betaFeedback

WorksMathematics

Structural latent dimension: deterministic geometry and local Fisher structure

Abstract

We study the structural dimension of a second-order object: the minimum latent dimension that remains necessary relative to a declared class of competing dependence. Building on minimum-rank factor analysis and low-rank-plus-sparse decomposition, we use distances to lower-dimensional classes to describe inference near structural ambiguity. General distance and tangent-cone facts yield confidence statements and local margins. For rank-zero competitors with bounded residual degree, we derive exact maximum-weight bb-matching formulas, support-slack tangent cones and sparse-spike angles; a shared residual support across a field can produce separation absent pointwise. We give positive-semidefinite brackets, a dimension-free Gaussian likelihood expansion, an exact quotient by the tangent of a low-rank background, and a detection floor from the margin. Equal-angle configurations with different branch multiplicities motivate static and spectral detection conjectures. Joint low-rank/residual profiling and the sharp composite information boundaries remain open.

Introduction

Factor and latent-variable models represent high-dimensional second-order objects by a low-rank component and a structured remainder. Factor-number criteria relate dimension to signal strength under their model assumptions [3]. Minimum-rank factor analysis studies the least common rank compatible with diagonal noise (ten Berge and Kiers, 1991; Shapiro and ten Berge, 2002; Saunderson et al., 2012). Low-rank-plus-sparse decomposition jointly recovers components with unknown rank and support, using geometric conditions for identifiability and exact recovery [4, 5]. These theories already recognise that a low-rank direction can also have a sparse explanation. Strength and necessity can then diverge: a direction with a large eigenvalue may be reproducible by sparse residual dependence, while a much weaker diffuse direction may not be (Example 4.3).

We consider inference for the minimum dimension as such explanations approach one another. The distance to a lower-dimensional class quantifies the separation even when a particular decomposition is nonunique. Its tangent geometry gives a local first-order scale; the number and organisation of competing branches can create additional statistical difficulty. A further issue is whether observation preserves the distinguishing differences between explanations.

Sections 2–3 collect general metric and variational tools in this setting. Sections 4–5 develop explicit degree-constrained prototypes and local Gaussian formulas. Section 6 formulates separated testing problems and conjectured rate benchmarks. All statements presented as results have proofs; the composite boundaries are open questions.

Results and organisation.

  1. Section 2 defines structural dimension and the margin Δq\Delta_q and shows that certified bounds on Δq\Delta_q and simultaneous bands on the distances yield honest confidence statements (Propositions 2.4 and 2.5).

  1. Section 3 identifies the first-order behaviour of the margin with the distance of the entering direction to a tangent cone (Proposition 3.1).

  1. Section 4 treats competitors with bounded residual degree: the tangent cone is a union of subspaces indexed by the support slack (Proposition 4.4); distances and angles are maximum-weight bb-matchings (Theorem 4.2); equal-magnitude spikes have an exact angle (Theorem 4.7); a shared residual support across a field creates separation absent pointwise (Proposition 4.9); positive semidefiniteness is bracketed by computable certificates (Proposition 4.10).

  1. Section 5 gives a dimension-free local likelihood expansion for Gaussian covariance experiments (Lemma 5.1), an exact formula for the Fisher information that remains after projecting out the tangent space of a rank-rr background (Proposition 5.3), and a detection floor in terms of the margin (Proposition 5.5).

Notation. Sp\mathbb{S}^{p} is the space of real symmetric p×pp \times p matrices with Frobenius inner product ⟨A,B⟩=tr⁡(AB)\langle A,B\rangle= \operatorname{tr}(AB) and norm ∥⋅∥F\lVert\cdot\rVert_{F}; ∥⋅∥2\lVert\cdot\rVert_{2} is the operator norm, A⪰0A \succeq0 (A≻0A \succ0) means positive semidefinite (definite), λmin⁡\lambda_{\min} is the smallest eigenvalue and [x]+=max⁡{x,0}[x]_{+} = \max\{x,0\}. For a set CC and a point YY in a normed space, d(Y,C)=inf⁡Z∈C∥Y−Z∥d(Y,C) = \inf_{Z \in C}\lVert Y-Z\rVert. Unordered pairs {i,j}\{i,j\}, i≠ji \ne j, of [p]={1,…,p}[p] = \{1,\ldots,p\} are called edges; for an edge set EE, deg⁡E(i)\operatorname{deg}_{E}(i) is the number of edges of EE containing ii.

Structural dimension and its margin

Let Y∈SpY \in\mathbb{S}^{p} be an observable second-order object, typically a covariance matrix, and let ∥⋅∥\lVert\cdot\rVert be a norm on Sp\mathbb{S}^{p}. A competing class S⊆Sp\mathcal{S} \subseteq\mathbb{S}^{p} is a closed set containing 00, declared on scientific grounds before any rank is inferred. Fix a magnitude bound ρ>0\rho> 0 and, for r≥0r \ge0, let Lr={L∈Sp:L⪰0, rank⁡L≤r, ∥L∥≤ρ}\mathcal{L}_{r} = \{L \in\mathbb{S}^{p}: L \succeq0,\ \operatorname{rank}L \le r,\ \lVert L\rVert\le\rho\}, a compact set. The nested classes

M≤r=Lr+S,M≤0⊆M≤1⊆⋯ ,\mathcal{M}_{\le r} = \mathcal{L}_{r} + \mathcal{S}, \qquad\mathcal{M}_{\le0} \subseteq\mathcal{M}_{\le1} \subseteq\cdots,

are closed, as sums of a compact and a closed set. Assume Y∈M≤pY \in\mathcal{M}_{\le p}.

Definition 2.1. The structural dimension of YY is qstr(Y)=min⁡{r:Y∈M≤r}q_{\mathrm{str}}(Y) = \min\{r : Y \in\mathcal{M}_{\le r}\}. For r≥0r \ge0 let Dr(Y)=d(Y,M≤r)D_{r}(Y) = d(Y,\mathcal{M}_{\le r}), and for q≥1q \ge1 define the margin Δq(Y)=Dq−1(Y)\Delta_{q}(Y) = D_{q-1}(Y), the amount by which the explanation with q−1q-1 latent directions fails.

Since the classes are closed, qstr(Y)=min⁡{r:Dr(Y)=0}q_{\mathrm{str}}(Y) = \min\{r : D_{r}(Y) = 0\}, and YY requires a qq-th direction exactly when Δq(Y)>0\Delta_{q}(Y) > 0.

Remark 2.2. The bound ρ\rho cannot be replaced by a closure of the unbounded sum without changing the meaning of qstrq_{\mathrm{str}}. Let S\mathcal{S} be the diagonal matrices (of either sign) and wn=(n,1/n,1/n)w_n = (n,1/n,1/n). Then wnwnT−diag⁡(wnwnT)w_n w_n^{\mathsf{T}} - \operatorname{diag}(w_n w_n^{\mathsf{T}}) converges to the matrix YY with Y12=Y21=Y13=Y31=1Y_{12} = Y_{21} = Y_{13} = Y_{31} = 1 and all other entries 00, so YY lies in the closure of the unbounded rank-one class; yet YY has no exact rank-one-plus-diagonal decomposition, which would require w1w2=w1w3=1w_1w_2 = w_1w_3 = 1 and w2w3=0w_2w_3 = 0. Its closure dimension 1 would be realised only by diverging components, a Heywood-type phenomenon. If the competing class consists of positive semidefinite matrices, no bound is needed: two positive semidefinite summands of a bounded sum are bounded.

Lemma 2.3. (i) r↦Dr(Y)r \mapsto D_{r}(Y) is nonincreasing.

(ii) ∣Dr(Y)−Dr(Y′)∣≤∥Y−Y′∥\lvert D_{r}(Y) - D_{r}(Y')\rvert\le\lVert Y-Y'\rVert for all Y,Y′Y,Y'.

(iii) If M′\mathcal{M}' is a closed set at Hausdorff distance at most hh from M≤r\mathcal{M}_{\leq r}, then ∣d(Y,M′)−Dr(Y)∣≤h\lvert d(Y,\mathcal{M}')-D_r(Y)\rvert\leq h.

Proof. (i) follows from nestedness. (ii) For any ZZ, ∥Y−Z∥≤∥Y−Y′∥+∥Y′−Z∥\lVert Y-Z\rVert\leq\lVert Y-Y'\rVert+\lVert Y'-Z\rVert; take the infimum over Z∈M≤rZ \in\mathcal{M}_{\leq r} and exchange Y,Y′Y,Y'. (iii) Every Z∈M≤rZ \in\mathcal{M}_{\leq r} has a point of M′\mathcal{M}' within hh, so d(Y,M′)≤Dr(Y)+hd(Y,\mathcal{M}') \leq D_r(Y)+h, and symmetrically.

Lemma 2.3 and Propositions 2.4–2.5 below use only nestedness and the triangle inequality; they hold verbatim in any normed space, provided the nested classes are closed. There closedness is an assumption: in L2(μ;Sp)L^2(\mu;S^p) bounded field classes need not be compact, so M≤r\mathcal{M}_{\leq r}, r≥1r \geq1, need not be closed; the fields of Section 4 use only M≤0\mathcal{M}_{\leq0}, which is closed. Finite dimension is used only in Proposition 3.1.

Certificates. The exact margin is generally the value of a nonconvex problem. A lower certificate is a computable Δ‾q(Z)≤Δq(Z)\underline{\Delta}_q(Z) \leq\Delta_q(Z), typically the distance to a superset of M≤q−1\mathcal{M}_{\leq q-1} or a dual bound; an upper certificate is Δ‾q(Z)=∥Z−Z0∥\overline{\Delta}_q(Z)=\lVert Z-Z_0\rVert for an explicit feasible Z0∈M≤q−1Z_0 \in\mathcal{M}_{\leq q-1}.

Proposition 2.4. Let Y^\widehat{Y} be an estimate of YY and let C\mathcal{C} be an event on which ∥Y^−Y∥≤ε\lVert\widehat{Y}-Y\rVert\leq\varepsilon. If Δ‾q(Y^)≤Δq(Y^)≤Δ‾q(Y^)\underline{\Delta}_q(\widehat{Y}) \leq\Delta_q(\widehat{Y}) \leq\overline{\Delta}_q(\widehat{Y}), then on C\mathcal{C}

[Δ‾q(Y^)−ε]+≤Δq(Y)≤Δ‾q(Y^)+ε.\left[\underline{\Delta}_q(\widehat{Y})-\varepsilon\right]_+ \leq\Delta_q(Y) \leq\overline{\Delta}_q(\widehat{Y})+\varepsilon.

In particular, if P(C)≥1−αP(\mathcal{C}) \geq1-\alpha, the interval is a confidence interval of level 1−α1-\alpha for Δq(Y)\Delta_q(Y), and Δ‾q(Y^)>ε\underline{\Delta}_q(\widehat{Y})>\varepsilon certifies at that level that qstr(Y)≥qq_{\mathrm{str}}(Y)\geq q.

Proof. Lemma 2.3(ii) with Y′=Y^Y'=\widehat{Y} gives ∣Δq(Y)−Δq(Y^)∣≤ε\lvert\Delta_q(Y)-\Delta_q(\widehat{Y})\rvert\leq\varepsilon on C\mathcal{C}; combine with the certificates and with Δq≥0\Delta_q\geq0.

Proposition 2.5. Let [Br−,Br+][B_r^-,B_r^+], r=0,…,pr=0,\ldots,p, be intervals with P(⋂r=0p{Br−≤Dr(Y)≤Br+})≥1−αP\left(\bigcap_{r=0}^{p}\{B_r^-\leq D_r(Y)\leq B_r^+\}\right)\geq1-\alpha. Set B‾r−=max⁡j≥rBj−\overline{B}_r^-=\max_{j\geq r}B_j^-, B‾r+=min⁡j≤rBj+\underline{B}_r^+=\min_{j\leq r}B_j^+ and B‾−1+=+∞\underline{B}_{-1}^+=+\infty. For a tolerance τ≥0\tau\geq0 let r∗(τ)=min⁡{r:Dr(Y)≤τ}r^*(\tau)=\min\{r:D_r(Y)\leq\tau\} and

Q^τ={r∈{0,…,p}:B‾r−1+>τ, B‾r−≤τ}.\widehat{\mathcal{Q}}_\tau=\{r\in\{0,\ldots,p\}:\underline{B}_{r-1}^+>\tau,\ \overline{B}_r^-\leq\tau\}.

Then P(r∗(τ)∈Q^τ)≥1−αP(r^*(\tau)\in\widehat{\mathcal{Q}}_\tau)\geq1-\alpha, and r∗(0)=qstr(Y)r^*(0)=q_{\mathrm{str}}(Y).

Proof. On the coverage event, monotonicity (Lemma 2.3(i)) gives Dr≥Dj≥Bj−D_r\geq D_j\geq B_j^- for j≥rj\geq r and Dr≤Dj≤Bj+D_r\leq D_j\leq B_j^+ for j≤rj\leq r, hence B‾r−≤Dr≤B‾r+\overline{B}_r^-\leq D_r\leq\underline{B}_r^+. For r=r∗r=r^*, B‾r−≤Dr≤τ\overline{B}_{r}^-\leq D_r\leq\tau and, if r≥1r\geq1, B‾r−1+≥Dr−1>τ\underline{B}_{r-1}^+\geq D_{r-1}>\tau.

Remark 2.6. Both tightened bands are nonincreasing in rr, so Q^τ\widehat{\mathcal{Q}}_\tau is the integer interval [r−,r+][r_-,r_+] with r−=min⁡{r:B‾r−≤τ}r_-=\min\{r:\overline{B}_r^-\leq\tau\} and r+=max⁡{r:B‾r−1+>τ}r_+=\max\{r:\underline{B}_{r-1}^+>\tau\}, with both extrema restricted to {0,…,p}\{0,\ldots,p\} and min⁡∅=p+1\min\varnothing=p+1. The interval is empty if r−>r+r_->r_+, which may occur outside the simultaneous coverage event. Otherwise it is a set of compatible dimensions rather than a point. The lower band gives the honest lower confidence bound r−r_-, the upper band the upper bound r+r_+. At τ=0\tau=0, r−r_- is a lower confidence bound for qstr(Y)q_{\mathrm{str}}(Y), while r+r_+ is informative only when an upper certificate vanishes exactly; a positive tolerance makes both ends informative.

The structural angle

For a closed set CC and x∈Cx\in C, the tangent (contingent) cone is

Tan⁡C(x)={h:∃ tn↓0, hn→h with x+tnhn∈C}.\operatorname{Tan}_C(x)=\{h:\exists\,t_n\downarrow0,\ h_n\to h\text{ with }x+t_nh_n\in C\}.

[2, 8].

Proposition 3.1. Let CC be closed in a finite-dimensional normed space, x∈Cx\in C and h≠0h\neq0. Then

lim inf⁡t↓0d(x+th,C)t=d(h,Tan⁡C(x)).\liminf_{t\downarrow0}\frac{d(x+th,C)}{t}=d(h,\operatorname{Tan}_C(x)).

The same holds for d(Yt,C)d(Y_t,C) along any path Yt=x+th+o(t)Y_t=x+th+o(t).

Proof. Write ℓ\ell for the left side and T=Tan⁡C(x)T = \operatorname{Tan}_{C}(x). Upper bound: TT is closed, so let w∈Tw \in T attain d(h,T)d(h,T), with tn↓0t_n \downarrow0, wn→ww_n \to w, x+tnwn∈Cx+t_nw_n \in C. Then d(x+tnh,C)/tn≤∥h−wn∥→∥h−w∥d(x+t_nh,C)/t_n \le\lVert h-w_n\rVert\to\lVert h-w\rVert, so ℓ≤d(h,T)\ell\le d(h,T). Lower bound: let tn↓0t_n \downarrow0 realise ℓ\ell and let cn∈Cc_n \in C be nearest to x+tnhx+t_nh. With wn=(cn−x)/tnw_n=(c_n-x)/t_n, ∥h−wn∥→ℓ\lVert h-w_n\rVert\to\ell, so (wn)(w_n) is bounded and has a limit point ww, which lies in TT by definition; hence d(h,T)≤∥h−w∥=ℓd(h,T) \le\lVert h-w\rVert= \ell. The path statement follows from Lemma 2.3(ii) applied to CC. □

Definition 3.2. For Y0∈M≤q−1Y_0 \in\mathcal{M}_{\le q-1} and a direction h≠0h \ne0, the structural angle is κ(Y0,h)=d(h,Tan⁡M≤q−1(Y0))/∥h∥\kappa(Y_0,h)=d(h,\operatorname{Tan}_{\mathcal{M}_{\le q-1}}(Y_0))/\lVert h\rVert. By Proposition 3.1, lim inf⁡t↓0Δq(Y0+th)/t=∥h∥κ(Y0,h)\liminf_{t\downarrow0}\Delta_q(Y_0+th)/t=\lVert h\rVert\kappa(Y_0,h): the margin grows at least at rate ∥h∥κ\lVert h\rVert\kappa as t↓0t\downarrow0, and exactly at that rate along some sequence.

Equality of the full limit requires regularity of CC at xx. The following case covers all the examples below.

Lemma 3.3. Suppose that, in a neighbourhood UU of xx, C∩U=(⋃k=1K(x+Vk))∩UC\cap U=\left(\bigcup_{k=1}^{K}(x+V_k)\right)\cap U for linear subspaces V1,…,VKV_1,\ldots,V_K. Then Tan⁡C(x)=⋃kVk\operatorname{Tan}_{C}(x)=\bigcup_k V_k and d(x+th,C)=t d(h,Tan⁡C(x))d(x+th,C)=t\,d(h,\operatorname{Tan}_{C}(x)) for all sufficiently small t>0t>0.

Proof. The tangent cone is local, and that of a finite union of affine subspaces through xx is the union of their directions. Since d(x+th,C)≤t∥h∥d(x+th,C)\le t\lVert h\rVert, nearest points lie within 2t∥h∥2t\lVert h\rVert of xx, hence in UU for small tt; there d(x+th,C)=min⁡kd(th,Vk)=tmin⁡kd(h,Vk)d(x+th,C)=\min_k d(th,V_k)=t\min_k d(h,V_k). □

Competitors with bounded residual degree

Throughout this section V=Sp\mathcal{V}=\mathbb{S}^{p} with the Frobenius norm; the working-norm normalisation p−1∥⋅∥Fp^{-1}\lVert\cdot\rVert_F multiplies all distances by p−1p^{-1} and leaves angles unchanged. For an edge set EE let

ΩE={S∈Sp:Sij=0 for i≠j, {i,j}∉E}.\Omega_E=\{S\in\mathbb{S}^{p}:S_{ij}=0\text{ for }i\ne j,\ \{i,j\}\notin E\}.

a linear subspace with free diagonal. For a degree budget dˉ∈Z≥0\bar d\in\mathbb{Z}_{\ge0} the competing class of residuals with at most dˉ\bar d off-diagonal entries per row is

Sdˉ=⋃E∈BdˉΩE,Bdˉ={E:deg⁡E(i)≤dˉ for all i},\mathcal{S}_{\bar d}=\bigcup_{E\in\mathcal{B}_{\bar d}}\Omega_E,\qquad\mathcal{B}_{\bar d}=\{E:\deg_E(i)\le\bar d\text{ for all }i\},

a closed cone (a finite union of subspaces). Magnitude bounds on the residual entries are omitted: all local statements below concern points at which such bounds are slack. With this competitor M≤0=Sdˉ\mathcal{M}_{\le0}=\mathcal{S}_{\bar d}.

Lemma 4.1. For Y∈SpY\in\mathbb{S}^{p}, dF(Y,ΩE)2=2∑{i,j}∉EYij2d_F(Y,\Omega_E)^2=2\sum_{\{i,j\}\notin E}Y_{ij}^2, attained by the matrix that copies YY on the diagonal and on $E and vanishes elsewhere.

Proof. ΩE\Omega_E is a coordinate subspace of Sp\mathbb{S}^{p}; orthogonal projection keeps the free coordinates. Each off-diagonal pair contributes twice to the Frobenius norm. □

Distances as bb-matchings

Given nonnegative edge weights ww and vertex capacities bi∈Z≥0b_i\in\mathbb{Z}_{\ge0}, a maximum-weight bb-matching is an edge set EE maximising ∑e∈Ewe\sum_{e\in E}w_e subject to deg⁡E(i)≤bi\deg_E(i)\le b_i. It is solvable in polynomial time [12 Chapters 31–33].

Theorem 4.2. For Y∈SpY\in\mathbb{S}^{p},

dF(Y,Sdˉ)2=W(Y)−max⁡E∈Bdˉ∑{i,j}∈Ewij(Y),wij(Y)=2Yij2,W(Y)=∑i<jwij(Y).d_F(Y,\mathcal{S}_{\bar d})^2=W(Y)-\max_{E\in\mathcal{B}_{\bar d}}\sum_{\{i,j\}\in E}w_{ij}(Y),\qquad w_{ij}(Y)=2Y_{ij}^2,\qquad W(Y)=\sum_{i<j}w_{ij}(Y).

so the rank-zero margin Δ1(Y)\Delta_1(Y) is computable exactly in polynomial time. In the working norm, Δ1(Y)2=W−max⁡E∈Bdˉ∑Ewij\Delta_1(Y)^2=W-\max_{E\in\mathcal{B}_{\bar d}}\sum_Ew_{ij} with wij=2p−2Yij2w_{ij}=2p^{-2}Y_{ij}^2.

Proof. dF(Y,Sdˉ)2=min⁡E∈BdˉdF(Y,ΩE)2d_F(Y,\mathcal{S}_{\bar d})^2=\min_{E\in\mathcal{B}_{\bar d}}d_F(Y,\Omega_E)^2, and by Lemma 4.1 this equals min⁡E(W−∑Ewij)\min_E\left(W-\sum_E w_{ij}\right).

Example 4.3 (strong but redundant, weak but necessary). Let dˉ≥1\bar d\ge1. For v=(e1+e2)/2v=(e_1+e_2)/\sqrt{2} and any a>0a>0, Y=avvTY=avv^{\mathsf T} has a single nonzero off-diagonal pair, which a one-edge EE covers: Δ1(Y)=0\Delta_1(Y)=0 however large the eigenvalue aa. For v=p−1/21v=p^{-1/2}\mathbf{1} and a>0a>0, all p(p−1)/2p(p-1)/2 pairs carry weight 2a2/p22a^2/p^2 and at most ⌊pdˉ/2⌋\lfloor p\bar d/2\rfloor can be covered, so dF(Y,Sdˉ)2≥a2(1−1p−dˉp)>0d_F(Y,\mathcal{S}_{\bar d})^2\ge a^2\left(1-\frac{1}{p}-\frac{\bar d}{p}\right)>0 for dˉ<p−1\bar d<p-1, however small aa is.

Tangent cones and support slack

Let S0∈SdˉS_0\in\mathcal{S}_{\bar d} have active off-diagonal support E0={{i,j}:(S0)ij≠0}E_0=\{\{i,j\}:(S_0)_{ij}\ne0\} and define the support slack ci=dˉ−deg⁡E0(i)≥0c_i=\bar d-\deg_{E_0}(i)\ge0 and

E(E0,c)={E′ disjoint from E0:deg⁡E′(i)≤ci for all i}.\mathcal{E}(E_0,c)=\{E'\text{ disjoint from }E_0:\deg_{E'}(i)\le c_i\text{ for all }i\}.

Proposition 4.4. Tan⁡Sdˉ(S0)=⋃E′∈E(E0,c)ΩE0∪E′\operatorname{Tan}_{\mathcal{S}_{\bar d}}(S_0)=\bigcup_{E'\in\mathcal{E}(E_0,c)}\Omega_{E_0\cup E'}, and for every hh, dF(S0+th,Sdˉ)=t dF(h,Tan⁡Sdˉ(S0))d_F(S_0+th,\mathcal{S}_{\bar d})=t\,d_F(h,\operatorname{Tan}_{\mathcal{S}_{\bar d}}(S_0)) for small t>0t>0.

Proof. Let δ=min⁡E0∣(S0)ij∣>0\delta=\min_{E_0}|(S_0)_{ij}|>0 (or δ=∞\delta=\infty if E0=∅E_0=\varnothing). If max⁡ij∣Sij−(S0)ij∣<δ\max_{ij}|S_{ij}-(S_0)_{ij}|<\delta, the off-diagonal support of SS contains E0E_0, so S∈SdˉS\in\mathcal{S}_{\bar d} if and only if its support is E0∪E′E_0\cup E' with E′∈E(E0,c)E'\in\mathcal{E}(E_0,c), i.e. S∈ΩE0∪E′S\in\Omega_{E_0\cup E'}. Conversely each ΩE0∪E′⊆Sdˉ\Omega_{E_0\cup E'}\subseteq\mathcal{S}_{\bar d} and contains S0S_0. Apply Lemma 3.3.

Saturated supports (c≡0c\equiv0) are regular points with a single tangent subspace; an empty support (c≡dˉc\equiv\bar d) is maximally singular, with tangent cone the whole of Sdˉ\mathcal{S}_{\bar d}.

Corollary 4.5. Let h=vvTh=vv^{\mathsf T} with ∥v∥2=1\|v\|_2=1, so ∥h∥F=1\|h\|_F=1. Then

κ(S0,h)2=1−∑ivi4−2∑{i,j}∈E0vi2vj2−max⁡E′∈E(E0,c)2∑{i,j}∈E′vi2vj2.\kappa(S_0,h)^2=1-\sum_i v_i^4-2\sum_{\{i,j\}\in E_0}v_i^2v_j^2-\max_{E'\in\mathcal{E}(E_0,c)}2\sum_{\{i,j\}\in E'}v_i^2v_j^2.

the last term being a maximum-weight bb-matching with capacities bi=cib_i=c_i on the complement of E0E_0.

Proof. Combine Proposition 4.4 with Lemma 4.1 and 2∑i<jvi2vj2=(∑ivi2)2−∑ivi4=1−∑ivi42\sum_{i<j}v_i^2v_j^2=(\sum_i v_i^2)^2-\sum_i v_i^4=1-\sum_i v_i^4.

The exact angle of equal-magnitude spikes

Lemma 4.6. For integers k≥1k\ge1 and 0≤d≤k−10\le d\le k-1, the complete graph KkK_k has a subgraph of maximum degree at most dd with exactly ⌊kd/2⌋\lfloor kd/2\rfloor edges, and no such subgraph has more.

Proof. The upper bound is the handshake identity 2∣E∣=∑ideg⁡E(i)≤kd2|E|=\sum_i\deg_E(i)\le kd. For the construction label the vertices by Zk\mathbb{Z}_k and let GsG_s be the circulant graph with edges {i,i±1},…,{i,i±s}\{i,i\pm1\},\ldots,\{i,i\pm s\}. If d=k−1d=k-1 take KkK_k. If d<k−1d<k-1 is even, Gd/2G_{d/2} is dd-regular since d/2<k/2d/2<k/2. If dd is odd and kk even, add to G(d−1)/2G_{(d-1)/2} the perfect matching {i,i+k/2}\{i,i+k/2\}, which is disjoint from it because (d−1)/2<k/2(d-1)/2<k/2; the result is dd-regular. If dd is odd and kk odd, then d≤k−2d\le k-2; the edges {i,i+(k−1)/2}\{i,i+(k-1)/2\} are not in G(d−1)/2G_{(d-1)/2} (their difference ±(k−1)/2\pm(k-1)/2 exceeds (d−1)/2(d-1)/2 in absolute value modulo kk) and form a Hamiltonian cycle, because gcd⁡((k−1)/2,k)=1\gcd((k-1)/2,k)=1; it contains a matching with (k−1)/2(k-1)/2 edges, and adding it to G(d−1)/2G_{(d-1)/2} gives k(d−1)/2+(k−1)/2=⌊kd/2⌋k(d-1)/2+(k-1)/2=\lfloor kd/2\rfloor edges with maximum degree dd.

Theorem 4.7. Let vv have support K⊆[p]K\subseteq[p] with ∣K∣=k|K|=k and ∣vi∣=k−1/2|v_i|=k^{-1/2} on KK, let 0≤dˉ≤k−10\le\bar d\le k-1, and g=k−1−dˉg=k-1-\bar d. At a residual S0S_0 with empty off-diagonal support (for instance S0=0S_0=0 diagonal), the structural angle of h=vvTh=vv^{\mathsf T} with respect to M≤0=Sdˉ\mathcal{M}_{\le0}=\mathcal{S}_{\bar d} is

κk,dˉ2=2∣kg/2∣k2=gk+O(k−2).\kappa_{k,\bar d}^2=\frac{2\left\lvert kg/2\right\rvert}{k^2}=\frac{g}{k}+O(k^{-2}).

Proof. By Corollary 4.5 with E0=∅E_0=\varnothing and c≡dˉc\equiv\bar d, the weights 2vi2vj22v_i^2v_j^2 equal 2/k22/k^2 on pairs inside KK and vanish elsewhere, and 1−∑ivi4=1−1/k=(2/k2)(k2)1-\sum_i v_i^4=1-1/k=(2/k^2)\binom{k}{2}. By Lemma 4.6 at most ⌊kdˉ/2⌋\lfloor k\bar d/2\rfloor pairs inside KK can be covered, and this is attained. Hence κ2=(2/k2)((k2)−⌊kdˉ/2⌋)\kappa^2=(2/k^2)(\binom{k}{2}-\lfloor k\bar d/2\rfloor). Since k(k−1)k(k-1) is even, kg=k(k−1)−kdˉkg=k(k-1)-k\bar d has the parity of kdˉk\bar d, and (k2)−⌊kdˉ/2⌋=∣kg/2∣\binom{k}{2}-\lfloor k\bar d/2\rfloor=\lvert kg/2\rvert.

Proposition 4.8 (equal angle, different branching). Let k≥2k \ge2 be even, dˉ=k−2\bar{d}=k-2 and vv as in Theorem 4.7.

(i) At an empty support, κ2=1/k\kappa^2=1/k, and the minimising tangent subspaces are exactly the ΩE\Omega_E whose restriction E∩(K2)E\cap\binom{K}{2} is the complement of a perfect matching of KK: there are (k−1)!!=(k−1)(k−3)⋯1(k-1)!!=(k-1)(k-3)\cdots1 such restrictions.

(ii) At a residual S0S_0 whose support on KK is the complete graph minus a fixed perfect matching P\mathcal{P} (every vertex of KK saturated), κ2=1/k\kappa^2=1/k and the tangent cone restricted to KK is the single subspace ΩE0\Omega_{E_0}.

Proof. (i) Theorem 4.7 with g=1g=1 gives κ2=2(k/2)/k2=1/k\kappa^2=2(k/2)/k^2=1/k. An optimal EE inside KK has k(k−2)/2k(k-2)/2 edges and maximum degree k−2k-2, hence is (k−2)(k-2)-regular on KK, and its complement in KkK_k is 1-regular, i.e. a perfect matching; there are (k−1)!!(k-1)!! perfect matchings of KkK_k. (ii) The slack vanishes on KK, so by Proposition 4.4 no edge inside KK can be added, and the uncovered pairs are those of P\mathcal{P}: κ2=2⋅12⋅1k2=1k\kappa^2=2\cdot\frac{1}{2}\cdot\frac{1}{k^2}=\frac{1}{k}.

The two backgrounds share degree budget and angle; they differ in the number of nuisance branches through which the spike can be explained. This is the reason for Conjecture 6.1.

Fields with a shared residual support

In this subsection the ambient space is L2(μ;Sp)L^2(\mu;\mathbb{S}^p), possibly infinite-dimensional; only M≤0\mathcal{M}_{\le0} and Lemma 2.3 are used. Let Y(u)∈SpY(u)\in\mathbb{S}^p, u∈Uu\in\mathcal{U}, be measurable and square-integrable with respect to a probability measure μ\mu, with norm ∥Y∥2=∫∥Y(u)∥F2 dμ(u)\lVert Y\rVert^2=\int\lVert Y(u)\rVert_F^2\,d\mu(u). Compare the pointwise competitor, in which S(u)∈SdˉS(u)\in\mathcal{S}_{\bar d} for μ\mu-almost every uu with a support that may depend on uu, with the shared-support competitor Sdˉsh=⋃E∈Bdˉ{S:S(u)∈ΩE for μ-almost every u}\mathcal{S}_{\bar d}^{\mathrm{sh}}=\bigcup_{E\in\mathcal{B}_{\bar d}}\{S:S(u)\in\Omega_E\text{ for }\mu\text{-almost every }u\}.

Proposition 4.9. With wˉij=2∫Yij(u)2 dμ(u)\bar w_{ij}=2\int Y_{ij}(u)^2\,d\mu(u) and Wˉ=∑i<jwˉij\bar W=\sum_{i<j}\bar w_{ij},

d(Y,Sdˉsh)2=Wˉ−max⁡E∈Bdˉ∑{i,j}∈Ewˉij,d(Y,pointwise)2=∫dF(Y(u),Sdˉ)2 dμ(u)≤d(Y,Sdˉsh)2.d(Y,\mathcal{S}_{\bar d}^{\mathrm{sh}})^2=\bar W-\max_{E\in\mathcal{B}_{\bar d}}\sum_{\{i,j\}\in E}\bar w_{ij},\qquad d(Y,\text{pointwise})^2=\int d_F(Y(u),\mathcal{S}_{\bar d})^2\,d\mu(u)\le d(Y,\mathcal{S}_{\bar d}^{\mathrm{sh}})^2.

and the inequality can be strict with the left side equal to 0.

Proof. For fixed EE the projection acts pointwise (Lemma 4.1), giving ∫2∑{i,j}∉EYij2 dμ=Wˉ−∑Ewˉij\int2\sum_{\{i,j\}\notin E}Y_{ij}^2\,d\mu=\bar W-\sum_E\bar w_{ij}; minimise over EE. The pointwise problem decouples across uu (a measurable selection of optimal supports exists because Bdˉ\mathcal{B}_{\bar d} is finite), and the shared class is contained in the pointwise one. Example: p=3p=3, dˉ=1\bar d=1, μ\mu uniform on {u1,u2}\{u_1,u_2\}, Y(u1)Y(u_1) supported on the pair {1,2}\{1,2\} and Y(u2)Y(u_2) on {1,3}\{1,3\} with unit entries. Pointwise each matrix is covered by one edge, so the pointwise distance is 0; a shared EE with deg⁡E(1)≤1\deg_E(1)\le1 cannot contain both pairs, so d(Y,S1sh)2=min⁡(wˉ12,wˉ13)=1>0d(Y,\mathcal{S}_1^{\mathrm{sh}})^2=\min(\bar w_{12},\bar w_{13})=1>0.

Positive structural dimension can therefore be certified from the whole field even when every pointwise problem is explained by the competitor.

Certifying positive semidefiniteness

Covariance residuals are positive semidefinite. Let Sdˉ+=Sdˉ∩{S⪰0}\mathcal{S}_{\bar d}^{+}=\mathcal{S}_{\bar d}\cap\{S\succeq0\}, and for a fixed EE write Drel,E=dF(Y,ΩΣE)D_{\mathrm{rel},E}=d_F(Y,\Omega_{\Sigma_E}), DPSD,E=dF(Y,ΩΣE∩{S⪰0})D_{\mathrm{PSD},E}=d_F(Y,\Omega_{\Sigma_E}\cap\{S\succeq0\}), SErelS_E^{\mathrm{rel}} for the projection of Lemma 4.1 and ζE=[−λmin⁡(SErel)]+\zeta_E=[-\lambda_{\min}(S_E^{\mathrm{rel}})]_+.

Proposition 4.10. Drel,E≤DPSD,E≤(Drel,E2+p ζE2)1/2D_{\mathrm{rel},E}\le D_{\mathrm{PSD},E}\le(D_{\mathrm{rel},E}^2+p\,\zeta_E^2)^{1/2}. Consequently

dF(Y,Sdˉ)≤dF(Y,Sdˉ+)≤min⁡E∈Bdˉ(Drel,E2+p ζE2)1/2,d_F(Y,\mathcal{S}_{\bar d})\le d_F(Y,\mathcal{S}_{\bar d}^{+})\le\min_{E\in\mathcal{B}_{\bar d}}(D_{\mathrm{rel},E}^2+p\,\zeta_E^2)^{1/2},

where the left side is the bb-matching value of Theorem 4.2 and the right side can be evaluated at any computed support, for instance the bb-matching optimiser. In the working norm the correction term becomes p−1ζE2p^{-1}\zeta_E^2.

Proof. The lower bounds hold because the PSD classes are subsets. SErel+ζEIS^{\mathrm{rel}}_E+\zeta_E I is positive semidefinite, belongs to ΩE\Omega_E and differs from SErelS^{\mathrm{rel}}_E only on the diagonal, where Y−SErelY-S^{\mathrm{rel}}_E vanishes; hence ∥Y−SErel−ζEI∥F2=Drel,E2+pζE2\lVert Y-S^{\mathrm{rel}}_E-\zeta_E I\rVert_F^2=D_{\mathrm{rel},E}^2+p\zeta_E^2. Multiplying distances by p−1p^{-1} turns pζE2p\zeta_E^2 into p−1ζE2p^{-1}\zeta_E^2.

The same argument shows DPSD,E2≤Drel,E2+∥Λ∥F2D_{\mathrm{PSD},E}^2\le D_{\mathrm{rel},E}^2+\lVert\Lambda\rVert_F^2 for any diagonal Λ\Lambda with SErel+Λ⪰0S^{\mathrm{rel}}_E+\Lambda\succeq0. The scalar shift is the simplest choice and can be conservative when the eigenvalue deficit is localised on few coordinates; minimising ∥Λ∥F\lVert\Lambda\rVert_F under this constraint is a semidefinite programme.

Local Fisher structure

Let z1,…,zTz_1,\ldots,z_T be independent N(0,Σ)N(0,\Sigma) vectors in Rp\mathbb{R}^p and PΣ⊗TP_\Sigma^{\otimes T} their joint law. The Fisher inner product at Σ0≻0\Sigma_0\succ0 is

⟨A,B⟩I,0=12tr⁡(Σ0−1AΣ0−1B),∥A∥I,02=⟨A,A⟩I,0.\langle A,B\rangle_{I,0}=\frac{1}{2}\operatorname{tr}\left(\Sigma_0^{-1}A\Sigma_0^{-1}B\right),\qquad\lVert A\rVert_{I,0}^2=\langle A,A\rangle_{I,0}.

Lemma 5.1 (dimension-free likelihood expansion). Let Σ0=I\Sigma_0=I and let A=Ap,T∈SpA=A_{p,T}\in\mathbb{S}^p satisfy T∥A∥F2=O(1)T\lVert A\rVert_F^2=O(1) (hence ∥A∥2≤∥A∥F→0\lVert A\rVert_2\le\lVert A\rVert_F\to0), with p=pTp=p_T arbitrary. Under PI⊗TP_I^{\otimes T},

log⁡dPI+A⊗TdPI⊗T=12∑t=1T(zt⊤Azt−tr⁡A)−T4tr⁡(A2)+oP(1).\log\frac{dP_{I+A}^{\otimes T}}{dP_I^{\otimes T}}=\frac{1}{2}\sum_{t=1}^T\left(z_t^\top A z_t-\operatorname{tr}A\right)-\frac{T}{4}\operatorname{tr}(A^2)+o_P(1).

If moreover T2tr⁡(A2)→σ2\frac{T}{2}\operatorname{tr}(A^2)\to\sigma^2, the right side converges in law to N(−σ2/2,σ2)N(-\sigma^2/2,\sigma^2), and PI+A⊗TP_{I+A}^{\otimes T} and PI⊗TP_I^{\otimes T} are mutually contiguous.

Proof. Assume ∥A∥2≤1/2\lVert A\rVert_2\le1/2 and let R=(I+A)−1R=(I+A)^{-1}, so ∥R∥2≤2\lVert R\rVert_2\le2 and R−I=−A+A2RR-I=-A+A^2R. The log-likelihood ratio is

ℓT=−T2log⁡det⁡(I+A)+12∑tzt⊤Azt−12∑tzt⊤A2Rzt.\ell_T=-\frac{T}{2}\log\det(I+A)+\frac{1}{2}\sum_t z_t^\top A z_t-\frac{1}{2}\sum_t z_t^\top A^2Rz_t.

For ∣λ∣≤1/2|\lambda|\le1/2, ∣log⁡(1+λ)−λ+λ2/2∣≤∣λ∣3|\log(1+\lambda)-\lambda+\lambda^2/2|\le|\lambda|^3, so log⁡det⁡(I+A)=tr⁡A−12tr⁡A2+r1\log\det(I+A)=\operatorname{tr}A-\frac{1}{2}\operatorname{tr}A^2+r_1 with ∣r1∣≤∥A∥2∥A∥F2|r_1|\le\lVert A\rVert_2\lVert A\rVert_F^2, and T∣r1∣→0T|r_1|\to0. The matrix A2RA^2R is symmetric with ∥A2R∥F≤2∥A∥2∥A∥F\lVert A^2R\rVert_F\le2\lVert A\rVert_2\lVert A\rVert_F; hence ∑tzt⊤A2Rzt\sum_t z_t^\top A^2Rz_t has mean Ttr⁡(A2R)=Ttr⁡A2−Ttr⁡(A3R)T\operatorname{tr}(A^2R)=T\operatorname{tr}A^2-T\operatorname{tr}(A^3R), with T∣tr⁡(A3R)∣≤2T∥A∥2∥A∥F2→0T|\operatorname{tr}(A^3R)|\le2T\lVert A\rVert_2\lVert A\rVert_F^2\to0, and variance 2T∥A2R∥F2≤8∥A∥22T∥A∥F2→02T\lVert A^2R\rVert_F^2\le8\lVert A\rVert_2^2T\lVert A\rVert_F^2\to0. Collecting terms gives the expansion. For the limit law, Xt=12(zt⊤Azt−tr⁡A)X_t=\frac{1}{2}(z_t^\top A z_t-\operatorname{tr}A) are i.i.d., centred, with variance 12tr⁡A2\frac{1}{2}\operatorname{tr}A^2; as second-order Gaussian chaoses they satisfy E∣Xt∣3≤C(tr⁡A2)3/2E|X_t|^3\le C(\operatorname{tr}A^2)^{3/2} by hypercontractivity ([6]), so for σ2>0\sigma^2>0 the Lyapunov ratio is O(T−1/2)O(T^{-1/2}) and ∑tXt⇒N(0,σ2)\sum_tX_t\Rightarrow N(0,\sigma^2); for σ2=0\sigma^2=0 the variance T2tr⁡A2\frac{T}{2}\operatorname{tr}A^2 of ∑tXt\sum_tX_t tends to 00. Contiguity follows from Le Cam’s first lemma ([7]).

For general Σ0\Sigma_0, whitening with Σ0−1/2\Sigma_0^{-1/2} maps a perturbation GG to A=Σ0−1/2GΣ0−1/2A=\Sigma_0^{-1/2}G\Sigma_0^{-1/2}, and T2tr⁡(A2)=T∥G∥I,02\frac{T}{2}\operatorname{tr}(A^2)=T\lVert G\rVert_{I,0}^2. In particular a perturbation is locally detectable at scale (T∥G∥I,0)−1(\sqrt{T}\lVert G\rVert_{I,0})^{-1}.

Corollary 5.2. If M=Σ−1/2GΣ−1/2M=\Sigma^{-1/2}G\Sigma^{-1/2} satisfies ∥M∥2→0\lVert M\rVert_2\to0, then KL⁡(PΣ⊗T∥PΣ+G⊗T)=T4∥M∥F2{1+O(∥M∥2)}=T2∥G∥I,Σ2{1+O(∥M∥2)}\operatorname{KL}(P_\Sigma^{\otimes T}\mathbin\Vert P_{\Sigma+G}^{\otimes T})=\frac{T}{4}\lVert M\rVert_F^2\{1+O(\lVert M\rVert_2)\}=\frac{T}{2}\lVert G\rVert_{I,\Sigma}^2\{1+O(\lVert M\rVert_2)\}.

Proof. With eigenvalues λj\lambda_j of MM, KL⁡=T2∑jϕ(λj)\operatorname{KL}=\frac{T}{2}\sum_j\phi(\lambda_j) where ϕ(λ)=11+λ−1+log⁡(1+λ)=λ22+O(∣λ∣3)\phi(\lambda)=\frac{1}{1+\lambda}-1+\log(1+\lambda)=\frac{\lambda^2}{2}+O(|\lambda|^3).

The Fisher quotient at a rank-rr background

When the null model contains a rank-rr component, a perturbation is informative only through the part not absorbed by moving that component. Let Σ0=L0+S0\Sigma_0=L_0+S_0 with L0=UΛU⊤L_0=U\Lambda U^\top, U∈Rp×rU\in\mathbb{R}^{p\times r} with orthonormal columns, Λ≻0\Lambda\succ0 diagonal and S0≻0S_0\succ0, so L0⊥LL_0\perp L have orthogonal columns spanning the orthogonal complement of the range of UU. Assume ∥L0∥≤ρ\lVert L_0\rVert\le\rho, so that the magnitude bound of Section 2 is slack; then the tangent cone of Lr\mathcal{L}_r at L0L_0 is the tangent space of the rank-rr positive semidefinite matrices,

Tan⁡U={UB⊤+BU⊤:B∈Rp×r}={Q∈Sp:U⊥⊤QU⊥=0}.\operatorname{Tan}_U=\{UB^\top+BU^\top:B\in\mathbb{R}^{p\times r}\}=\{Q\in\mathbb{S}^p:U_\perp^\top Q U_\perp=0\}.

([1]).

Proposition 5.3. For G∈SpG \in\mathbb{S}^{p}, with G22=U⊥⊤GU⊥G_{22}=U_{\perp}^{\top}GU_{\perp} and S⊥=U⊥⊤S0U⊥≻0S_{\perp}=U_{\perp}^{\top}S_{0}U_{\perp}\succ0,

min⁡Q∈Tan⁡U∥G−Q∥I,02=12∥S⊥−1/2G22S⊥−1/2∥F2.\min_{Q\in\operatorname{Tan}_{U}}\|G-Q\|_{I,0}^{2}=\frac{1}{2}\left\|S_{\perp}^{-1/2}G_{22}S_{\perp}^{-1/2}\right\|_{F}^{2}.

In particular the Fisher information left after projecting out Tan⁡U\operatorname{Tan}_{U} does not depend on the strength Λ\Lambda of the background.

Proof. Write M=G−QM=G-Q. As QQ ranges over Tan⁡U\operatorname{Tan}_{U}, MM ranges over {M∈Sp:U⊥⊤MU⊥=G22}\{M\in\mathbb{S}^{p}:U_{\perp}^{\top}MU_{\perp}=G_{22}\}, so we minimise the convex function 12tr⁡(Σ0−1MΣ0−1M)\frac{1}{2}\operatorname{tr}(\Sigma_{0}^{-1}M\Sigma_{0}^{-1}M) under a linear constraint. Stationarity gives Σ0−1MΣ0−1=U⊥ΓU⊥⊤\Sigma_{0}^{-1}M\Sigma_{0}^{-1}=U_{\perp}\Gamma U_{\perp}^{\top} for a symmetric multiplier Γ\Gamma, i.e. M=Σ0U⊥ΓU⊥⊤Σ0M=\Sigma_{0}U_{\perp}\Gamma U_{\perp}^{\top}\Sigma_{0}. Since U⊥⊤L0=0U_{\perp}^{\top}L_{0}=0, U⊥⊤Σ0U⊥=S⊥U_{\perp}^{\top}\Sigma_{0}U_{\perp}=S_{\perp}, and the constraint becomes S⊥ΓS⊥=G22S_{\perp}\Gamma S_{\perp}=G_{22}, so Γ=S⊥−1G22S⊥−1\Gamma=S_{\perp}^{-1}G_{22}S_{\perp}^{-1}. The value is 12tr⁡(U⊥ΓU⊥⊤Σ0U⊥ΓU⊥⊤Σ0)=12tr⁡(ΓS⊥ΓS⊥)=12tr⁡(S⊥−1G22S⊥−1G22)\frac{1}{2}\operatorname{tr}(U_{\perp}\Gamma U_{\perp}^{\top}\Sigma_{0}U_{\perp}\Gamma U_{\perp}^{\top}\Sigma_{0})=\frac{1}{2}\operatorname{tr}(\Gamma S_{\perp}\Gamma S_{\perp})=\frac{1}{2}\operatorname{tr}(S_{\perp}^{-1}G_{22}S_{\perp}^{-1}G_{22}).

Lemma 5.4. Let GG be supported on a coordinate set I\mathcal{I}, i.e. G=EIGIEI⊤G=E_{\mathcal{I}}G_{\mathcal{I}}E_{\mathcal{I}}^{\top} with EIE_{\mathcal{I}} the coordinate embedding, and let ϑ=∥U⊤EI∥22≤1\vartheta=\|U^{\top}E_{\mathcal{I}}\|_{2}^{2}\leq1. Then

(1−ϑ)∥G∥F≤∥U⊥⊤GU⊥∥F≤∥G∥F,(1-\vartheta)\|G\|_{F}\leq\|U_{\perp}^{\top}GU_{\perp}\|_{F}\leq\|G\|_{F},

and therefore

(1−ϑ)22∥S⊥∥22∥G∥F2≤min⁡Q∈Tan⁡U∥G−Q∥I,02≤12λmin⁡(S⊥)2∥G∥F2.\frac{(1-\vartheta)^{2}}{2\|S_{\perp}\|_{2}^{2}}\|G\|_{F}^{2}\leq\min_{Q\in\operatorname{Tan}_{U}}\|G-Q\|_{I,0}^{2}\leq\frac{1}{2\lambda_{\min}(S_{\perp})^{2}}\|G\|_{F}^{2}.

Proof. Let N=EI⊤U⊥U⊥⊤EI=I−EI⊤UU⊤EIN=E_{\mathcal{I}}^{\top}U_{\perp}U_{\perp}^{\top}E_{\mathcal{I}}=I-E_{\mathcal{I}}^{\top}UU^{\top}E_{\mathcal{I}}, whose eigenvalues lie in [1−ϑ,1][1-\vartheta,1]. Then ∥U⊥⊤GU⊥∥F2=tr⁡(GINGIN)=∥N1/2GIN1/2∥F2∈[(1−ϑ)2,1]⋅∥GI∥F2\|U_{\perp}^{\top}GU_{\perp}\|_{F}^{2}=\operatorname{tr}(G_{\mathcal{I}}NG_{\mathcal{I}}N)=\|N^{1/2}G_{\mathcal{I}}N^{1/2}\|_{F}^{2}\in[(1-\vartheta)^{2},1]\cdot\|G_{\mathcal{I}}\|_{F}^{2}. The last display follows from Proposition 5.3 and ∥S⊥−1/2CS⊥−1/2∥F∈[∥C∥F/∥S⊥∥2,∥C∥F/λmin⁡(S⊥)]\|S_{\perp}^{-1/2}CS_{\perp}^{-1/2}\|_{F}\in[\|C\|_{F}/\|S_{\perp}\|_{2},\|C\|_{F}/\lambda_{\min}(S_{\perp})].

For coordinate-supported perturbations, Lemma 5.4 controls the Frobenius norm retained after compression onto the orthogonal complement of the latent space. The corresponding Fisher information is weighted by S⊥−1S_{\perp}^{-1}, as specified in Proposition 5.3. For perturbations through the fixed decomposition Σ0=L0+S0\Sigma_{0}=L_{0}+S_{0}, allowing the residual to move along a branch ΩE\Omega_{E} gives the Fisher distance to Tan⁡U+ΩE\operatorname{Tan}_{U}+\Omega_{E}, minimised over the branches through S0S_{0}. The sandwich S⊥−1/2U⊥⊤(⋅)U⊥S⊥−1/2S_{\perp}^{-1/2}U_{\perp}^{\top}(\cdot)U_{\perp}S_{\perp}^{-1/2} couples the residual coordinate subspaces. Identifying this branchwise profile with the tangent geometry of the full structural null also requires control of other decompositions and of the projection. These joint questions remain open.

A detection floor from the margin

The margin bounds from below what any test can detect. Let N⊂{Σ≻0}\mathcal{N}\subset\{\Sigma\succ0\} be a null family and, for a test ψ\psi, write R(ψ;Σ0,Σ1)=EΣ0ψ+EΣ1(1−ψ)\mathcal{R}(\psi;\Sigma_{0},\Sigma_{1})=E_{\Sigma_{0}}\psi+E_{\Sigma_{1}}(1-\psi) for the sum of its two error probabilities under PΣ0⊗TP_{\Sigma_{0}}^{\otimes T} and PΣ1⊗TP_{\Sigma_{1}}^{\otimes T}.

Proposition 5.5 (margin floor). Let Σ1≻0\Sigma_{1}\succ0 and let Σ0∈N\Sigma_{0}\in\mathcal{N} satisfy λmin⁡(Σ0)≥c>0\lambda_{\min}(\Sigma_{0})\geq c>0 and ∥Σ1−Σ0∥F≤c/2\|\Sigma_{1}-\Sigma_{0}\|_{F}\leq c/2. Then for every test ψ\psi,

R(ψ;Σ0,Σ1)≥1−T∥Σ1−Σ0∥F2c.\mathcal{R}(\psi;\Sigma_{0},\Sigma_{1})\geq1-\frac{\sqrt{T}\|\Sigma_{1}-\Sigma_{0}\|_{F}}{\sqrt{2}c}.

In particular, let N=Sdˉ∩{Σ≻0}\mathcal{N}=\mathcal{S}_{\bar d}\cap\{\Sigma\succ0\} and measure Δ1\Delta_{1} in the unnormalised Frobenius norm. If a nearest point Σ0\Sigma_{0} of Sdˉ\mathcal{S}_{\bar d} to Σ1\Sigma_{1} satisfies λmin⁡(Σ0)≥c\lambda_{\min}(\Sigma_{0})\geq c and Δ1(Σ1)≤c/2\Delta_{1}(\Sigma_{1})\leq c/2, the bound reads 1−T/2Δ1(Σ1)/c1-\sqrt{T}/2\Delta_{1}(\Sigma_{1})/c: no test separates Σ1\Sigma_{1} from the null once TΔ1(Σ1)→0\sqrt{T}\Delta_{1}(\Sigma_{1})\to0, whatever pp.

Proof. Let G=Σ1−Σ0G=\Sigma_{1}-\Sigma_{0} and M=Σ0−1/2GΣ0−1/2M=\Sigma_{0}^{-1/2}G\Sigma_{0}^{-1/2}, so ∥M∥2≤∥M∥F≤∥G∥F/c≤12\|M\|_{2}\leq\|M\|_{F}\leq\|G\|_{F}/c\leq\frac{1}{2}. For ∣λ∣≤12|\lambda|\leq\frac{1}{2} the function ϕ\phi in the proof of Corollary 5.2 satisfies 0≤ϕ(λ)≤2λ20\leq\phi(\lambda)\leq2\lambda^{2}, hence KL⁡(PΣ0⊗T∥PΣ1⊗T)≤T∥M∥F2≤T∥G∥F2/c2\operatorname{KL}(P_{\Sigma_{0}}^{\otimes T}\|P_{\Sigma_{1}}^{\otimes T})\leq T\|M\|_{F}^{2}\leq T\|G\|_{F}^{2}/c^{2}. By Pinsker’s inequality the total variation distance is at most T/2∥G∥F/c\sqrt{T/2}\|G\|_{F}/c, and R(ψ;Σ0,Σ1)≥1−TV\mathcal{R}(\psi;\Sigma_{0},\Sigma_{1})\geq1-\mathrm{TV} for every test.

Corollary 5.6. In the setting of Proposition 4.8, let Σ0∈Sk−2\Sigma_0 \in\mathcal{S}_{k-2} be either residual of parts (i)–(ii), with λmin⁡(Σ0)≥c>0\lambda_{\min}(\Sigma_0) \ge c > 0, and let Σ1=Σ0+θvvT\Sigma_1 = \Sigma_0 + \theta vv^{\mathsf{T}} with θ>0\theta> 0. If θ/k≤c/4\theta/\sqrt{k} \le c/4, there is Σ0∈Sk−2\Sigma_0 \in\mathcal{S}_{k-2}, Σ0≻0\Sigma_0 \succ0, such that every test satisfies

R(ψ;Σ0,Σ1)≥1−2cθT/k.\mathcal{R}(\psi;\Sigma_0,\Sigma_1) \ge1-\frac{\sqrt{2}}{c}\theta\sqrt{T/k}.

In particular no test separates Σ1\Sigma_1 from Sk−2\mathcal{S}_{k-2} when θT/k→0\theta\sqrt{T/k}\to0, uniformly in pp and kk for fixed cc.

Proof. Let EE be the support of a minimising subspace in Proposition 4.8: in case (i) any EE whose restriction to KK is the complement of a perfect matching P\mathcal{P} of KK, in case (ii) E=E0E=E_0 with the matching P\mathcal{P} of that statement. Let Σ0\Sigma_0 be the orthogonal projection of Σ1\Sigma_1 onto ΩE\Omega_E; since Σ0∈ΩE\Sigma_0\in\Omega_E in both cases, Σ0∈ΩE⊆Sk−2\Sigma_0\in\Omega_E\subseteq\mathcal{S}_{k-2} and G=Σ1−Σ0G=\Sigma_1-\Sigma_0 is supported on the pairs of P\mathcal{P}, with entries ±θ/k\pm\theta/k. Hence ∥G∥F=θ/k\lVert G\rVert_F=\theta/\sqrt{k} and ∥G∥2=θ/k\lVert G\rVert_2=\theta/k, for every θ>0\theta>0. By Weyl’s inequality λmin⁡(Σ0)≥λmin⁡(Σ1)−θ/k≥c−c/4≥c/2\lambda_{\min}(\Sigma_0)\ge\lambda_{\min}(\Sigma_1)-\theta/k\ge c-c/4\ge c/2, using λmin⁡(Σ1)≥λmin⁡(Σ0)\lambda_{\min}(\Sigma_1)\ge\lambda_{\min}(\Sigma_0). Apply Proposition 5.5 with c/2c/2 in place of cc, noting ∥G∥F≤c/4\lVert G\rVert_F\le c/4.

For small θ\theta, ∥G∥F=θ/k=θκ\lVert G\rVert_F=\theta/\sqrt{k}=\theta\kappa is exactly the margin Δ1(Σ1)\Delta_1(\Sigma_1) (Propositions 4.4 and 4.8); the construction above avoids any condition on the size of the active entries of Σ0\Sigma_0. The floor depends on the background only through the structural angle and is therefore the same at the empty and at the saturated support. It is a single-branch bound: it ignores that, at an empty support, the null can be approached through any of (k−1)!!(k-1)!! branches.

Remark 5.7 (two scales). Proposition 2.4 resolves margins above the estimation error ϵ≍∥Σ^−Σ∥F\epsilon\asymp\lVert\widehat{\Sigma}-\Sigma\rVert_F, which for the sample covariance with tr⁡Σ≍p\operatorname{tr}\Sigma\asymp p is of order p/Tp/\sqrt{T} in the unnormalised Frobenius norm, since E∥Σ^−Σ∥F2={(tr⁡Σ)2+∥Σ∥F2}/T\mathbb{E}\lVert\widehat{\Sigma}-\Sigma\rVert_F^2=\{(\operatorname{tr}\Sigma)^2+\lVert\Sigma\rVert_F^2\}/T. Proposition 5.5 places the information floor at Δ1≍1/T\Delta_1\asymp1/\sqrt{T}. Lipschitz certificates are therefore valid but may be conservative by a factor up to pp; certificates that exploit the structure of the competitor rather than the global norm of the estimation error are needed to close this gap, and are not treated here.

Conjectures and open problems

The degree-constrained distance and tangent formulas above account exactly for unions of support branches. The Gaussian bounds, however, do not determine the sharp testing cost of those unions. We specify the structural testing target before stating the rate predictions.

Separated hypotheses and local benchmarks

Let 0<c<C<∞0<c<C<\infty be fixed, and restrict the covariance classes to cI⪯Σ⪯CIcI\preceq\Sigma\preceq CI. For a declared neighbourhood UT\mathcal{U}_T and margin δ>0\delta>0, define

H0=M≤r∩UT,H1(δ)={Σ∈M≤r+1∩UT:dF(Σ,M≤r)≥δ}.\mathcal{H}_0=\mathcal{M}_{\le r}\cap\mathcal{U}_T,\qquad\mathcal{H}_1(\delta)=\{\Sigma\in\mathcal{M}_{\le r+1}\cap\mathcal{U}_T:d_F(\Sigma,\mathcal{M}_{\le r})\ge\delta\}.

The classes here use positive-semidefinite residuals, the degree budget and any magnitude caps specified for the experiment. The margin is measured in the unnormalised Frobenius norm. For a test ψ\psi with values in [0,1][0,1], its minimax sum of errors is

RT(δ)=inf⁡ψ{sup⁡Σ∈H0EΣψ+sup⁡Σ∈H1(δ)EΣ(1−ψ)}.R_T(\delta)=\inf_{\psi}\left\{\sup_{\Sigma\in\mathcal{H}_0}\mathbb{E}_{\Sigma}\psi+\sup_{\Sigma\in\mathcal{H}_1(\delta)}\mathbb{E}_{\Sigma}(1-\psi)\right\}.

This excludes a covariance that has both a spike representation and a lower-rank representation from the alternatives. A nonzero spike coefficient alone would not ensure structural separation.

The rank-zero amplitude benchmarks use Σ=S+θvvT\Sigma=S+\theta vv^{\mathsf{T}}, with S∈SpS\in\mathcal{S}_p positive definite and vv an equal-magnitude unit vector on kk coordinates. For an amplitude parameter a>0a>0, restrict a≤θ≤2aa\le\theta\le2a and require

c⋆θκk,d~≤dF(Σ,M≤0)≤C⋆θκk,d~,0<c⋆<1<C⋆<∞.c_\star\theta\kappa_{k,\tilde d}\le d_F(\Sigma,\mathcal{M}_{\le0})\le C_\star\theta\kappa_{k,\tilde d},\qquad0<c_\star<1<C_\star<\infty.

Thus amplitude and structural margin are linked on the declared alternatives. Both hypotheses are localised around the specified residual background, with uniform spectral bounds. In a saturated-support experiment the neighbourhood preserves the active edges; at an empty background it permits all degree-admissible support branches. Neighbourhood widths must contain the signal-scale perturbations. The precise ranges of localisation and growth for which sharp matching is possible form part of the rate problem.

For these amplitude-indexed families, a boundary aTa_T means that the minimax error tends to zero when a/aT→∞a/a_T \to\infty, and stays away from zero when a/aT→0a/a_T \to0, within the declared local regime. Unknown-direction experiments take the union over the permitted sparse directions and their corresponding separated alternatives.

Static and spectral predictions

Conjecture 6.1 (static branching benchmark). Consider the separated rank-zero local experiments above with a known balanced signed direction on an even number kk of coordinates and dˉ=k−2\bar d = k-2. Under suitable localisation and growth conditions and up to logarithmic factors, the boundary is aT≳k/Ta_T \gtrsim k/\sqrt{T} at an empty residual support and aT≳k/Ta_T \gtrsim\sqrt{k/T} at a saturated clique-minus-matching support: branch uncertainty costs a factor of order k\sqrt{k} despite the equal structural angles of Proposition 4.8. For an unknown direction and general 0≤dˉ≤k−20 \le\bar d \le k-2, writing g=k−1−dˉg = k-1-\bar d, the boundary is of order

max⁡{klog⁡(ep/k)T,k(dˉ+1)gT}.\max\left\{\sqrt{\frac{k\log(ep/k)}{T}},\sqrt{\frac{k(\bar d+1)}{gT}}\right\}.

Corollary 5.6 establishes a k/T\sqrt{k/T} floor in both endpoint configurations. It does not prove the larger branching rate, its achievability, or uniformity over the local nuisance families. Establishing or revising the proposed orders requires composite lower bounds and matching tests; the count of branches alone does not determine their statistical cost.

For the dynamic benchmark, the data are TT consecutive observations of a real stationary Gaussian process. The spectral classes are Hermitian, conjugate-symmetric and uniformly bounded between cIcI and CICI. On [0,π][0,\pi] the partition into mm equal bands is known; the residual support may be independently reconfigured on those bands and is reflected to negative frequencies. For Nm\mathcal{N}_m the corresponding rank-zero spectral class, measure separation by

dF,2(F,Nm)=inf⁡H∈Nm{1π∫0π∥F(ω)−H(ω)∥F2 dω}1/2.d_{F,2}(F,\mathcal{N}_m)=\inf_{H\in\mathcal{N}_m}\left\{\frac{1}{\pi}\int_0^\pi\lVert F(\omega)-H(\omega)\rVert_F^2\,\mathrm{d}\omega\right\}^{1/2}.

The alternatives F(ω)=S(ω)+θvvTF(\omega)=S(\omega)+\theta vv^{\mathsf T} have a≤θ≤2aa\le\theta\le2a and margin between c∗θ/kc_*\theta/\sqrt{k} and C∗θ/kC_*\theta/\sqrt{k}. The minimax risk is defined as above using the exact finite-sample Gaussian law with block Toeplitz covariance. Unknown band locations and globally smooth support transitions define additional experiments.

Conjecture 6.2 (switching benchmark). For the separated stationary-Gaussian benchmark with known balanced direction on an even number kk of coordinates, dˉ=k−2\bar d=k-2, and m=J+1m=J+1 equal bands with independently reconfigured perfect matchings, the boundary is, up to logarithmic factors,

aT≳km1/4T,k2mT→0,km5/2log⁡(eT)T→0,a_T\gtrsim\frac{km^{1/4}}{\sqrt{T}},\qquad\frac{k^2\sqrt{m}}{T}\to0,\qquad\frac{km^{5/2}\log(eT)}{T}\to0,

in the exact Toeplitz experiment, under suitable localisation conditions.

The fourth-root penalty is a prediction to be tested, not an implication of the deterministic switching geometry. A Whittle approximation, if used, must be justified uniformly over the declared bands and nuisance alternatives; the independent-observation expansion of Lemma 5.1 does not itself provide that justification.

Further open problems are joint profiling of the low-rank and residual tangents, separation bounds for growing rank, information loss under discretisation and spectral smoothing, and computational certificates narrow enough at the statistical resolution. A characterization that changes a conjectured exponent or identifies an additional necessary condition resolves part of these questions.

References

  1. [1]Absil, P.-A., Mahony, R. and Sepulchre, R. (2008). Optimization Algorithms on Matrix Manifolds. Princeton University Press.DOI
  2. [2]Aubin, J.-P. and Frankowska, H. (1990). Set-Valued Analysis. Birkhäuser.
  3. [3]Bai, J. and Ng, S. (2002). Determining the number of factors in approximate factor models. Econometrica 70, 191–221.DOI
  4. [4]Chandrasekaran, V., Sanghavi, S., Parrilo, P. A. and Willsky, A. S. (2011). Rank–sparsity incoherence for matrix decomposition. SIAM Journal on Optimization 21, 572–596.
  5. [5]Chandrasekaran, V., Parrilo, P. A. and Willsky, A. S. (2012). Latent variable graphical model selection via convex optimization. Annals of Statistics 40, 1935–1967.
  6. [6]Janson, S. (1997). Gaussian Hilbert Spaces. Cambridge University Press.DOI
  7. [7]Le Cam, L. and Yang, G. L. (2000). Asymptotics in Statistics: Some Basic Concepts, 2nd ed. Springer.
  8. [8]Rockafellar, R. T. and Wets, R. J.-B. (1998). Variational Analysis. Springer.DOI
  9. [12]Schrijver, A. (2003). Combinatorial Optimization: Polyhedra and Efficiency. Springer.

Paper details

vlnc:W_qyg4te22
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents