < ∞ \Delta_{\mathrm{within}} < \infty Δ within < ∞ . We may therefore use the auxiliary lemmas above. For brevity, write ℓ n ( a , b ) : = log det ( Σ ^ a : b + ε I q ) \ell_n(a,b) := \log\det(\widehat{\Sigma}_{a:b}+\varepsilon I_q) ℓ n ( a , b ) := log det ( Σ a : b + ε I q ) .
Fixed-candidate bounds. Fix ν > θ \nu> \theta ν > θ and recall k θ = ⌊ θ n ⌋ k_\theta= \lfloor\theta n\rfloor k θ = ⌊ θ n ⌋ , k ν = ⌊ ν n ⌋ k_\nu= \lfloor\nu n\rfloor k ν = ⌊ ν n ⌋ . For all large enough n n n , (19) gives
Σ ^ 0 : ν = k θ k ν Σ ^ 0 : θ + k ν − k θ k ν Σ ^ θ : ν , Σ ^ θ : 1 = k ν − k θ n − k θ Σ ^ θ : ν + n − k ν n − k θ Σ ^ ν : 1 . (40) \widehat{\Sigma}_{0:\nu}=\frac{k_\theta}{k_\nu}\widehat{\Sigma}_{0:\theta}+\frac{k_\nu-k_\theta}{k_\nu}\widehat{\Sigma}_{\theta:\nu},\qquad
\widehat{\Sigma}_{\theta:1}=\frac{k_\nu-k_\theta}{n-k_\theta}\widehat{\Sigma}_{\theta:\nu}+\frac{n-k_\nu}{n-k_\theta}\widehat{\Sigma}_{\nu:1}. \tag*{(40)} Σ 0 : ν = k ν k θ Σ 0 : θ + k ν k ν − k θ Σ θ : ν , Σ θ : 1 = n − k θ k ν − k θ Σ θ : ν + n − k θ n − k ν Σ ν : 1 . (40)
Define
A : = Σ ^ 0 : θ + ε I q , B : = Σ ^ θ : ν + ε I q , C : = Σ ^ ν : 1 + ε I q , A := \widehat{\Sigma}_{0:\theta} + \varepsilon I_q,\qquad B := \widehat{\Sigma}_{\theta:\nu} + \varepsilon I_q,\qquad C := \widehat{\Sigma}_{\nu:1} + \varepsilon I_q, A := Σ 0 : θ + ε I q , B := Σ θ : ν + ε I q , C := Σ ν : 1 + ε I q ,
λ 1 : = k ν − k θ k ν , λ 2 : = n − k ν n − k θ , \lambda_1 := \frac{k_\nu-k_\theta}{k_\nu},\qquad\lambda_2 := \frac{n-k_\nu}{n-k_\theta}, λ 1 := k ν k ν − k θ , λ 2 := n − k θ n − k ν ,
so that Σ ^ 0 : ν + ε I q = ( 1 − λ 1 ) A + λ 1 B \widehat{\Sigma}_{0:\nu}+\varepsilon I_q=(1-\lambda_1)A+\lambda_1B Σ 0 : ν + ε I q = ( 1 − λ 1 ) A + λ 1 B and Σ ^ θ : 1 + ε I q = ( 1 − λ 2 ) B + λ 2 C \widehat{\Sigma}_{\theta:1}+\varepsilon I_q=(1-\lambda_2)B+\lambda_2C Σ θ : 1 + ε I q = ( 1 − λ 2 ) B + λ 2 C .
Using the Jensen gap G G G of Lemma 44, direct expansion yields
J n ( ν ) − J n ( θ ) = k ν n G ( A , B ; λ 1 ) − n − k θ n G ( B , C ; λ 2 ) + r n ( ν ) , (41) J_n(\nu)-J_n(\theta)=\frac{k_\nu}{n}G(A,B;\lambda_1)-\frac{n-k_\theta}{n}G(B,C;\lambda_2)+r_n(\nu), \tag*{(41)} J n ( ν ) − J n ( θ ) = n k ν G ( A , B ; λ 1 ) − n n − k θ G ( B , C ; λ 2 ) + r n ( ν ) , (41)
where the floor-weight remainder is
r n ( ν ) : = ( ν − k ν n ) ( ℓ n ( 0 , ν ) − ℓ n ( ν , 1 ) ) − ( θ − k θ n ) ( ℓ n ( 0 , θ ) − ℓ n ( θ , 1 ) ) . (42) r_n(\nu):=\left(\nu-\frac{k_\nu}{n}\right)\bigl(\ell_n(0,\nu)-\ell_n(\nu,1)\bigr)-\left(\theta-\frac{k_\theta}{n}\right)\bigl(\ell_n(0,\theta)-\ell_n(\theta,1)\bigr). \tag*{(42)} r n ( ν ) := ( ν − n k ν ) ( ℓ n ( 0 , ν ) − ℓ n ( ν , 1 ) ) − ( θ − n k θ ) ( ℓ n ( 0 , θ ) − ℓ n ( θ , 1 ) ) . (42)
For fixed ν \nu ν , Lemma 36 supplies a common finite spectral ceiling for the four blocks in (42) . The ridge supplies a positive floor when ε > 0 \varepsilon>0 ε > 0 ; when ε = 0 \varepsilon=0 ε = 0 , (PD), Definition 5, and the decomposition (19) of the crossing block supply one. Thus there are constants c ν , C ν > 0 c_\nu,C_\nu>0 c ν , C ν > 0 such that, for all large enough n n n ,
c ν I q ⪯ Σ ^ a : b + ε I q ⪯ C ν I q , ( a , b ) ∈ { ( 0 , ν ) , ( ν , 1 ) , ( 0 , θ ) , ( θ , 1 ) } . (43) c_\nu I_q\preceq\widehat{\Sigma}_{a:b}+\varepsilon I_q\preceq C_\nu I_q,\qquad(a,b)\in\{(0,\nu),(\nu,1),(0,\theta),(\theta,1)\}. \tag*{(43)} c ν I q ⪯ Σ a : b + ε I q ⪯ C ν I q , ( a , b ) ∈ {( 0 , ν ) , ( ν , 1 ) , ( 0 , θ ) , ( θ , 1 )} . (43)
The sandwich (43) bounds the four log-determinants. Since 0 ≤ ν − k ν / n < 1 / n 0\le\nu-k_\nu/n<1/n 0 ≤ ν − k ν / n < 1/ n and 0 ≤ θ − k θ / n < 1 / n 0\le\theta-k_\theta/n<1/n 0 ≤ θ − k θ / n < 1/ n in (42) ,
r n ( ν ) ⟶ 0 for every fixed ν ∈ ( 0 , 1 ) . (44) r_n(\nu)\longrightarrow0\qquad\text{for every fixed }\nu\in(0,1). \tag*{(44)} r n ( ν ) ⟶ 0 for every fixed ν ∈ ( 0 , 1 ) . (44)
Since the macroscopic intervals underlying A A A and B B B belong to I θ b r \mathcal{I}^{\mathrm{br}}_\theta I θ br , while those underlying B B B and C C C belong to I θ c u t \mathcal{I}^{\mathrm{cut}}_\theta I θ cut , equation (2) gives
lim sup n → ∞ ( λ max ( A ) ∨ λ max ( B ) ) ≤ M + ε , lim inf n → ∞ ( λ min ( B ) ∧ λ min ( C ) ) ≥ m + ε > 0. \limsup_{n\to\infty}\bigl(\lambda_{\max}(A)\vee\lambda_{\max}(B)\bigr)\le M+\varepsilon,\qquad
\liminf_{n\to\infty}\bigl(\lambda_{\min}(B)\wedge\lambda_{\min}(C)\bigr)\ge m+\varepsilon>0. n → ∞ lim sup ( λ m a x ( A ) ∨ λ m a x ( B ) ) ≤ M + ε , n → ∞ lim inf ( λ m i n ( B ) ∧ λ m i n ( C ) ) ≥ m + ε > 0.
Here m + ε > 0 m+\varepsilon>0 m + ε > 0 follows from the ridge when ε > 0 \varepsilon>0 ε > 0 , or from (PD) when ε = 0 \varepsilon=0 ε = 0 . The complementary bounds required in (33), a floor for A , B A,B A , B and a ceiling for B , C B,C B , C , follow from the ridge or (PD) and from Lemma 36, respectively. Since the blocks underlying A A A and B B B lie in opposite regimes, whereas those underlying B B B and C C C lie in the same regime, equations (3) and (4) give
lim inf n → ∞ ∥ A − B ∥ F ≥ Δ b e t w e e n , lim sup n → ∞ ∥ B − C ∥ F ≤ Δ w i t h i n , \liminf_{n\to\infty}\lVert A-B\rVert_F\ge\Delta_{\mathrm{between}},\qquad
\limsup_{n\to\infty}\lVert B-C\rVert_F\le\Delta_{\mathrm{within}}, n → ∞ lim inf ∥ A − B ∥ F ≥ Δ between , n → ∞ lim sup ∥ B − C ∥ F ≤ Δ within ,
where the ε I q \varepsilon I_q ε I q terms cancel in both differences. Lemma 44 now yields the two estimates below.
Therefore
lim inf n → ∞ G ( A , B ; λ 1 ) ≥ θ ( ν − θ ) 2 ν 2 ( M + ε ) 2 Δ b e t w e e n 2 , \liminf_{n\to\infty}G(A,B;\lambda_1)\ge\frac{\theta(\nu-\theta)}{2\nu^2(M+\varepsilon)^2}\Delta_{\mathrm{between}}^2, n → ∞ lim inf G ( A , B ; λ 1 ) ≥ 2 ν 2 ( M + ε ) 2 θ ( ν − θ ) Δ between 2 ,
lim sup n → ∞ G ( B , C ; λ 2 ) ≤ ( 1 − ν ) ( ν − θ ) 2 ( 1 − θ ) 2 ( m + ε ) 2 Δ w i t h i n 2 . \limsup_{n\to\infty}G(B,C;\lambda_2)\le\frac{(1-\nu)(\nu-\theta)}{2(1-\theta)^2(m+\varepsilon)^2}\Delta_{\mathrm{within}}^2. n → ∞ lim sup G ( B , C ; λ 2 ) ≤ 2 ( 1 − θ ) 2 ( m + ε ) 2 ( 1 − ν ) ( ν − θ ) Δ within 2 .
Introduce the normalized signal and noise levels S : = Δ b e t w e e n 2 / ( M + ε ) 2 S:=\Delta_{\mathrm{between}}^2/(M+\varepsilon)^2 S := Δ between 2 / ( M + ε ) 2 and W : = Δ w i t h i n 2 / ( m + ε ) 2 W:=\Delta_{\mathrm{within}}^2/(m+\varepsilon)^2 W := Δ within 2 / ( m + ε ) 2 . Inserting the last two estimates into (41) and using (44) gives
lim inf n → ∞ ( J n ( ν ) − J n ( θ ) ) ≥ 1 2 T + ( ν ) ( ν − θ ) , T + ( ν ) : = θ ν S − 1 − ν 1 − θ W , ν > θ . (45) \liminf_{n\to\infty}\bigl(J_n(\nu)-J_n(\theta)\bigr)\ge\frac{1}{2}T_+(\nu)(\nu-\theta),\qquad
T_+(\nu):=\frac{\theta}{\nu}S-\frac{1-\nu}{1-\theta}W,\qquad\nu>\theta. \tag*{(45)} n → ∞ lim inf ( J n ( ν ) − J n ( θ ) ) ≥ 2 1 T + ( ν ) ( ν − θ ) , T + ( ν ) := ν θ S − 1 − θ 1 − ν W , ν > θ . (45)
Applying the same calculation to the time-reversed sequence, equivalently interchanging left and right and replacing ( ν , θ ) (\nu,\theta) ( ν , θ ) by ( 1 − ν , 1 − θ ) (1-\nu,1-\theta) ( 1 − ν , 1 − θ ) , yields
lim inf n → ∞ ( J n ( ν ) − J n ( θ ) ) ≥ 1 2 T − ( ν ) ( θ − ν ) , T − ( ν ) : = 1 − θ 1 − ν S − ν θ W , ν < θ . (46) \liminf_{n\to\infty}\left(J_n(\nu)-J_n(\theta)\right)\ge\frac{1}{2}T_-(\nu)(\theta-\nu),\qquad T_-(\nu):=\frac{1-\theta}{1-\nu}S-\frac{\nu}{\theta}W,\qquad\nu<\theta. \tag*{(46)} n → ∞ lim inf ( J n ( ν ) − J n ( θ ) ) ≥ 2 1 T − ( ν ) ( θ − ν ) , T − ( ν ) := 1 − ν 1 − θ S − θ ν W , ν < θ . (46)
Argument for (i). Assumption (A3) is exactly S > W S>W S > W . Moreover, T + ( θ ) = T − ( θ ) = S − W > 0 T_+(\theta)=T_-(\theta)=S-W>0 T + ( θ ) = T − ( θ ) = S − W > 0 , so by continuity there is δ ˉ > 0 \bar{\delta}>0 δ ˉ > 0 such that min { T − ( ν ) , T + ( ν ) } ≥ ( S − W ) / 2 \min\{T_-(\nu),T_+(\nu)\}\ge(S-W)/2 min { T − ( ν ) , T + ( ν )} ≥ ( S − W ) /2 whenever ∣ ν − θ ∣ ≤ δ ˉ |\nu-\theta|\le\bar{\delta} ∣ ν − θ ∣ ≤ δ ˉ . The two inequalities (45) and (46) prove part (i) with
c = S − W 4 . (47) c=\frac{S-W}{4}. \tag*{(47)} c = 4 S − W . (47)
Argument for (ii). Assumption (A4) is equivalent to
4 θ ( 1 − θ ) S > W . (48) 4\theta(1-\theta)S>W. \tag*{(48)} 4 θ ( 1 − θ ) S > W . (48)
If W = 0 W=0 W = 0 , then S > 0 S>0 S > 0 and T + ( ν ) ≥ θ S T_+(\nu)\ge\theta S T + ( ν ) ≥ θ S , T − ( ν ) ≥ ( 1 − θ ) S T_-(\nu)\ge(1-\theta)S T − ( ν ) ≥ ( 1 − θ ) S . Suppose now that W > 0 W>0 W > 0 and set d : = 2 θ ( 1 − θ ) S W − W d:=2\sqrt{\theta(1-\theta)SW}-W d := 2 θ ( 1 − θ ) S W − W , which is positive by (48) . The arithmetic-geometric mean inequality gives, for ν ≥ θ \nu\ge\theta ν ≥ θ and for ν ≤ θ \nu\le\theta ν ≤ θ respectively,
T + ( ν ) = θ S 1 − ν + ν W 1 − θ − W 1 − θ ≥ d 1 − θ , T_+(\nu)=\frac{\theta S}{1-\nu}+\frac{\nu W}{1-\theta}-\frac{W}{1-\theta}\ge\frac{d}{1-\theta}, T + ( ν ) = 1 − ν θ S + 1 − θ ν W − 1 − θ W ≥ 1 − θ d ,
T − ( ν ) = ( 1 − θ ) S 1 − ν + ( 1 − ν ) W θ − W θ ≥ d θ . T_-(\nu)=\frac{(1-\theta)S}{1-\nu}+\frac{(1-\nu)W}{\theta}-\frac{W}{\theta}\ge\frac{d}{\theta}. T − ( ν ) = 1 − ν ( 1 − θ ) S + θ ( 1 − ν ) W − θ W ≥ θ d .
Thus inequalities (45) and (46) prove (9). One may take
c = { 1 2 min { θ , 1 − θ } S , W = 0 , d 2 max { θ , 1 − θ } , W > 0. (49) c=
\begin{cases}
\frac{1}{2}\min\{\theta,1-\theta\}S, & W=0,\\
\frac{d}{2\max\{\theta,1-\theta\}}, & W>0.
\end{cases} \tag*{(49)} c = { 2 1 min { θ , 1 − θ } S , 2 m a x { θ , 1 − θ } d , W = 0 , W > 0. (49)
Consistency of the minimizers. Let θ n : = ⌊ θ n ⌋ / n \theta_n:=\lfloor\theta n\rfloor/n θ n := ⌊ θ n ⌋ / n . Then θ n ∈ D n \theta_n\in\mathcal{D}_n θ n ∈ D n for all large enough n n n and θ n → θ \theta_n\to\theta θ n → θ . Lemma 41 gives
J n ( θ n ) − J n ( θ ) = o ( 1 ) . (50) J_n(\theta_n)-J_n(\theta)=o(1). \tag*{(50)} J n ( θ n ) − J n ( θ ) = o ( 1 ) . (50)
First consider the trimmed estimator, which is the only assertion made when ε = 0 \varepsilon=0 ε = 0 . Fix 0 < δ < min { θ , 1 − θ } 0<\delta<\min\{\theta,1-\theta\} 0 < δ < min { θ , 1 − θ } and let θ ^ n , δ ∈ arg min ν ∈ D n ( δ ) J n ( ν ) \widehat{\theta}_{n,\delta}\in\arg\min_{\nu\in\mathcal{D}_n(\delta)}J_n(\nu) θ n , δ ∈ arg min ν ∈ D n ( δ ) J n ( ν ) . For all large enough n n n , also θ n ∈ D n ( δ ) \theta_n\in\mathcal{D}_n(\delta) θ n ∈ D n ( δ ) . Every subsequence has a further subsequence, again denoted by θ ^ n , δ \widehat{\theta}_{n,\delta} θ n , δ , converging to some ν ∈ [ δ , 1 − δ ] \nu\in[\delta,1-\delta] ν ∈ [ δ , 1 − δ ] . If ν ≠ θ \nu\ne\theta ν = θ , then Lemma 41, (9), and (50) give
lim inf n → ∞ ( J n ( θ ^ n , δ ) − J n ( θ n ) ) = lim inf n → ∞ ( J n ( ν ) − J n ( θ ) ) ≥ c ∣ ν − θ ∣ > 0. \liminf_{n\to\infty}\left(J_n(\widehat{\theta}_{n,\delta})-J_n(\theta_n)\right)=\liminf_{n\to\infty}\left(J_n(\nu)-J_n(\theta)\right)\ge c|\nu-\theta|>0. n → ∞ lim inf ( J n ( θ n , δ ) − J n ( θ n ) ) = n → ∞ lim inf ( J n ( ν ) − J n ( θ ) ) ≥ c ∣ ν − θ ∣ > 0.
This contradicts the minimality of θ ^ n , δ \widehat{\theta}_{n,\delta} θ n , δ . Hence every subsequential limit equals θ \theta θ , and θ ^ n , δ → θ \widehat{\theta}_{n,\delta}\to\theta θ n , δ → θ .
Now assume ε > 0 \varepsilon>0 ε > 0 and let θ ^ n ∈ arg min ν ∈ D n J n ( ν ) \widehat{\theta}_n\in\arg\min_{\nu\in\mathcal{D}_n}J_n(\nu) θ n ∈ arg min ν ∈ D n J n ( ν ) . The preceding argument excludes every interior subsequential limit other than θ \theta θ . It remains to exclude the endpoints. We treat 1 1 1 ; the case 0 0 0 is symmetric. Suppose along a subsequence that θ ^ n → 1 \widehat{\theta}_n \to1 θ n → 1 . For all large enough n n n , θ ^ n > θ \widehat{\theta}_n > \theta θ n > θ . Set
A n : = Σ ^ 0 : θ + ε I q , B n : = Σ ^ θ : θ ^ n + ε I q , C n : = Σ ^ θ ^ n : 1 + ε I q , A_n := \widehat{\Sigma}_{0:\theta} + \varepsilon I_q,\qquad B_n := \widehat{\Sigma}_{\theta:\widehat{\theta}_n} + \varepsilon I_q,\qquad C_n := \widehat{\Sigma}_{\widehat{\theta}_n:1} + \varepsilon I_q, A n := Σ 0 : θ + ε I q , B n := Σ θ : θ n + ε I q , C n := Σ θ n : 1 + ε I q ,
λ 1 , n : = k θ ^ n − k θ k θ ^ n , λ 2 , n : = n − k θ ^ n n − k θ . \lambda_{1,n} := \frac{k_{\widehat{\theta}_n}-k_\theta}{k_{\widehat{\theta}_n}},\qquad\lambda_{2,n} := \frac{n-k_{\widehat{\theta}_n}}{n-k_\theta}. λ 1 , n := k θ n k θ n − k θ , λ 2 , n := n − k θ n − k θ n .
Applying equation (41) with ν = θ ^ n \nu=\widehat{\theta}_n ν = θ n gives
J n ( θ ^ n ) − J n ( θ ) = k θ ^ n n G ( A n , B n ; λ 1 , n ) − n − k θ ^ n n G ( B n , C n ; λ 2 , n ) + r n ( θ ^ n ) . (51) J_n(\widehat{\theta}_n)-J_n(\theta)=\frac{k_{\widehat{\theta}_n}}{n}G(A_n,B_n;\lambda_{1,n})-\frac{n-k_{\widehat{\theta}_n}}{n}G(B_n,C_n;\lambda_{2,n})+r_n(\widehat{\theta}_n). \tag*{(51)} J n ( θ n ) − J n ( θ ) = n k θ n G ( A n , B n ; λ 1 , n ) − n n − k θ n G ( B n , C n ; λ 2 , n ) + r n ( θ n ) . (51)
By (42),
r n ( θ ^ n ) = ( θ ^ n − k θ ^ n n ) ( ℓ n ( 0 , θ ^ n ) − ℓ n ( θ ^ n , 1 ) ) − ( θ − k θ n ) ( ℓ n ( 0 , θ ) − ℓ n ( θ , 1 ) ) . r_n(\widehat{\theta}_n)=\left(\widehat{\theta}_n-\frac{k_{\widehat{\theta}_n}}{n}\right)\bigl(\ell_n(0,\widehat{\theta}_n)-\ell_n(\widehat{\theta}_n,1)\bigr)-\left(\theta-\frac{k_\theta}{n}\right)\bigl(\ell_n(0,\theta)-\ell_n(\theta,1)\bigr). r n ( θ n ) = ( θ n − n k θ n ) ( ℓ n ( 0 , θ n ) − ℓ n ( θ n , 1 ) ) − ( θ − n k θ ) ( ℓ n ( 0 , θ ) − ℓ n ( θ , 1 ) ) .
Since θ ^ n ∈ D n \widehat{\theta}_n\in\mathcal{D}_n θ n ∈ D n , the first floor error vanishes, θ ^ n − k θ ^ n / n = 0 \widehat{\theta}_n-k_{\widehat{\theta}_n}/n=0 θ n − k θ n / n = 0 , and hence
r n ( θ ^ n ) = − ( θ − k θ n ) ( ℓ n ( 0 , θ ) − ℓ n ( θ , 1 ) ) = o ( 1 ) r_n(\widehat{\theta}_n)=-\left(\theta-\frac{k_\theta}{n}\right)\bigl(\ell_n(0,\theta)-\ell_n(\theta,1)\bigr)=o(1) r n ( θ n ) = − ( θ − n k θ ) ( ℓ n ( 0 , θ ) − ℓ n ( θ , 1 ) ) = o ( 1 )
by Assumption (A1) and the ridge lower bound.
By Lemma 42 and ∥ ⋅ ∥ F ≤ q ∥ ⋅ ∥ \lVert\cdot\rVert_F\le\sqrt{q}\lVert\cdot\rVert ∥ ⋅ ∥ F ≤ q ∥ ⋅ ∥ ,
∥ Σ ^ θ : θ ^ n − Σ ^ θ : 1 ∥ F ⟶ 0. (52) \left\lVert\widehat{\Sigma}_{\theta:\widehat{\theta}_n}-\widehat{\Sigma}_{\theta:1}\right\rVert_F\longrightarrow0. \tag*{(52)} Σ θ : θ n − Σ θ : 1 F ⟶ 0. (52)
For G ( A n , B n ; λ 1 , n ) G(A_n,B_n;\lambda_{1,n}) G ( A n , B n ; λ 1 , n ) , equation (52), the triangle inequality, and the definition of M M M first give
lim sup n → ∞ ∥ Σ ^ θ : θ ^ n ∥ ≤ lim sup n → ∞ ∥ Σ ^ θ : 1 ∥ ≤ M . \limsup_{n\to\infty}\left\lVert\widehat{\Sigma}_{\theta:\widehat{\theta}_n}\right\rVert\le\limsup_{n\to\infty}\left\lVert\widehat{\Sigma}_{\theta:1}\right\rVert\le M. n → ∞ lim sup Σ θ : θ n ≤ n → ∞ lim sup Σ θ : 1 ≤ M .
For the Frobenius gap, the ridge terms cancel and A n − B n = ( Σ ^ 0 : θ − Σ ^ θ : 1 ) − ( Σ ^ θ : θ ^ n − Σ ^ θ : 1 ) A_n-B_n=(\widehat{\Sigma}_{0:\theta}-\widehat{\Sigma}_{\theta:1})-(\widehat{\Sigma}_{\theta:\widehat{\theta}_n}-\widehat{\Sigma}_{\theta:1}) A n − B n = ( Σ 0 : θ − Σ θ : 1 ) − ( Σ θ : θ n − Σ θ : 1 ) . (52) , the reverse triangle inequality and the definition of Δ b e t w e e n \Delta_{\mathrm{between}} Δ between therefore yield
lim inf n → ∞ ∥ A n − B n ∥ F ≥ lim inf n → ∞ ∥ Σ ^ 0 : θ − Σ ^ θ : 1 ∥ F ≥ Δ b e t w e e n . \liminf_{n\to\infty}\lVert A_n-B_n\rVert_F\ge\liminf_{n\to\infty}\lVert\widehat{\Sigma}_{0:\theta}-\widehat{\Sigma}_{\theta:1}\rVert_F\ge\Delta_{\mathrm{between}}. n → ∞ lim inf ∥ A n − B n ∥ F ≥ n → ∞ lim inf ∥ Σ 0 : θ − Σ θ : 1 ∥ F ≥ Δ between .
Moreover, k θ ^ n / n → 1 k_{\widehat{\theta}_n}/n\to1 k θ n / n → 1 and λ 1 , n → 1 − θ \lambda_{1,n}\to1-\theta λ 1 , n → 1 − θ . Lemma 44 therefore gives
lim inf n → ∞ k θ ^ n n G ( A n , B n ; λ 1 , n ) ≥ θ ( 1 − θ ) 2 ( M + ε ) 2 Δ b e t w e e n 2 > 0. (53) \liminf_{n\to\infty}\frac{k_{\widehat{\theta}_n}}{n}G(A_n,B_n;\lambda_{1,n})\ge\frac{\theta(1-\theta)}{2(M+\varepsilon)^2}\Delta_{\mathrm{between}}^2>0. \tag*{(53)} n → ∞ lim inf n k θ n G ( A n , B n ; λ 1 , n ) ≥ 2 ( M + ε ) 2 θ ( 1 − θ ) Δ between 2 > 0. (53)
For G ( B n , C n ; λ 2 , n ) G(B_n,C_n;\lambda_{2,n}) G ( B n , C n ; λ 2 , n ) , identity (19) gives ( 1 − λ 2 , n ) B n + λ 2 , n C n = Σ ^ θ : 1 + ε I q (1-\lambda_{2,n})B_n+\lambda_{2,n}C_n=\widehat{\Sigma}_{\theta:1}+\varepsilon I_q ( 1 − λ 2 , n ) B n + λ 2 , n C n = Σ θ : 1 + ε I q , and hence
G ( B n , C n ; λ 2 , n ) = ℓ n ( θ , 1 ) − log det B n + λ 2 , n ( log det B n − log det C n ) . G(B_n,C_n;\lambda_{2,n})=\ell_n(\theta,1)-\log\det B_n+\lambda_{2,n}(\log\det B_n-\log\det C_n). G ( B n , C n ; λ 2 , n ) = ℓ n ( θ , 1 ) − log det B n + λ 2 , n ( log det B n − log det C n ) .
The first difference is o ( 1 ) o(1) o ( 1 ) by Lemma 42, and λ 2 , n log det B n = o ( 1 ) \lambda_{2,n}\log\det B_n=o(1) λ 2 , n log det B n = o ( 1 ) because λ 2 , n → 0 \lambda_{2,n}\to0 λ 2 , n → 0 and log det B n \log\det B_n log det B n is bounded. Finally, for a constant independent of n n n ,
λ 2 , n ∣ log det C n ∣ ≤ C ( 1 − θ ^ n ) ∣ log det ( Σ ^ θ ^ n : 1 + ε I q ) ∣ ⟶ 0 \lambda_{2,n}\lvert\log\det C_n\rvert\le C(1-\widehat{\theta}_n)\left\lvert\log\det\left(\widehat{\Sigma}_{\widehat{\theta}_n:1}+\varepsilon I_q\right)\right\rvert\longrightarrow0 λ 2 , n ∣ log det C n ∣ ≤ C ( 1 − θ n ) log det ( Σ θ n : 1 + ε I q ) ⟶ 0
by Lemma 39. Thus G ( B n , C n ; λ 2 , n ) = o ( 1 ) G(B_n,C_n;\lambda_{2,n})=o(1) G ( B n , C n ; λ 2 , n ) = o ( 1 ) .
Together with (51) and (53), this gives lim inf n → ∞ ( J n ( θ ^ n ) − J n ( θ ) ) > 0 \liminf_{n\to\infty}(J_n(\widehat{\theta}_n)-J_n(\theta))>0 lim inf n → ∞ ( J n ( θ n ) − J n ( θ )) > 0 . (50) then implies J n ( θ ^ n ) > J n ( θ n ) J_n(\widehat{\theta}_n)>J_n(\theta_n) J n ( θ n ) > J n ( θ n ) for all large enough n n n , contradicting minimality over D n \mathcal{D}_n D n . Hence neither endpoint is a subsequential limit, so θ ^ n → θ \widehat{\theta}_n\to\theta θ n → θ . □ \square □
Proof of Proposition 24. Let us define a = 1 + K + r a = 1 + K + r a = 1 + K + r , b = 1 + K b = 1 + K b = 1 + K , and c = 1 c = 1 c = 1 and ν ⋆ = 0.5 \nu_{\star} = 0.5 ν ⋆ = 0.5 .
Argument for (11) and (12): To compute Δ w i t h i n \Delta_{\mathrm{within}} Δ within and Δ b e t w e e n \Delta_{\mathrm{between}} Δ between we observe that for q = 1 q = 1 q = 1 the block second moment on a macroscopic interval is just the block average of y 2 y^2 y 2 , and the Frobenius norm ∥ ⋅ ∥ F \lVert\cdot\rVert_F ∥ ⋅ ∥ F is just the absolute value. On the left regime [ 0 , θ ] [0,\theta] [ 0 , θ ] the profile is constant with second moment equal to a a a , so all left-block averages equal a a a and therefore the within-left fluctuation is 0. On the right regime ( θ , 1 ] (\theta,1] ( θ , 1 ] , the profile takes the two macroscopic values b b b and c c c on ( θ , ν ⋆ ] (\theta,\nu_{\star}] ( θ , ν ⋆ ] and ( ν ⋆ , 1 ] (\nu_{\star},1] ( ν ⋆ , 1 ] , respectively. Choosing one block entirely in the b b b -part and one entirely in the c c c -part yields
Δ w i t h i n = ∣ b − c ∣ = K . \Delta_{\mathrm{within}} = \lvert b - c \rvert= K. Δ within = ∣ b − c ∣ = K .
For Δ b e t w e e n \Delta_{\mathrm{between}} Δ between , any left-block average equals a a a , while any right-block average lies in [ c , b ] [c,b] [ c , b ] (since it is an average of b b b and c c c ). Hence the minimal separation is attained by a block inside the b b b -part:
Δ b e t w e e n = ∣ a − b ∣ = r . \Delta_{\mathrm{between}} = \lvert a - b \rvert= r. Δ between = ∣ a − b ∣ = r .
At the candidate ν ⋆ = 1 2 \nu_{\star} = \frac{1}{2} ν ⋆ = 2 1 , every left-block average lies in [ b , a ] [b,a] [ b , a ] , with both endpoints attained by blocks contained in the b b b - and a a a -parts, while every right-block average equals c c c . Hence
Δ w i t h i n ( ν ⋆ ) = a − b = r , Δ b e t w e e n ( ν ⋆ ) = b − c = K . \Delta_{\mathrm{within}}(\nu_{\star}) = a - b = r,\qquad\Delta_{\mathrm{between}}(\nu_{\star}) = b - c = K. Δ within ( ν ⋆ ) = a − b = r , Δ between ( ν ⋆ ) = b − c = K .
Together with the preceding calculation, this proves (11) and (12).
In the next step, we explicitly calculate J ∞ J_{\infty} J ∞ . Here J ∞ J_{\infty} J ∞ denotes the pointwise limit of J n J_n J n for the three-level series (10). Since Assumption (A2) fails for this series, this is not the two-regime function of Definition 20; it is, however, given by the same formula, with Σ 0 : ν \Sigma_{0:\nu} Σ 0 : ν and Σ ν : 1 \Sigma_{\nu:1} Σ ν : 1 replaced by the block limits computed next. As our data is piecewise constant, direct averaging yields the following formulas:
Σ 0 : ν = { a , 0 < ν ≤ θ , θ a + ( ν − θ ) b ν , θ < ν ≤ ν ⋆ , θ a + ( ν ⋆ − θ ) b + ( ν − ν ⋆ ) c ν , ν ⋆ < ν < 1. \Sigma_{0:\nu} =
\begin{cases}
a, & 0 < \nu\le\theta,\\
\frac{\theta a + (\nu- \theta)b}{\nu}, & \theta< \nu\le\nu_{\star},\\
\frac{\theta a + (\nu_{\star} - \theta)b + (\nu- \nu_{\star})c}{\nu}, & \nu_{\star} < \nu< 1.
\end{cases} Σ 0 : ν = ⎩ ⎨ ⎧ a , ν θ a + ( ν − θ ) b , ν θ a + ( ν ⋆ − θ ) b + ( ν − ν ⋆ ) c , 0 < ν ≤ θ , θ < ν ≤ ν ⋆ , ν ⋆ < ν < 1.
and
Σ ν : 1 = { ( θ − ν ) a + ( ν ⋆ − θ ) b + ( 1 − ν ⋆ ) c 1 − ν , 0 < ν < θ , ( ν ⋆ − ν ) b + ( 1 − ν ⋆ ) c 1 − ν , θ ≤ ν < ν ⋆ , c , ν ⋆ ≤ ν < 1. \Sigma_{\nu:1} =
\begin{cases}
\frac{(\theta- \nu)a + (\nu_{\star} - \theta)b + (1 - \nu_{\star})c}{1 - \nu}, & 0 < \nu< \theta,\\
\frac{(\nu_{\star} - \nu)b + (1 - \nu_{\star})c}{1 - \nu}, & \theta\le\nu< \nu_{\star},\\
c, & \nu_{\star} \le\nu< 1.
\end{cases} Σ ν : 1 = ⎩ ⎨ ⎧ 1 − ν ( θ − ν ) a + ( ν ⋆ − θ ) b + ( 1 − ν ⋆ ) c , 1 − ν ( ν ⋆ − ν ) b + ( 1 − ν ⋆ ) c , c , 0 < ν < θ , θ ≤ ν < ν ⋆ , ν ⋆ ≤ ν < 1.
Using the definition φ ( x ) : = log ( x + ε ) \varphi(x) := \log(x + \varepsilon) φ ( x ) := log ( x + ε ) , the limiting objective function is therefore explicitly
J ∞ ( ν ) = ν φ ( Σ 0 : ν ) + ( 1 − ν ) φ ( Σ ν : 1 ) , ν ∈ ( 0 , 1 ) , J_{\infty}(\nu) = \nu\varphi(\Sigma_{0:\nu}) + (1 - \nu)\varphi(\Sigma_{\nu:1}),\qquad\nu\in(0,1), J ∞ ( ν ) = ν φ ( Σ 0 : ν ) + ( 1 − ν ) φ ( Σ ν : 1 ) , ν ∈ ( 0 , 1 ) ,
where Σ 0 : ν \Sigma_{0:\nu} Σ 0 : ν and Σ ν : 1 \Sigma_{\nu:1} Σ ν : 1 are given by the piecewise expressions above. In particular, at the two relevant points one has
J ∞ ( θ ) = θ φ ( a ) + ( 1 − θ ) φ ( ( ν ⋆ − θ ) b + ( 1 − ν ⋆ ) c 1 − θ ) , J_{\infty}(\theta) = \theta\varphi(a) + (1 - \theta)\varphi\left(\frac{(\nu_{\star} - \theta)b + (1 - \nu_{\star})c}{1 - \theta}\right), J ∞ ( θ ) = θ φ ( a ) + ( 1 − θ ) φ ( 1 − θ ( ν ⋆ − θ ) b + ( 1 − ν ⋆ ) c ) ,
J ∞ ( ν ⋆ ) = ν ⋆ φ ( θ a + ( ν ⋆ − θ ) b ν ⋆ ) + ( 1 − ν ⋆ ) φ ( c ) . J_{\infty}(\nu_{\star})=\nu_{\star}\varphi\left(\frac{\theta a+(\nu_{\star}-\theta)b}{\nu_{\star}}\right)+(1-\nu_{\star})\varphi(c). J ∞ ( ν ⋆ ) = ν ⋆ φ ( ν ⋆ θ a + ( ν ⋆ − θ ) b ) + ( 1 − ν ⋆ ) φ ( c ) .
We now show, for an arbitrary ε > 0 \varepsilon>0 ε > 0 , that the only candidates for a local minimizer of J ∞ J_{\infty} J ∞ are θ \theta θ and ν ⋆ = 0.5 \nu_{\star}=0.5 ν ⋆ = 0.5 and that the global minimum is attained at ν ⋆ \nu_{\star} ν ⋆ . As J ∞ J_{\infty} J ∞ is piecewise smooth, we consider several cases.
Case 1: If ν ∈ ( ν ⋆ , 1 ) \nu\in(\nu_{\star},1) ν ∈ ( ν ⋆ , 1 ) , then by definition
Σ 0 : ν = θ a + ( ν ⋆ − θ ) b + ( ν − ν ⋆ ) c ν > c = Σ ν : 1 . \Sigma_{0:\nu}=\frac{\theta a+(\nu_{\star}-\theta)b+(\nu-\nu_{\star})c}{\nu}>c=\Sigma_{\nu:1}. Σ 0 : ν = ν θ a + ( ν ⋆ − θ ) b + ( ν − ν ⋆ ) c > c = Σ ν : 1 .
(The inequality is strict because the displayed mixture assigns positive weight to a , b > c a,b>c a , b > c .) Hence
J ∞ ( ν ) = ν φ ( Σ 0 : ν ) + ( 1 − ν ) φ ( c ) , for ν ∈ ( ν ⋆ , 1 ) , J_{\infty}(\nu)=\nu\varphi(\Sigma_{0:\nu})+(1-\nu)\varphi(c),\qquad\text{for }\nu\in(\nu_{\star},1), J ∞ ( ν ) = ν φ ( Σ 0 : ν ) + ( 1 − ν ) φ ( c ) , for ν ∈ ( ν ⋆ , 1 ) ,
and differentiating gives
J ∞ ′ ( ν ) = φ ( Σ 0 : ν ) − φ ( c ) + ν φ ′ ( Σ 0 : ν ) Σ 0 : ν ′ , J_{\infty}'(\nu)=\varphi(\Sigma_{0:\nu})-\varphi(c)+\nu\varphi'(\Sigma_{0:\nu})\Sigma'_{0:\nu}, J ∞ ′ ( ν ) = φ ( Σ 0 : ν ) − φ ( c ) + ν φ ′ ( Σ 0 : ν ) Σ 0 : ν ′ ,
where
Σ 0 : ν ′ = − θ a + ( ν ⋆ − θ ) b − ν ⋆ c ν 2 = − Σ 0 : ν ν + c ν . \Sigma'_{0:\nu}=-\frac{\theta a+(\nu_{\star}-\theta)b-\nu_{\star}c}{\nu^{2}}=-\frac{\Sigma_{0:\nu}}{\nu}+\frac{c}{\nu}. Σ 0 : ν ′ = − ν 2 θ a + ( ν ⋆ − θ ) b − ν ⋆ c = − ν Σ 0 : ν + ν c .
Using strict concavity of φ \varphi φ ,
φ ( Σ 0 : ν ) − φ ( c ) > φ ′ ( Σ 0 : ν ) ( Σ 0 : ν − c ) , \varphi(\Sigma_{0:\nu})-\varphi(c)>\varphi'(\Sigma_{0:\nu})(\Sigma_{0:\nu}-c), φ ( Σ 0 : ν ) − φ ( c ) > φ ′ ( Σ 0 : ν ) ( Σ 0 : ν − c ) ,
we obtain for all ν ∈ ( ν ⋆ , 1 ) \nu\in(\nu_{\star},1) ν ∈ ( ν ⋆ , 1 ) that
J ∞ ′ ( ν ) > φ ′ ( Σ 0 : ν ) ( Σ 0 : ν − c ) + ν φ ′ ( Σ 0 : ν ) Σ 0 : ν ′ = φ ′ ( Σ 0 : ν ) [ ( Σ 0 : ν − c ) + ν Σ 0 : ν ′ ] = 0. \begin{aligned}
J_{\infty}'(\nu)&>\varphi'(\Sigma_{0:\nu})(\Sigma_{0:\nu}-c)+\nu\varphi'(\Sigma_{0:\nu})\Sigma'_{0:\nu}\\
&=\varphi'(\Sigma_{0:\nu})\left[(\Sigma_{0:\nu}-c)+\nu\Sigma'_{0:\nu}\right]=0.
\end{aligned} J ∞ ′ ( ν ) > φ ′ ( Σ 0 : ν ) ( Σ 0 : ν − c ) + ν φ ′ ( Σ 0 : ν ) Σ 0 : ν ′ = φ ′ ( Σ 0 : ν ) [ ( Σ 0 : ν − c ) + ν Σ 0 : ν ′ ] = 0.
Thus J ∞ J_{\infty} J ∞ is strictly increasing on ( ν ⋆ , 1 ) (\nu_{\star},1) ( ν ⋆ , 1 ) , so there is no local minimizer in ( ν ⋆ , 1 ) (\nu_{\star},1) ( ν ⋆ , 1 ) and ν ⋆ \nu_{\star} ν ⋆ is a (one-sided) local minimizer from the right.
Case 2: The argument for ν ∈ ( 0 , θ ) \nu\in(0,\theta) ν ∈ ( 0 , θ ) is similar to Case 1. For the convenience of the reader we carry out all details. For ν ∈ ( 0 , θ ) \nu\in(0,\theta) ν ∈ ( 0 , θ ) ,
Σ 0 : ν = a and Σ ν : 1 = ( θ − ν ) a + ( ν ⋆ − θ ) b + ( 1 − ν ⋆ ) c 1 − ν . \Sigma_{0:\nu}=a\qquad\text{and}\qquad\Sigma_{\nu:1}=\frac{(\theta-\nu)a+(\nu_{\star}-\theta)b+(1-\nu_{\star})c}{1-\nu}. Σ 0 : ν = a and Σ ν : 1 = 1 − ν ( θ − ν ) a + ( ν ⋆ − θ ) b + ( 1 − ν ⋆ ) c .
Hence J ∞ J_{\infty} J ∞ is smooth on ( 0 , θ ) (0,\theta) ( 0 , θ ) and given by (recalling φ ( x ) = log ( x + ε ) \varphi(x)=\log(x+\varepsilon) φ ( x ) = log ( x + ε ) )
J ∞ ( ν ) = ν φ ( a ) + ( 1 − ν ) φ ( Σ ν : 1 ) , ν ∈ ( 0 , θ ) . J_{\infty}(\nu)=\nu\varphi(a)+(1-\nu)\varphi(\Sigma_{\nu:1}),\qquad\nu\in(0,\theta). J ∞ ( ν ) = ν φ ( a ) + ( 1 − ν ) φ ( Σ ν : 1 ) , ν ∈ ( 0 , θ ) .
Its derivative is given by
J ∞ ′ ( ν ) = φ ( a ) − φ ( Σ ν : 1 ) + ( 1 − ν ) φ ′ ( Σ ν : 1 ) Σ ν : 1 ′ . (54) J_{\infty}'(\nu)=\varphi(a)-\varphi(\Sigma_{\nu:1})+(1-\nu)\varphi'(\Sigma_{\nu:1})\Sigma'_{\nu:1}. \tag*{(54)} J ∞ ′ ( ν ) = φ ( a ) − φ ( Σ ν : 1 ) + ( 1 − ν ) φ ′ ( Σ ν : 1 ) Σ ν : 1 ′ . (54)
where
Σ ν : 1 ′ = − a 1 − ν + Σ ν : 1 1 − ν . (55) \Sigma'_{\nu:1}=-\frac{a}{1-\nu}+\frac{\Sigma_{\nu:1}}{1-\nu}. \tag*{(55)} Σ ν : 1 ′ = − 1 − ν a + 1 − ν Σ ν : 1 . (55)
Using the strict concavity of φ \varphi φ in combination with a > Σ ν : 1 a>\Sigma_{\nu:1} a > Σ ν : 1 (a strict mixture with positive weight on b , c < a b,c<a b , c < a ), we obtain
φ ( a ) − φ ( Σ ν : 1 ) < φ ′ ( Σ ν : 1 ) ( a − Σ ν : 1 ) . (56) \varphi(a)-\varphi(\Sigma_{\nu:1})<\varphi'(\Sigma_{\nu:1})(a-\Sigma_{\nu:1}). \tag*{(56)} φ ( a ) − φ ( Σ ν : 1 ) < φ ′ ( Σ ν : 1 ) ( a − Σ ν : 1 ) . (56)
Inserting (56) into (54) and using (55), we obtain
J ∞ ′ ( ν ) < φ ′ ( Σ ν : 1 ) ( a − Σ ν : 1 ) + φ ′ ( Σ ν : 1 ) ( 1 − ν ) Σ ν : 1 ′ = 0. J_{\infty}'(\nu)<\varphi'(\Sigma_{\nu:1})(a-\Sigma_{\nu:1})+\varphi'(\Sigma_{\nu:1})(1-\nu)\Sigma'_{\nu:1}=0. J ∞ ′ ( ν ) < φ ′ ( Σ ν : 1 ) ( a − Σ ν : 1 ) + φ ′ ( Σ ν : 1 ) ( 1 − ν ) Σ ν : 1 ′ = 0.
Thus J ∞ ′ ( ν ) < 0 J_{\infty}'(\nu)<0 J ∞ ′ ( ν ) < 0 for all ν ∈ ( 0 , θ ) \nu\in(0,\theta) ν ∈ ( 0 , θ ) , and in particular there cannot be a local minimizer in ( 0 , θ ) (0,\theta) ( 0 , θ ) .
Case 3: For ν ∈ ( θ , ν ⋆ ) \nu\in(\theta,\nu_\star) ν ∈ ( θ , ν ⋆ ) , we will show below that
J ∞ ′ ′ ( ν ) < 0 , (57) J_\infty''(\nu) < 0, \tag*{(57)} J ∞ ′′ ( ν ) < 0 , (57)
so J ∞ J_\infty J ∞ is strictly concave on ( θ , ν ⋆ ) (\theta,\nu_\star) ( θ , ν ⋆ ) and therefore has at most one critical point there; if it exists, it is necessarily a strict local maximum. In particular, there is no local minimizer inside ( θ , ν ⋆ ) (\theta,\nu_\star) ( θ , ν ⋆ ) . Therefore let us move to the verification of (57) . For ν ∈ ( θ , ν ⋆ ) \nu\in(\theta,\nu_\star) ν ∈ ( θ , ν ⋆ ) we have
Σ 0 : ν = b + θ ( a − b ) ν , Σ ν : 1 = c + ( ν ⋆ − ν ) ( b − c ) 1 − ν . \Sigma_{0:\nu}=b+\frac{\theta(a-b)}{\nu}, \qquad\Sigma_{\nu:1}=c+\frac{(\nu_\star-\nu)(b-c)}{1-\nu}. Σ 0 : ν = b + ν θ ( a − b ) , Σ ν : 1 = c + 1 − ν ( ν ⋆ − ν ) ( b − c ) .
Recalling φ ( x ) = log ( x + ε ) \varphi(x)=\log(x+\varepsilon) φ ( x ) = log ( x + ε ) , we can write
J ∞ ( ν ) = ν φ ( Σ 0 : ν ) + ( 1 − ν ) φ ( Σ ν : 1 ) , ν ∈ ( θ , ν ⋆ ) . J_\infty(\nu)=\nu\varphi(\Sigma_{0:\nu})+(1-\nu)\varphi(\Sigma_{\nu:1}), \qquad\nu\in(\theta,\nu_\star). J ∞ ( ν ) = ν φ ( Σ 0 : ν ) + ( 1 − ν ) φ ( Σ ν : 1 ) , ν ∈ ( θ , ν ⋆ ) .
A direct differentiation gives
J ∞ ′ ( ν ) = φ ( Σ 0 : ν ) + ν φ ′ ( Σ 0 : ν ) Σ 0 : ν ′ − φ ( Σ ν : 1 ) + ( 1 − ν ) φ ′ ( Σ ν : 1 ) Σ ν : 1 ′ . J_\infty'(\nu)=\varphi(\Sigma_{0:\nu})+\nu\varphi'(\Sigma_{0:\nu})\Sigma'_{0:\nu}-\varphi(\Sigma_{\nu:1})+(1-\nu)\varphi'(\Sigma_{\nu:1})\Sigma'_{\nu:1}. J ∞ ′ ( ν ) = φ ( Σ 0 : ν ) + ν φ ′ ( Σ 0 : ν ) Σ 0 : ν ′ − φ ( Σ ν : 1 ) + ( 1 − ν ) φ ′ ( Σ ν : 1 ) Σ ν : 1 ′ .
Differentiating once more,
J ∞ ′ ′ ( ν ) = 2 φ ′ ( Σ 0 : ν ) Σ 0 : ν ′ + ν ( φ ′ ′ ( Σ 0 : ν ) ( Σ 0 : ν ′ ) 2 + φ ′ ( Σ 0 : ν ) Σ 0 : ν ′ ′ ) − 2 φ ′ ( Σ ν : 1 ) Σ ν : 1 ′ + ( 1 − ν ) ( φ ′ ′ ( Σ ν : 1 ) ( Σ ν : 1 ′ ) 2 + φ ′ ( Σ ν : 1 ) Σ ν : 1 ′ ′ ) . \begin{aligned}
J_\infty''(\nu)={}&2\varphi'(\Sigma_{0:\nu})\Sigma'_{0:\nu}+\nu\left(\varphi''(\Sigma_{0:\nu})(\Sigma'_{0:\nu})^2+\varphi'(\Sigma_{0:\nu})\Sigma''_{0:\nu}\right)\\
&-2\varphi'(\Sigma_{\nu:1})\Sigma'_{\nu:1}+(1-\nu)\left(\varphi''(\Sigma_{\nu:1})(\Sigma'_{\nu:1})^2+\varphi'(\Sigma_{\nu:1})\Sigma''_{\nu:1}\right).
\end{aligned} J ∞ ′′ ( ν ) = 2 φ ′ ( Σ 0 : ν ) Σ 0 : ν ′ + ν ( φ ′′ ( Σ 0 : ν ) ( Σ 0 : ν ′ ) 2 + φ ′ ( Σ 0 : ν ) Σ 0 : ν ′′ ) − 2 φ ′ ( Σ ν : 1 ) Σ ν : 1 ′ + ( 1 − ν ) ( φ ′′ ( Σ ν : 1 ) ( Σ ν : 1 ′ ) 2 + φ ′ ( Σ ν : 1 ) Σ ν : 1 ′′ ) .
Direct computation shows
Σ 0 : ν ′ = − θ ( a − b ) ν 2 , Σ 0 : ν ′ ′ = 2 θ ( a − b ) ν 3 , Σ ν : 1 ′ = − ( 1 − ν ⋆ ) ( b − c ) ( 1 − ν ) 2 , Σ ν : 1 ′ ′ = − 2 ( 1 − ν ⋆ ) ( b − c ) ( 1 − ν ) 3 . \Sigma'_{0:\nu}=-\frac{\theta(a-b)}{\nu^2}, \qquad\Sigma''_{0:\nu}=\frac{2\theta(a-b)}{\nu^3}, \qquad
\Sigma'_{\nu:1}=-\frac{(1-\nu_\star)(b-c)}{(1-\nu)^2}, \qquad\Sigma''_{\nu:1}=-\frac{2(1-\nu_\star)(b-c)}{(1-\nu)^3}. Σ 0 : ν ′ = − ν 2 θ ( a − b ) , Σ 0 : ν ′′ = ν 3 2 θ ( a − b ) , Σ ν : 1 ′ = − ( 1 − ν ) 2 ( 1 − ν ⋆ ) ( b − c ) , Σ ν : 1 ′′ = − ( 1 − ν ) 3 2 ( 1 − ν ⋆ ) ( b − c ) .
Hence we obtain the cancellations
2 Σ 0 : ν ′ + ν Σ 0 : ν ′ ′ = 0 , − 2 Σ ν : 1 ′ + ( 1 − ν ) Σ ν : 1 ′ ′ = 0. 2\Sigma'_{0:\nu}+\nu\Sigma''_{0:\nu}=0, \qquad-2\Sigma'_{\nu:1}+(1-\nu)\Sigma''_{\nu:1}=0. 2 Σ 0 : ν ′ + ν Σ 0 : ν ′′ = 0 , − 2 Σ ν : 1 ′ + ( 1 − ν ) Σ ν : 1 ′′ = 0.
This allows us to simplify J ∞ ′ ′ J_\infty'' J ∞ ′′ as the φ ′ \varphi' φ ′ -terms drop out and we are left with
J ∞ ′ ′ ( ν ) = ν φ ′ ′ ( Σ 0 : ν ) ( Σ 0 : ν ′ ) 2 + ( 1 − ν ) φ ′ ′ ( Σ ν : 1 ) ( Σ ν : 1 ′ ) 2 . J_\infty''(\nu)=\nu\varphi''(\Sigma_{0:\nu})(\Sigma'_{0:\nu})^2+(1-\nu)\varphi''(\Sigma_{\nu:1})(\Sigma'_{\nu:1})^2. J ∞ ′′ ( ν ) = ν φ ′′ ( Σ 0 : ν ) ( Σ 0 : ν ′ ) 2 + ( 1 − ν ) φ ′′ ( Σ ν : 1 ) ( Σ ν : 1 ′ ) 2 .
Finally, since φ ( x ) = log ( x + ε ) \varphi(x)=\log(x+\varepsilon) φ ( x ) = log ( x + ε ) satisfies
φ ′ ′ ( x ) = − ( x + ε ) − 2 < 0 , \varphi''(x)=-(x+\varepsilon)^{-2}<0, φ ′′ ( x ) = − ( x + ε ) − 2 < 0 ,
and since a − b = r > 0 a-b=r>0 a − b = r > 0 , the identity Σ 0 : ν ′ = − θ ( a − b ) / ν 2 \Sigma'_{0:\nu}=-\theta(a-b)/\nu^2 Σ 0 : ν ′ = − θ ( a − b ) / ν 2 shows that Σ 0 : ν ′ ≠ 0 \Sigma'_{0:\nu}\ne0 Σ 0 : ν ′ = 0 . Together with ν ∈ ( θ , ν ⋆ ) ⊂ ( 0 , 1 ) \nu\in(\theta,\nu_\star)\subset(0,1) ν ∈ ( θ , ν ⋆ ) ⊂ ( 0 , 1 ) , this makes the first summand in the preceding display strictly negative, and hence we get
J ∞ ′ ′ ( ν ) < 0 for all ν ∈ ( θ , ν ⋆ ) , J_\infty''(\nu)<0 \qquad\text{for all }\nu\in(\theta,\nu_\star), J ∞ ′′ ( ν ) < 0 for all ν ∈ ( θ , ν ⋆ ) ,
which is (57) .
Overall, the only candidates for a local minimizer of J ∞ J_\infty J ∞ on [ 0 , 1 ] [0,1] [ 0 , 1 ] are the kink points θ \theta θ and ν ⋆ \nu_\star ν ⋆ : by Case 2 the objective is strictly decreasing on ( 0 , θ ) (0,\theta) ( 0 , θ ) , by Case 3 strictly concave on ( θ , ν ⋆ ) (\theta,\nu_\star) ( θ , ν ⋆ ) , and by Case 1 strictly increasing on ( ν ⋆ , 1 ) (\nu_\star,1) ( ν ⋆ , 1 ) . Consequently the global minimum over [ 0 , 1 ] [0,1] [ 0 , 1 ] is attained at θ \theta θ or at ν ⋆ \nu_\star ν ⋆ . To identify which, it remains to show that
J ∞ ( ν ⋆ ) < J ∞ ( θ ) . (58) J_\infty(\nu_\star)<J_\infty(\theta). \tag*{(58)} J ∞ ( ν ⋆ ) < J ∞ ( θ ) . (58)
Verification of (58): The argument shares similarities to the calculations in the proof of Theorem 21 and Theorem 23. At the two points ν = θ \nu=\theta ν = θ and ν = ν ⋆ \nu=\nu_{\star} ν = ν ⋆ we have
Σ 0 : θ = a , Σ θ : 1 = ( ν ⋆ − θ ) b + ( 1 − ν ⋆ ) c 1 − θ , and \Sigma_{0:\theta}=a,\qquad\Sigma_{\theta:1}=\frac{(\nu_{\star}-\theta)b+(1-\nu_{\star})c}{1-\theta},\qquad\text{and} Σ 0 : θ = a , Σ θ : 1 = 1 − θ ( ν ⋆ − θ ) b + ( 1 − ν ⋆ ) c , and
Σ 0 : ν ⋆ = θ a + ( ν ⋆ − θ ) b ν ⋆ , Σ ν ⋆ : 1 = c . \Sigma_{0:\nu_{\star}}=\frac{\theta a+(\nu_{\star}-\theta)b}{\nu_{\star}},\qquad\Sigma_{\nu_{\star}:1}=c. Σ 0 : ν ⋆ = ν ⋆ θ a + ( ν ⋆ − θ ) b , Σ ν ⋆ : 1 = c .
To simplify the computation, we introduce the weights
λ : = θ ν ⋆ = 2 θ ∈ ( 0 , 1 ) , μ : = 1 − ν ⋆ 1 − θ = 1 2 ( 1 − θ ) ∈ ( 0 , 1 ) . \lambda:=\frac{\theta}{\nu_{\star}}=2\theta\in(0,1),\qquad\mu:=\frac{1-\nu_{\star}}{1-\theta}=\frac{1}{2(1-\theta)}\in(0,1). λ := ν ⋆ θ = 2 θ ∈ ( 0 , 1 ) , μ := 1 − θ 1 − ν ⋆ = 2 ( 1 − θ ) 1 ∈ ( 0 , 1 ) .
Then, direct calculation gives
Σ 0 : ν ⋆ = ( 1 − λ ) b + λ a , and Σ θ : 1 = ( 1 − μ ) b + μ c \Sigma_{0:\nu_{\star}}=(1-\lambda)b+\lambda a,\qquad\text{and}\qquad\Sigma_{\theta:1}=(1-\mu)b+\mu c Σ 0 : ν ⋆ = ( 1 − λ ) b + λa , and Σ θ : 1 = ( 1 − μ ) b + μ c
Hence
J ∞ ( ν ⋆ ) = ν ⋆ φ ( ( 1 − λ ) b + λ a ) + ( 1 − ν ⋆ ) φ ( c ) , J_{\infty}(\nu_{\star})=\nu_{\star}\varphi\left((1-\lambda)b+\lambda a\right)+(1-\nu_{\star})\varphi(c), J ∞ ( ν ⋆ ) = ν ⋆ φ ( ( 1 − λ ) b + λa ) + ( 1 − ν ⋆ ) φ ( c ) ,
J ∞ ( θ ) = θ φ ( a ) + ( 1 − θ ) φ ( ( 1 − μ ) b + μ c ) . J_{\infty}(\theta)=\theta\varphi(a)+(1-\theta)\varphi\left((1-\mu)b+\mu c\right). J ∞ ( θ ) = θ φ ( a ) + ( 1 − θ ) φ ( ( 1 − μ ) b + μ c ) .
Using θ = λ ν ⋆ \theta=\lambda\nu_{\star} θ = λ ν ⋆ and 1 − ν ⋆ = μ ( 1 − θ ) 1-\nu_{\star}=\mu(1-\theta) 1 − ν ⋆ = μ ( 1 − θ ) , we can regroup the difference as
J ∞ ( ν ⋆ ) − J ∞ ( θ ) = ν ⋆ ( φ ( ( 1 − λ ) b + λ a ) − λ φ ( a ) ) − ( 1 − θ ) ( φ ( ( 1 − μ ) b + μ c ) − μ φ ( c ) ) = ν ⋆ ( φ ( ( 1 − λ ) b + λ a ) − ( 1 − λ ) φ ( b ) − λ φ ( a ) ) − ( 1 − θ ) ( φ ( ( 1 − μ ) b + μ c ) − ( 1 − μ ) φ ( b ) − μ φ ( c ) ) . (59) \begin{aligned}
J_{\infty}(\nu_{\star})-J_{\infty}(\theta)
={}&\nu_{\star}\left(\varphi\left((1-\lambda)b+\lambda a\right)-\lambda\varphi(a)\right)\\
&-(1-\theta)\left(\varphi\left((1-\mu)b+\mu c\right)-\mu\varphi(c)\right)\\
={}&\nu_{\star}\left(\varphi\left((1-\lambda)b+\lambda a\right)-(1-\lambda)\varphi(b)-\lambda\varphi(a)\right)\\
&-(1-\theta)\left(\varphi\left((1-\mu)b+\mu c\right)-(1-\mu)\varphi(b)-\mu\varphi(c)\right). \tag*{(59)}
\end{aligned} J ∞ ( ν ⋆ ) − J ∞ ( θ ) = = ν ⋆ ( φ ( ( 1 − λ ) b + λa ) − λ φ ( a ) ) − ( 1 − θ ) ( φ ( ( 1 − μ ) b + μ c ) − μ φ ( c ) ) ν ⋆ ( φ ( ( 1 − λ ) b + λa ) − ( 1 − λ ) φ ( b ) − λ φ ( a ) ) − ( 1 − θ ) ( φ ( ( 1 − μ ) b + μ c ) − ( 1 − μ ) φ ( b ) − μ φ ( c ) ) . (59)
where we added and subtracted the φ ( b ) \varphi(b) φ ( b ) term in the last step, observing that by definition ν ⋆ ( 1 − λ ) = ( 1 − θ ) ( 1 − μ ) \nu_{\star}(1-\lambda)=(1-\theta)(1-\mu) ν ⋆ ( 1 − λ ) = ( 1 − θ ) ( 1 − μ ) . Identifying positive scalars with 1 × 1 1\times1 1 × 1 positive-definite matrices and recalling the Jensen gap G G G from Lemma 44, equation (59) becomes
J ∞ ( ν ⋆ ) − J ∞ ( θ ) = ν ⋆ G ( b + ε , a + ε ; λ ) − ( 1 − θ ) G ( b + ε , c + ε ; μ ) . J_{\infty}(\nu_{\star})-J_{\infty}(\theta)=\nu_{\star}G(b+\varepsilon,a+\varepsilon;\lambda)-(1-\theta)G(b+\varepsilon,c+\varepsilon;\mu). J ∞ ( ν ⋆ ) − J ∞ ( θ ) = ν ⋆ G ( b + ε , a + ε ; λ ) − ( 1 − θ ) G ( b + ε , c + ε ; μ ) .
For the first Jensen gap, the interpolation lies between b + ε b+\varepsilon b + ε and a + ε a+\varepsilon a + ε . Since a > b a>b a > b , we may take μ min = b + ε \mu_{\min}=b+\varepsilon μ m i n = b + ε . The upper bound in Lemma 44 therefore gives
G ( b + ε , a + ε ; λ ) ≤ λ ( 1 − λ ) 2 ( b + ε ) 2 ( a − b ) 2 = λ ( 1 − λ ) 2 ( b + ε ) 2 r 2 . (60) G(b+\varepsilon,a+\varepsilon;\lambda)\leq\frac{\lambda(1-\lambda)}{2(b+\varepsilon)^2}(a-b)^2=\frac{\lambda(1-\lambda)}{2(b+\varepsilon)^2}r^2. \tag*{(60)} G ( b + ε , a + ε ; λ ) ≤ 2 ( b + ε ) 2 λ ( 1 − λ ) ( a − b ) 2 = 2 ( b + ε ) 2 λ ( 1 − λ ) r 2 . (60)
For the second Jensen gap, the interpolation lies between c + ε c+\varepsilon c + ε and b + ε b+\varepsilon b + ε . Since b > c b>c b > c , we may take μ max = b + ε \mu_{\max}=b+\varepsilon μ m a x = b + ε . The lower bound in Lemma 44 gives
G ( b + ε , c + ε ; μ ) ≥ μ ( 1 − μ ) 2 ( b + ε ) 2 ( b − c ) 2 = μ ( 1 − μ ) 2 ( b + ε ) 2 K 2 . (61) G(b+\varepsilon,c+\varepsilon;\mu)\geq\frac{\mu(1-\mu)}{2(b+\varepsilon)^2}(b-c)^2=\frac{\mu(1-\mu)}{2(b+\varepsilon)^2}K^2. \tag*{(61)} G ( b + ε , c + ε ; μ ) ≥ 2 ( b + ε ) 2 μ ( 1 − μ ) ( b − c ) 2 = 2 ( b + ε ) 2 μ ( 1 − μ ) K 2 . (61)
Combining (59), (60), and (61) yields
J ∞ ( ν ⋆ ) − J ∞ ( θ ) ≤ 1 2 ( b + ε ) 2 ( ν ⋆ λ ( 1 − λ ) r 2 − ( 1 − θ ) μ ( 1 − μ ) K 2 ) . J_{\infty}(\nu_{\star})-J_{\infty}(\theta)\leq\frac{1}{2(b+\varepsilon)^2}\left(\nu_{\star}\lambda(1-\lambda)r^2-(1-\theta)\mu(1-\mu)K^2\right). J ∞ ( ν ⋆ ) − J ∞ ( θ ) ≤ 2 ( b + ε ) 2 1 ( ν ⋆ λ ( 1 − λ ) r 2 − ( 1 − θ ) μ ( 1 − μ ) K 2 ) .
Now we plug in ν ⋆ = 1 2 \nu_{\star}=\frac{1}{2} ν ⋆ = 2 1 , λ = 2 θ \lambda=2\theta λ = 2 θ , and μ = 1 2 ( 1 − θ ) \mu=\frac{1}{2(1-\theta)} μ = 2 ( 1 − θ ) 1 and obtain
ν ⋆ λ ( 1 − λ ) = 1 2 ( 2 θ ) ( 1 − 2 θ ) = θ ( 1 − 2 θ ) , ( 1 − θ ) μ ( 1 − μ ) = 1 − 2 θ 4 ( 1 − θ ) . \nu_{\star}\lambda(1-\lambda)=\frac{1}{2}(2\theta)(1-2\theta)=\theta(1-2\theta),\qquad(1-\theta)\mu(1-\mu)=\frac{1-2\theta}{4(1-\theta)}. ν ⋆ λ ( 1 − λ ) = 2 1 ( 2 θ ) ( 1 − 2 θ ) = θ ( 1 − 2 θ ) , ( 1 − θ ) μ ( 1 − μ ) = 4 ( 1 − θ ) 1 − 2 θ .
Hence
J ∞ ( ν ⋆ ) − J ∞ ( θ ) ≤ 1 − 2 θ 2 ( b + ε ) 2 ( θ r 2 − K 2 4 ( 1 − θ ) ) . (62) J_{\infty}(\nu_{\star})-J_{\infty}(\theta)\leq\frac{1-2\theta}{2(b+\varepsilon)^2}\left(\theta r^2-\frac{K^2}{4(1-\theta)}\right). \tag*{(62)} J ∞ ( ν ⋆ ) − J ∞ ( θ ) ≤ 2 ( b + ε ) 2 1 − 2 θ ( θ r 2 − 4 ( 1 − θ ) K 2 ) . (62)
Since θ ∈ ( 0 , 1 2 ) \theta\in\left(0,\frac{1}{2}\right) θ ∈ ( 0 , 2 1 ) we have 1 − 2 θ > 0 1-2\theta>0 1 − 2 θ > 0 . Therefore (62) is strictly negative as soon as
θ r 2 < K 2 4 ( 1 − θ ) ⟺ r K < 1 2 θ ( 1 − θ ) . \theta r^2<\frac{K^2}{4(1-\theta)}
\qquad\Longleftrightarrow\qquad
\frac{r}{K}<\frac{1}{2\sqrt{\theta(1-\theta)}}. θ r 2 < 4 ( 1 − θ ) K 2 ⟺ K r < 2 θ ( 1 − θ ) 1 .
This holds due to (13), and we conclude
J ∞ ( ν ⋆ ) − J ∞ ( θ ) < 0 , i.e. J ∞ ( ν ⋆ ) < J ∞ ( θ ) , J_{\infty}(\nu_{\star})-J_{\infty}(\theta)<0,\qquad\text{i.e.}\qquad J_{\infty}(\nu_{\star})<J_{\infty}(\theta), J ∞ ( ν ⋆ ) − J ∞ ( θ ) < 0 , i.e. J ∞ ( ν ⋆ ) < J ∞ ( θ ) ,
which proves (14). Together with the monotonicity and concavity properties established above — J ∞ J_{\infty} J ∞ is strictly decreasing on ( 0 , θ ) (0,\theta) ( 0 , θ ) , strictly increasing on ( ν ⋆ , 1 ) (\nu_{\star},1) ( ν ⋆ , 1 ) , and strictly concave on ( θ , ν ⋆ ) (\theta,\nu_{\star}) ( θ , ν ⋆ ) , so that on the middle interval it exceeds its smaller endpoint value J ∞ ( ν ⋆ ) J_{\infty}(\nu_{\star}) J ∞ ( ν ⋆ ) — this shows that J ∞ J_{\infty} J ∞ has its unique minimizer at ν ⋆ = 0.5 \nu_{\star}=0.5 ν ⋆ = 0.5 . □
Proof of Proposition 31
The proof has three steps: we first show that the error and cross terms are negligible on the pervasive scale; we then use the order-p p p eigengap and the Davis–Kahan theorem to control the empirical projection; finally, we transfer the regime separation through that projection to identify Δ w i t h i n \Delta_{\mathrm{within}} Δ within and Δ b e t w e e n \Delta_{\mathrm{between}} Δ between .
Let us recall the notation of Section 2 and some basic facts. First, we want to mention that p = p n p=p_n p = p n is a sequence in n n n , and we suppressed the dependency on n n n for convenience. Since y t = 1 p U ^ T x t y_t=\frac{1}{\sqrt{p}}\widehat{U}^{\mathsf T}x_t y t = p 1 U T x t , we have for the macroscopic interval I = I ( a , b ) I=I(a,b) I = I ( a , b ) by definition
Σ ^ a : b = 1 p U ^ T ( 1 ∣ I ∣ ∑ t ∈ I x t x t T ) U ^ . \widehat{\Sigma}_{a:b}=\frac{1}{p}\widehat{U}^{\mathsf T}\left(\frac{1}{|I|}\sum_{t\in I}x_tx_t^{\mathsf T}\right)\widehat{U}. Σ a : b = p 1 U T ( ∣ I ∣ 1 t ∈ I ∑ x t x t T ) U .
We recall the decomposition x t = s t + e t x_t=s_t+e_t x t = s t + e t into signal s t s_t s t and noise e t e_t e t . For a macroscopic interval I = I ( a , b ) I=I(a,b) I = I ( a , b ) we have defined
S a : b = 1 ∣ I ∣ ∑ t ∈ I s t s t T , E a : b = 1 ∣ I ∣ ∑ t ∈ I e t e t T , C a : b = 1 ∣ I ∣ ∑ t ∈ I ( s t e t T + e t s t T ) . S_{a:b}=\frac{1}{|I|}\sum_{t\in I}s_ts_t^{\mathsf T},\qquad
E_{a:b}=\frac{1}{|I|}\sum_{t\in I}e_te_t^{\mathsf T},\qquad
C_{a:b}=\frac{1}{|I|}\sum_{t\in I}(s_te_t^{\mathsf T}+e_ts_t^{\mathsf T}). S a : b = ∣ I ∣ 1 t ∈ I ∑ s t s t T , E a : b = ∣ I ∣ 1 t ∈ I ∑ e t e t T , C a : b = ∣ I ∣ 1 t ∈ I ∑ ( s t e t T + e t s t T ) .
This yields the decomposition
Σ ^ a : b = 1 p U ^ T ( S a : b + C a : b + E a : b ) U ^ . \widehat{\Sigma}_{a:b}=\frac{1}{p}\widehat{U}^{\mathsf T}(S_{a:b}+C_{a:b}+E_{a:b})\widehat{U}. Σ a : b = p 1 U T ( S a : b + C a : b + E a : b ) U .
We use throughout the contraction ∥ U ^ T A U ^ ∥ ≤ ∥ A ∥ \|\widehat{U}^{\mathsf T}A\widehat{U}\|\leq\|A\| ∥ U T A U ∥ ≤ ∥ A ∥ , since U ^ \widehat{U} U has orthonormal columns. We also use the operator Cauchy–Schwarz inequality
∥ 1 ∣ I ∣ ∑ t ∈ I s t e t T ∥ ≤ ∥ S a : b ∥ 1 / 2 ∥ E a : b ∥ 1 / 2 , \left\|\frac{1}{|I|}\sum_{t\in I}s_te_t^{\mathsf T}\right\|\leq\|S_{a:b}\|^{1/2}\|E_{a:b}\|^{1/2}, ∣ I ∣ 1 t ∈ I ∑ s t e t T ≤ ∥ S a : b ∥ 1/2 ∥ E a : b ∥ 1/2 ,
which gives the estimate
∥ C a : b ∥ ≤ 2 ∥ S a : b ∥ 1 / 2 ∥ E a : b ∥ 1 / 2 . (63) \|C_{a:b}\|\leq2\|S_{a:b}\|^{1/2}\|E_{a:b}\|^{1/2}. \tag*{(63)} ∥ C a : b ∥ ≤ 2∥ S a : b ∥ 1/2 ∥ E a : b ∥ 1/2 . (63)
We continue with the review of notation and fix a macroscopic interval I = I ( a , b ) I=I(a,b) I = I ( a , b ) in the regime ∙ \bullet ∙ . For t ∈ I ( a , b ) t\in I(a,b) t ∈ I ( a , b ) , we have s t = Λ ∙ f t s_t=\Lambda_{\bullet}f_t s t = Λ ∙ f t . Therefore,
S a : b = Λ ∙ G a : b Λ ∙ T , G a : b : = 1 ∣ I ∣ ∑ t ∈ I f t f t T . S_{a:b}=\Lambda_{\bullet}G_{a:b}\Lambda_{\bullet}^{\mathsf T},\qquad
G_{a:b}:=\frac{1}{|I|}\sum_{t\in I}f_tf_t^{\mathsf T}. S a : b = Λ ∙ G a : b Λ ∙ T , G a : b := ∣ I ∣ 1 t ∈ I ∑ f t f t T .
We also have
Σ ∙ s i g : = Λ ∙ Σ F ∙ Λ ∙ T . \Sigma_{\bullet}^{\mathrm{sig}}:=\Lambda_{\bullet}\Sigma_{F_{\bullet}}\Lambda_{\bullet}^{\mathsf T}. Σ ∙ sig := Λ ∙ Σ F ∙ Λ ∙ T .
by definition.
Operator-norm estimates. We start by providing auxiliary estimates on operator norms. By pervasiveness (F1),
1 p ∥ Λ ∙ ∥ 2 = O ( 1 ) . (64) \frac{1}{p}\lVert\Lambda_{\bullet}\rVert^2 = O(1). \tag*{(64)} p 1 ∥ Λ ∙ ∥ 2 = O ( 1 ) . (64)
We also have, by operator-norm submultiplicativity and (64),
1 p ∥ Σ ∙ s i g ∥ ≤ 1 p ∥ Λ ∙ ∥ 2 ∥ Σ F , ∙ ∥ = O ( 1 ) . (65) \frac{1}{p}\lVert\Sigma^{\mathrm{sig}}_{\bullet}\rVert\leq\frac{1}{p}\lVert\Lambda_{\bullet}\rVert^2\lVert\Sigma_{F,\bullet}\rVert= O(1). \tag*{(65)} p 1 ∥ Σ ∙ sig ∥ ≤ p 1 ∥ Λ ∙ ∥ 2 ∥ Σ F , ∙ ∥ = O ( 1 ) . (65)
We now deduce that
1 p ∥ S a : b ∥ = O ( 1 ) . (66) \frac{1}{p}\lVert S_{a:b}\rVert= O(1). \tag*{(66)} p 1 ∥ S a : b ∥ = O ( 1 ) . (66)
Indeed, for intervals contained in the regime ∙ \bullet ∙ ,
1 p ∥ S a : b − Σ ∙ s i g ∥ = 1 p ∥ Λ ∙ ( G a : b − Σ F , ∙ ) Λ ∙ T ∥ ≤ ( 1 p ∥ Λ ∙ ∥ 2 ) ∥ G a : b − Σ F , ∙ ∥ → n → ∞ 0 , (67) \frac{1}{p}\lVert S_{a:b}-\Sigma^{\mathrm{sig}}_{\bullet}\rVert
= \frac{1}{p}\lVert\Lambda_{\bullet}(G_{a:b}-\Sigma_{F,\bullet})\Lambda_{\bullet}^{\mathsf T}\rVert
\leq\left(\frac{1}{p}\lVert\Lambda_{\bullet}\rVert^2\right)\lVert G_{a:b}-\Sigma_{F,\bullet}\rVert
\xrightarrow[n\to\infty]{} 0, \tag*{(67)} p 1 ∥ S a : b − Σ ∙ sig ∥ = p 1 ∥ Λ ∙ ( G a : b − Σ F , ∙ ) Λ ∙ T ∥ ≤ ( p 1 ∥ Λ ∙ ∥ 2 ) ∥ G a : b − Σ F , ∙ ∥ n → ∞ 0 , (67)
by a combination of (64) and Assumption (F2). If an interval crosses the breakpoint, the same O ( 1 ) O(1) O ( 1 ) bound (66) follows by decomposing it into its left- and right-regime parts.
Allowing from now on arbitrary fixed macroscopic intervals I ( a , b ) I(a,b) I ( a , b ) , we observe that Assumption (F4) controls the noise, i.e.
1 p ∥ E a : b ∥ → 0. (68) \frac{1}{p}\lVert E_{a:b}\rVert\to0. \tag*{(68)} p 1 ∥ E a : b ∥ → 0. (68)
Additionally, the cross terms vanish by (63) and (68):
1 p ∥ C a : b ∥ ≤ 2 ( 1 p ∥ S a : b ∥ ) 1 / 2 ( 1 p ∥ E a : b ∥ ) 1 / 2 → n → ∞ 0. (69) \frac{1}{p}\lVert C_{a:b}\rVert\leq2\left(\frac{1}{p}\lVert S_{a:b}\rVert\right)^{1/2}\left(\frac{1}{p}\lVert E_{a:b}\rVert\right)^{1/2}\xrightarrow[n\to\infty]{}0. \tag*{(69)} p 1 ∥ C a : b ∥ ≤ 2 ( p 1 ∥ S a : b ∥ ) 1/2 ( p 1 ∥ E a : b ∥ ) 1/2 n → ∞ 0. (69)
Eigengap and consistency of the projection. Writing
Λ : = [ Λ ≤ Λ > ] \Lambda:= [\Lambda_{\leq}\ \Lambda_{>}] Λ := [ Λ ≤ Λ > ]
and
D ˉ : = diag ( θ Σ F ≤ , ( 1 − θ ) Σ F > ) ≻ 0 , \bar{D} := \operatorname{diag}(\theta\Sigma_{F_{\leq}},(1-\theta)\Sigma_{F_{>}})\succ0, D ˉ := diag ( θ Σ F ≤ , ( 1 − θ ) Σ F > ) ≻ 0 ,
the global signal moment is
Σ s i g : = θ Σ ≤ s i g + ( 1 − θ ) Σ > s i g = Λ D ˉ Λ T . \Sigma^{\mathrm{sig}} := \theta\Sigma^{\mathrm{sig}}_{\leq}+(1-\theta)\Sigma^{\mathrm{sig}}_{>}=\Lambda\bar{D}\Lambda^{\mathsf T}. Σ sig := θ Σ ≤ sig + ( 1 − θ ) Σ > sig = Λ D ˉ Λ T .
The non-zero eigenvalues of 1 p Σ s i g \frac{1}{p}\Sigma^{\mathrm{sig}} p 1 Σ sig coincide with those of
D ˉ 1 / 2 ( 1 p Λ T Λ ) D ˉ 1 / 2 . \bar{D}^{1/2}\left(\frac{1}{p}\Lambda^{\mathsf T}\Lambda\right)\bar{D}^{1/2}. D ˉ 1/2 ( p 1 Λ T Λ ) D ˉ 1/2 .
By Assumption (F1),
D ˉ 1 / 2 ( 1 p Λ T Λ ) D ˉ 1 / 2 ⟶ D ˉ 1 / 2 Σ Λ D ˉ 1 / 2 . \bar{D}^{1/2}\left(\frac{1}{p}\Lambda^{\mathsf T}\Lambda\right)\bar{D}^{1/2}\longrightarrow\bar{D}^{1/2}\Sigma_{\Lambda}\bar{D}^{1/2}. D ˉ 1/2 ( p 1 Λ T Λ ) D ˉ 1/2 ⟶ D ˉ 1/2 Σ Λ D ˉ 1/2 .
Since D ˉ ≻ 0 \bar{D}\succ0 D ˉ ≻ 0 and rank ( Σ Λ ) = q \operatorname{rank}(\Sigma_{\Lambda})=q rank ( Σ Λ ) = q , the limit D ˉ 1 / 2 Σ Λ D ˉ 1 / 2 \bar{D}^{1/2}\Sigma_{\Lambda}\bar{D}^{1/2} D ˉ 1/2 Σ Λ D ˉ 1/2 has rank q q q . Hence the leading q q q eigenvalues of 1 p Σ s i g \frac{1}{p}\Sigma^{\mathrm{sig}} p 1 Σ sig are bounded away from 0 in the limit whereas the rest are 0. Therefore, for n n n large, there is an eigengap of order p p p at level q q q , i.e.
λ q ( Σ s i g ) − λ q + 1 ( Σ s i g ) ≥ γ 0 p for some γ 0 > 0. (70) \lambda_q(\Sigma^{\mathrm{sig}})-\lambda_{q+1}(\Sigma^{\mathrm{sig}})\geq\gamma_0p\qquad\text{for some }\gamma_0>0. \tag*{(70)} λ q ( Σ sig ) − λ q + 1 ( Σ sig ) ≥ γ 0 p for some γ 0 > 0. (70)
Let P s i g P_{\mathrm{sig}} P sig denote the orthogonal projection onto the leading q q q -dimensional eigenspace of Σ s i g \Sigma^{\mathrm{sig}} Σ sig , and let P ^ : = U ^ U ^ T \widehat{P}:=\widehat{U}\widehat{U}^{\mathsf{T}} P := U U T denote the orthogonal projection onto span ( U ^ ) \operatorname{span}(\widehat{U}) span ( U ) , where U ^ \widehat{U} U contains the leading q q q eigenvectors of the full-data second moment
1 n ∑ t = 1 n x t x t T = S 0 : 1 + C 0 : 1 + E 0 : 1 . \frac{1}{n}\sum_{t=1}^{n}x_tx_t^{\mathsf{T}}=S_{0:1}+C_{0:1}+E_{0:1}. n 1 t = 1 ∑ n x t x t T = S 0 : 1 + C 0 : 1 + E 0 : 1 .
Set w n : = τ / n w_n:=\tau/n w n := τ / n . Lemma 35, applied to the signal vectors s t ∈ R p s_t \in\mathbb{R}^p s t ∈ R p , and the definition of Σ s i g \Sigma^{\mathrm{sig}} Σ sig give the exact weight-mismatch decomposition
S 0 : 1 − Σ s i g = w n ( S 0 : θ − Σ ≤ s i g ) + ( 1 − w n ) ( S θ : 1 − Σ > s i g ) + ( w n − θ ) ( Σ ≤ s i g − Σ > s i g ) , S_{0:1}-\Sigma^{\mathrm{sig}}=w_n\left(S_{0:\theta}-\Sigma_{\leq}^{\mathrm{sig}}\right)+(1-w_n)\left(S_{\theta:1}-\Sigma_{>}^{\mathrm{sig}}\right)+(w_n-\theta)\left(\Sigma_{\leq}^{\mathrm{sig}}-\Sigma_{>}^{\mathrm{sig}}\right), S 0 : 1 − Σ sig = w n ( S 0 : θ − Σ ≤ sig ) + ( 1 − w n ) ( S θ : 1 − Σ > sig ) + ( w n − θ ) ( Σ ≤ sig − Σ > sig ) ,
1 p ∥ S 0 : 1 − Σ s i g ∥ ≤ w n 1 p ∥ S 0 : θ − Σ ≤ s i g ∥ + ( 1 − w n ) 1 p ∥ S θ : 1 − Σ > s i g ∥ + ∣ w n − θ ∣ 1 p ∥ Σ ≤ s i g − Σ > s i g ∥ → n → ∞ 0. \frac{1}{p}\left\|S_{0:1}-\Sigma^{\mathrm{sig}}\right\|\leq w_n\frac{1}{p}\left\|S_{0:\theta}-\Sigma_{\leq}^{\mathrm{sig}}\right\|+(1-w_n)\frac{1}{p}\left\|S_{\theta:1}-\Sigma_{>}^{\mathrm{sig}}\right\|+|w_n-\theta|\frac{1}{p}\left\|\Sigma_{\leq}^{\mathrm{sig}}-\Sigma_{>}^{\mathrm{sig}}\right\|\xrightarrow{n\to\infty}0. p 1 S 0 : 1 − Σ sig ≤ w n p 1 S 0 : θ − Σ ≤ sig + ( 1 − w n ) p 1 S θ : 1 − Σ > sig + ∣ w n − θ ∣ p 1 Σ ≤ sig − Σ > sig n → ∞ 0.
Here we used (67), w n → θ w_n\to\theta w n → θ , and the O ( 1 ) O(1) O ( 1 ) signal bounds (65). Together with (68) and (69), this gives
1 p ∥ ( S 0 : 1 + C 0 : 1 + E 0 : 1 ) − Σ s i g ∥ ≤ 1 p ∥ S 0 : 1 − Σ s i g ∥ + 1 p ∥ C 0 : 1 ∥ + 1 p ∥ E 0 : 1 ∥ → n → ∞ 0. \frac{1}{p}\left\|(S_{0:1}+C_{0:1}+E_{0:1})-\Sigma^{\mathrm{sig}}\right\|
\leq\frac{1}{p}\left\|S_{0:1}-\Sigma^{\mathrm{sig}}\right\|+\frac{1}{p}\left\|C_{0:1}\right\|+\frac{1}{p}\left\|E_{0:1}\right\|\xrightarrow{n\to\infty}0. p 1 ( S 0 : 1 + C 0 : 1 + E 0 : 1 ) − Σ sig ≤ p 1 S 0 : 1 − Σ sig + p 1 ∥ C 0 : 1 ∥ + p 1 ∥ E 0 : 1 ∥ n → ∞ 0.
Let P ^ \widehat{P} P and P P P denote the orthogonal projections onto the leading eigenspaces, of dimension q q q , of symmetric matrices Σ ^ \widehat{\Sigma} Σ and Σ \Sigma Σ , and let Θ \Theta Θ denote the diagonal matrix of the principal angles between their ranges. If γ : = λ q ( Σ ) − λ q + 1 ( Σ ) > 0 \gamma:=\lambda_q(\Sigma)-\lambda_{q+1}(\Sigma)>0 γ := λ q ( Σ ) − λ q + 1 ( Σ ) > 0 , then the Davis–Kahan theorem in the form of Yu, Wang, and Samworth [30 ] (a variant of the classical Davis–Kahan theorem [16 ] ) gives
∥ P ^ − P ∥ = ∥ sin Θ ∥ ≤ ∥ sin Θ ∥ F ≤ 2 q γ ∥ Σ ^ − Σ ∥ . \left\|\widehat{P}-P\right\|=\left\|\sin\Theta\right\|\leq\left\|\sin\Theta\right\|_F\leq\frac{2\sqrt{q}}{\gamma}\left\|\widehat{\Sigma}-\Sigma\right\|. P − P = ∥ sin Θ ∥ ≤ ∥ sin Θ ∥ F ≤ γ 2 q Σ − Σ .
With the gap (70) this yields
∥ P ^ − P s i g ∥ ≤ 2 q γ 0 1 p ∥ ( S 0 : 1 + C 0 : 1 + E 0 : 1 ) − Σ s i g ∥ → n → ∞ 0. (71) \left\|\widehat{P}-P_{\mathrm{sig}}\right\|\leq\frac{2\sqrt{q}}{\gamma_0}\frac{1}{p}\left\|(S_{0:1}+C_{0:1}+E_{0:1})-\Sigma^{\mathrm{sig}}\right\|\xrightarrow{n\to\infty}0. \tag*{(71)} P − P sig ≤ γ 0 2 q p 1 ( S 0 : 1 + C 0 : 1 + E 0 : 1 ) − Σ sig n → ∞ 0. (71)
We shall also use the following consequence of pervasiveness. For every fixed matrix A ∈ R ( q ≤ + q > ) × ( q ≤ + q > ) A\in\mathbb{R}^{(q_{\leq}+q_{>})\times(q_{\leq}+q_{>})} A ∈ R ( q ≤ + q > ) × ( q ≤ + q > ) ,
1 p ∥ Λ A Λ T − P s i g Λ A Λ T P s i g ∥ F → n → ∞ 0. (72) \frac{1}{p}\left\|\Lambda A\Lambda^{\mathsf{T}}-P_{\mathrm{sig}}\Lambda A\Lambda^{\mathsf{T}}P_{\mathrm{sig}}\right\|_F\xrightarrow{n\to\infty}0. \tag*{(72)} p 1 Λ A Λ T − P sig Λ A Λ T P sig F n → ∞ 0. (72)
Indeed, since P s i g P_{\mathrm{sig}} P sig is the projection onto the leading left singular space of Λ D ‾ 1 / 2 \Lambda\overline{D}^{1/2} Λ D 1/2 , we have
∥ ( I − P s i g ) Λ D ‾ 1 / 2 ∥ F 2 = ∑ j > q λ j ( Σ s i g ) = o ( p ) , \left\|(I-P_{\mathrm{sig}})\Lambda\overline{D}^{1/2}\right\|_F^2=\sum_{j>q}\lambda_j(\Sigma^{\mathrm{sig}})=o(p), ( I − P sig ) Λ D 1/2 F 2 = j > q ∑ λ j ( Σ sig ) = o ( p ) ,
where we used that only finitely many eigenvalues are involved and that λ q + 1 ( Σ s i g ) = o ( p ) \lambda_{q+1}(\Sigma^{\mathrm{sig}})=o(p) λ q + 1 ( Σ sig ) = o ( p ) . Since D ‾ − 1 / 2 \overline{D}^{-1/2} D − 1/2 is bounded, this implies
∥ ( I − P s i g ) Λ ∥ F ≤ ∥ ( I − P s i g ) Λ D ‾ 1 / 2 ∥ F ∥ D ‾ − 1 / 2 ∥ = o ( p ) . \left\|(I-P_{\mathrm{sig}})\Lambda\right\|_F\leq\left\|(I-P_{\mathrm{sig}})\Lambda\overline{D}^{1/2}\right\|_F\left\|\overline{D}^{-1/2}\right\|=o(\sqrt{p}). ∥ ( I − P sig ) Λ ∥ F ≤ ( I − P sig ) Λ D 1/2 F D − 1/2 = o ( p ) .
Moreover, by (F1),
∥ Λ ∥ = O ( p ) . \left\|\Lambda\right\|=O(\sqrt{p}). ∥ Λ ∥ = O ( p ) .
Therefore,
∥ ( I − P s i g ) Λ A Λ T ∥ F ≤ ∥ ( I − P s i g ) Λ ∥ F ∥ A ∥ ∥ Λ ∥ = o ( p ) , \left\|(I-P_{\mathrm{sig}})\Lambda A\Lambda^{\mathsf T}\right\|_F \le\left\|(I-P_{\mathrm{sig}})\Lambda\right\|_F\|A\|\|\Lambda\| = o(p), ( I − P sig ) Λ A Λ T F ≤ ∥ ( I − P sig ) Λ ∥ F ∥ A ∥∥Λ∥ = o ( p ) ,
∥ Λ A Λ T ( I − P s i g ) ∥ F ≤ ∥ Λ ∥ ∥ A ∥ ∥ Λ T ( I − P s i g ) ∥ F = o ( p ) . \left\|\Lambda A\Lambda^{\mathsf T}(I-P_{\mathrm{sig}})\right\|_F \le\|\Lambda\|\|A\|\left\|\Lambda^{\mathsf T}(I-P_{\mathrm{sig}})\right\|_F = o(p). Λ A Λ T ( I − P sig ) F ≤ ∥Λ∥∥ A ∥ Λ T ( I − P sig ) F = o ( p ) .
Since
Λ A Λ T − P s i g Λ A Λ T P s i g = ( I − P s i g ) Λ A Λ T + P s i g Λ A Λ T ( I − P s i g ) , \Lambda A\Lambda^{\mathsf T}-P_{\mathrm{sig}}\Lambda A\Lambda^{\mathsf T}P_{\mathrm{sig}}=(I-P_{\mathrm{sig}})\Lambda A\Lambda^{\mathsf T}+P_{\mathrm{sig}}\Lambda A\Lambda^{\mathsf T}(I-P_{\mathrm{sig}}), Λ A Λ T − P sig Λ A Λ T P sig = ( I − P sig ) Λ A Λ T + P sig Λ A Λ T ( I − P sig ) ,
we obtain (72) .
Argument for (i). By the contraction and the estimates (66), (68), and (69), every macroscopic block satisfies
∥ Σ ^ a : b ∥ = 1 p ∥ U ^ T ( S a : b + C a : b + E a : b ) U ^ ∥ ≤ 1 p ∥ S a : b ∥ + 1 p ∥ C a : b ∥ + 1 p ∥ E a : b ∥ = O ( 1 ) . \begin{aligned}
\left\|\widehat{\Sigma}_{a:b}\right\| = \frac{1}{p}\left\|\widehat{U}^{\mathsf T}(S_{a:b}+C_{a:b}+E_{a:b})\widehat{U}\right\| \\
&\le\frac{1}{p}\|S_{a:b}\|+\frac{1}{p}\|C_{a:b}\|+\frac{1}{p}\|E_{a:b}\|=O(1).
\end{aligned} Σ a : b = p 1 U T ( S a : b + C a : b + E a : b ) U ≤ p 1 ∥ S a : b ∥ + p 1 ∥ C a : b ∥ + p 1 ∥ E a : b ∥ = O ( 1 ) .
Set
C 0 : = max ∙ ∈ { ≤ , > } lim sup n → ∞ 1 p ∥ Σ ^ ∙ s i g ∥ < ∞ , C_0:=\max_{\mathord{\bullet}\in\{\leq,>\}}\limsup_{n\to\infty}\frac{1}{p}\left\|\widehat{\Sigma}^{\mathrm{sig}}_{\mathord{\bullet}}\right\|<\infty, C 0 := ∙ ∈ { ≤ , > } max n → ∞ lim sup p 1 Σ ∙ sig < ∞ ,
where finiteness follows from (65). Every break-anchored block lies in one of the two regimes. For each fixed I ( a , b ) ∈ I ∙ b r I(a,b)\in\mathcal{I}^{\mathrm{br}}_{\mathord{\bullet}} I ( a , b ) ∈ I ∙ br in regime ∙ \mathord{\bullet} ∙ , (67), (68), and (69) give
lim sup n → ∞ ∥ Σ ^ a : b ∥ ≤ lim sup n → ∞ 1 p ∥ Σ ^ ∙ s i g ∥ ≤ C 0 . \limsup_{n\to\infty}\left\|\widehat{\Sigma}_{a:b}\right\|\leq\limsup_{n\to\infty}\frac{1}{p}\left\|\widehat{\Sigma}^{\mathrm{sig}}_{\mathord{\bullet}}\right\|\leq C_0. n → ∞ lim sup Σ a : b ≤ n → ∞ lim sup p 1 Σ ∙ sig ≤ C 0 .
Since C 0 C_0 C 0 is independent of the particular block,
sup I ( a , b ) ∈ I ∙ b r lim sup n → ∞ ∥ Σ ^ a : b ∥ ≤ C 0 , \sup_{I(a,b)\in\mathcal{I}^{\mathrm{br}}_{\mathord{\bullet}}}\limsup_{n\to\infty}\left\|\widehat{\Sigma}_{a:b}\right\|\leq C_0, I ( a , b ) ∈ I ∙ br sup n → ∞ lim sup Σ a : b ≤ C 0 ,
which is precisely M < ∞ M<\infty M < ∞ , and hence Assumption (A1).
Argument for (ii). We fix two macroscopic intervals I 1 , I 2 I_1,I_2 I 1 , I 2 . Using the identity
Σ ^ I = 1 p U ^ T ( S I + C I + E I ) U ^ \widehat{\Sigma}_I=\frac{1}{p}\widehat{U}^{\mathsf T}(S_I+C_I+E_I)\widehat{U} Σ I = p 1 U T ( S I + C I + E I ) U
for I = I 1 I=I_1 I = I 1 and I = I 2 I=I_2 I = I 2 , together with (68) and (69), we get
∥ Σ ^ I 1 − Σ ^ I 2 ∥ ≤ 1 p ∥ S I 1 − S I 2 ∥ + 1 p ( ∥ C I 1 ∥ + ∥ C I 2 ∥ + ∥ E I 1 ∥ + ∥ E I 2 ∥ ) , \left\|\widehat{\Sigma}_{I_1}-\widehat{\Sigma}_{I_2}\right\|\leq\frac{1}{p}\left\|S_{I_1}-S_{I_2}\right\|
+\frac{1}{p}\left(\|C_{I_1}\|+\|C_{I_2}\|+\|E_{I_1}\|+\|E_{I_2}\|\right), Σ I 1 − Σ I 2 ≤ p 1 ∥ S I 1 − S I 2 ∥ + p 1 ( ∥ C I 1 ∥ + ∥ C I 2 ∥ + ∥ E I 1 ∥ + ∥ E I 2 ∥ ) ,
where the second term converges to 0 0 0 as n → ∞ n\to\infty n → ∞ .
Since the projected matrices are q × q q\times q q × q and q q q is fixed, the same estimates imply the corresponding convergence in Frobenius norm.
If I 1 I_1 I 1 and I 2 I_2 I 2 lie in the same regime ∙ \mathord{\bullet} ∙ , then by (67) both
1 p ∥ S I j − Σ ∙ s i g ∥ → 0 , j = 1 , 2. \frac{1}{p}\left\|S_{I_j}-\Sigma^{\mathrm{sig}}_{\mathord{\bullet}}\right\|\to0,\qquad j=1,2. p 1 S I j − Σ ∙ sig → 0 , j = 1 , 2.
Hence
1 p ∥ S I 1 − S I 2 ∥ → 0 , \frac{1}{p}\left\|S_{I_1}-S_{I_2}\right\|\to0, p 1 ∥ S I 1 − S I 2 ∥ → 0 ,
and therefore
∥ Σ ^ I 1 − Σ ^ I 2 ∥ F → 0 \left\|\widehat{\Sigma}_{I_1}-\widehat{\Sigma}_{I_2}\right\|_F\to0 Σ I 1 − Σ I 2 F → 0
for each such pair. As Δ w i t h i n \Delta_{\mathrm{within}} Δ within takes the limit in n n n before the supremum over pairs,
Δ w i t h i n = sup same regime lim sup n → ∞ ∥ Σ ^ I 1 − Σ ^ I 2 ∥ F = 0. \Delta_{\mathrm{within}}=\sup_{\text{same regime}}\limsup_{n\to\infty}\left\|\widehat{\Sigma}_{I_1}-\widehat{\Sigma}_{I_2}\right\|_F=0. Δ within = same regime sup n → ∞ lim sup Σ I 1 − Σ I 2 F = 0.
Now assume that I 1 ⊂ [ 0 , θ ] I_1 \subset[0,\theta] I 1 ⊂ [ 0 , θ ] and I 2 ⊂ [ θ , 1 ] I_2 \subset[\theta,1] I 2 ⊂ [ θ , 1 ] , and set
Ξ : = Σ ≤ s i g − Σ > s i g . \Xi:= \Sigma^{\mathrm{sig}}_{\leq} - \Sigma^{\mathrm{sig}}_{>}. Ξ := Σ ≤ sig − Σ > sig .
By (67),
1 p ∥ ( S I 1 − S I 2 ) − Ξ ∥ → 0. \frac{1}{p}\left\|(S_{I_1}-S_{I_2})-\Xi\right\| \to0. p 1 ∥ ( S I 1 − S I 2 ) − Ξ ∥ → 0.
Therefore, in Frobenius norm,
Σ ^ I 1 − Σ ^ I 2 = 1 p U ^ ⊤ Ξ U ^ + o ( 1 ) . (73) \widehat{\Sigma}_{I_1}-\widehat{\Sigma}_{I_2}=\frac{1}{p}\widehat{U}^{\top}\Xi\widehat{U}+o(1). \tag*{(73)} Σ I 1 − Σ I 2 = p 1 U ⊤ Ξ U + o ( 1 ) . (73)
Since U ^ \widehat{U} U has orthonormal columns,
1 p ∥ U ^ ⊤ Ξ U ^ ∥ F = 1 p ∥ P ^ Ξ P ^ ∥ F . (74) \frac{1}{p}\left\|\widehat{U}^{\top}\Xi\widehat{U}\right\|_F=\frac{1}{p}\left\|\widehat{P}\Xi\widehat{P}\right\|_F. \tag*{(74)} p 1 U ⊤ Ξ U F = p 1 P Ξ P F . (74)
We now compare P ^ Ξ P ^ \widehat{P}\Xi\widehat{P} P Ξ P with Ξ \Xi Ξ . First observe that
Ξ = Λ A Ξ Λ ⊤ , A Ξ : = diag ( Σ F ≤ , − Σ F > ) . \Xi=\Lambda A_{\Xi}\Lambda^{\top},\qquad A_{\Xi}:=\operatorname{diag}(\Sigma_{F_{\leq}},-\Sigma_{F_{>}}). Ξ = Λ A Ξ Λ ⊤ , A Ξ := diag ( Σ F ≤ , − Σ F > ) .
Thus, applying (72) with A = A Ξ A=A_{\Xi} A = A Ξ gives
1 p ∥ Ξ − P s i g Ξ P s i g ∥ F → 0. (75) \frac{1}{p}\left\|\Xi-P_{\mathrm{sig}}\Xi P_{\mathrm{sig}}\right\|_F\to0. \tag*{(75)} p 1 ∥ Ξ − P sig Ξ P sig ∥ F → 0. (75)
Moreover,
1 p ∥ Ξ ∥ F = O ( 1 ) , \frac{1}{p}\|\Xi\|_F=O(1), p 1 ∥Ξ ∥ F = O ( 1 ) ,
because Ξ \Xi Ξ has rank at most q ≤ + q > q_{\leq}+q_{>} q ≤ + q > and 1 p ∥ Ξ ∥ = O ( 1 ) \frac{1}{p}\|\Xi\|=O(1) p 1 ∥Ξ∥ = O ( 1 ) by (65). Hence, using (71),
1 p ∥ P ^ Ξ P ^ − P s i g Ξ P s i g ∥ F ≤ 1 p ∥ ( P ^ − P s i g ) Ξ P ^ ∥ F + 1 p ∥ P s i g Ξ ( P ^ − P s i g ) ∥ F \frac{1}{p}\left\|\widehat{P}\Xi\widehat{P}-P_{\mathrm{sig}}\Xi P_{\mathrm{sig}}\right\|_F
\leq\frac{1}{p}\left\|(\widehat{P}-P_{\mathrm{sig}})\Xi\widehat{P}\right\|_F+\frac{1}{p}\left\|P_{\mathrm{sig}}\Xi(\widehat{P}-P_{\mathrm{sig}})\right\|_F p 1 P Ξ P − P sig Ξ P sig F ≤ p 1 ( P − P sig ) Ξ P F + p 1 P sig Ξ ( P − P sig ) F
≤ 2 ( 1 p ∥ Ξ ∥ F ) ∥ P ^ − P s i g ∥ → n → ∞ 0. \leq2\left(\frac{1}{p}\|\Xi\|_F\right)\|\widehat{P}-P_{\mathrm{sig}}\|\xrightarrow[n\to\infty]{}0. ≤ 2 ( p 1 ∥Ξ ∥ F ) ∥ P − P sig ∥ n → ∞ 0.
Combining this estimate with (75), we obtain
1 p ∣ ∥ P ^ Ξ P ^ ∥ F − ∥ Ξ ∥ F ∣ ≤ 1 p ∥ P ^ Ξ P ^ − Ξ ∥ F → n → ∞ 0. \frac{1}{p}\left|\|\widehat{P}\Xi\widehat{P}\|_F-\|\Xi\|_F\right|
\leq\frac{1}{p}\left\|\widehat{P}\Xi\widehat{P}-\Xi\right\|_F\xrightarrow[n\to\infty]{}0. p 1 ∥ P Ξ P ∥ F − ∥Ξ ∥ F ≤ p 1 P Ξ P − Ξ F n → ∞ 0.
A combination of the last statement, (74), and (73) yields that for every between-regime pair
∥ Σ ^ I 1 − Σ ^ I 2 ∥ F = 1 p ∥ Σ ≤ s i g − Σ > s i g ∥ F + o ( 1 ) . \left\|\widehat{\Sigma}_{I_1}-\widehat{\Sigma}_{I_2}\right\|_F
=\frac{1}{p}\left\|\Sigma^{\mathrm{sig}}_{\leq}-\Sigma^{\mathrm{sig}}_{>}\right\|_F+o(1). Σ I 1 − Σ I 2 F = p 1 Σ ≤ sig − Σ > sig F + o ( 1 ) .
The leading term is independent of the chosen between-regime pair. Taking the infimum over between-regime pairs and then lim infn _n n gives
Δ b e t w e e n = lim inf n → ∞ 1 p ∥ Σ ≤ s i g − Σ > s i g ∥ F , \Delta_{\mathrm{between}}=\liminf_{n\to\infty}\frac{1}{p}\left\|\Sigma^{\mathrm{sig}}_{\leq}-\Sigma^{\mathrm{sig}}_{>}\right\|_F, Δ between = n → ∞ lim inf p 1 Σ ≤ sig − Σ > sig F ,
which is positive by (F3). This proves Proposition 31. □ \square □
Acknowledgments
We thank Alec Kercheval and Mihai Cucuringu for valuable discussions on the topics of this work. We acknowledge the use of the AI assistants Claude (Anthropic, via the Claude Code command-line interface) and ChatGPT (OpenAI) for assistance with editing the manuscript and with developing the code for the numerical examples; the authors reviewed all such output and take full responsibility for the content of this paper. We thank the UCLA Department of Mathematics for making these tools available, and the Olga Radko Endowed Math Circle at UCLA for financial support.
Appendix A. Code
The code that generates the figures in this paper is included with the manuscript source as ancillary files (see the anc/ directory of the arXiv submission). The equity data for Figure 4 are daily adjusted close prices from Yahoo Finance, retrieved with the yfinance package; the ticker universe is defined inside the generating script, and names with insufficient price coverage over 2005–2010 are dropped by the filter documented there, leaving p = 200 p = 200 p = 200 names. Table 1 lists, for each figure, the source that generates it.
Table 1. Correspondence between the figures of this paper and the ancillary scripts that generate them.
References