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 b-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.
Section 2 defines structural dimension and the margin Δq and shows that certified bounds on Δq and simultaneous bands on the distances yield honest confidence statements (Propositions 2.4 and 2.5).
Section 3 identifies the first-order behaviour of the margin with the distance of the entering direction to a tangent cone (Proposition 3.1).
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 b-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).
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-r background (Proposition 5.3), and a detection floor in terms of the margin (Proposition 5.5).
Notation.Sp is the space of real symmetric p×p matrices with Frobenius inner product ⟨A,B⟩=tr(AB) and norm ∥⋅∥F; ∥⋅∥2 is the operator norm, A⪰0 (A≻0) means positive semidefinite (definite), λmin is the smallest eigenvalue and [x]+=max{x,0}. For a set C and a point Y in a normed space, d(Y,C)=infZ∈C∥Y−Z∥. Unordered pairs {i,j}, i=j, of [p]={1,…,p} are called edges; for an edge set E, degE(i) is the number of edges of E containing i.
Structural dimension and its margin
Let Y∈Sp be an observable second-order object, typically a covariance matrix, and let ∥⋅∥ be a norm on Sp. A competing classS⊆Sp is a closed set containing 0, declared on scientific grounds before any rank is inferred. Fix a magnitude bound ρ>0 and, for r≥0, let Lr={L∈Sp:L⪰0,rankL≤r,∥L∥≤ρ}, a compact set. The nested classes
M≤r=Lr+S,M≤0⊆M≤1⊆⋯,
are closed, as sums of a compact and a closed set. Assume Y∈M≤p.
Definition 2.1. The structural dimension of Y is qstr(Y)=min{r:Y∈M≤r}. For r≥0 let Dr(Y)=d(Y,M≤r), and for q≥1 define the marginΔq(Y)=Dq−1(Y), the amount by which the explanation with q−1 latent directions fails.
Since the classes are closed, qstr(Y)=min{r:Dr(Y)=0}, and Y requires a q-th direction exactly when Δq(Y)>0.
Remark 2.2. The bound ρ cannot be replaced by a closure of the unbounded sum without changing the meaning of qstr. Let S be the diagonal matrices (of either sign) and wn=(n,1/n,1/n). Then wnwnT−diag(wnwnT) converges to the matrix Y with Y12=Y21=Y13=Y31=1 and all other entries 0, so Y lies in the closure of the unbounded rank-one class; yet Y has no exact rank-one-plus-diagonal decomposition, which would require w1w2=w1w3=1 and w2w3=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) is nonincreasing.
(ii) ∣Dr(Y)−Dr(Y′)∣≤∥Y−Y′∥ for all Y,Y′.
(iii) If M′ is a closed set at Hausdorff distance at most h from M≤r, then ∣d(Y,M′)−Dr(Y)∣≤h.
Proof. (i) follows from nestedness. (ii) For any Z, ∥Y−Z∥≤∥Y−Y′∥+∥Y′−Z∥; take the infimum over Z∈M≤r and exchange Y,Y′. (iii) Every Z∈M≤r has a point of M′ within h, so d(Y,M′)≤Dr(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) bounded field classes need not be compact, so M≤r, r≥1, need not be closed; the fields of Section 4 use only M≤0, 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 Δq(Z)≤Δq(Z), typically the distance to a superset of M≤q−1 or a dual bound; an upper certificate is Δq(Z)=∥Z−Z0∥ for an explicit feasible Z0∈M≤q−1.
Proposition 2.4.LetYbe an estimate ofYand letCbe an event on which∥Y−Y∥≤ε. IfΔq(Y)≤Δq(Y)≤Δq(Y), then onC
[Δq(Y)−ε]+≤Δq(Y)≤Δq(Y)+ε.
In particular, ifP(C)≥1−α, the interval is a confidence interval of level1−αforΔq(Y), andΔq(Y)>εcertifies at that level thatqstr(Y)≥q.
Proof. Lemma 2.3(ii) with Y′=Y gives ∣Δq(Y)−Δq(Y)∣≤ε on C; combine with the certificates and with Δq≥0.
Proposition 2.5.Let[Br−,Br+], r=0,…,p, be intervals withP(⋂r=0p{Br−≤Dr(Y)≤Br+})≥1−α. SetBr−=maxj≥rBj−, Br+=minj≤rBj+andB−1+=+∞. For a toleranceτ≥0letr∗(τ)=min{r:Dr(Y)≤τ}and
Qτ={r∈{0,…,p}:Br−1+>τ,Br−≤τ}.
ThenP(r∗(τ)∈Qτ)≥1−α, andr∗(0)=qstr(Y).
Proof. On the coverage event, monotonicity (Lemma 2.3(i)) gives Dr≥Dj≥Bj− for j≥r and Dr≤Dj≤Bj+ for j≤r, hence Br−≤Dr≤Br+. For r=r∗, Br−≤Dr≤τ and, if r≥1, Br−1+≥Dr−1>τ.
Remark 2.6. Both tightened bands are nonincreasing in r, so Qτ is the integer interval [r−,r+] with r−=min{r:Br−≤τ} and r+=max{r:Br−1+>τ}, with both extrema restricted to {0,…,p} and min∅=p+1. The interval is empty if 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−, the upper band the upper bound r+. At τ=0, r− is a lower confidence bound for qstr(Y), while r+ is informative only when an upper certificate vanishes exactly; a positive tolerance makes both ends informative.
The structural angle
For a closed set C and x∈C, the tangent (contingent) cone is
Proposition 3.1.LetCbe closed in a finite-dimensional normed space,x∈Candh=0. Then
t↓0liminftd(x+th,C)=d(h,TanC(x)).
The same holds ford(Yt,C)along any pathYt=x+th+o(t).
Proof. Write ℓ for the left side and T=TanC(x). Upper bound: T is closed, so let w∈T attain d(h,T), with tn↓0, wn→w, x+tnwn∈C. Then d(x+tnh,C)/tn≤∥h−wn∥→∥h−w∥, so ℓ≤d(h,T). Lower bound: let tn↓0 realise ℓ and let cn∈C be nearest to x+tnh. With wn=(cn−x)/tn, ∥h−wn∥→ℓ, so (wn) is bounded and has a limit point w, which lies in T by definition; hence d(h,T)≤∥h−w∥=ℓ. The path statement follows from Lemma 2.3(ii) applied to C. □
Definition 3.2. For Y0∈M≤q−1 and a direction h=0, the structural angle is κ(Y0,h)=d(h,TanM≤q−1(Y0))/∥h∥. By Proposition 3.1, liminft↓0Δq(Y0+th)/t=∥h∥κ(Y0,h): the margin grows at least at rate ∥h∥κ as t↓0, and exactly at that rate along some sequence.
Equality of the full limit requires regularity of C at x. The following case covers all the examples below.
Lemma 3.3.Suppose that, in a neighbourhood U of x, C∩U=(⋃k=1K(x+Vk))∩U for linear subspaces V1,…,VK. Then TanC(x)=⋃kVk and d(x+th,C)=td(h,TanC(x)) for all sufficiently small t>0.
Proof. The tangent cone is local, and that of a finite union of affine subspaces through x is the union of their directions. Since d(x+th,C)≤t∥h∥, nearest points lie within 2t∥h∥ of x, hence in U for small t; there d(x+th,C)=minkd(th,Vk)=tminkd(h,Vk). □
Competitors with bounded residual degree
Throughout this section V=Sp with the Frobenius norm; the working-norm normalisation p−1∥⋅∥F multiplies all distances by p−1 and leaves angles unchanged. For an edge set E let
ΩE={S∈Sp:Sij=0 for i=j,{i,j}∈/E}.
a linear subspace with free diagonal. For a degree budget dˉ∈Z≥0 the competing class of residuals with at most dˉ off-diagonal entries per row is
Sdˉ=E∈Bdˉ⋃ΩE,Bdˉ={E:degE(i)≤dˉ 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.
Lemma 4.1. For Y∈Sp, dF(Y,ΩE)2=2∑{i,j}∈/EYij2, attained by the matrix that copies Y on the diagonal and on $E and vanishes elsewhere.
Proof.ΩE is a coordinate subspace of Sp; orthogonal projection keeps the free coordinates. Each off-diagonal pair contributes twice to the Frobenius norm. □
Distances as b-matchings
Given nonnegative edge weights w and vertex capacities bi∈Z≥0, a maximum-weight b-matching is an edge set E maximising ∑e∈Ewe subject to degE(i)≤bi. It is solvable in polynomial time [12 Chapters 31–33].
so the rank-zero margin Δ1(Y) is computable exactly in polynomial time. In the working norm, Δ1(Y)2=W−maxE∈Bdˉ∑Ewij with wij=2p−2Yij2.
Proof.dF(Y,Sdˉ)2=minE∈BdˉdF(Y,ΩE)2, and by Lemma 4.1 this equals minE(W−∑Ewij).
Example 4.3 (strong but redundant, weak but necessary). Let dˉ≥1. For v=(e1+e2)/2 and any a>0, Y=avvT has a single nonzero off-diagonal pair, which a one-edge E covers: Δ1(Y)=0 however large the eigenvalue a. For v=p−1/21 and a>0, all p(p−1)/2 pairs carry weight 2a2/p2 and at most ⌊pdˉ/2⌋ can be covered, so dF(Y,Sdˉ)2≥a2(1−p1−pdˉ)>0 for dˉ<p−1, however small a is.
Tangent cones and support slack
Let S0∈Sdˉ have active off-diagonal support E0={{i,j}:(S0)ij=0} and define the support slackci=dˉ−degE0(i)≥0 and
E(E0,c)={E′ disjoint from E0:degE′(i)≤ci for all i}.
Proposition 4.4.TanSdˉ(S0)=⋃E′∈E(E0,c)ΩE0∪E′, and for every h, dF(S0+th,Sdˉ)=tdF(h,TanSdˉ(S0)) for small t>0.
Proof. Let δ=minE0∣(S0)ij∣>0 (or δ=∞ if E0=∅). If maxij∣Sij−(S0)ij∣<δ, the off-diagonal support of S contains E0, so S∈Sdˉ if and only if its support is E0∪E′ with E′∈E(E0,c), i.e. S∈ΩE0∪E′. Conversely each ΩE0∪E′⊆Sdˉ and contains S0. Apply Lemma 3.3.
Saturated supports (c≡0) are regular points with a single tangent subspace; an empty support (c≡dˉ) is maximally singular, with tangent cone the whole of Sdˉ.
Corollary 4.5. Let h=vvT with ∥v∥2=1, so ∥h∥F=1. Then
the last term being a maximum-weight b-matching with capacities bi=ci on the complement of E0.
Proof. Combine Proposition 4.4 with Lemma 4.1 and 2∑i<jvi2vj2=(∑ivi2)2−∑ivi4=1−∑ivi4.
The exact angle of equal-magnitude spikes
Lemma 4.6. For integers k≥1 and 0≤d≤k−1, the complete graph Kk has a subgraph of maximum degree at most d with exactly ⌊kd/2⌋ edges, and no such subgraph has more.
Proof. The upper bound is the handshake identity 2∣E∣=∑idegE(i)≤kd. For the construction label the vertices by Zk and let Gs be the circulant graph with edges {i,i±1},…,{i,i±s}. If d=k−1 take Kk. If d<k−1 is even, Gd/2 is d-regular since d/2<k/2. If d is odd and k even, add to G(d−1)/2 the perfect matching {i,i+k/2}, which is disjoint from it because (d−1)/2<k/2; the result is d-regular. If d is odd and k odd, then d≤k−2; the edges {i,i+(k−1)/2} are not in G(d−1)/2 (their difference ±(k−1)/2 exceeds (d−1)/2 in absolute value modulo k) and form a Hamiltonian cycle, because gcd((k−1)/2,k)=1; it contains a matching with (k−1)/2 edges, and adding it to G(d−1)/2 gives k(d−1)/2+(k−1)/2=⌊kd/2⌋ edges with maximum degree d.
Theorem 4.7. Let v have support K⊆[p] with ∣K∣=k and ∣vi∣=k−1/2 on K, let 0≤dˉ≤k−1, and g=k−1−dˉ. At a residual S0 with empty off-diagonal support (for instance S0=0 diagonal), the structural angle of h=vvT with respect to M≤0=Sdˉ is
κk,dˉ2=k22∣kg/2∣=kg+O(k−2).
Proof. By Corollary 4.5 with E0=∅ and c≡dˉ, the weights 2vi2vj2 equal 2/k2 on pairs inside K and vanish elsewhere, and 1−∑ivi4=1−1/k=(2/k2)(2k). By Lemma 4.6 at most ⌊kdˉ/2⌋ pairs inside K can be covered, and this is attained. Hence κ2=(2/k2)((2k)−⌊kdˉ/2⌋). Since k(k−1) is even, kg=k(k−1)−kdˉ has the parity of kdˉ, and (2k)−⌊kdˉ/2⌋=∣kg/2∣.
Proposition 4.8 (equal angle, different branching). Let k≥2 be even, dˉ=k−2 and v as in Theorem 4.7.
(i) At an empty support, κ2=1/k, and the minimising tangent subspaces are exactly the ΩE whose restriction E∩(2K) is the complement of a perfect matching of K: there are (k−1)!!=(k−1)(k−3)⋯1 such restrictions.
(ii) At a residual S0 whose support on K is the complete graph minus a fixed perfect matching P (every vertex of K saturated), κ2=1/k and the tangent cone restricted to K is the single subspace ΩE0.
Proof. (i) Theorem 4.7 with g=1 gives κ2=2(k/2)/k2=1/k. An optimal E inside K has k(k−2)/2 edges and maximum degree k−2, hence is (k−2)-regular on K, and its complement in Kk is 1-regular, i.e. a perfect matching; there are (k−1)!! perfect matchings of Kk. (ii) The slack vanishes on K, so by Proposition 4.4 no edge inside K can be added, and the uncovered pairs are those of P: κ2=2⋅21⋅k21=k1.
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), possibly infinite-dimensional; only M≤0 and Lemma 2.3 are used. Let Y(u)∈Sp, u∈U, be measurable and square-integrable with respect to a probability measure μ, with norm ∥Y∥2=∫∥Y(u)∥F2dμ(u). Compare the pointwise competitor, in which S(u)∈Sdˉ for μ-almost every u with a support that may depend on u, with the shared-support competitorSdˉsh=⋃E∈Bdˉ{S:S(u)∈ΩE for μ-almost every u}.
Proposition 4.9. With wˉij=2∫Yij(u)2dμ(u) and Wˉ=∑i<jwˉij,
and the inequality can be strict with the left side equal to 0.
Proof. For fixed E the projection acts pointwise (Lemma 4.1), giving ∫2∑{i,j}∈/EYij2dμ=Wˉ−∑Ewˉij; minimise over E. The pointwise problem decouples across u (a measurable selection of optimal supports exists because Bdˉ is finite), and the shared class is contained in the pointwise one. Example: p=3, dˉ=1, μ uniform on {u1,u2}, Y(u1) supported on the pair {1,2} and Y(u2) on {1,3} with unit entries. Pointwise each matrix is covered by one edge, so the pointwise distance is 0; a shared E with degE(1)≤1 cannot contain both pairs, so d(Y,S1sh)2=min(wˉ12,wˉ13)=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}, and for a fixed E write Drel,E=dF(Y,ΩΣE), DPSD,E=dF(Y,ΩΣE∩{S⪰0}), SErel for the projection of Lemma 4.1 and ζE=[−λmin(SErel)]+.
where the left side is the b-matching value of Theorem 4.2 and the right side can be evaluated at any computed support, for instance the b-matching optimiser. In the working norm the correction term becomes p−1ζE2.
Proof. The lower bounds hold because the PSD classes are subsets. SErel+ζEI is positive semidefinite, belongs to ΩE and differs from SErel only on the diagonal, where Y−SErel vanishes; hence ∥Y−SErel−ζEI∥F2=Drel,E2+pζE2. Multiplying distances by p−1 turns pζE2 into p−1ζE2.
The same argument shows DPSD,E2≤Drel,E2+∥Λ∥F2 for any diagonal Λ with SErel+Λ⪰0. The scalar shift is the simplest choice and can be conservative when the eigenvalue deficit is localised on few coordinates; minimising ∥Λ∥F under this constraint is a semidefinite programme.
Local Fisher structure
Let z1,…,zT be independent N(0,Σ) vectors in Rp and PΣ⊗T their joint law. The Fisher inner product at Σ0≻0 is
⟨A,B⟩I,0=21tr(Σ0−1AΣ0−1B),∥A∥I,02=⟨A,A⟩I,0.
Lemma 5.1 (dimension-free likelihood expansion). Let Σ0=I and let A=Ap,T∈Sp satisfy T∥A∥F2=O(1) (hence ∥A∥2≤∥A∥F→0), with p=pT arbitrary. Under PI⊗T,
For ∣λ∣≤1/2, ∣log(1+λ)−λ+λ2/2∣≤∣λ∣3, so logdet(I+A)=trA−21trA2+r1 with ∣r1∣≤∥A∥2∥A∥F2, and T∣r1∣→0. The matrix A2R is symmetric with ∥A2R∥F≤2∥A∥2∥A∥F; hence ∑tzt⊤A2Rzt has mean Ttr(A2R)=TtrA2−Ttr(A3R), with T∣tr(A3R)∣≤2T∥A∥2∥A∥F2→0, and variance 2T∥A2R∥F2≤8∥A∥22T∥A∥F2→0. Collecting terms gives the expansion. For the limit law, Xt=21(zt⊤Azt−trA) are i.i.d., centred, with variance 21trA2; as second-order Gaussian chaoses they satisfy E∣Xt∣3≤C(trA2)3/2 by hypercontractivity ([6]), so for σ2>0 the Lyapunov ratio is O(T−1/2) and ∑tXt⇒N(0,σ2); for σ2=0 the variance 2TtrA2 of ∑tXt tends to 0. Contiguity follows from Le Cam’s first lemma ([7]).
For general Σ0, whitening with Σ0−1/2 maps a perturbation G to A=Σ0−1/2GΣ0−1/2, and 2Ttr(A2)=T∥G∥I,02. In particular a perturbation is locally detectable at scale (T∥G∥I,0)−1.
Corollary 5.2. If M=Σ−1/2GΣ−1/2 satisfies ∥M∥2→0, then KL(PΣ⊗T∥PΣ+G⊗T)=4T∥M∥F2{1+O(∥M∥2)}=2T∥G∥I,Σ2{1+O(∥M∥2)}.
Proof. With eigenvalues λj of M, KL=2T∑jϕ(λj) where ϕ(λ)=1+λ1−1+log(1+λ)=2λ2+O(∣λ∣3).
The Fisher quotient at a rank-r background
When the null model contains a rank-r component, a perturbation is informative only through the part not absorbed by moving that component. Let Σ0=L0+S0 with L0=UΛU⊤, U∈Rp×r with orthonormal columns, Λ≻0 diagonal and S0≻0, so L0⊥L have orthogonal columns spanning the orthogonal complement of the range of U. Assume ∥L0∥≤ρ, so that the magnitude bound of Section 2 is slack; then the tangent cone of Lr at L0 is the tangent space of the rank-r positive semidefinite matrices,
Proposition 5.3. For G∈Sp, with G22=U⊥⊤GU⊥ and S⊥=U⊥⊤S0U⊥≻0,
Q∈TanUmin∥G−Q∥I,02=21S⊥−1/2G22S⊥−1/2F2.
In particular the Fisher information left after projecting out TanU does not depend on the strength Λ of the background.
Proof. Write M=G−Q. As Q ranges over TanU, M ranges over {M∈Sp:U⊥⊤MU⊥=G22}, so we minimise the convex function 21tr(Σ0−1MΣ0−1M) under a linear constraint. Stationarity gives Σ0−1MΣ0−1=U⊥ΓU⊥⊤ for a symmetric multiplier Γ, i.e. M=Σ0U⊥ΓU⊥⊤Σ0. Since U⊥⊤L0=0, U⊥⊤Σ0U⊥=S⊥, and the constraint becomes S⊥ΓS⊥=G22, so Γ=S⊥−1G22S⊥−1. The value is 21tr(U⊥ΓU⊥⊤Σ0U⊥ΓU⊥⊤Σ0)=21tr(ΓS⊥ΓS⊥)=21tr(S⊥−1G22S⊥−1G22).
Lemma 5.4. Let G be supported on a coordinate set I, i.e. G=EIGIEI⊤ with EI the coordinate embedding, and let ϑ=∥U⊤EI∥22≤1. Then
Proof. Let N=EI⊤U⊥U⊥⊤EI=I−EI⊤UU⊤EI, whose eigenvalues lie in [1−ϑ,1]. Then ∥U⊥⊤GU⊥∥F2=tr(GINGIN)=∥N1/2GIN1/2∥F2∈[(1−ϑ)2,1]⋅∥GI∥F2. The last display follows from Proposition 5.3 and ∥S⊥−1/2CS⊥−1/2∥F∈[∥C∥F/∥S⊥∥2,∥C∥F/λmin(S⊥)].
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⊥−1, as specified in Proposition 5.3. For perturbations through the fixed decomposition Σ0=L0+S0, allowing the residual to move along a branch ΩE gives the Fisher distance to TanU+ΩE, minimised over the branches through S0. The sandwich S⊥−1/2U⊥⊤(⋅)U⊥S⊥−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} be a null family and, for a test ψ, write R(ψ;Σ0,Σ1)=EΣ0ψ+EΣ1(1−ψ) for the sum of its two error probabilities under PΣ0⊗T and PΣ1⊗T.
Proposition 5.5 (margin floor). Let Σ1≻0 and let Σ0∈N satisfy λmin(Σ0)≥c>0 and ∥Σ1−Σ0∥F≤c/2. Then for every test ψ,
R(ψ;Σ0,Σ1)≥1−2cT∥Σ1−Σ0∥F.
In particular, let N=Sdˉ∩{Σ≻0} and measure Δ1 in the unnormalised Frobenius norm. If a nearest point Σ0 of Sdˉ to Σ1 satisfies λmin(Σ0)≥c and Δ1(Σ1)≤c/2, the bound reads 1−T/2Δ1(Σ1)/c: no test separates Σ1 from the null once TΔ1(Σ1)→0, whatever p.
Proof. Let G=Σ1−Σ0 and M=Σ0−1/2GΣ0−1/2, so ∥M∥2≤∥M∥F≤∥G∥F/c≤21. For ∣λ∣≤21 the function ϕ in the proof of Corollary 5.2 satisfies 0≤ϕ(λ)≤2λ2, hence KL(PΣ0⊗T∥PΣ1⊗T)≤T∥M∥F2≤T∥G∥F2/c2. By Pinsker’s inequality the total variation distance is at most T/2∥G∥F/c, and R(ψ;Σ0,Σ1)≥1−TV for every test.
Corollary 5.6.In the setting ofProposition 4.8, letΣ0∈Sk−2be either residual of parts (i)–(ii), withλmin(Σ0)≥c>0, and letΣ1=Σ0+θvvTwithθ>0. Ifθ/k≤c/4, there isΣ0∈Sk−2, Σ0≻0, such that every test satisfies
R(ψ;Σ0,Σ1)≥1−c2θT/k.
In particular no test separatesΣ1fromSk−2whenθT/k→0, uniformly inpandkfor fixedc.
Proof. Let E be the support of a minimising subspace in Proposition 4.8: in case (i) any E whose restriction to K is the complement of a perfect matching P of K, in case (ii) E=E0 with the matching P of that statement. Let Σ0 be the orthogonal projection of Σ1 onto ΩE; since Σ0∈ΩE in both cases, Σ0∈ΩE⊆Sk−2 and G=Σ1−Σ0 is supported on the pairs of P, with entries ±θ/k. Hence ∥G∥F=θ/k and ∥G∥2=θ/k, for every θ>0. By Weyl’s inequality λmin(Σ0)≥λmin(Σ1)−θ/k≥c−c/4≥c/2, using λmin(Σ1)≥λmin(Σ0). Apply Proposition 5.5 with c/2 in place of c, noting ∥G∥F≤c/4.
For small θ, ∥G∥F=θ/k=θκ is exactly the margin Δ1(Σ1) (Propositions 4.4 and 4.8); the construction above avoids any condition on the size of the active entries of Σ0. 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)!! branches.
Remark 5.7 (two scales). Proposition 2.4 resolves margins above the estimation error ϵ≍∥Σ−Σ∥F, which for the sample covariance with trΣ≍p is of order p/T in the unnormalised Frobenius norm, since E∥Σ−Σ∥F2={(trΣ)2+∥Σ∥F2}/T. Proposition 5.5 places the information floor at Δ1≍1/T. Lipschitz certificates are therefore valid but may be conservative by a factor up to p; 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<∞ be fixed, and restrict the covariance classes to cI⪯Σ⪯CI. For a declared neighbourhood UT and margin δ>0, define
H0=M≤r∩UT,H1(δ)={Σ∈M≤r+1∩UT:dF(Σ,M≤r)≥δ}.
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 with values in [0,1], its minimax sum of errors is
RT(δ)=ψinf{Σ∈H0supEΣψ+Σ∈H1(δ)supEΣ(1−ψ)}.
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, with S∈Sp positive definite and v an equal-magnitude unit vector on k coordinates. For an amplitude parameter a>0, restrict a≤θ≤2a and require
c⋆θκk,d~≤dF(Σ,M≤0)≤C⋆θκk,d~,0<c⋆<1<C⋆<∞.
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 aT means that the minimax error tends to zero when a/aT→∞, and stays away from zero when a/aT→0, 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 k of coordinates and dˉ=k−2. Under suitable localisation and growth conditions and up to logarithmic factors, the boundary is aT≳k/T at an empty residual support and aT≳k/T at a saturated clique-minus-matching support: branch uncertainty costs a factor of order k despite the equal structural angles ofProposition 4.8. For an unknown direction and general 0≤dˉ≤k−2, writing g=k−1−dˉ, the boundary is of order
max⎩⎨⎧Tklog(ep/k),gTk(dˉ+1)⎭⎬⎫.
Corollary 5.6 establishes a 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 T consecutive observations of a real stationary Gaussian process. The spectral classes are Hermitian, conjugate-symmetric and uniformly bounded between cI and CI. On [0,π] the partition into m equal bands is known; the residual support may be independently reconfigured on those bands and is reflected to negative frequencies. For Nm the corresponding rank-zero spectral class, measure separation by
The alternatives F(ω)=S(ω)+θvvT have a≤θ≤2a and margin between c∗θ/k and C∗θ/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 k of coordinates, dˉ=k−2, and m=J+1 equal bands with independently reconfigured perfect matchings, the boundary is, up to logarithmic factors,
aT≳Tkm1/4,Tk2m→0,Tkm5/2log(eT)→0,
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]Absil, P.-A., Mahony, R. and Sepulchre, R. (2008). Optimization Algorithms on Matrix Manifolds. Princeton University Press.DOI
[2]Aubin, J.-P. and Frankowska, H. (1990). Set-Valued Analysis. Birkhäuser.
[3]Bai, J. and Ng, S. (2002). Determining the number of factors in approximate factor models. Econometrica 70, 191–221.DOI
[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]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]Janson, S. (1997). Gaussian Hilbert Spaces. Cambridge University Press.DOI
[7]Le Cam, L. and Yang, G. L. (2000). Asymptotics in Statistics: Some Basic Concepts, 2nd ed. Springer.
[8]Rockafellar, R. T. and Wets, R. J.-B. (1998). Variational Analysis. Springer.DOI
[12]Schrijver, A. (2003). Combinatorial Optimization: Polyhedra and Efficiency. Springer.