Introduction
Let A , B A,B A , B be fixed Hilbert–Schmidt operators on H = ℓ 2 ( N ; F ) H=\ell^2(\mathbb{N};\mathbb{F}) H = ℓ 2 ( N ; F ) , where F = R \mathbb{F}=\mathbb{R} F = R or C \mathbb{C} C , and let G G G be a standard Gaussian array over the same field. Write s ( T ) = ( s n ( T ) ) n ≥ 1 s(T)=(s_n(T))_{n\geq1} s ( T ) = ( s n ( T ) ) n ≥ 1 for the singular-value sequence of a compact operator T T T , and I ( T ) \mathcal{I}(T) I ( T ) for the algebraic two-sided principal ideal that it generates. The Gaussian product satisfies
I ( A G B ) = I ( s ( A ) ⊙ s ( B ) ) (1) \mathcal{I}(AGB)=\mathcal{I}(s(A)\mathbin{\odot}s(B)) \tag*{(1)} I ( A GB ) = I ( s ( A ) ⊙ s ( B )) (1)
almost surely. Here I ( a ) = I ( diag ( a n ) ) \mathcal{I}(a)=\mathcal{I}(\operatorname{diag}(a_n)) I ( a ) = I ( diag ( a n )) for a nonnegative nonincreasing null sequence a a a . The operation ⊙ \odot ⊙ is defined by multiplication of square tails: if a , b a,b a , b are nonnegative, nonincreasing, and square summable, then
F a ( n ) = ∑ k ≥ n a k 2 , F a ⊙ b ( n ) = F a ( n ) F b ( n ) , (2) F_a(n)=\sum_{k\geq n}a_k^2,\qquad F_{a\odot b}(n)=F_a(n)F_b(n), \tag*{(2)} F a ( n ) = k ≥ n ∑ a k 2 , F a ⊙ b ( n ) = F a ( n ) F b ( n ) , (2)
( a ⊙ b ) n 2 = a n 2 F b ( n ) + b n 2 F a ( n + 1 ) = ∑ min ( i , j ) = n a i 2 b j 2 . (3) (a\odot b)_n^2=a_n^2F_b(n)+b_n^2F_a(n+1)=\sum_{\min(i,j)=n}a_i^2b_j^2. \tag*{(3)} ( a ⊙ b ) n 2 = a n 2 F b ( n ) + b n 2 F a ( n + 1 ) = m i n ( i , j ) = n ∑ a i 2 b j 2 . (3)
Thus, on one event for the fixed pair A , B A,B A , B , membership of A G B AGB A GB in every algebraic two-sided operator ideal is determined by the same profile. Principal-ideal equality permits constant factors and finite index dilations in comparing singular values; the individual singular values may vary with the outcome. The notation A G B AGB A GB refers to a Hilbert–Schmidt-valued Gaussian series, defined in (23) ; it does not require the unweighted array G G G to define a bounded operator.
The deterministic operation in (1) raises two questions: which numerical arrays realize it uniformly, and which families of factors retain the inclusion order of the input ideals? Theorems Theorem A and Theorem B answer these questions. Theorem Theorem C then combines their consequences with an operator-valued Wick identity: in a specified Gaussian realization, one fixed Wick composition restores the inclusion relations within the local common fibres constructed in Theorem B .
2020 Mathematics Subject Classification. Primary 47B10, 47L20; Secondary 60B20, 60H07, 06F05.
Key words and phrases. Gaussian products; principal ideals; rectangular tests; Wick products; order recovery; Boolean fibres.
Uniform spectral comparison and deterministic realization
We use the standard orthonormal basis ( e n ) (e_n) ( e n ) of H H H . The entries of G G G are independent standard real Gaussians when F = R \mathbb{F}=\mathbb{R} F = R . In the complex case they have the form ( ξ + i η ) / 2 (\xi+i\eta)/\sqrt{2} ( ξ + i η ) / 2 , where ξ , η \xi,\eta ξ , η are independent standard real Gaussians. Singular values are arranged in nonincreasing order and extended by zero beyond the rank. Write B ( H ) B(H) B ( H ) for the bounded operators, ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ for the operator norm, and
S q = { T compact : ∑ n ≥ 1 s n ( T ) q < ∞ } , ∥ T ∥ q = ( ∑ n ≥ 1 s n ( T ) q ) 1 / q ( 0 < q < ∞ ) . \mathcal{S}_q=\left\{T\text{ compact}:\sum_{n\geq1}s_n(T)^q<\infty\right\},\qquad\lVert T\rVert_q=\left(\sum_{n\geq1}s_n(T)^q\right)^{1/q}\qquad(0<q<\infty). S q = { T compact : n ≥ 1 ∑ s n ( T ) q < ∞ } , ∥ T ∥ q = ( n ≥ 1 ∑ s n ( T ) q ) 1/ q ( 0 < q < ∞ ) .
In particular, S 1 \mathcal{S}_1 S 1 and S 2 \mathcal{S}_2 S 2 are the trace-class and Hilbert–Schmidt operators; we also write ∥ ⋅ ∥ H S \lVert\cdot\rVert_{\mathrm{HS}} ∥ ⋅ ∥ HS for ∥ ⋅ ∥ 2 \lVert\cdot\rVert_2 ∥ ⋅ ∥ 2 . The inner product is linear in its first variable, and x ⊗ y x\otimes y x ⊗ y denotes the operator z ↦ ⟨ z , y ⟩ x z\mapsto\langle z,y\rangle x z ↦ ⟨ z , y ⟩ x . Explicitly,
I ( T ) = { ∑ j = 1 m B j T C j : m ∈ N , B j , C j ∈ B ( H ) } . \mathcal{I}(T)=\left\{\sum_{j=1}^{m}B_jTC_j:m\in\mathbb{N},\ B_j,C_j\in B(H)\right\}. I ( T ) = { j = 1 ∑ m B j T C j : m ∈ N , B j , C j ∈ B ( H ) } .
No norm closure is taken. We put D a = diag ( a n ) D_a=\operatorname{diag}(a_n) D a = diag ( a n ) . The singular-value decomposition gives I ( T ) = I ( s ( T ) ) \mathcal{I}(T)=\mathcal{I}(s(T)) I ( T ) = I ( s ( T )) . Products, powers, inequalities, minima, and maxima of sequences are understood coordinatewise unless stated otherwise.
Let C \mathcal{C} C be the set of nonnegative nonincreasing sequences in ℓ 2 \ell^2 ℓ 2 , and let C + \mathcal{C}_+ C + consist of its strictly positive members. § 2 verifies that a ∘ b ∈ C a\mathbin{\circ} b\in\mathcal{C} a ∘ b ∈ C for a , b ∈ C a,b\in\mathcal{C} a , b ∈ C . On squared sequences this operation is convolution on the semilattice ( N , min ) (\mathbb{N},\min) ( N , min ) , with the multiplicative square-tail transform in (2) [4 Example 2.5 and Lemma 2.8] .
Uniform singular-value comparison for weighted Gaussian matrices. Corollary 4.1 gives one event of probability one, determined by G G G , on which every a , b ∈ C a,b\in\mathcal{C} a , b ∈ C defines a Hilbert–Schmidt matrix Z G ( a , b ) = ( a i g i j b j ) Z_G(a,b)=(a_i g_{ij}b_j) Z G ( a , b ) = ( a i g ij b j ) and, with ρ = a ∘ b \rho=a\mathbin{\circ} b ρ = a ∘ b ,
s n ( Z G ( a , b ) ) ≤ C ω ρ ⌈ n / 2 ⌉ , s ⌈ n / λ ⌉ ( Z G ( a , b ) ) ≥ c ω , λ ρ n ( n ≥ 1 , λ > 1 ) . s_n(Z_G(a,b))\leq C_\omega\rho_{\lceil n/2\rceil},\qquad s_{\lceil n/\lambda\rceil}(Z_G(a,b))\geq c_{\omega,\lambda}\rho_n\qquad(n\geq1,\ \lambda>1). s n ( Z G ( a , b )) ≤ C ω ρ ⌈ n /2 ⌉ , s ⌈ n / λ ⌉ ( Z G ( a , b )) ≥ c ω , λ ρ n ( n ≥ 1 , λ > 1 ) .
The positive constants are independent of a , b a,b a , b ; the lower constant may depend on λ \lambda λ . In particular, on this same event,
D a G D b ∈ S q ⟺ a ∘ b ∈ ℓ q ( a , b ∈ C , 0 < q < ∞ ) . D_aGD_b\in\mathcal{S}_q\Longleftrightarrow a\mathbin{\circ} b\in\ell^q\qquad(a,b\in\mathcal{C},\ 0<q<\infty). D a G D b ∈ S q ⟺ a ∘ b ∈ ℓ q ( a , b ∈ C , 0 < q < ∞ ) .
Corollary 4.2 gives the corresponding Schatten estimates and their uniform versions for leading finite rectangles. For example, if d n = n − 1 d_n=n^{-1} d n = n − 1 , then ( d ∘ d ) n ∼ 2 n − 3 / 2 (d\mathbin{\circ}d)_n\sim\sqrt{2}\,n^{-3/2} ( d ∘ d ) n ∼ 2 n − 3/2 , so
s n ( D d G D d ) ≍ n − 3 / 2 , D d G D d ∈ S q ⟺ q > 2 / 3. s_n(D_dGD_d)\asymp n^{-3/2},\qquad D_dGD_d\in\mathcal{S}_q\Longleftrightarrow q>2/3. s n ( D d G D d ) ≍ n − 3/2 , D d G D d ∈ S q ⟺ q > 2/3.
Here ≍ \asymp ≍ denotes two-sided comparison by positive constants.
The common event concerns all diagonal weights in the fixed coordinate basis. For each fixed pair A , B A,B A , B , Gaussian invariance under changes of singular-vector coordinates gives (1) , with an exceptional set that may depend on A , B A,B A , B . For a measurable random pair ( A , B ) (A,B) ( A , B ) independent of G G G , the same identity holds conditionally on the pair, by Corollary 4.1 ; the two factors need not be independent of one another.
The uniform realization is governed by a deterministic rectangular criterion. A rank flag is a strictly increasing sequence of integers 0 = N 0 < N 1 < ⋯ 0=N_0<N_1<\cdots 0 = N 0 < N 1 < ⋯ . Let C N \mathcal{C}_{\mathcal{N}} C N consist of the a ∈ C a\in\mathcal{C} a ∈ C that are constant on each set { N j − 1 + 1 , … , N j } \{N_{j-1}+1,\ldots,N_j\} { N j − 1 + 1 , … , N j } , and put C N , + = C N ∩ C + \mathcal{C}_{\mathcal{N},+}=\mathcal{C}_{\mathcal{N}}\cap\mathcal{C}_+ C N , + = C N ∩ C + . For a numerical array M = ( m i j ) M=(m_{ij}) M = ( m ij ) over F \mathbb{F} F , let M [ m , k ] M[m,k] M [ m , k ] denote its leading m × k m\times k m × k submatrix, and write Z M ( a , b ) = ( a i m i j b j ) Z_M(a,b)=(a_i m_{ij}b_j) Z M ( a , b ) = ( a i m ij b j ) . This notation initially denotes a matrix; its existence as a Hilbert–Schmidt operator is part of the following statement.
Theorem A (Rectangular characterization). Let P , Q P,Q P , Q be arbitrary rank flags. For a numerical array M M M , the following conditions are equivalent.
(i) There exist C ≥ 1 C\geq1 C ≥ 1 and L ∈ N L\in\mathbb{N} L ∈ N such that, for every a ∈ C P a\in\mathcal{C}_P a ∈ C P and b ∈ C Q b\in\mathcal{C}_Q b ∈ C Q , the matrix Z M ( a , b ) Z_M(a,b) Z M ( a , b ) defines a Hilbert–Schmidt operator and, with ρ = a ∘ b \rho=a\mathbin{\circ} b ρ = a ∘ b ,
s n ( Z M ( a , b ) ) ≤ C ρ ⌈ n / L ⌉ , ρ n ≤ C s ⌈ n / L ⌉ ( Z M ( a , b ) ) ( n ≥ 1 ) . (4) s_n(Z_M(a,b))\leq C\rho_{\lceil n/L\rceil},\qquad\rho_n\leq C s_{\lceil n/L\rceil}(Z_M(a,b))\qquad(n\geq1). \tag*{(4)} s n ( Z M ( a , b )) ≤ C ρ ⌈ n / L ⌉ , ρ n ≤ C s ⌈ n / L ⌉ ( Z M ( a , b )) ( n ≥ 1 ) . (4)
(ii) There exist K < ∞ K<\infty K < ∞ and κ > 0 \kappa>0 κ > 0 such that
∥ M [ P j , Q k ] ∥ 2 ≤ K max ( P j , Q k ) , ∥ M [ P j , Q k ] ∥ H S 2 ≥ κ P j Q k ( j , k ≥ 1 ) . (5) \|M[P_j,Q_k]\|^2 \le K\max(P_j,Q_k),\qquad\|M[P_j,Q_k]\|_{\mathrm{HS}}^2\ge\kappa P_jQ_k\qquad(j,k\ge1). \tag*{(5)} ∥ M [ P j , Q k ] ∥ 2 ≤ K max ( P j , Q k ) , ∥ M [ P j , Q k ] ∥ HS 2 ≥ κ P j Q k ( j , k ≥ 1 ) . (5)
Under (ii), the upper comparison in (4) holds with dilation 2 2 2 , and the lower comparison holds with every real dilation λ > K / κ \lambda>K/\kappa λ > K / κ , with constants depending only on K , κ , λ K,\kappa,\lambda K , κ , λ . No upper dilation less than 2 2 2 is valid uniformly over all flag pairs and all arrays satisfying (ii). No relation between the flags or bound on their gaps is assumed.
The principal-ideal criterion permits a constant factor and a finite repetition of the singular-value sequence [3 , 12 ] . Consequently, Theorem A implies
I ( Z M ( a , b ) ) = I ( a ⊙ b ) ( a ∈ C P , b ∈ C Q ) . \mathcal{I}(Z_M(a,b))=\mathcal{I}(a\mathbin{\odot} b)\qquad(a\in\mathcal{C}_P,\ b\in\mathcal{C}_Q). I ( Z M ( a , b )) = I ( a ⊙ b ) ( a ∈ C P , b ∈ C Q ) .
The full-prefix tests P j = Q j = j P_j=Q_j=j P j = Q j = j hold almost surely for the Gaussian array. Section 4 applies the criterion and proves the Gaussian statements above, including the lower comparisons for every λ > 1 \lambda>1 λ > 1 .
Order determination and local common fibres
Let P 2 = { I ( a ) : a ∈ C } \mathfrak{P}_2=\{\mathcal{I}(a):a\in\mathcal{C}\} P 2 = { I ( a ) : a ∈ C } . The square-tail product induces the well-defined operation
I ( a ) ⊙ I ( b ) = I ( a ⊙ b ) on P 2 . \mathcal{I}(a)\mathbin{\odot}\mathcal{I}(b)=\mathcal{I}(a\mathbin{\odot}b)\quad\text{on }\mathfrak{P}_2. I ( a ) ⊙ I ( b ) = I ( a ⊙ b ) on P 2 .
This operation preserves inclusion in each variable and distributes over ideal sums; these facts are proved in Section 2. For c ∈ C + c\in\mathcal{C}_+ c ∈ C + , define Φ c ( I ( a ) ) = I ( a ⊙ c ) \Phi_c(\mathcal{I}(a))=\mathcal{I}(a\mathbin{\odot}c) Φ c ( I ( a )) = I ( a ⊙ c ) , and call c c c a probe. The numerical identity F a = F a ⊙ c / F c F_a=F_{a\mathbin{\odot}c}/F_c F a = F a ⊙ c / F c gives cancellation for the sequence operation a ↦ a ⊙ c a\mapsto a\mathbin{\odot}c a ↦ a ⊙ c . Passing to principal ideals can nevertheless identify distinct inputs. The following theorem characterizes the families of probes that retain all inclusion relations, and describes the common fibres when they do not. The determining quantities are the tail-rank profiles
r c ( n ) = F c ( n ) c n 2 . (6) r_c(n)=\frac{F_c(n)}{c_n^2}. \tag*{(6)} r c ( n ) = c n 2 F c ( n ) . (6)
For each n n n , this is the stable rank of the diagonal operator with diagonal ( c n , c n + 1 , … ) (c_n,c_{n+1},\ldots) ( c n , c n + 1 , … ) .
For functions f , h : N → [ 1 , ∞ ) f,h:\mathbb{N}\to[1,\infty) f , h : N → [ 1 , ∞ ) , write f ≤ ∗ h f\leq^* h f ≤ ∗ h if f ( n ) ≤ h ( n ) f(n)\leq h(n) f ( n ) ≤ h ( n ) for all sufficiently large n n n . A family of profiles has a common eventual upper bound if there is a finite-valued function h : N → [ 1 , ∞ ) h:\mathbb{N}\to[1,\infty) h : N → [ 1 , ∞ ) such that r c ≤ ∗ h r_c\leq^*h r c ≤ ∗ h for every profile in the family; the starting index may depend on c c c . For a rank flag N N N and d ∈ C N , + d\in\mathcal{C}_{N,+} d ∈ C N , + , equip
C N [ d ] = { a ∈ C N , + : a ≤ d } \mathcal{C}_N[d]=\{a\in\mathcal{C}_{N,+}:a\leq d\} C N [ d ] = { a ∈ C N , + : a ≤ d }
with the relative ℓ 2 \ell^2 ℓ 2 topology. A local flag region means a nonempty relatively open subset of such a set. We identify 2 N 2^{\mathbb{N}} 2 N , with its product topology, with the subsets of N \mathbb{N} N , and write S ⊆ ∗ T S\subseteq^*T S ⊆ ∗ T if S ∖ T S\setminus T S ∖ T is finite. Here Fin \operatorname{Fin} Fin denotes the ideal of finite subsets of N \mathbb{N} N ; the quotient P ( N ) / Fin \mathcal{P}(\mathbb{N})/\operatorname{Fin} P ( N ) / Fin identifies sets with finite symmetric difference.
Theorem B (Order determination and local Boolean fibres). Let N N N be a rank flag, and let ∅ ≠ Q ⊆ C + \varnothing\ne Q\subseteq\mathcal{C}_+ ∅ = Q ⊆ C + . The following conditions are equivalent.
(i) The profiles { r c : c ∈ Q } \{r_c:c\in Q\} { r c : c ∈ Q } have no common finite-valued eventual upper bound.
(ii) For every a , b ∈ C a,b\in\mathcal{C} a , b ∈ C ,
I ( a ) ⊆ I ( b ) ⟺ I ( a ⊙ c ) ⊆ I ( b ⊙ c ) for every c ∈ Q . \mathcal{I}(a)\subseteq\mathcal{I}(b)\quad\Longleftrightarrow\quad\mathcal{I}(a\mathbin{\odot}c)\subseteq\mathcal{I}(b\mathbin{\odot}c)\text{ for every }c\in Q. I ( a ) ⊆ I ( b ) ⟺ I ( a ⊙ c ) ⊆ I ( b ⊙ c ) for every c ∈ Q . (iii) Equivalence (1.7) holds for every a , b ∈ C N a,b\in\mathcal{C}_N a , b ∈ C N .
If the profiles have a common eventual upper bound, then for every d ∈ C N , + d\in\mathcal{C}_{N,+} d ∈ C N , + and every nonempty relatively open U ⊆ C N [ d ] U\subseteq\mathcal{C}_N[d] U ⊆ C N [ d ] , there is a continuous embedding S ↦ a S S\mapsto a^S S ↦ a S of 2 N 2^{\mathbb{N}} 2 N into U U U such that, on writing J S = I ( a S ) \mathcal{J}_S=\mathcal{I}(a^S) J S = I ( a S ) ,
J S ⊆ J T ⟺ S ⊆ ∗ T . (7) \mathcal{J}_S\subseteq\mathcal{J}_T\quad\Longleftrightarrow\quad S\subseteq^*T. \tag*{(7)} J S ⊆ J T ⟺ S ⊆ ∗ T . (7) J S + J T = J S ∪ T , J S ∩ J T = J S ∩ T . (8) \mathcal{J}_S+\mathcal{J}_T=\mathcal{J}_{S\cup T},\qquad\mathcal{J}_S\cap\mathcal{J}_T=\mathcal{J}_{S\cap T}. \tag*{(8)} J S + J T = J S ∪ T , J S ∩ J T = J S ∩ T . (8) For every c ∈ Q c\in Q c ∈ Q , there is an ideal J c \mathcal{J}_c J c such that I ( a S ⊙ c ) = J c \mathcal{I}(a^S\mathbin{\odot}c)=\mathcal{J}_c I ( a S ⊙ c ) = J c for all S ⊆ N S\subseteq\mathbb{N} S ⊆ N .
Thus the generators form a topological copy of Cantor space, and their principal ideals form a sublattice isomorphic to P ( N ) / F i n \mathcal{P}(\mathbb{N})/\mathrm{Fin} P ( N ) / Fin . This sublattice is contained in one common fibre of the maps { Φ c : c ∈ Q } \{\Phi_c : c \in\mathcal{Q}\} { Φ c : c ∈ Q } . Section 5 gives equivalent formulations in terms of equality and prescribed countable families of input ideals. The local construction in Proposition 5.2 also permits a prescribed finite prefix, a common summable envelope, and quantitative comparison of the output sequences. No measurability or topological assumption on Q \mathcal{Q} Q is required.
Wick composition and order reflection
In this subsection the scalar field is real. Let g = ( g i ) g=(g_i) g = ( g i ) be a sequence of independent standard Gaussian variables and let G = ( γ i j ) G=(\gamma_{ij}) G = ( γ ij ) be an independent array of independent standard Gaussian variables. For a , b ∈ C + a,b\in\mathcal{C}_+ a , b ∈ C + , define
u a = ( a i g i ) i , W a , b = u a ⊗ u b − D a b , p n = n − 1 , E = W p , p . u_a=(a_i g_i)_i,\qquad W_{a,b}=u_a\otimes u_b-D_{ab},\qquad p_n=n^{-1},\qquad E=W_{p,p}. u a = ( a i g i ) i , W a , b = u a ⊗ u b − D ab , p n = n − 1 , E = W p , p .
The random vectors u a u_a u a belong to L 2 ( Ω ; H ) L^2(\Omega;H) L 2 ( Ω ; H ) . The random operator E E E depends on the same Gaussian sequence g g g as W a , b W_{a,b} W a , b . All chaos projections and Wick products refer to the joint Gaussian space generated by g g g and G G G . Let J j J_j J j denote the orthogonal projection onto the j j j th homogeneous Wiener chaos, also acting on Hilbert–Schmidt-valued random variables. For X , Y ∈ L 2 ( Ω ; S 2 ) X,Y\in L^2(\Omega;S_2) X , Y ∈ L 2 ( Ω ; S 2 ) with finite chaos expansions, define their Wick composition by
X ⋄ o p Y = ∑ j , k ≥ 0 J j + k ( ( J j X ) ( J k Y ) ) . X\mathbin{\diamond_{\mathrm{op}}}Y=\sum_{j,k\geq0}J_{j+k}\bigl((J_jX)(J_kY)\bigr). X ⋄ op Y = j , k ≥ 0 ∑ J j + k ( ( J j X ) ( J k Y ) ) .
The sum is finite. The required L 2 L^2 L 2 bounds for operator products and Wick composition are given in Section 6.
Two random operators can generate the same principal ideal while their Wick compositions with E E E generate distinct ideals. Fix a nonzero real rank-one operator B B B , and set
r n = n − 1 , s n = n − 2 , Y r = W r , r G B , Y s = W s , s G B . r_n=n^{-1},\qquad s_n=n^{-2},\qquad Y_r=W_{r,r}GB,\qquad Y_s=W_{s,s}GB. r n = n − 1 , s n = n − 2 , Y r = W r , r GB , Y s = W s , s GB .
By Proposition 6.2, there is a common probability-one event on which both Y r Y_r Y r and Y s Y_s Y s have rank one and
I ( Y r ) = I ( Y s ) = F ( H ) , \mathcal{I}(Y_r)=\mathcal{I}(Y_s)=\mathcal{F}(H), I ( Y r ) = I ( Y s ) = F ( H ) ,
where F ( H ) \mathcal{F}(H) F ( H ) denotes the ideal of all finite-rank operators. On the same event, Wick composition with the same random operator E = W p , p E=W_{p,p} E = W p , p gives
I ( Y r ⋄ o p E ) = I ( ( n − 2 ) n ) , I ( Y s ⋄ o p E ) = I ( ( n − 3 ) n ) . \mathcal{I}(Y_r\mathbin{\diamond_{\mathrm{op}}}E)=\mathcal{I}((n^{-2})_n),\qquad
\mathcal{I}(Y_s\mathbin{\diamond_{\mathrm{op}}}E)=\mathcal{I}((n^{-3})_n). I ( Y r ⋄ op E ) = I (( n − 2 ) n ) , I ( Y s ⋄ op E ) = I (( n − 3 ) n ) .
These ideals are distinct. Thus the ideal generated before Wick composition does not determine the ideal generated afterwards.
For fixed a , b ∈ C + a,b\in\mathcal{C}_+ a , b ∈ C + and nonzero real B ∈ S 2 B\in S_2 B ∈ S 2 , the proof gives the normal form
( W a , b G B ) ⋄ o p E = F + D a ( λ I + K ) D p , (W_{a,b}GB)\mathbin{\diamond_{\mathrm{op}}}E=F+D_a(\lambda I+K)D_p, ( W a , b GB ) ⋄ op E = F + D a ( λ I + K ) D p ,
where, almost surely, λ ≠ 0 \lambda\ne0 λ = 0 , K K K is compact and F F F has finite rank. Lemma 6.1 then identifies its principal ideal as I ( a p ) \mathcal{I}(ap) I ( a p ) .
Theorem C (Order reflection under a fixed Wick composition). Fix r ∈ C + r\in\mathcal{C}_+ r ∈ C + and a nonzero real operator B ∈ S 2 B\in S_2 B ∈ S 2 , and set Y r , B = W r , r G B Y_{r,B}=W_{r,r}GB Y r , B = W r , r GB . Then Y r , B Y_{r,B} Y r , B and Y r , B ⋄ o p E Y_{r,B}\mathbin{\diamond_{\mathrm{op}}}E Y r , B ⋄ op E belong to L 2 ( Ω ; S 2 ) L^2(\Omega;S_2) L 2 ( Ω ; S 2 ) , and
I ( Y r , B ) = I ( r 2 ∘ s ( B ) ) , I ( Y r , B ⋄ o p E ) = I ( p r ) almost surely . (9) \mathcal{I}(Y_{r,B})=\mathcal{I}(r^2\circ s(B)),\qquad
\mathcal{I}(Y_{r,B}\mathbin{\diamond_{\mathrm{op}}}E)=\mathcal{I}(pr)
\quad\text{almost surely}. \tag*{(9)} I ( Y r , B ) = I ( r 2 ∘ s ( B )) , I ( Y r , B ⋄ op E ) = I ( p r ) almost surely . (9) Moreover, for every r , s ∈ C + r,s\in\mathcal{C}_+ r , s ∈ C + ,
I ( p r ) ⊆ I ( p s ) ⟺ I ( r 2 ) ⊆ I ( s 2 ) . \mathcal{I}(pr)\subseteq\mathcal{I}(ps)
\quad\Longleftrightarrow\quad
\mathcal{I}(r^2)\subseteq\mathcal{I}(s^2). I ( p r ) ⊆ I ( p s ) ⟺ I ( r 2 ) ⊆ I ( s 2 ) . Let ∅ ≠ Q ⊆ C + \varnothing\ne\mathcal{Q}\subseteq\mathcal{C}_+ ∅ = Q ⊆ C + have profiles with a common eventual upper bound, and let U U U be a local flag region as in Theorem B. The family in that theorem may be chosen with a common ℓ 1 \ell^1 ℓ 1 envelope. For this family, put
X S = W a S , a S , K S = I ( p a S ) . X_S=W_{\sqrt{a^S},\sqrt{a^S}},\qquad K_S=\mathcal{I}(p\sqrt{a^S}). X S = W a S , a S , K S = I ( p a S ) . For each fixed S ⊆ N S\subseteq\mathbb{N} S ⊆ N and c ∈ Q c\in\mathcal{Q} c ∈ Q ,
I ( X S G D c ) = J c , I ( ( X S G D c ) ⋄ o p E ) = K S almost surely , \mathcal{I}(X_SG D_c)=J_c,\qquad
\mathcal{I}((X_SG D_c)\mathbin{\diamond_{\mathrm{op}}}E)=K_S
\quad\text{almost surely}, I ( X S G D c ) = J c , I (( X S G D c ) ⋄ op E ) = K S almost surely ,
where J c J_c J c is the common value in Theorem B . The deterministic ideals K S K_S K S satisfy
K S ⊆ K T ⟺ S ⊆ ∗ T , (10) K_S \subseteq K_T \quad\Longleftrightarrow\quad S \subseteq^{*} T, \tag*{(10)} K S ⊆ K T ⟺ S ⊆ ∗ T , (10)
K S + K T = K S ∪ T , K S ∩ K T = K S ∩ T . (11) K_S+K_T=K_{S\cup T},\qquad K_S\cap K_T=K_{S\cap T}. \tag*{(11)} K S + K T = K S ∪ T , K S ∩ K T = K S ∩ T . (11)
The same fixed random operator E E E is used for all parameter values. The exceptional null set in each random ideal identity may depend on its parameters. For any prescribed countable collection of parameter pairs in either assertion, all corresponding identities hold on one event of probability one.
The operators W r , r W_{r,r} W r , r are almost surely trace class. Wick composition is defined using the joint Gaussian expansions of the random operators. Theorem C therefore concerns the specified operators W r , r G B W_{r,r}GB W r , r GB and their Gaussian expansions; it does not define an operation on their principal ideals alone. Equations (10) –(11) are relations among deterministic principal ideals. Proposition 6.2 proves the more general ideal identities for W a , b G B W_{a,b}GB W a , b GB .
Context and organization
The description of operator ideals by singular-value sequences goes back to Calkin [3 ] ; see also [12 , 14 ] for their order and lattice structure. The convolution algebra of the semilattice ( N , min ) (\mathbb{N},\min) ( N , min ) and its character transform are described in [4 Example 2.5 and Lemma 2.8] . This supplies the algebraic interpretation of the square-tail identity. Theorem A addresses its spectral realization: it characterizes the numerical arrays that realize the resulting profile uniformly for all weights on two independently prescribed flags. Its proof combines positive Gram expansions with the variational principles for eigenvalue sums [6 ] .
The estimates of Davidson and Szarek [5 Theorem II.13] and Vershynin [15 Corollary 5.35] control the extreme singular values of a finite standard Gaussian matrix. In the present setting, finite rectangular estimates in both scalar fields first give one probability-one event for the infinite array. The deterministic characterization then compares the entire singular-value sequence of D a G D b D_aGD_b D a G D b with a ∘ b a\circ b a ∘ b , simultaneously for all nonnegative, nonincreasing square-summable diagonal weights. The covering arguments used to obtain the required finite estimates are standard; see [15 Section 5.2.2] . A related use of square-tail quantities occurs in the work of Hinrichs, Krieg, Novak, Prochno and Ullrich [8 ] , who estimate radii of random sections of ellipsoids and the corresponding optimal worst-case recovery errors from Gaussian linear information. Their estimates involve square tails of the semiaxes. The output considered here is the singular-value sequence and principal ideal of a product with two diagonal weights.
For arithmetic-mean operations on ideals, Kaftal and Weiss [12 Section 6] characterize equality cancellation for principal ideals in terms of arithmetic-mean stability and exhibit failures of inclusion cancellation. They also treat arithmetic means at infinity. Theorem B concerns the probe-indexed maps I ( a ) ↦ I ( a ∘ c ) \mathcal{I}(a)\mapsto\mathcal{I}(a\circ c) I ( a ) ↦ I ( a ∘ c ) : it characterizes joint order reflection for an arbitrary family of positive probes by an eventual bound on their tail-rank profiles. When such a bound exists, the conclusion includes a continuous Boolean family in every prescribed local flag region, with a common output ideal for each probe. The eventual-domination preorder itself is classical; see [2 ] .
Hu and Yan [9 Section 4] develop Wick-product identities and integrability criteria for scalar Gaussian functionals, including expansions in Malliavin derivatives. We apply this calculus to explicit operator-valued chaoses. The required bounds use the Wiener–Itô decomposition [10 , 11 ] , hypercontractivity [7 , 11 ] , and Gaussian differentiation [13 Chapter 1] . The operator-ideal conclusion in Theorem C comes from a normal form whose middle factor is a nonzero random scalar multiple of the identity plus a compact operator. It yields order reflection after the specified Wick composition, including for families whose preceding ideal observations agree. Under regular variation, the same normal form gives exact singular-value asymptotics; the scalar asymptotic theory is developed in [1 ] .
Section 2 develops the square-tail calculus. Sections 3 and 4 establish the deterministic and Gaussian spectral realizations, respectively. Sections 5 and 6 prove the order and Wick results, including their refinements. Appendix A proves the finite Gaussian matrix estimates used in Section 4.
Principal ideals and square tails
For M ∈ N M \in\mathbb{N} M ∈ N , define the dilation ( Δ M a ) n = a ⌈ n / M ⌉ (\Delta_Ma)_n=a_{\lceil n/M\rceil} ( Δ M a ) n = a ⌈ n / M ⌉ . For nonnegative, nonincreasing null sequences a , b a,b a , b , the principal-ideal criterion is
I ( a ) ⊆ I ( b ) ⟺ a ≤ C Δ M b for some C < ∞ , M ∈ N . (12) \mathcal{I}(a)\subseteq\mathcal{I}(b)\quad\Longleftrightarrow\quad a\leq C\Delta_Mb\text{ for some }C<\infty,\ M\in\mathbb{N}. \tag*{(12)} I ( a ) ⊆ I ( b ) ⟺ a ≤ C Δ M b for some C < ∞ , M ∈ N . (12)
Indeed, if D a = ∑ j = 1 m B j D b C j D_a=\sum_{j=1}^{m}B_jD_bC_j D a = ∑ j = 1 m B j D b C j , the singular-value inequality for sums gives
a m ( n − 1 ) + 1 ≤ ∑ j = 1 m ∥ B j ∥ ∥ C j ∥ b n . a_{m(n-1)+1}\leq\sum_{j=1}^{m}\lVert B_j\rVert\lVert C_j\rVert b_n. a m ( n − 1 ) + 1 ≤ j = 1 ∑ m ∥ B j ∥ ∥ C j ∥ b n .
Monotonicity extends this estimate to each block of m m m coordinates. Conversely, suppose that a ≤ C Δ M b a\leq C\Delta_Mb a ≤ C Δ M b . For 1 ≤ j ≤ M 1\leq j\leq M 1 ≤ j ≤ M , let V j e n = e M ( n − 1 ) + j V_je_n=e_{M(n-1)+j} V j e n = e M ( n − 1 ) + j and define t j , n = a M ( n − 1 ) + j / b n t_{j,n}=a_{M(n-1)+j}/b_n t j , n = a M ( n − 1 ) + j / b n when b n > 0 b_n>0 b n > 0 , with t j , n = 0 t_{j,n}=0 t j , n = 0 when b n = 0 b_n=0 b n = 0 . Then ∥ D t j ∥ ≤ C \lVert Dt_j\rVert\leq C ∥ D t j ∥ ≤ C and
D a = ∑ j = 1 M V j D t j D b V j ∗ , D_a=\sum_{j=1}^{M}V_jD_{t_j}D_bV_j^*, D a = j = 1 ∑ M V j D t j D b V j ∗ ,
which proves the reverse implication. The singular-value decomposition gives the corresponding criterion for arbitrary compact operators [3 , 12 , 14 ] .
For nonnegative, nonincreasing null sequences,
I ( a ) + I ( b ) = I ( max ( a , b ) ) , I ( a ) ∩ I ( b ) = I ( min ( a , b ) ) . (13) \mathcal{I}(a)+\mathcal{I}(b)=\mathcal{I}(\max(a,b)),\qquad\mathcal{I}(a)\cap\mathcal{I}(b)=\mathcal{I}(\min(a,b)). \tag*{(13)} I ( a ) + I ( b ) = I ( max ( a , b )) , I ( a ) ∩ I ( b ) = I ( min ( a , b )) . (13)
The sum identity follows from max ( a , b ) ≤ a + b ≤ 2 max ( a , b ) \max(a,b)\leq a+b\leq2\max(a,b) max ( a , b ) ≤ a + b ≤ 2 max ( a , b ) and D a + b = D a + D b D_{a+b}=D_a+D_b D a + b = D a + D b . For the intersection, if a compact operator belongs to both ideals, its two singular-value comparisons can be enlarged to a common dilation and a common constant; their pointwise minimum then gives membership in I ( min ( a , b ) ) \mathcal{I}(\min(a,b)) I ( min ( a , b )) . The reverse inclusion is immediate. All nonzero finite-rank operators generate the ideal of finite-rank operators, which is contained in every nonzero principal ideal. Thus, if T T T has infinite rank and F F F has finite rank, both T T T and T + F T+F T + F have infinite rank and I ( T + F ) = I ( T ) \mathcal{I}(T+F)=\mathcal{I}(T) I ( T + F ) = I ( T ) . For strictly positive a , b a,b a , b , it suffices that the comparison in (12) hold eventually, since a larger constant absorbs the finite initial segment.
For a , b ∈ C a,b\in\mathcal{C} a , b ∈ C , formula (3) has the equivalent expansion
( a ∘ b ) n 2 = a n 2 b n 2 + a n 2 F b ( n + 1 ) + b n 2 F a ( n + 1 ) . (14) (a\mathbin{\circ}b)_n^2=a_n^2b_n^2+a_n^2F_b(n+1)+b_n^2F_a(n+1). \tag*{(14)} ( a ∘ b ) n 2 = a n 2 b n 2 + a n 2 F b ( n + 1 ) + b n 2 F a ( n + 1 ) . (14)
Each term on the right is nonincreasing in n n n . Telescoping gives F a ∘ b = F a F b F_{a\mathbin{\circ}b}=F_aF_b F a ∘ b = F a F b and ∥ a ∘ b ∥ 2 = ∥ a ∥ 2 ∥ b ∥ 2 \lVert a\mathbin{\circ}b\rVert_2=\lVert a\rVert_2\lVert b\rVert_2 ∥ a ∘ b ∥ 2 = ∥ a ∥ 2 ∥ b ∥ 2 , so a ∘ b ∈ C a\mathbin{\circ}b\in\mathcal{C} a ∘ b ∈ C . Its support has cardinality equal to the smaller of the two input support cardinalities. The product is coordinatewise monotone and homogeneous in each variable.
To verify that the induced operation on P 2 \mathcal{P}_2 P 2 is well defined, write n = M ( k − 1 ) + r n=M(k-1)+r n = M ( k − 1 ) + r , where 1 ≤ r ≤ M 1\leq r\leq M 1 ≤ r ≤ M . Direct substitution gives
( Δ M a ∘ Δ M b ) n 2 = ( 2 M − 2 r + 1 ) a k 2 b k 2 + M ( a k 2 F b ( k + 1 ) + b k 2 F a ( k + 1 ) ) . (\Delta_Ma\mathbin{\circ}\Delta_Mb)_n^2=(2M-2r+1)a_k^2b_k^2+M\bigl(a_k^2F_b(k+1)+b_k^2F_a(k+1)\bigr). ( Δ M a ∘ Δ M b ) n 2 = ( 2 M − 2 r + 1 ) a k 2 b k 2 + M ( a k 2 F b ( k + 1 ) + b k 2 F a ( k + 1 ) ) .
Consequently,
Δ M ( a ∘ b ) ≤ Δ M a ∘ Δ M b ≤ 2 M Δ M ( a ∘ b ) . (15) \Delta_M(a\mathbin{\circ}b)\leq\Delta_Ma\mathbin{\circ}\Delta_Mb\leq\sqrt{2M}\,\Delta_M(a\mathbin{\circ}b). \tag*{(15)} Δ M ( a ∘ b ) ≤ Δ M a ∘ Δ M b ≤ 2 M Δ M ( a ∘ b ) . (15)
Given inclusions between the two pairs of input ideals, (12) permits a common dilation for the two comparisons. Monotonicity, homogeneity, and (15) then give the corresponding output inclusion. In particular, I ( a ) ∘ I ( b ) = I ( a ∘ b ) \mathcal{I}(a)\mathbin{\circ}\mathcal{I}(b)=\mathcal{I}(a\mathbin{\circ}b) I ( a ) ∘ I ( b ) = I ( a ∘ b ) is independent of the chosen generators.
For the auxiliary operation ( a ⊕ b ) n = ( a n 2 + b n 2 ) 1 / 2 (a\oplus b)_n=(a_n^2+b_n^2)^{1/2} ( a ⊕ b ) n = ( a n 2 + b n 2 ) 1/2 , one has I ( a ⊕ b ) = I ( a ) + I ( b ) \mathcal{I}(a\oplus b)=\mathcal{I}(a)+\mathcal{I}(b) I ( a ⊕ b ) = I ( a ) + I ( b ) . The injective map a ↦ F a a\mapsto F_a a ↦ F a transforms ⊕ \oplus ⊕ and ∘ \circ ∘ into addition and multiplication, respectively. It follows that ∘ \circ ∘ is associative and commutative, and that its induced operation on P 2 \mathcal{P}_2 P 2 distributes over ideal sums.
For summable sequences x , w x,w x , w , put R x ( n ) = ∑ k ≥ n x k R_x(n)=\sum_{k\geq n}x_k R x ( n ) = ∑ k ≥ n x k and define
( L w x ) n = x n R w ( n ) + w n R x ( n + 1 ) . (16) (L_wx)_n=x_nR_w(n)+w_nR_x(n+1). \tag*{(16)} ( L w x ) n = x n R w ( n ) + w n R x ( n + 1 ) . (16)
The map ( w , x ) ↦ L w x (w,x)\mapsto L_wx ( w , x ) ↦ L w x is bilinear, and absolute summation gives ∥ L w x ∥ 1 ≤ ∥ w ∥ 1 ∥ x ∥ 1 \lVert L_wx\rVert_1\leq\lVert w\rVert_1\lVert x\rVert_1 ∥ L w x ∥ 1 ≤ ∥ w ∥ 1 ∥ x ∥ 1 . For a ∈ C a\in\mathcal{C} a ∈ C , c ∈ C + c\in\mathcal{C}_+ c ∈ C + , and x ∈ ℓ 1 x\in\ell^1 x ∈ ℓ 1 ,
L c 2 ( a 2 ) = ( a ∘ c ) 2 , ( L c 2 x ) n = c n 2 [ R x ( n ) + ( r c ( n ) − 1 ) x n ] . (17) L_{c^2}(a^2)=(a\mathbin{\circ}c)^2,\qquad(L_{c^2}x)_n=c_n^2\left[R_x(n)+(r_c(n)-1)x_n\right]. \tag*{(17)} L c 2 ( a 2 ) = ( a ∘ c ) 2 , ( L c 2 x ) n = c n 2 [ R x ( n ) + ( r c ( n ) − 1 ) x n ] . (17)
Since F c ( n ) > 0 F_c(n)>0 F c ( n ) > 0 for every n n n , equality a ∘ c = b ∘ c a\circ c=b\circ c a ∘ c = b ∘ c implies a = b a=b a = b by (2) . Theorem B characterizes the corresponding inclusion and equality implications after passing to principal ideals.
Spectral realization on independent flags
We prove Theorem A by expanding the squared weights as nonnegative combinations of prefix indicators. The lower comparison follows from positive Gram expansions and the variational formula for sums of the smallest eigenvalues. For the upper comparison, write ρ = a ∘ b \rho=a\circ b ρ = a ∘ b . Deleting the first n − 1 n-1 n − 1 rows and columns changes Z M ( a , b ) Z_M(a,b) Z M ( a , b ) by an operator of rank at most 2 n − 2 2n-2 2 n − 2 , but the resulting corner estimate contains 2 ( n − 1 ) a n 2 b n 2 2(n-1)a_{n}^{2}b_{n}^{2} 2 ( n − 1 ) a n 2 b n 2 , which need not be bounded by a constant multiple of ρ n 2 \rho_n^2 ρ n 2 . When this term is too large, block constancy places the next flag endpoints close to n n n , and the rectangular conditions give a lower bound for the singular values on suitable left and right singular subspaces. A correction of rank r r r reduces each of the two prefix ranks by r r r . This simultaneous reduction permits the additional truncation needed to bound the remainder by a constant multiple of ρ n \rho_n ρ n , while keeping the total approximation rank at most 2 n − 2 2n-2 2 n − 2 .
Proof of Theorem A. Necessity of the rectangular conditions. Suppose that (4) holds. Fix flag endpoints m = P j m=P_j m = P j and k = Q ℓ k=Q_\ell k = Q ℓ , and take a = 1 [ 1 , m ] a=\mathbf{1}_{[1,m]} a = 1 [ 1 , m ] and b = 1 [ 1 , k ] b=\mathbf{1}_{[1,k]} b = 1 [ 1 , k ] . The weighted matrix is M [ m , k ] M[m,k] M [ m , k ] , extended by zero. If d = min ( m , k ) d=\min(m,k) d = min ( m , k ) and N = max ( m , k ) N=\max(m,k) N = max ( m , k ) , then
ρ t 2 = m + k − 2 t + 1 ( 1 ≤ t ≤ d ) , ρ t = 0 ( t > d ) . \rho_t^2=m+k-2t+1 \quad(1\le t\le d),\qquad\rho_t=0 \quad(t>d). ρ t 2 = m + k − 2 t + 1 ( 1 ≤ t ≤ d ) , ρ t = 0 ( t > d ) .
The upper estimate at the first singular value gives ∥ M [ m , k ] ∥ 2 ≤ C 2 ( m + k − 1 ) ≤ 2 C 2 N \lVert M[m,k]\rVert^2\le C^2(m+k-1)\le2C^2N ∥ M [ m , k ] ∥ 2 ≤ C 2 ( m + k − 1 ) ≤ 2 C 2 N . For t = ⌈ d / 2 ⌉ t=\lceil d/2\rceil t = ⌈ d /2 ⌉ and q = ⌊ t / L ⌋ q=\lfloor t/L\rfloor q = ⌊ t / L ⌋ , one has ρ t 2 ≥ N \rho_t^2\ge N ρ t 2 ≥ N and q ≥ d / ( 2 L ) q\ge d/(2L) q ≥ d / ( 2 L ) . The lower estimate therefore implies
∥ M [ m , k ] ∥ H S 2 ≥ q s q ( M [ m , k ] ) 2 ≥ q N C 2 ≥ m k 2 L C 2 . \lVert M[m,k]\rVert_{\mathrm{HS}}^2\ge q\,s_q(M[m,k])^2\ge\frac{qN}{C^2}\ge\frac{mk}{2LC^2}. ∥ M [ m , k ] ∥ HS 2 ≥ q s q ( M [ m , k ] ) 2 ≥ C 2 q N ≥ 2 L C 2 mk .
Thus (5) holds with K = 2 C 2 K=2C^2 K = 2 C 2 and κ = ( 2 L C 2 ) − 1 \kappa=(2LC^2)^{-1} κ = ( 2 L C 2 ) − 1 .
Positive expansions and preliminary estimates. Conversely, assume (5) , and fix a ∈ C P a\in\mathcal{C}_P a ∈ C P and b ∈ C Q b\in\mathcal{C}_Q b ∈ C Q . Put Z = Z M ( a , b ) Z=Z_M(a,b) Z = Z M ( a , b ) , ρ = a ∘ b \rho=a\circ b ρ = a ∘ b , and
α j = a P j 2 − a P j + 1 2 , β k = b Q k 2 − b Q k + 1 2 . \alpha_j=a_{P_j}^2-a_{P_{j+1}}^2,\qquad\beta_k=b_{Q_k}^2-b_{Q_{k+1}}^2. α j = a P j 2 − a P j + 1 2 , β k = b Q k 2 − b Q k + 1 2 .
These coefficients are nonnegative. Since the weights are constant on their flag intervals and tend to zero,
a i 2 = ∑ P j ≥ i α j , b i 2 = ∑ Q k ≥ i β k . a_i^2=\sum_{P_j\ge i}\alpha_j,\qquad b_i^2=\sum_{Q_k\ge i}\beta_k. a i 2 = P j ≥ i ∑ α j , b i 2 = Q k ≥ i ∑ β k .
In particular, ∑ j P j α j = ∥ a ∥ 2 2 \sum_jP_j\alpha_j=\lVert a\rVert_2^2 ∑ j P j α j = ∥ a ∥ 2 2 , ∑ j α j = a 1 2 \sum_j\alpha_j=a_1^2 ∑ j α j = a 1 2 , and the corresponding identities hold for b b b and β \beta β . The rank of a P j × Q k P_j\times Q_k P j × Q k matrix is at most min ( P j , Q k ) \min(P_j,Q_k) min ( P j , Q k ) , so the upper rectangular condition also yields
κ P j Q k ≤ ∥ M [ P j , Q k ] ∥ H S 2 ≤ K P j Q k . \kappa P_jQ_k\le\lVert M[P_j,Q_k]\rVert_{\mathrm{HS}}^2\le KP_jQ_k. κ P j Q k ≤ ∥ M [ P j , Q k ] ∥ HS 2 ≤ K P j Q k .
It follows in particular that 0 < κ ≤ K 0<\kappa\le K 0 < κ ≤ K . Tonelli’s theorem now gives
κ ∥ a ∥ 2 2 ∥ b ∥ 2 2 ≤ ∥ Z ∥ H S 2 = ∑ j , k α j β k ∥ M [ P j , Q k ] ∥ H S 2 ≤ K ∥ a ∥ 2 2 ∥ b ∥ 2 2 . (18) \kappa\lVert a\rVert_2^2\lVert b\rVert_2^2\le\lVert Z\rVert_{\mathrm{HS}}^2=\sum_{j,k}\alpha_j\beta_k\lVert M[P_j,Q_k]\rVert_{\mathrm{HS}}^2\le K\lVert a\rVert_2^2\lVert b\rVert_2^2. \tag*{(18)} κ ∥ a ∥ 2 2 ∥ b ∥ 2 2 ≤ ∥ Z ∥ HS 2 = j , k ∑ α j β k ∥ M [ P j , Q k ] ∥ HS 2 ≤ K ∥ a ∥ 2 2 ∥ b ∥ 2 2 . (18)
Thus the weighted entries define a Hilbert–Schmidt operator, and their rectangular truncations converge in Hilbert–Schmidt norm. This argument applies to every wedged pair of weights, including pairs in which one weight is a prefix indicator.
Let Π p \Pi_p Π p and Θ q \Theta_q Θ q be the projections onto the first p p p row coordinates and the first q q q column coordinates, respectively, with Π 0 = Θ 0 = 0 \Pi_0=\Theta_0=0 Π 0 = Θ 0 = 0 . Expanding the row weights in Z ∗ Z Z^*Z Z ∗ Z , and then the column weights in the corresponding row Gram matrices, gives
∥ Z ∥ 2 ≤ ∑ j α j ∥ Π P j M D b ∥ 2 ≤ ∑ j , k α j β k ∥ M [ P j , Q k ] ∥ 2 . \lVert Z\rVert^2\le\sum_j\alpha_j\lVert\Pi_{P_j}MD_b\rVert^2\le\sum_{j,k}\alpha_j\beta_k\lVert M[P_j,Q_k]\rVert^2. ∥ Z ∥ 2 ≤ j ∑ α j ∥ Π P j M D b ∥ 2 ≤ j , k ∑ α j β k ∥ M [ P j , Q k ] ∥ 2 .
Here the matrices Π P j M D b \Pi_{P_j}MD_b Π P j M D b are bounded by (18) . Each positive Gram expansion converges in trace norm: the sum of the traces of its nonnegative terms is the squared Hilbert–Schmidt norm of the matrix being expanded. Alternatively, the inequalities follow first for finite prefix sums and then by Hilbert–Schmidt convergence. Since max ( P j , Q k ) ≤ P j + Q k \max(P_j,Q_k)\le P_j+Q_k max ( P j , Q k ) ≤ P j + Q k , we obtain
∥ Z ∥ 2 ≤ ∑ j , k α j β k ∥ M [ P j , Q k ] ∥ 2 ≤ K ( ∥ a ∥ 2 2 b 1 2 + a 1 2 ∥ b ∥ 2 2 ) . (19) \|Z\|^2\le\sum_{j,k}\alpha_j\beta_k\|M[P_j,Q_k]\|^2\le K(\|a\|_2^2b_1^2+a_1^2\|b\|_2^2). \tag*{(19)} ∥ Z ∥ 2 ≤ j , k ∑ α j β k ∥ M [ P j , Q k ] ∥ 2 ≤ K ( ∥ a ∥ 2 2 b 1 2 + a 1 2 ∥ b ∥ 2 2 ) . (19)
For arbitrary integers p , q ≥ 0 p,q\ge0 p , q ≥ 0 , define capped weights by a ~ i = min ( a i , a p + 1 ) \widetilde a_i=\min(a_i,a_{p+1}) a i = min ( a i , a p + 1 ) and b ~ i = min ( b i , b q + 1 ) \widetilde b_i=\min(b_i,b_{q+1}) b i = min ( b i , b q + 1 ) . They remain constant on their respective flag intervals, and their weighted matrix has the same ( p , q ) (p,q) ( p , q ) tail corner as Z Z Z . Moreover,
∥ a ~ ∥ 2 2 = p a p + 1 2 + F a ( p + 1 ) , ∥ b ~ ∥ 2 2 = q b q + 1 2 + F b ( q + 1 ) . \|\widetilde a\|_2^2=pa_{p+1}^2+F_a(p+1),\qquad\|\widetilde b\|_2^2=qb_{q+1}^2+F_b(q+1). ∥ a ∥ 2 2 = p a p + 1 2 + F a ( p + 1 ) , ∥ b ∥ 2 2 = q b q + 1 2 + F b ( q + 1 ) .
Applying (19) to these weights and compressing gives
∥ ( I − Π p ) Z ( I − Θ q ) ∥ 2 ≤ K { ( p + q ) a p + 1 2 b q + 1 2 + F a ( p + 1 ) b q + 1 2 + a p + 1 2 F b ( q + 1 ) } . (20) \|(I-\Pi_p)Z(I-\Theta_q)\|^2\le K\{(p+q)a_{p+1}^2b_{q+1}^2+F_a(p+1)b_{q+1}^2+a_{p+1}^2F_b(q+1)\}. \tag*{(20)} ∥ ( I − Π p ) Z ( I − Θ q ) ∥ 2 ≤ K {( p + q ) a p + 1 2 b q + 1 2 + F a ( p + 1 ) b q + 1 2 + a p + 1 2 F b ( q + 1 )} . (20)
No assumption that p p p or q q q is a flag endpoint is needed.
The lower singular-value estimate. Fix n ≥ 1 n\ge1 n ≥ 1 and λ > K / κ \lambda>K/\kappa λ > K / κ , and put
q = ⌈ n / λ ⌉ , r = q − 1 , c λ = κ λ − K λ − 1 > 0. q=\lceil n/\lambda\rceil,\qquad r=q-1,\qquad c_\lambda=\frac{\kappa\lambda-K}{\lambda-1}>0. q = ⌈ n / λ ⌉ , r = q − 1 , c λ = λ − 1 κλ − K > 0.
We first prove the estimate involving b n 2 F a ( n ) b_n^2F_a(n) b n 2 F a ( n ) . If b n = 0 b_n=0 b n = 0 , that estimate is immediate. Otherwise, let D D D be the least endpoint of the column flag with D ≥ n D\ge n D ≥ n . Block constancy gives b D = b n > 0 b_D=b_n>0 b D = b n > 0 . For a row endpoint N = P j ≥ n N=P_j\ge n N = P j ≥ n , set d = min ( N , D ) d=\min(N,D) d = min ( N , D ) and H = max ( N , D ) H=\max(N,D) H = max ( N , D ) . Singular values of the rectangular matrix are padded with zeros up to length D D D . The rectangular conditions imply
∑ i = q D s i ( M [ N , D ] ) 2 ≥ ∥ M [ N , D ] ∥ H S 2 − r ∥ M [ N , D ] ∥ 2 ≥ H ( κ d − K r ) . \sum_{i=q}^{D}s_i(M[N,D])^2\ge\|M[N,D]\|_{\mathrm{HS}}^2-r\|M[N,D]\|^2\ge H(\kappa d-Kr). i = q ∑ D s i ( M [ N , D ] ) 2 ≥ ∥ M [ N , D ] ∥ HS 2 − r ∥ M [ N , D ] ∥ 2 ≥ H ( κ d − K r ) .
Since r < n / λ ≤ d / λ r<n/\lambda\le d/\lambda r < n / λ ≤ d / λ and the function t ↦ ( κ − K t ) / ( 1 − t ) t\mapsto(\kappa-Kt)/(1-t) t ↦ ( κ − K t ) / ( 1 − t ) is nonincreasing on [ 0 , 1 ) [0,1) [ 0 , 1 ) ,
H ( κ d − K r ) ≥ c λ H ( d − r ) ≥ c λ ( D − r ) ( N − n + 1 ) . H(\kappa d-Kr)\ge c_\lambda H(d-r)\ge c_\lambda(D-r)(N-n+1). H ( κ d − K r ) ≥ c λ H ( d − r ) ≥ c λ ( D − r ) ( N − n + 1 ) .
For the last inequality, if N ≥ D N\ge D N ≥ D , then H ( d − r ) / ( D − r ) = N H(d-r)/(D-r)=N H ( d − r ) / ( D − r ) = N . If N < D N<D N < D , then
H ( d − r ) D − r = D ( N − r ) D − r ≥ N − r ≥ N − n + 1. \frac{H(d-r)}{D-r}=\frac{D(N-r)}{D-r}\ge N-r\ge N-n+1. D − r H ( d − r ) = D − r D ( N − r ) ≥ N − r ≥ N − n + 1.
Let T D T_D T D denote the matrix consisting of the first D D D columns of D a M D_aM D a M . It is Hilbert–Schmidt, because
T D = ( Z ∣ F D ) diag ( b 1 , … , b D ) − 1 . T_D=(Z|_{F_D})\operatorname{diag}(b_1,\ldots,b_D)^{-1}. T D = ( Z ∣ F D ) diag ( b 1 , … , b D ) − 1 .
Its Gram matrix has the positive expansion
Γ D = T D ∗ T D = ∑ j α j M [ P j , D ] ∗ M [ P j , D ] . \Gamma_D=T_D^*T_D=\sum_j\alpha_jM[P_j,D]^*M[P_j,D]. Γ D = T D ∗ T D = j ∑ α j M [ P j , D ] ∗ M [ P j , D ] .
The series converges in trace norm, since Tonelli’s theorem gives
∑ j α j ∥ M [ P j , D ] ∥ H S 2 = ∥ T D ∥ H S 2 < ∞ . \sum_j\alpha_j\|M[P_j,D]\|_{\mathrm{HS}}^2=\|T_D\|_{\mathrm{HS}}^2<\infty. j ∑ α j ∥ M [ P j , D ] ∥ HS 2 = ∥ T D ∥ HS 2 < ∞.
For a positive D × D D\times D D × D matrix A A A , list its eigenvalues in nonincreasing order. The variational identity for eigenvalue sums [6 ]
∑ i = q D λ i ( A ) = min { tr ( E A ) : E = E ∗ = E 2 , rank E = D − r } \sum_{i=q}^{D}\lambda_i(A)=\min\{\operatorname{tr}(EA):E=E^*=E^2,\ \operatorname{rank}E=D-r\} i = q ∑ D λ i ( A ) = min { tr ( E A ) : E = E ∗ = E 2 , rank E = D − r }
shows that this eigenvalue sum is superadditive and nonnegative. Applying it to the partial sums defining Γ D \Gamma_D Γ D and then taking the trace-norm limit, we obtain
( D − r ) λ q ( Γ D ) ≥ ∑ i = q D λ i ( Γ D ) ≥ ∑ j α j ∑ i = q D s i ( M [ P j , D ] ) 2 ≥ c λ ( D − r ) ∑ P j ≥ n ( P j − n + 1 ) α j = c λ ( D − r ) F a ( n ) . \begin{aligned}
(D-r)\lambda_q(\Gamma_D) \ge\sum_{i=q}^{D}\lambda_i(\Gamma_D) \\
&\ge\sum_j \alpha_j\sum_{i=q}^{D}s_i(M[P_j,D])^2 \\
&\ge c_\lambda(D-r)\sum_{P_j\ge n}(P_j-n+1)\alpha_j \\
&= c_\lambda(D-r)F_a(n).
\end{aligned} ( D − r ) λ q ( Γ D ) ≥ i = q ∑ D λ i ( Γ D ) ≥ j ∑ α j i = q ∑ D s i ( M [ P j , D ] ) 2 ≥ c λ ( D − r ) P j ≥ n ∑ ( P j − n + 1 ) α j = c λ ( D − r ) F a ( n ) .
In the third line we discarded the nonnegative terms with P j < n P_j<n P j < n . The last equality follows from the prefix expansion of a i 2 a_i^2 a i 2 . Restriction to the first D D D columns does not increase singular values, while multiplication by diag ( b 1 , … , b D ) \operatorname{diag}(b_1,\ldots,b_D) diag ( b 1 , … , b D ) decreases them by at most the factor b D b_D b D . Consequently,
s q ( Z ) 2 ≥ b D 2 λ q ( Γ D ) ≥ c λ b n 2 F a ( n ) . s_q(Z)^2 \ge b_D^2\lambda_q(\Gamma_D) \ge c_\lambda b_n^2F_a(n). s q ( Z ) 2 ≥ b D 2 λ q ( Γ D ) ≥ c λ b n 2 F a ( n ) .
Apply the same argument to the adjoint array, with the two flags interchanged, to obtain s q ( Z ) 2 ≥ c λ a n 2 F b ( n ) s_q(Z)^2\ge c_\lambda a_n^2F_b(n) s q ( Z ) 2 ≥ c λ a n 2 F b ( n ) . Averaging these estimates and using (3) yields
s ⌈ n / λ ⌉ ( Z ) 2 ≥ c λ 2 ( b n 2 F a ( n ) + a n 2 F b ( n ) ) ≥ c λ 2 ρ n 2 . s_{\lceil n/\lambda\rceil}(Z)^2 \ge\frac{c_\lambda}{2}\left(b_n^2F_a(n)+a_n^2F_b(n)\right)\ge\frac{c_\lambda}{2}\rho_n^2. s ⌈ n / λ ⌉ ( Z ) 2 ≥ 2 c λ ( b n 2 F a ( n ) + a n 2 F b ( n ) ) ≥ 2 c λ ρ n 2 .
The upper singular-value estimate. We will prove
s 2 n − 1 ( Z ) 2 ≤ 156 K 3 κ 2 ρ n 2 ( n ≥ 1 ) . s_{2n-1}(Z)^2\le156\frac{K^3}{\kappa^2}\rho_n^2 \qquad(n\ge1). s 2 n − 1 ( Z ) 2 ≤ 156 κ 2 K 3 ρ n 2 ( n ≥ 1 ) .
If a n b n = 0 a_nb_n=0 a n b n = 0 , one of the weights has fewer than n n n nonzero coordinates, and hence rank Z < n \operatorname{rank}Z<n rank Z < n . We may therefore suppose that a n b n > 0 a_nb_n>0 a n b n > 0 . Set
τ = K / κ ≥ 1 , E n = F a ( n ) b n 2 + a n 2 F b ( n ) = ρ n 2 + a n 2 b n 2 ≤ 2 ρ n 2 . \tau=K/\kappa\ge1,\qquad E_n=F_a(n)b_n^2+a_n^2F_b(n)=\rho_n^2+a_n^2b_n^2\le2\rho_n^2. τ = K / κ ≥ 1 , E n = F a ( n ) b n 2 + a n 2 F b ( n ) = ρ n 2 + a n 2 b n 2 ≤ 2 ρ n 2 .
First suppose that E n ≥ n a n 2 b n 2 / ( 4 τ ) E_n\ge n a_n^2b_n^2/(4\tau) E n ≥ n a n 2 b n 2 / ( 4 τ ) . The difference between Z Z Z and its tail corner ( I − Π n − 1 ) Z ( I − Θ n − 1 ) (I-\Pi_{n-1})Z(I-\Theta_{n-1}) ( I − Π n − 1 ) Z ( I − Θ n − 1 ) has rank at most 2 n − 2 2n-2 2 n − 2 . Equation (20) gives
∥ ( I − Π n − 1 ) Z ( I − Θ n − 1 ) ∥ 2 ≤ K ( 2 ( n − 1 ) a n 2 b n 2 + E n ) \|(I-\Pi_{n-1})Z(I-\Theta_{n-1})\|^2\le K\left(2(n-1)a_n^2b_n^2+E_n\right) ∥ ( I − Π n − 1 ) Z ( I − Θ n − 1 ) ∥ 2 ≤ K ( 2 ( n − 1 ) a n 2 b n 2 + E n )
≤ 9 K τ E n ≤ 18 K τ ρ n 2 . \le9K\tau E_n\le18K\tau\rho_n^2. ≤ 9 K τ E n ≤ 18 K τ ρ n 2 .
The claimed estimate follows from the variational characterization of singular values and τ ≥ 1 \tau\ge1 τ ≥ 1 .
It remains to consider E n < n a n 2 b n 2 / ( 4 τ ) E_n<na_n^2b_n^2/(4\tau) E n < n a n 2 b n 2 / ( 4 τ ) . Let P P P and Q Q Q be the least endpoints of the row and column flags, respectively, that are at least n n n . Taking adjoints and interchanging the weights if necessary, assume Q ≤ P Q\le P Q ≤ P . Define
r = ⌈ Q 2 τ ⌉ , h = P + Q − 2 n + 2. r=\left\lceil\frac{Q}{2\tau}\right\rceil,\qquad h=P+Q-2n+2. r = ⌈ 2 τ Q ⌉ , h = P + Q − 2 n + 2.
The weights are constant from n n n to the corresponding endpoints, so
h ≤ F a ( n ) a n 2 + F b ( n ) b n 2 < n 4 τ ≤ r 2 . h\le\frac{F_a(n)}{a_n^2}+\frac{F_b(n)}{b_n^2}<\frac{n}{4\tau}\le\frac{r}{2}. h ≤ a n 2 F a ( n ) + b n 2 F b ( n ) < 4 τ n ≤ 2 r .
For A = M [ P , Q ] A=M[P,Q] A = M [ P , Q ] , the rectangular conditions and r − 1 < Q / ( 2 τ ) r-1<Q/(2\tau) r − 1 < Q / ( 2 τ ) imply
Q s r ( A ) 2 ≥ ∥ A ∥ H S 2 − ( r − 1 ) ∥ A ∥ 2 ≥ κ P Q 2 . Q s_r(A)^2\ge\|A\|_{\mathrm{HS}}^2-(r-1)\|A\|^2\ge\frac{\kappa P Q}{2}. Q s r ( A ) 2 ≥ ∥ A ∥ HS 2 − ( r − 1 ) ∥ A ∥ 2 ≥ 2 κ P Q .
Choose isometries U : C r → C P U:\mathbb{C}^r\to\mathbb{C}^P U : C r → C P and V : C r → C Q V:\mathbb{C}^r\to\mathbb{C}^Q V : C r → C Q whose columns are the corresponding left and right singular vectors. Over the real field, use real isometries instead. With S S S the diagonal matrix of these positive singular values,
A V = U S , U ∗ A = S V ∗ , ∥ S − 1 ∥ 2 ≤ 2 κ P . AV=US,\qquad U^*A=SV^*,\qquad\|S^{-1}\|^2\le\frac{2}{\kappa P}. A V = U S , U ∗ A = S V ∗ , ∥ S − 1 ∥ 2 ≤ κ P 2 .
Extend U U U and V V V by zero outside their finite coordinate spaces, and put
C = D a M V S − 1 U ∗ M D b , B 0 = Z − C . C=D_aMVS^{-1}U^*MD_b,\qquad B_0=Z-C. C = D a M V S − 1 U ∗ M D b , B 0 = Z − C .
The factors D a M V D_aMV D a M V and U ∗ M D b U^{*}MD_b U ∗ M D b are bounded: apply (18) to ( a , 1 [ 1 , Q ] ) (a,1_{[1,Q]}) ( a , 1 [ 1 , Q ] ) and ( 1 [ 1 , P ] , b ) (1_{[1,P]},b) ( 1 [ 1 , P ] , b ) , respectively. In particular, the expression for C C C does not require the unweighted array to define a bounded operator. The singular-vector identities give
Π P B 0 = D a ( Π P − U U ∗ ) M D b , B 0 Θ Q = D a M ( Θ Q − V V ∗ ) D b . \Pi_PB_0=D_a(\Pi_P-UU^{*})MD_b,\qquad B_0\Theta_Q=D_aM(\Theta_Q-VV^{*})D_b. Π P B 0 = D a ( Π P − U U ∗ ) M D b , B 0 Θ Q = D a M ( Θ Q − V V ∗ ) D b .
Thus rank ( Π P B 0 ) ≤ P − r \operatorname{rank}(\Pi_PB_0)\le P-r rank ( Π P B 0 ) ≤ P − r and rank ( B 0 Θ Q ) ≤ Q − r \operatorname{rank}(B_0\Theta_Q)\le Q-r rank ( B 0 Θ Q ) ≤ Q − r .
Set δ = ⌈ r / 2 ⌉ \delta=\lceil r/2\rceil δ = ⌈ r /2 ⌉ , t = P + δ + 1 t=P+\delta+1 t = P + δ + 1 , and
R = ( I − Π t − 1 ) B 0 ( I − Θ Q ) . R=(I-\Pi_{t-1})B_0(I-\Theta_Q). R = ( I − Π t − 1 ) B 0 ( I − Θ Q ) .
Since h h h is an integer and h < r / 2 h<r/2 h < r /2 , one has h + δ ≤ r h+\delta\le r h + δ ≤ r . Furthermore,
Z − R = C + Π t − 1 B 0 + ( I − Π t − 1 ) B 0 Θ Q , Z-R=C+\Pi_{t-1}B_0+(I-\Pi_{t-1})B_0\Theta_Q, Z − R = C + Π t − 1 B 0 + ( I − Π t − 1 ) B 0 Θ Q ,
and rank ( Π t − 1 B 0 ) ≤ P − r + δ \operatorname{rank}(\Pi_{t-1}B_0)\le P-r+\delta rank ( Π t − 1 B 0 ) ≤ P − r + δ . It follows that
rank ( Z − R ) ≤ r + ( P − r + δ ) + ( Q − r ) = 2 n − 2 + h + δ − r ≤ 2 n − 2. (21) \operatorname{rank}(Z-R)\le r+(P-r+\delta)+(Q-r)=2n-2+h+\delta-r\le2n-2. \tag*{(21)} rank ( Z − R ) ≤ r + ( P − r + δ ) + ( Q − r ) = 2 n − 2 + h + δ − r ≤ 2 n − 2. (21)
To estimate R R R , write F = F a ( n ) F=F_a(n) F = F a ( n ) and define
R 0 = ( I − Π t − 1 ) Z ( I − Θ Q ) , X = ( I − Π t − 1 ) D a M V , Y = U ∗ M D b ( I − Θ Q ) . R_0=(I-\Pi_{t-1})Z(I-\Theta_Q),\qquad X=(I-\Pi_{t-1})D_aMV,\qquad Y=U^{*}MD_b(I-\Theta_Q). R 0 = ( I − Π t − 1 ) Z ( I − Θ Q ) , X = ( I − Π t − 1 ) D a M V , Y = U ∗ M D b ( I − Θ Q ) .
Then R = R 0 − X S − 1 Y R=R_0-XS^{-1}Y R = R 0 − X S − 1 Y . Monotonicity of a a a gives F ≥ ( t − n + 1 ) a t 2 ≥ ( r / 2 ) a t 2 F\ge(t-n+1)a_t^2\ge(r/2)a_t^2 F ≥ ( t − n + 1 ) a t 2 ≥ ( r /2 ) a t 2 . The present case assumption, together with n ≤ Q ≤ 2 τ r n\le Q\le2\tau r n ≤ Q ≤ 2 τ r , also gives
a t 2 ≤ 2 F r , F b ( Q + 1 ) ≤ F b ( n ) < r 4 τ b n 2 ≤ r 2 b n 2 . a_t^2\le\frac{2F}{r},\qquad F_b(Q+1)\le F_b(n)<\frac{r}{4\tau}b_n^2\le\frac{r}{2}b_n^2. a t 2 ≤ r 2 F , F b ( Q + 1 ) ≤ F b ( n ) < 4 τ r b n 2 ≤ 2 r b n 2 .
Finally, h + δ ≤ r h+\delta\le r h + δ ≤ r implies
P + Q + δ ≤ 2 ( n − 1 ) + r ≤ 5 τ r . P+Q+\delta\le2(n-1)+r\le5\tau r. P + Q + δ ≤ 2 ( n − 1 ) + r ≤ 5 τ r .
Apply (20) to the weights ( a , 1 [ 1 , Q ] ) (a,1_{[1,Q]}) ( a , 1 [ 1 , Q ] ) for X X X and ( 1 [ 1 , P ] , b ) (1_{[1,P]},b) ( 1 [ 1 , P ] , b ) for Y Y Y . Since U U U and V V V are isometries, Q ≤ P Q\le P Q ≤ P , and r ≤ Q r\le Q r ≤ Q , these estimates give
∥ X ∥ 2 ≤ K ( F a ( t ) + ( P + Q + δ ) a t 2 ) ≤ 11 K τ F , \lVert X\rVert^2\le K\bigl(F_a(t)+(P+Q+\delta)a_t^2\bigr)\le11K\tau F, ∥ X ∥ 2 ≤ K ( F a ( t ) + ( P + Q + δ ) a t 2 ) ≤ 11 K τ F ,
∥ Y ∥ 2 ≤ K ( ( P + Q ) b Q + 1 2 + F b ( Q + 1 ) ) ≤ 3 K P b n 2 . \lVert Y\rVert^2\le K\bigl((P+Q)b_{Q+1}^2+F_b(Q+1)\bigr)\le3KPb_n^2. ∥ Y ∥ 2 ≤ K ( ( P + Q ) b Q + 1 2 + F b ( Q + 1 ) ) ≤ 3 K P b n 2 .
Applying the same corner estimate directly to Z Z Z gives
∥ R 0 ∥ 2 ≤ K ( F a ( t ) b Q + 1 2 + a t 2 ( ( P + Q + δ ) b Q + 1 2 + F b ( Q + 1 ) ) ) \lVert R_0\rVert^2\le K\bigl(F_a(t)b_{Q+1}^2+a_t^2((P+Q+\delta)b_{Q+1}^2+F_b(Q+1))\bigr) ∥ R 0 ∥ 2 ≤ K ( F a ( t ) b Q + 1 2 + a t 2 (( P + Q + δ ) b Q + 1 2 + F b ( Q + 1 )) )
≤ 12 K τ F b n 2 . \le12K\tau Fb_n^2. ≤ 12 K τ F b n 2 .
Therefore,
∥ R ∥ 2 ≤ 2 ∥ R 0 ∥ 2 + 2 ∥ X ∥ 2 ∥ S − 1 ∥ 2 ∥ Y ∥ 2 \lVert R\rVert^2\le2\lVert R_0\rVert^2+2\lVert X\rVert^2\lVert S^{-1}\rVert^2\lVert Y\rVert^2 ∥ R ∥ 2 ≤ 2 ∥ R 0 ∥ 2 + 2 ∥ X ∥ 2 ∥ S − 1 ∥ 2 ∥ Y ∥ 2
≤ 24 K τ F b n 2 + 132 K 2 τ κ F b n 2 \le24K\tau Fb_n^2+132\frac{K^2\tau}{\kappa}Fb_n^2 ≤ 24 K τ F b n 2 + 132 κ K 2 τ F b n 2
≤ 156 K 3 κ 2 ρ n 2 , \le156\frac{K^3}{\kappa^2}\rho_n^2, ≤ 156 κ 2 K 3 ρ n 2 ,
where the last line uses τ = K / κ ≥ 1 \tau=K/\kappa\ge1 τ = K / κ ≥ 1 and F a ( n ) b n 2 ≤ ρ n 2 F_a(n)b_n^2\le\rho_n^2 F a ( n ) b n 2 ≤ ρ n 2 . Together with (21) , this proves the upper estimate at 2 n − 1 2n-1 2 n − 1 . Monotonicity of the singular values then yields
s m ( Z ) ≤ 156 K 3 / 2 κ ρ ⌈ m / 2 ⌉ ( m ≥ 1 ) . s_m(Z)\le\sqrt{156}\frac{K^{3/2}}{\kappa}\rho_{\lceil m/2\rceil}\qquad(m\ge1). s m ( Z ) ≤ 156 κ K 3/2 ρ ⌈ m /2 ⌉ ( m ≥ 1 ) .
Choose an integer L ≥ max { 2 , λ } L\ge\max\{2,\lambda\} L ≥ max { 2 , λ } and enlarge the comparison constant if necessary. Monotonicity of ρ \rho ρ and of the singular values gives both inequalities in (4) with this common dilation. This proves the equivalence and the asserted quantitative comparisons. □
Corollary 3.1 (Full-prefix realization). For P j = Q j = j P_j=Q_j=j P j = Q j = j , the rectangular conditions are equivalent to
∥ M [ N , N ] ∥ 2 ≤ K N ( N ≥ 1 ) , ∥ M [ m , k ] ∥ H S 2 ≥ κ m k ( m , k ≥ 1 ) . (22) \lVert M[N,N]\rVert^2\le KN\qquad(N\ge1),\qquad\lVert M[m,k]\rVert_{\mathrm{HS}}^2\ge\kappa mk\qquad(m,k\ge1). \tag*{(22)} ∥ M [ N , N ] ∥ 2 ≤ K N ( N ≥ 1 ) , ∥ M [ m , k ] ∥ HS 2 ≥ κmk ( m , k ≥ 1 ) . (22) An array satisfying these conditions for some K < ∞ K<\infty K < ∞ and κ > 0 \kappa>0 κ > 0 is called admissible. Every such array satisfies
I ( Z M ( a , b ) ) = I ( a ⊙ b ) ( a , b ∈ C ) . \mathcal{I}(Z_M(a,b))=\mathcal{I}(a\mathbin{\odot}b)\qquad(a,b\in\mathcal{C}). I ( Z M ( a , b )) = I ( a ⊙ b ) ( a , b ∈ C ) .
Proof. Every leading m × k m \times k m × k rectangle is a compression of the leading max ( m , k ) \max(m,k) max ( m , k ) -square. Hence the square norm estimates imply all the rectangular norm estimates with the same constant K K K ; the reverse implication follows by taking m = k m=k m = k . The energy conditions are identical in the two formulations. The ideal equality follows from Theorem A and (12) . □
Proposition 3.2 (Dilation bounds and independent rectangular tests). The following statements hold.
(i) For every admissible array M M M and every 0 < μ < 2 0<\mu<2 0 < μ < 2 , there is no constant C < ∞ C<\infty C < ∞ such that
s n ( Z M ( a , b ) ) ≤ C ( a ⊙ b ) ⌈ n / μ ⌉ ( a , b ∈ C , n ≥ 1 ) . s_n(Z_M(a,b)) \le C(a\mathbin{\odot}b)_{\lceil n/\mu\rceil}\qquad(a,b\in\mathcal{C},\ n\ge1). s n ( Z M ( a , b )) ≤ C ( a ⊙ b ) ⌈ n / μ ⌉ ( a , b ∈ C , n ≥ 1 ) . (ii) For each integer r ≥ 1 r\ge1 r ≥ 1 , there is an admissible array with valid rectangular constants K = 2 r K=2r K = 2 r and κ = 1 \kappa=1 κ = 1 for which no uniform lower comparison can have dilation λ < r \lambda<r λ < r .
(iii) For arbitrary pairs of flags, the upper rectangular conditions cannot be replaced by conditions only at matched flag levels, even if all the lower rectangular energy conditions are retained.
Proof. Upper dilation. Fix an admissible array and constants K , κ K,\kappa K , κ as in (22) . For an integer R ≥ 1 R\ge1 R ≥ 1 , let a = b a=b a = b have first coordinate 1 1 1 , the next R R R coordinates equal to ε = R − 1 / 4 \varepsilon=R^{-1/4} ε = R − 1/4 , and all remaining coordinates zero. Decompose Z M ( a , a ) Z_M(a,a) Z M ( a , a ) into its off-diagonal blocks relative to the first coordinate and its block-diagonal part. The off-diagonal matrix is
C R = ( 0 ε ( m 1 j ) j = 2 R + 1 ε ( m i 1 ) i = 2 R + 1 0 ) . C_R=\begin{pmatrix}
0 & \varepsilon(m_{1j})_{j=2}^{R+1}\\
\varepsilon(m_{i1})_{i=2}^{R+1} & 0
\end{pmatrix}. C R = ( 0 ε ( m i 1 ) i = 2 R + 1 ε ( m 1 j ) j = 2 R + 1 0 ) .
Its two potentially nonzero singular values are the Euclidean norms of the displayed row and column. The energy estimates for the 1 × ( R + 1 ) 1\times(R+1) 1 × ( R + 1 ) and ( R + 1 ) × 1 (R+1)\times1 ( R + 1 ) × 1 rectangles show that both are at least
ε [ κ ( R + 1 ) − ∣ m 11 ∣ 2 ] 1 / 2 \varepsilon\left[\kappa(R+1)-|m_{11}|^2\right]^{1/2} ε [ κ ( R + 1 ) − ∣ m 11 ∣ 2 ] 1/2
whenever the expression under the square root is positive. The block-diagonal remainder has norm at most
max { ∣ m 11 ∣ , ε 2 ∥ M [ R + 1 , R + 1 ] ∥ } ≤ 2 K . \max\{|m_{11}|,\varepsilon^2\|M[R+1,R+1]\|\}\le\sqrt{2K}. max { ∣ m 11 ∣ , ε 2 ∥ M [ R + 1 , R + 1 ] ∥ } ≤ 2 K .
Consequently, for all sufficiently large R R R ,
s 2 ( Z M ( a , a ) ) ≥ κ / 2 R 1 / 4 − 2 K ⟶ ∞ , ( a ⊙ a ) 2 = 2 − 1 R . s_2(Z_M(a,a))\ge\sqrt{\kappa/2}\,R^{1/4}-\sqrt{2K}\longrightarrow\infty,\qquad(a\mathbin{\odot}a)_2=2-\frac{1}{R}. s 2 ( Z M ( a , a )) ≥ κ /2 R 1/4 − 2 K ⟶ ∞ , ( a ⊙ a ) 2 = 2 − R 1 .
For every 0 < μ < 2 0<\mu<2 0 < μ < 2 , one has ⌈ 2 / μ ⌉ ≥ 2 \lceil2/\mu\rceil\ge2 ⌈ 2/ μ ⌉ ≥ 2 . Monotonicity, including the zero padding after the support, gives ( a ⊙ a ) ⌈ 2 / μ ⌉ ≤ ( a ⊙ a ) 2 (a\mathbin{\odot}a)_{\lceil2/\mu\rceil}\le(a\mathbin{\odot}a)_2 ( a ⊙ a ) ⌈ 2/ μ ⌉ ≤ ( a ⊙ a ) 2 . Thus a uniform upper comparison with dilation μ < 2 \mu<2 μ < 2 is impossible. This establishes optimality over the class of all flag pairs; it does not assert optimality for each prescribed sparse pair of flags.
Lower dilation. Let W i j = ( − 1 ) ⟨ i − 1 , j − 1 ⟩ 2 W_{ij}=(-1)^{\langle i-1,j-1\rangle_2} W ij = ( − 1 ) ⟨ i − 1 , j − 1 ⟩ 2 , where the exponent is the scalar product modulo two of the binary digit sequences. Its leading 2 k 2^k 2 k -square is a Walsh matrix and satisfies
W [ 2 k , 2 k ] W [ 2 k , 2 k ] ∗ = 2 k I . W[2^k,2^k]W[2^k,2^k]^*=2^k I. W [ 2 k , 2 k ] W [ 2 k , 2 k ] ∗ = 2 k I .
For the given integer r r r , set
V i j ( r ) = W ⌊ i / r ⌋ , ⌊ j / r ⌋ . V_{ij}^{(r)}=W_{\lfloor i/r\rfloor,\lfloor j/r\rfloor}. V ij ( r ) = W ⌊ i / r ⌋ , ⌊ j / r ⌋ .
Every entry has modulus one, so every rectangular energy equals the area of the rectangle. If m > r m>r m > r , choose k k k minimally so that r 2 k ≥ m r2^k\ge m r 2 k ≥ m . The leading m m m -square is a compression of W [ 2 k , 2 k ] ⊗ J r W[2^k,2^k]\otimes J_r W [ 2 k , 2 k ] ⊗ J r , where J r J_r J r is the r × r r\times r r × r all-ones matrix. Therefore
∥ V ( r ) [ m , m ] ∥ 2 ≤ 2 k r 2 < 2 r m . \|V^{(r)}[m,m]\|^2\le2^kr^2<2rm. ∥ V ( r ) [ m , m ] ∥ 2 ≤ 2 k r 2 < 2 r m .
For m ≤ r m\le r m ≤ r , the leading square is J m J_m J m , and ∥ J m ∥ 2 = m 2 ≤ r m \|J_m\|^2=m^2\le rm ∥ J m ∥ 2 = m 2 ≤ r m . Thus K = 2 r K=2r K = 2 r and κ = 1 \kappa=1 κ = 1 are valid. For every N ≥ 1 N\ge1 N ≥ 1 , the leading r N rN r N -square has rank at most N N N , because its rows are repeated in groups of r r r . With both weights equal to 1 [ 1 , r N ] \mathbf{1}_{[1,rN]} 1 [ 1 , r N ] , one has ρ r N = 1 \rho_{rN}=1 ρ r N = 1 . If λ < r \lambda<r λ < r , then ⌊ r N / λ ⌋ > N \lfloor rN/\lambda\rfloor>N ⌊ r N / λ ⌋ > N , and the singular value at that index is zero. The lower comparison is therefore impossible for any amplitude constant. In particular, the necessary dilation in these examples is at least r = K / ( 2 κ ) r=K/(2\kappa) r = K / ( 2 κ ) .
Matched flag levels. Take P j = 2 2 j P_j=2^{2^j} P j = 2 2 j and Q j = j Q_j=j Q j = j for j ≥ 1 j\ge1 j ≥ 1 . Starting from the Walsh array, change its first row to
M 1 j = ( 1 + P j ) W 1 j , M_{1j}=(1+\sqrt{P_j})W_{1j}, M 1 j = ( 1 + P j ) W 1 j ,
and leave all other entries unchanged. Every rectangular energy is at least its area. Since P j P_j P j is a power of two and j ≤ P j j\le P_j j ≤ P j , the Walsh rectangle W [ P j , Q j ] W[P_j,Q_j] W [ P j , Q j ] has norm at most P j \sqrt{P_j} P j . Also ∑ ℓ ≤ j P ℓ ≤ 2 P j \sum_{\ell\le j}P_\ell\le2P_j ∑ ℓ ≤ j P ℓ ≤ 2 P j . It follows that
∥ M [ P j , Q j ] ∥ ≤ P j + ( ∑ ℓ ≤ j P ℓ ) 1 / 2 ≤ 2 2 P j . \lVert M[P_j,Q_j]\rVert\le\sqrt{P_j}+\left(\sum_{\ell\le j}P_\ell\right)^{1/2}\le2\sqrt{2P_j}. ∥ M [ P j , Q j ]∥ ≤ P j + ℓ ≤ j ∑ P ℓ 1/2 ≤ 2 2 P j .
Thus the upper tests at matched levels hold with K = 8 K=8 K = 8 . In contrast,
∥ M [ P 1 , Q j ] ∥ 2 ≥ ∣ M 1 j ∣ 2 ≥ P j . \lVert M[P_1,Q_j]\rVert^2\ge|M_{1j}|^2\ge P_j. ∥ M [ P 1 , Q j ] ∥ 2 ≥ ∣ M 1 j ∣ 2 ≥ P j .
The allowed flat weights of lengths P 1 = 4 P_1=4 P 1 = 4 and Q j = j Q_j=j Q j = j satisfy ρ 1 2 = j + 3 \rho_1^2=j+3 ρ 1 2 = j + 3 . Since P j / ( j + 3 ) → ∞ P_j/(j+3)\to\infty P j / ( j + 3 ) → ∞ , even the comparison at the first singular value fails for any uniform constant. This proves the last assertion. □
Remark. The characterization also applies to two exhaustive orthogonal flags in separable Hilbert spaces. Choose orthonormal bases adapted to the two flags and repeat each scalar weight according to the dimension of the corresponding orthogonal increment. The compressed matrices then have dimensions given by the ranks of the flag projections. Their operator norms, Hilbert–Schmidt norms, and singular values are invariant under changes of adapted orthonormal bases. The rectangular conditions and their conclusions therefore have the same intrinsic meaning for these flags.
Gaussian realization and conditioning
Fix orthonormal coordinate bases and let G = ( g i j ) i , j ≥ 1 G=(g_{ij})_{i,j\ge1} G = ( g ij ) i , j ≥ 1 be an array of independent standard Gaussian variables over F = R \mathbb{F}=\mathbb{R} F = R or C \mathbb{C} C . In the complex case, standard means circular complex Gaussian with E ∣ g i j ∣ 2 = 1 \mathbb{E}|g_{ij}|^2=1 E ∣ g ij ∣ 2 = 1 ; equivalently, g i j = ( ξ i j + i η i j ) / 2 g_{ij}=(\xi_{ij}+i\eta_{ij})/\sqrt{2} g ij = ( ξ ij + i η ij ) / 2 , where all the real variables ξ i j , η i j \xi_{ij},\eta_{ij} ξ ij , η ij are independent N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) variables. The Hilbert spaces and operators below use the same scalar field. No bounded operator is associated with the unweighted array G G G . For A , B ∈ S 2 A,B\in S_2 A , B ∈ S 2 , define the Gaussian product by
A G B = lim N → ∞ ∑ 1 ≤ i , j ≤ N g i j ( A e i ) ⊗ ( B ∗ e j ) . (23) AGB=\lim_{N\to\infty}\sum_{1\le i,j\le N}g_{ij}(Ae_i)\otimes(B^*e_j). \tag*{(23)} A GB = N → ∞ lim 1 ≤ i , j ≤ N ∑ g ij ( A e i ) ⊗ ( B ∗ e j ) . (23)
The convergence asserted below is in Hilbert–Schmidt norm. We use ( x ⊗ y ) u = ⟨ u , y ⟩ x (x\otimes y)u=\langle u,y\rangle x ( x ⊗ y ) u = ⟨ u , y ⟩ x , with the inner product linear in its first variable.
Corollary 4.1 (Uniform Gaussian realization and conditional products). For every fixed A , B ∈ S 2 A,B\in S_2 A , B ∈ S 2 , the square partial sums in (23) converge almost surely and in L 2 ( Ω ; S 2 ) L^2(\Omega;S_2) L 2 ( Ω ; S 2 ) , and
I ( A G B ) = I ( s ( A ) ∘ s ( B ) ) almost surely. (24) \mathcal{I}(AGB)=\mathcal{I}(s(A)\mathbin{\circ}s(B))
\quad\text{almost surely.} \tag*{(24)} I ( A GB ) = I ( s ( A ) ∘ s ( B )) almost surely. (24) More generally, let ( A , B ) (A,B) ( A , B ) be a measurable S 2 × S 2 S_2\times S_2 S 2 × S 2 -valued random pair independent of the entire array G G G . The same series converges almost surely, (24) holds almost surely, and
E ( ∥ A G B ∥ 2 2 ∣ A , B ) = ∥ A ∥ 2 2 ∥ B ∥ 2 2 . (25) \mathbb{E}(\lVert AGB\rVert_2^2\mid A,B)=\lVert A\rVert_2^2\lVert B\rVert_2^2. \tag*{(25)} E (∥ A GB ∥ 2 2 ∣ A , B ) = ∥ A ∥ 2 2 ∥ B ∥ 2 2 . (25) In this case the square partial sums converge in unconditional L 2 ( Ω ; S 2 ) L^2(\Omega;S_2) L 2 ( Ω ; S 2 ) if and only if E ( ∥ A ∥ 2 2 ∥ B ∥ 2 2 ) < ∞ \mathbb{E}(\lVert A\rVert_2^2\lVert B\rVert_2^2)<\infty E (∥ A ∥ 2 2 ∥ B ∥ 2 2 ) < ∞ . No independence between A A A and B B B is required.
There is also a single probability-one event, determined by G G G , on which the following assertions hold for every a , b ∈ C a,b\in\mathcal{C} a , b ∈ C . The matrix Z ( a , b ) = ( a i j b i j ) Z(a,b)=(a_{ij}b_{ij}) Z ( a , b ) = ( a ij b ij ) is Hilbert–Schmidt, its square coordinate truncations converge in Hilbert–Schmidt norm, and, with ρ = a ∘ b \rho=a\mathbin{\circ}b ρ = a ∘ b ,
s n ( Z ( a , b ) ) ≤ C ω ρ ⌈ n / 2 ⌉ , ρ n ≤ c ω , λ − 1 s ⌈ n / λ ⌉ ( Z ( a , b ) ) ( n ≥ 1 , λ > 1 ) . (26) s_n(Z(a,b))\le C_\omega\rho_{\lceil n/2\rceil},\qquad
\rho_n\le c_{\omega,\lambda}^{-1}s_{\lceil n/\lambda\rceil}(Z(a,b))
\quad(n\ge1,\ \lambda>1). \tag*{(26)} s n ( Z ( a , b )) ≤ C ω ρ ⌈ n /2 ⌉ , ρ n ≤ c ω , λ − 1 s ⌈ n / λ ⌉ ( Z ( a , b )) ( n ≥ 1 , λ > 1 ) . (26) Here C ω < ∞ C_\omega<\infty C ω < ∞ and, for each λ > 1 \lambda>1 λ > 1 , c ω , λ > 0 c_{\omega,\lambda}>0 c ω , λ > 0 are independent of a , b a,b a , b . In particular, on this event,
I ( Z ( a , b ) ) = I ( a ∘ b ) ( a , b ∈ C ) . (27) \mathcal{I}(Z(a,b))=\mathcal{I}(a\mathbin{\circ}b)
\quad(a,b\in\mathcal{C}). \tag*{(27)} I ( Z ( a , b )) = I ( a ∘ b ) ( a , b ∈ C ) . (27)
Almost surely, the lower estimate in (26) cannot hold at λ = 1 \lambda= 1 λ = 1 with any positive constant independent of the weights.
Proof. Construction and the Gaussian isometry. For deterministic A , B A,B A , B , let S N S_N S N denote the square partial sum in (23) . With respect to the filtration F N = σ ( g i j : i , j ≤ N ) \mathcal{F}_N = \sigma(g_{ij}: i,j \le N) F N = σ ( g ij : i , j ≤ N ) , the sequence ( S N ) (S_N) ( S N ) is a martingale in the separable Hilbert space S 2 \mathcal{S}_2 S 2 . Independence, centering and E ∣ g i j ∣ 2 = 1 \mathbb{E}|g_{ij}|^2 = 1 E ∣ g ij ∣ 2 = 1 give
E ∥ S N ∥ 2 2 = ( ∑ i ≤ N ∥ A e i ∥ 2 ) ( ∑ j ≤ N ∥ B ∗ e j ∥ 2 ) ↑ ∥ A ∥ 2 2 ∥ B ∥ 2 2 . \mathbb{E}\lVert S_N\rVert_2^2 =
\left(\sum_{i\le N}\lVert Ae_i\rVert^2\right)
\left(\sum_{j\le N}\lVert B^*e_j\rVert^2\right)
\uparrow\lVert A\rVert_2^2\lVert B\rVert_2^2. E ∥ S N ∥ 2 2 = ( i ≤ N ∑ ∥ A e i ∥ 2 ) j ≤ N ∑ ∥ B ∗ e j ∥ 2 ↑ ∥ A ∥ 2 2 ∥ B ∥ 2 2 .
The Hilbert-space martingale convergence theorem therefore gives almost sure and L 2 ( S 2 ) L^2(\mathcal{S}_2) L 2 ( S 2 ) convergence. In particular,
E ∥ A G B ∥ 2 2 = ∥ A ∥ 2 2 ∥ B ∥ 2 2 . \mathbb{E}\lVert AG B\rVert_2^2 = \lVert A\rVert_2^2\lVert B\rVert_2^2. E ∥ A GB ∥ 2 2 = ∥ A ∥ 2 2 ∥ B ∥ 2 2 .
These statements concern the series itself and do not require G G G to define an operator on the underlying Hilbert space.
A common event for diagonal factors. Put ν = 1 \nu= 1 ν = 1 in the real case and ν = 2 \nu= 2 ν = 2 in the complex case. Lemma A.1 records the finite-matrix estimates needed here, with their proofs in Appendix A. Taking q = N q = N q = N in (A.1) and applying Borel–Cantelli gives a finite random constant K K K such that ∥ G [ N , N ] ∥ 2 ≤ K N \lVert G[N,N]\rVert^2 \le KN ∥ G [ N , N ] ∥ 2 ≤ K N for every N N N . The probabilities in (A.2) are summable across all m , k ≥ 1 m,k \ge1 m , k ≥ 1 . Every finite prefix has positive Hilbert–Schmidt norm almost surely. Consequently, after absorbing the finitely many exceptional rectangles, there is a positive random κ \kappa κ such that ∥ G [ m , k ] ∥ 2 2 ≥ κ m k \lVert G[m,k]\rVert_2^2 \ge\kappa mk ∥ G [ m , k ] ∥ 2 2 ≥ κmk for every m , k m,k m , k . The full-prefix conditions (22) thus hold on a single event of probability one. On that event, Theorem A gives Hilbert–Schmidt existence for all diagonal weights, convergence of their coordinate truncations, and the upper estimate in (26) .
Lower estimates for every λ > 1 \lambda> 1 λ > 1 . First fix a rational λ > 1 \lambda> 1 λ > 1 and set q N = ⌈ N / λ ⌉ q_N = \lceil N/\lambda\rceil q N = ⌈ N / λ ⌉ . Choose α ∈ ( 1 / λ , 1 ) \alpha\in(1/\lambda,1) α ∈ ( 1/ λ , 1 ) , so that q N ≤ α N q_N \le\alpha N q N ≤ α N for all sufficiently large N N N . Choose t = t α ∈ ( 0 , 1 ) t = t_\alpha\in(0,1) t = t α ∈ ( 0 , 1 ) so small that 17 α 2 e t 1 − α < e − 1 17\alpha^2\sqrt{e}\,t^{1-\alpha} < e^{-1} 17 α 2 e t 1 − α < e − 1 . For q ≤ α N q \le\alpha N q ≤ α N , (A.3) then bounds P { s q ( G [ N , q ] ) < t N } \mathbb{P}\{s_q(G[N,q]) < t\sqrt{N}\} P { s q ( G [ N , q ]) < t N } by 2 e − 3 ν N + e − ν N 2e^{-3\nu N}+e^{-\nu N} 2 e − 3 ν N + e − ν N , which is summable in N N N . Every leading square is nonsingular almost surely: its determinant, or its squared modulus in the complex case, is a nonzero real polynomial in variables with a joint density. Hence every G [ N , q N ] G[N,q_N] G [ N , q N ] has full column rank almost surely. Borel–Cantelli and the finitely many remaining values of N N N give a random γ ω , λ > 0 \gamma_{\omega,\lambda} > 0 γ ω , λ > 0 such that
G [ N , q N ] ∗ G [ N , q N ] ≥ γ ω , λ N I q N ( N ≥ 1 ) . G[N,q_N]^*G[N,q_N] \ge\gamma_{\omega,\lambda}N I_{q_N}
\qquad(N \ge1). G [ N , q N ] ∗ G [ N , q N ] ≥ γ ω , λ N I q N ( N ≥ 1 ) .
Intersect these events over rational λ > 1 \lambda> 1 λ > 1 and for the two arrays G , G ∗ G,G^* G , G ∗ . For an arbitrary real λ > 1 \lambda> 1 λ > 1 , choose a rational λ ′ ∈ ( 1 , λ ) \lambda' \in(1,\lambda) λ ′ ∈ ( 1 , λ ) . Since ⌈ N / λ ⌉ ≤ ⌈ N / λ ′ ⌉ \lceil N/\lambda\rceil\le\lceil N/\lambda'\rceil ⌈ N / λ ⌉ ≤ ⌈ N / λ ′ ⌉ , compression to the smaller column space gives the same conclusion for λ \lambda λ . Thus, on one event, the displayed Gram lower bound holds for both orientations and every real λ > 1 \lambda> 1 λ > 1 . Below, decrease γ ω , λ \gamma_{\omega,\lambda} γ ω , λ if necessary so that it works for both orientations.
Fix an array in this event, arbitrary a , b ∈ C a,b \in\mathcal{C} a , b ∈ C , and d ≥ 1 d \ge1 d ≥ 1 . Write q = ⌈ d / λ ⌉ q = \lceil d/\lambda\rceil q = ⌈ d / λ ⌉ and let P q P_q P q be the projection onto the first q q q coordinates. If b d = 0 b_d = 0 b d = 0 , the following lower bound is trivial. Otherwise the positive prefix expansion of the row weights gives
P q Z ( a , b ) ∗ Z ( a , b ) P q = D b , [ q ] [ ∑ N ≥ 1 ( a N 2 − a N + 1 2 ) G [ N , q ] ∗ G [ N , q ] ] D b , [ q ] ≥ γ ω , λ b d 2 [ ∑ N ≥ d N ( a N 2 − a N + 1 2 ) ] I q . \begin{aligned}
P_q Z(a,b)^* Z(a,b)P_q
=
D_{b,[q]}
\left[
\sum_{N\ge1}(a_N^2-a_{N+1}^2)G[N,q]^*G[N,q]
\right]
D_{b,[q]} \\
&\ge\gamma_{\omega,\lambda}b_d^2
\left[
\sum_{N\ge d}N(a_N^2-a_{N+1}^2)
\right]I_q.
\end{aligned} P q Z ( a , b ) ∗ Z ( a , b ) P q = D b , [ q ] [ N ≥ 1 ∑ ( a N 2 − a N + 1 2 ) G [ N , q ] ∗ G [ N , q ] ] D b , [ q ] ≥ γ ω , λ b d 2 [ N ≥ d ∑ N ( a N 2 − a N + 1 2 ) ] I q .
Here D b , [ q ] D_{b,[q]} D b , [ q ] denotes the diagonal restriction to the first q q q coordinates. The positive matrix series is convergent because Z ( a , b ) Z(a,b) Z ( a , b ) is Hilbert–Schmidt; it can equivalently be obtained by monotone finite-prefix approximation. For N ≥ d N \ge d N ≥ d , the restriction to the first q q q columns is a compression of the Gram estimate at N N N , and b j ≥ b d b_j \ge b_d b j ≥ b d for j ≤ q j \le q j ≤ q . Since
∑ N ≥ d N ( a N 2 − a N + 1 2 ) = F a ( d ) + ( d − 1 ) a d 2 ≥ F a ( d ) , \sum_{N\ge d}N(a_N^2-a_{N+1}^2)=F_a(d)+(d-1)a_d^2\ge F_a(d), N ≥ d ∑ N ( a N 2 − a N + 1 2 ) = F a ( d ) + ( d − 1 ) a d 2 ≥ F a ( d ) ,
the variational principle yields
s q ( Z ( a , b ) ) 2 ≥ γ ω , λ b d 2 F a ( d ) . s_q(Z(a,b))^2 \ge\gamma_{\omega,\lambda} b_d^2 F_a(d). s q ( Z ( a , b ) ) 2 ≥ γ ω , λ b d 2 F a ( d ) .
Applying the same argument to Z ( a , b ) ∗ = D b G ∗ D a Z(a,b)^* = D_bG^*D_a Z ( a , b ) ∗ = D b G ∗ D a gives s q ( Z ( a , b ) ) 2 ≥ γ ω , λ a d 2 F b ( d ) s_q(Z(a,b))^2 \ge\gamma_{\omega,\lambda} a_d^2 F_b(d) s q ( Z ( a , b ) ) 2 ≥ γ ω , λ a d 2 F b ( d ) . Averaging these two inequalities and using (3) , we obtain
s q ( Z ( a , b ) ) 2 ≥ γ ω , λ 2 [ b d 2 F a ( d ) + a d 2 F b ( d ) ] ≥ γ ω , λ 2 ρ d 2 . s_q(Z(a,b))^2 \ge\frac{\gamma_{\omega,\lambda}}{2}\left[b_d^2F_a(d)+a_d^2F_b(d)\right] \ge\frac{\gamma_{\omega,\lambda}}{2}\rho_d^2. s q ( Z ( a , b ) ) 2 ≥ 2 γ ω , λ [ b d 2 F a ( d ) + a d 2 F b ( d ) ] ≥ 2 γ ω , λ ρ d 2 .
This proves the lower estimate in (26) , simultaneously for all weights and all λ > 1 \lambda>1 λ > 1 . The principal-ideal criterion (12) now gives (27) .
Fixed singular-vector coordinates and conditioning. For fixed nonzero A A A , B B B , choose singular expansions
A = ∑ k a k u k ⊗ v k , B = ∑ ℓ b ℓ r ℓ ⊗ s ℓ . A=\sum_k a_k u_k\otimes v_k,\qquad B=\sum_\ell b_\ell r_\ell\otimes s_\ell. A = k ∑ a k u k ⊗ v k , B = ℓ ∑ b ℓ r ℓ ⊗ s ℓ .
indexed by their positive singular values. The corresponding orthonormal families may be finite or infinite. The Gaussian variables
g ~ k ℓ = ∑ i , j ≥ 1 ⟨ e i , v k ⟩ ⟨ r ℓ , e j ⟩ g i j \widetilde{g}_{k\ell}=\sum_{i,j\geq1}\langle e_i,v_k\rangle\langle r_\ell,e_j\rangle g_{ij} g k ℓ = i , j ≥ 1 ∑ ⟨ e i , v k ⟩ ⟨ r ℓ , e j ⟩ g ij
are defined by L 2 L^2 L 2 limits. Orthogonality of the coefficient vectors shows that they form a standard Gaussian array over F \mathbb{F} F . In the complex case, circularity ensures that their pseudocovariances vanish. The Gaussian isometry proved above allows the finite singular expansions to pass to the limit and gives
A G B = ∑ k , ℓ a k b ℓ g ~ k ℓ u k ⊗ s ℓ in L 2 ( Ω ; S 2 ) . AGB=\sum_{k,\ell}a_kb_\ell\widetilde{g}_{k\ell}u_k\otimes s_\ell\quad\text{in }L^2(\Omega;S_2). A GB = k , ℓ ∑ a k b ℓ g k ℓ u k ⊗ s ℓ in L 2 ( Ω ; S 2 ) .
The coordinate isometries determined by ( u k ) (u_k) ( u k ) and ( s ℓ ) (s_\ell) ( s ℓ ) preserve all nonzero singular values. If either positive-singular-value index set is finite, extend this rectangular Gaussian family to an infinite standard array by independent coordinates and pad the corresponding weights with zeros. The resulting assertion depends only on the original coordinates. Applying the diagonal result therefore proves (24) in both finite and infinite rank; if either factor is zero, the conclusion is immediate. The Gaussian coordinates used here depend on the fixed factors, and no common event over arbitrary choices of singular vectors has been asserted.
For the conditional statement, regard the square partial sums as functions of ( A , B , G ) (A,B,G) ( A , B , G ) . They are jointly measurable, and the set on which they are Cauchy in S 2 S_2 S 2 is Borel. Their limit, defined to be zero off that set, is therefore jointly measurable. Singular values are continuous in the Hilbert–Schmidt topology, and the finite-dilation criterion (12) expresses ideal equality using only countably many inequalities. Thus the event of convergence and ideal equality is jointly measurable as well. For every deterministic pair it has Gaussian probability one. Independence of ( A , B ) (A,B) ( A , B ) and G G G , followed by Fubini’s theorem, proves almost sure convergence and (24) for the random pair. Applying the fixed-factor isometry conditionally gives (25) . If its right-hand side is integrable, the conditional second moments of the truncation errors tend to zero and are bounded by ∥ A ∥ 2 2 ∥ B ∥ 2 2 \|A\|_2^2\|B\|_2^2 ∥ A ∥ 2 2 ∥ B ∥ 2 2 , so dominated convergence gives unconditional L 2 L^2 L 2 convergence. Conversely, such convergence makes the limit square integrable, and integrating (25) gives the stated necessary condition.
Failure of the endpoint. Let D N D_N D N be the Euclidean distance of the last row of G [ N , N ] G[N,N] G [ N , N ] from the span of its first N − 1 N-1 N − 1 rows. Conditional on these rows, their orthogonal complement is one-dimensional almost surely, and D N D_N D N has the law of the modulus of a standard scalar Gaussian. This conditional law is constant. Moreover, the conditioning rows determine D 1 , … , D N − 1 D_1,\ldots,D_{N-1} D 1 , … , D N − 1 . Consequently ( D N ) N ≥ 1 (D_N)_{N\geq1} ( D N ) N ≥ 1 is an iid sequence, and inf N ≥ 1 D N = 0 \inf_{N\geq1}D_N=0 inf N ≥ 1 D N = 0 almost surely. The variational characterization of the least singular value gives s N ( G [ N , N ] ) ≤ D N s_N(G[N,N])\leq D_N s N ( G [ N , N ]) ≤ D N , so
inf N ≥ 1 s N ( G [ N , N ] ) = 0 almost surely . \inf_{N\geq1}s_N(G[N,N])=0\quad\text{almost surely}. N ≥ 1 inf s N ( G [ N , N ]) = 0 almost surely .
For a = b = 1 [ 1 , N ] a=b=\mathbf{1}_{[1,N]} a = b = 1 [ 1 , N ] , the operator Z ( a , b ) Z(a,b) Z ( a , b ) is the leading N N N -square, extended by zero, and ( a ∘ b ) N = 1 (a\mathbin{\circ}b)_N=1 ( a ∘ b ) N = 1 . A positive constant valid for all these weights at λ = 1 \lambda=1 λ = 1 would bound all the displayed least singular values away from zero. This is impossible. □
The common event in (26) concerns diagonal factors in the fixed coordinate bases. Since the assertions hold there for every a , b ∈ C a,b\in\mathcal{C} a , b ∈ C , they also hold for measurable weights chosen as functions of the array, and they restrict to every pair of rank flags. They do not assert a simultaneous result for arbitrary array-dependent singular-vector directions. The conditional assertion has a different quantifier: for a random pair independent of G G G , the factors are held fixed when the Gaussian probability is evaluated, and the result is then integrated over their joint law.
Corollary 4.2 (Schatten estimates and power weights). On the common event of Corollary 4.1 , for every a , b ∈ C a,b\in\mathcal{C} a , b ∈ C and every 0 < q < ∞ 0<q<\infty 0 < q < ∞ , put Z = Z G ( a , b ) Z=Z_G(a,b) Z = Z G ( a , b ) and ρ = a ∘ b \rho=a\mathbin{\circ}b ρ = a ∘ b . Then
2 − 1 / q c ω , 2 ∥ ρ ∥ ℓ q ≤ ∥ Z ∥ q ≤ 2 1 / q C ω ∥ ρ ∥ ℓ q , (28) 2^{-1/q}c_{\omega,2}\lVert\rho\rVert_{\ell^q}\leq\lVert Z\rVert_q\leq2^{1/q}C_\omega\lVert\rho\rVert_{\ell^q}, \tag*{(28)} 2 − 1/ q c ω , 2 ∥ ρ ∥ ℓ q ≤ ∥ Z ∥ q ≤ 2 1/ q C ω ∥ ρ ∥ ℓ q , (28) where ∥ ρ ∥ ℓ q = ( ∑ n ρ n q ) 1 / q \lVert\rho\rVert_{\ell^q}=\left(\sum_n \rho_n^q\right)^{1/q} ∥ ρ ∥ ℓ q = ( ∑ n ρ n q ) 1/ q and the quantities are understood in [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] . In particular, Z ∈ S q Z\in S_q Z ∈ S q if and only if ρ ∈ ℓ q \rho\in\ell^q ρ ∈ ℓ q .
Both (26) and (28) apply to every leading rectangular truncation ( a i g i j b j ) 1 ≤ i ≤ N , 1 ≤ j ≤ M (a_i g_{ij} b_j)_{1\leq i\leq N,\,1\leq j\leq M} ( a i g ij b j ) 1 ≤ i ≤ N , 1 ≤ j ≤ M , with the weights replaced by a ( N ) = a 1 [ 1 , N ] a^{(N)}=a\mathbf{1}_{[1,N]} a ( N ) = a 1 [ 1 , N ] and b ( M ) = b 1 [ 1 , M ] b^{(M)}=b\mathbf{1}_{[1,M]} b ( M ) = b 1 [ 1 , M ] , and the profile replaced by ρ ( N , M ) = a ( N ) ∘ b ( M ) \rho^{(N,M)}=a^{(N)}\mathbin{\circ}b^{(M)} ρ ( N , M ) = a ( N ) ∘ b ( M ) . The constants are independent of N , M N,M N , M as well as the weights.
For a n = n − α a_n=n^{-\alpha} a n = n − α and b n = n − β b_n=n^{-\beta} b n = n − β with α , β > 1 / 2 \alpha,\beta>1/2 α , β > 1/2 , set γ = α + β − 1 / 2 \gamma=\alpha+\beta-1/2 γ = α + β − 1/2 . Then
s n ( Z G ( a , b ) ) ≍ ω , α , β n − γ , Z G ( a , b ) ∈ S q ⟺ q γ > 1. s_n(Z_G(a,b))\asymp_{\omega,\alpha,\beta}n^{-\gamma},\qquad Z_G(a,b)\in S_q\quad\Longleftrightarrow\quad q\gamma>1. s n ( Z G ( a , b )) ≍ ω , α , β n − γ , Z G ( a , b ) ∈ S q ⟺ q γ > 1.
All these assertions hold on that same event for all the stated parameters. The notation ≍ \asymp ≍ denotes two-sided comparison by positive constants.
Proof. Fix an outcome in the event of Corollary 4.1 and write σ n = s n ( Z ) \sigma_n=s_n(Z) σ n = s n ( Z ) . Taking λ = 2 \lambda=2 λ = 2 in (26) , raising to the power q > 0 q>0 q > 0 , and summing gives
∑ n ≥ 1 σ n q ≤ 2 C ω q ∑ n ≥ 1 ρ n q , ∑ n ≥ 1 ρ n q ≤ 2 c ω , 2 − q ∑ n ≥ 1 σ n q . \sum_{n\geq1}\sigma_n^q\leq2C_\omega^q\sum_{n\geq1}\rho_n^q,\qquad
\sum_{n\geq1}\rho_n^q\leq2c_{\omega,2}^{-q}\sum_{n\geq1}\sigma_n^q. n ≥ 1 ∑ σ n q ≤ 2 C ω q n ≥ 1 ∑ ρ n q , n ≥ 1 ∑ ρ n q ≤ 2 c ω , 2 − q n ≥ 1 ∑ σ n q .
Each positive integer occurs twice as ⌈ n / 2 ⌉ \lceil n/2\rceil ⌈ n /2 ⌉ . These inequalities hold for extended nonnegative sums and imply (28) for every q > 0 q>0 q > 0 , including q < 1 q<1 q < 1 . The rectangular assertions follow by applying the same comparisons to a ( N ) , b ( M ) ∈ C a^{(N)},b^{(M)}\in\mathcal{C} a ( N ) , b ( M ) ∈ C and padding the finite matrix by zero.
For the power weights, the integral test gives
F a ( n ) ∼ n 1 − 2 α 2 α − 1 , F b ( n ) ∼ n 1 − 2 β 2 β − 1 . F_a(n)\sim\frac{n^{1-2\alpha}}{2\alpha-1},\qquad
F_b(n)\sim\frac{n^{1-2\beta}}{2\beta-1}. F a ( n ) ∼ 2 α − 1 n 1 − 2 α , F b ( n ) ∼ 2 β − 1 n 1 − 2 β .
Formula (3) therefore yields
ρ n ∼ ( 1 2 α − 1 + 1 2 β − 1 ) 1 / 2 n − γ . \rho_n\sim\left(\frac{1}{2\alpha-1}+\frac{1}{2\beta-1}\right)^{1/2}n^{-\gamma}. ρ n ∼ ( 2 α − 1 1 + 2 β − 1 1 ) 1/2 n − γ .
The comparisons c ω , 2 ρ 2 n ≤ σ n ≤ C ω ρ ⌈ n / 2 ⌉ c_{\omega,2}\rho_{2n}\leq\sigma_n\leq C_\omega\rho_{\lceil n/2\rceil} c ω , 2 ρ 2 n ≤ σ n ≤ C ω ρ ⌈ n /2 ⌉ give the asserted singular-value order. Finally, (28) and the convergence criterion for ∑ n n − q γ \sum_n n^{-q\gamma} ∑ n n − q γ give the stated Schatten threshold, with divergence also at q γ = 1 q\gamma=1 q γ = 1 . Everything after fixing the outcome is deterministic, so no additional exceptional set is needed for any of the parameters. □
Order recovery and common fibres
We first show that a fixed noninclusion between input ideals imposes a common eventual bound on the tail-rank profiles of all probes for which the corresponding output inclusion holds. Conversely, a common eventual profile bound permits a continuous local family of inputs with identical output ideals. We then use these statements to prove Theorem B.
For a finite-valued function h : N → [ 1 , ∞ ) h:\mathbb{N}\to[1,\infty) h : N → [ 1 , ∞ ) , define
B h = { c ∈ C + : r c ≤ ∗ h } , Bad ( a , b ) = { c ∈ C + : I ( a ∘ c ) ⊆ I ( b ∘ c ) } . \mathcal{B}_h=\{c\in\mathcal{C}_+:r_c\leq^* h\},\qquad
\operatorname{Bad}(a,b)=\{c\in\mathcal{C}_+:\mathcal{I}(a\mathbin{\circ}c)\subseteq\mathcal{I}(b\mathbin{\circ}c)\}. B h = { c ∈ C + : r c ≤ ∗ h } , Bad ( a , b ) = { c ∈ C + : I ( a ∘ c ) ⊆ I ( b ∘ c )} .
The starting index in the eventual inequality defining B h B_h B h may depend on c c c . When I ( a ) ⊈ I ( b ) \mathcal{I}(a) \not\subseteq\mathcal{I}(b) I ( a ) ⊆ I ( b ) , the set Bad ( a , b ) \operatorname{Bad}(a,b) Bad ( a , b ) consists precisely of the probes that do not detect this noninclusion.
Lemma 5.1 (An eventual bound for nonseparating probes). If a , b ∈ C + a,b \in\mathcal{C}_+ a , b ∈ C + and I ( a ) ⊈ I ( b ) \mathcal{I}(a) \not\subseteq\mathcal{I}(b) I ( a ) ⊆ I ( b ) , then Bad ( a , b ) ⊆ B h \operatorname{Bad}(a,b) \subseteq B_h Bad ( a , b ) ⊆ B h for some finite-valued h ≥ 1 h \ge1 h ≥ 1 .
Proof. Put x = a 2 x=a^2 x = a 2 and y = b 2 y=b^2 y = b 2 . For each integer t ≥ 1 t \ge1 t ≥ 1 , the principal-ideal criterion (12) implies that the ratios x n y ⌈ n / t ⌉ \frac{x_n}{y_{\lceil n/t\rceil}} y ⌈ n / t ⌉ x n are unbounded. Every finite collection of these ratios is finite because b b b is strictly positive. We may therefore choose n t n_t n t sufficiently large that
m t : = ⌈ n t t ⌉ ≥ t , x n t > 4 t y m t . m_t := \left\lceil\frac{n_t}{t} \right\rceil\ge t,\qquad x_{n_t} > 4t\,y_{m_t}. m t := ⌈ t n t ⌉ ≥ t , x n t > 4 t y m t .
Define
h ( t ) = max { 1 , 2 n t , 4 t R y ( m t + 1 ) x n t } . h(t)=\max\left\{1,2n_t,\frac{4tR_y(m_t+1)}{x_{n_t}}\right\}. h ( t ) = max { 1 , 2 n t , x n t 4 t R y ( m t + 1 ) } .
This is a finite-valued function, since x n t > 0 x_{n_t}>0 x n t > 0 and y ∈ ℓ 1 y\in\ell^1 y ∈ ℓ 1 .
Fix c ∈ C + c\in\mathcal{C}_+ c ∈ C + and write w = c 2 w=c^2 w = c 2 and W = R w W=R_w W = R w . Suppose that r c ( t ) > h ( t ) r_c(t)>h(t) r c ( t ) > h ( t ) . The indices satisfy t ≤ m t ≤ n t t\le m_t\le n_t t ≤ m t ≤ n t . Since w w w is nonincreasing and w t = W ( t ) / r c ( t ) w_t=W(t)/r_c(t) w t = W ( t ) / r c ( t ) ,
W ( n t ) = W ( t ) − ∑ k = t n t − 1 w k ≥ W ( t ) − ( n t − t ) w t > 1 2 W ( t ) , W(n_t)=W(t)-\sum_{k=t}^{n_t-1}w_k\ge W(t)-(n_t-t)w_t>\frac{1}{2}W(t), W ( n t ) = W ( t ) − k = t ∑ n t − 1 w k ≥ W ( t ) − ( n t − t ) w t > 2 1 W ( t ) ,
W ( m t ) ≤ W ( t ) , w m t ≤ w t = W ( t ) r c ( t ) . W(m_t)\le W(t),\qquad w_{m_t}\le w_t=\frac{W(t)}{r_c(t)}. W ( m t ) ≤ W ( t ) , w m t ≤ w t = r c ( t ) W ( t ) .
Using (2.5) and the definition of h ( t ) h(t) h ( t ) , we obtain
( L w x ) n t ≥ x n t W ( n t ) > 1 2 x n t W ( t ) , (L_wx)_{n_t}\ge x_{n_t}W(n_t)>\frac{1}{2}x_{n_t}W(t), ( L w x ) n t ≥ x n t W ( n t ) > 2 1 x n t W ( t ) ,
( L w y ) m t ≤ W ( t ) ( y m t + R y ( m t + 1 ) r c ( t ) ) < x n t W ( t ) 2 t . (L_wy)_{m_t}\le W(t)\left(y_{m_t}+\frac{R_y(m_t+1)}{r_c(t)}\right)<\frac{x_{n_t}W(t)}{2t}. ( L w y ) m t ≤ W ( t ) ( y m t + r c ( t ) R y ( m t + 1 ) ) < 2 t x n t W ( t ) .
Consequently ( L w x ) n t > t ( L w y ) m t (L_wx)_{n_t}>t(L_wy)_{m_t} ( L w x ) n t > t ( L w y ) m t .
If r c ̸ ≤ ∗ h r_c\not\le^* h r c ≤ ∗ h , the preceding estimates hold for infinitely many t t t . Fix any integer dilation M ≥ 1 M\ge1 M ≥ 1 and any constant K > 0 K>0 K > 0 . Choose such a t t t with t > max { M , K 2 } t>\max\{M,K^2\} t > max { M , K 2 } . Then m t ≤ ⌈ n t / M ⌉ m_t\le\lceil n_t/M\rceil m t ≤ ⌈ n t / M ⌉ . Moreover, L w y = ( b ∘ c ) 2 L_wy=(b\circ c)^2 L w y = ( b ∘ c ) 2 is nonincreasing. Hence
( L w x ) n t > K 2 ( L w y ) ⌈ n t / M ⌉ . (L_wx)_{n_t}>K^2(L_wy)_{\lceil n_t/M\rceil}. ( L w x ) n t > K 2 ( L w y ) ⌈ n t / M ⌉ .
Since L w x = ( a ∘ c ) 2 L_wx=(a\circ c)^2 L w x = ( a ∘ c ) 2 , taking square roots excludes the comparison a ∘ c ≤ K Δ M ( b ∘ c ) a\circ c\le K\Delta_M(b\circ c) a ∘ c ≤ K Δ M ( b ∘ c ) . The constants K K K and M M M were arbitrary, so (2.1) gives I ( a ∘ c ) ⊈ I ( b ∘ c ) \mathcal{I}(a\circ c)\not\subseteq\mathcal{I}(b\circ c) I ( a ∘ c ) ⊆ I ( b ∘ c ) . This proves Bad ( a , b ) ⊆ B h \operatorname{Bad}(a,b)\subseteq B_h Bad ( a , b ) ⊆ B h by contraposition. □ \square □
Noninclusions involving a zero or finite-rank input are detected by every positive probe. Indeed, rank ( a ∘ c ) = rank ( a ) \operatorname{rank}(a\circ c)=\operatorname{rank}(a) rank ( a ∘ c ) = rank ( a ) for c ∈ C + c\in\mathcal{C}_+ c ∈ C + , and all nonzero finite-rank operators generate the same ideal, contained in every nonzero principal ideal.
The local construction will use the following transfer of a relative-tail inequality. If x , y ∈ ℓ 1 x,y\in\ell^1 x , y ∈ ℓ 1 are nonnegative, η n ≥ 0 \eta_n\ge0 η n ≥ 0 , and R x ( n ) ≤ η n R y ( n ) R_x(n)\le\eta_nR_y(n) R x ( n ) ≤ η n R y ( n ) , then for c ∈ C + c\in\mathcal{C}_+ c ∈ C + ,
( L c 2 x ) n = c n 2 [ R x ( n ) + ( r c ( n ) − 1 ) x n ] ≤ η n r c ( n ) ( L c 2 y ) n . (29) (L_{c^2x})_n=c_n^2\left[R_x(n)+(r_c(n)-1)x_n\right]\le\eta_nr_c(n)(L_{c^2y})_n. \tag*{(29)} ( L c 2 x ) n = c n 2 [ R x ( n ) + ( r c ( n ) − 1 ) x n ] ≤ η n r c ( n ) ( L c 2 y ) n . (29)
To verify this, use r c ( n ) ≥ 1 r_c(n)\ge1 r c ( n ) ≥ 1 and x n ≤ R x ( n ) x_n\le R_x(n) x n ≤ R x ( n ) to bound the bracket by r c ( n ) R x ( n ) r_c(n)R_x(n) r c ( n ) R x ( n ) , and then use ( L c 2 y ) n ≥ c n 2 R y ( n ) (L_{c^2y})_n\ge c_n^2R_y(n) ( L c 2 y ) n ≥ c n 2 R y ( n ) .
Proposition 5.2 (Flag-local common fibres). Fix a flag N N N , an input a ∈ C N , + a\in\mathcal{C}_{N,+} a ∈ C N , + , a finite-valued h ≥ 1 h\ge1 h ≥ 1 , an integer m ≥ 1 m\ge1 m ≥ 1 , a number ρ > 0 \rho>0 ρ > 0 , and a positive sequence ε n → 0 \varepsilon_n\to0 ε n → 0 . There is a continuous embedding S ↦ z S S\mapsto z^S S ↦ z S of 2 N 2^N 2 N into C N , + \mathcal{C}_{N,+} C N , + such that, with b = z ∅ b=z^\varnothing b = z ∅ ,
b ≤ z S ≤ z N ≤ a , z n S = a n ( n ≤ m ) , sup S ∥ z S − a ∥ 2 < ρ . (30) b\le z^S\le z^N\le a,\qquad z_n^S=a_n\ (n\le m),\qquad\sup_S\lVert z^S-a\rVert_2<\rho. \tag*{(30)} b ≤ z S ≤ z N ≤ a , z n S = a n ( n ≤ m ) , S sup ∥ z S − a ∥ 2 < ρ . (30) The family preserves coordinatewise minima and maxima, and its principal ideals satisfy
I ( z S ) ⊆ I ( z T ) ⟺ S ⊆ ∗ T . \mathcal{I}(z^S)\subseteq\mathcal{I}(z^T)\quad\Longleftrightarrow\quad S\subseteq^*T. I ( z S ) ⊆ I ( z T ) ⟺ S ⊆ ∗ T .
For every c ∈ C + c \in\mathcal{C}_{+} c ∈ C + and every n n n with r c ( n ) ≤ h ( n ) r_{c}(n) \le h(n) r c ( n ) ≤ h ( n ) ,
1 ≤ ( z S ∘ c ) n ( b ∘ c ) n ≤ 1 + ε n ( S ⊆ N ) . (31) 1 \le\frac{(z^{S} \mathbin{\circ} c)_{n}}{(b \mathbin{\circ} c)_{n}} \le1+\varepsilon_{n} \qquad(S \subseteq\mathbb{N}). \tag*{(31)} 1 ≤ ( b ∘ c ) n ( z S ∘ c ) n ≤ 1 + ε n ( S ⊆ N ) . (31)
For all n ≥ 1 n \ge1 n ≥ 1 and S ⊆ N S \subseteq\mathbb{N} S ⊆ N , one also has
F b ( n ) ≤ F z S ( n ) ≤ ( 1 + ε n ) 2 F b ( n ) . (32) F_{b}(n) \le F_{z^{S}}(n) \le(1+\varepsilon_{n})^{2}F_{b}(n). \tag*{(32)} F b ( n ) ≤ F z S ( n ) ≤ ( 1 + ε n ) 2 F b ( n ) . (32)
Thus the ideals I ( F z S ) \mathcal{I}(\sqrt{F_{z^{S}}}) I ( F z S ) are independent of S S S , and for each c ∈ B h c \in B_{h} c ∈ B h the ideals I ( z S ∘ c ) \mathcal{I}(z^{S} \mathbin{\circ} c) I ( z S ∘ c ) are independent of S S S . The upper envelope z N z^{\mathbb{N}} z N can be chosen in ℓ 1 \ell^{1} ℓ 1 .
Proof. Put
q = a 2 , η n = ( 1 + ε n ) 2 − 1 h ( n ) > 0. q=a^{2}, \qquad\eta_{n}=\frac{(1+\varepsilon_{n})^{2}-1}{h(n)}>0. q = a 2 , η n = h ( n ) ( 1 + ε n ) 2 − 1 > 0.
Since a ∈ ℓ 2 a \in\ell^{2} a ∈ ℓ 2 and the flag endpoints tend to infinity, choose a flag endpoint P 0 ≥ m P_{0} \ge m P 0 ≥ m such that F a ( P 0 + 1 ) < ρ 2 F_{a}(P_{0}+1)<\rho^{2} F a ( P 0 + 1 ) < ρ 2 . For each j ≥ 1 j \ge1 j ≥ 1 , having chosen P j − 1 P_{j-1} P j − 1 , put k j = P j − 1 + 1 k_{j}=P_{j-1}+1 k j = P j − 1 + 1 . First choose a flag endpoint S j ≥ j k j S_{j} \ge jk_{j} S j ≥ j k j ; then choose a flag endpoint P j P_{j} P j sufficiently large that
L j : = P j − P j − 1 ≥ 2 j S j max 1 ≤ n ≤ S j { 1 , η n − 1 } . L_{j}:=P_{j}-P_{j-1}\ge2jS_{j}\max_{1\le n\le S_{j}}\{1,\eta_{n}^{-1}\}. L j := P j − P j − 1 ≥ 2 j S j 1 ≤ n ≤ S j max { 1 , η n − 1 } .
In particular, L j ≥ 2 S j L_{j}\ge2S_{j} L j ≥ 2 S j and P j > S j P_{j}>S_{j} P j > S j . Set β 0 = q P 0 \beta_{0}=q_{P_{0}} β 0 = q P 0 and, for j ≥ 1 j \ge1 j ≥ 1 , define
β j = 1 j + 1 min { q P j , β j − 1 , 4 − j L j 2 } . \beta_{j}=\frac{1}{j+1}\min\left\{q_{P_{j}},\beta_{j-1},\frac{4^{-j}}{L_{j}^{2}}\right\}. β j = j + 1 1 min { q P j , β j − 1 , L j 2 4 − j } .
These coefficients are positive and satisfy
( j + 1 ) β j ≤ q P j , ( j + 1 ) β j ≤ β j − 1 , L j ( j + 1 ) β j ≤ 2 − j . (j+1)\beta_{j}\le q_{P_{j}}, \qquad(j+1)\beta_{j}\le\beta_{j-1}, \qquad L_{j}\sqrt{(j+1)\beta_{j}}\le2^{-j}. ( j + 1 ) β j ≤ q P j , ( j + 1 ) β j ≤ β j − 1 , L j ( j + 1 ) β j ≤ 2 − j .
For the intervals J j = [ k j , P j ] ∩ N J_{j}=[k_{j},P_{j}]\cap\mathbb{N} J j = [ k j , P j ] ∩ N and I j = [ k j , S j ] ∩ N I_{j}=[k_{j},S_{j}]\cap\mathbb{N} I j = [ k j , S j ] ∩ N , define
y = q 1 [ 1 , P 0 ] + ∑ j ≥ 1 β j 1 J j , x S = ∑ j ∈ S j β j 1 I j , z S = y + x S . y=q1_{[1,P_{0}]}+\sum_{j\ge1}\beta_{j}1_{J_{j}}, \qquad x^{S}=\sum_{j\in S}j\beta_{j}1_{I_{j}}, \qquad z^{S}=\sqrt{y+x^{S}}. y = q 1 [ 1 , P 0 ] + j ≥ 1 ∑ β j 1 J j , x S = j ∈ S ∑ j β j 1 I j , z S = y + x S .
The intervals J j J_{j} J j partition the coordinates after P 0 P_{0} P 0 . On I j I_{j} I j , the sequence y + x S y+x^{S} y + x S has value either β j \beta_{j} β j or ( j + 1 ) β j (j+1)\beta_{j} ( j + 1 ) β j , and on J j ∖ I j J_{j}\setminus I_{j} J j ∖ I j it has value β j \beta_{j} β j . Thus it is nonincreasing within each J j J_{j} J j . The bound ( j + 1 ) β j ≤ β j − 1 (j+1)\beta_{j}\le\beta_{j-1} ( j + 1 ) β j ≤ β j − 1 gives the required comparison between consecutive blocks, including the comparison with the unchanged prefix when j = 1 j=1 j = 1 . Furthermore, on J j J_{j} J j ,
y n + x n S ≤ ( j + 1 ) β j ≤ q P j ≤ q n . y_{n}+x^{S}_{n}\le(j+1)\beta_{j}\le q_{P_{j}}\le q_{n}. y n + x n S ≤ ( j + 1 ) β j ≤ q P j ≤ q n .
All changes occur at flag endpoints. It follows that z S ∈ C N , + z^{S}\in\mathcal{C}_{\mathbb{N},+} z S ∈ C N , + and b ≤ z S ≤ z N ≤ a b\le z^{S}\le z^{\mathbb{N}}\le a b ≤ z S ≤ z N ≤ a . Since z S = a z^{S}=a z S = a through P 0 P_{0} P 0 ,
∥ z S − a ∥ 2 2 = ∑ n > P 0 ( a n − z n S ) 2 ≤ F a ( P 0 + 1 ) < ρ 2 , \lVert z^{S}-a\rVert_{2}^{2}=\sum_{n>P_{0}}(a_{n}-z^{S}_{n})^{2}\le F_{a}(P_{0}+1)<\rho^{2}, ∥ z S − a ∥ 2 2 = n > P 0 ∑ ( a n − z n S ) 2 ≤ F a ( P 0 + 1 ) < ρ 2 ,
uniformly in S S S . This proves (5.2).
We next compare the tails of x = x N x=x^{\mathbb{N}} x = x N and y y y . Fix n n n and j j j . If n ≤ S j n\le S_{j} n ≤ S j , then
R β j 1 J j ( n ) ≥ ( L j − S j ) β j ≥ 1 2 L j β j , R_{\beta_{j}1_{J_{j}}}(n)\ge(L_{j}-S_{j})\beta_{j}\ge\frac{1}{2}L_{j}\beta_{j}, R β j 1 J j ( n ) ≥ ( L j − S j ) β j ≥ 2 1 L j β j ,
R j β j 1 I j ( n ) ≤ j ∣ I j ∣ β j ≤ j S j β j ≤ 1 2 η n L j β j . R_{j\beta_{j}1_{I_{j}}}(n)\le j|I_{j}|\beta_{j}\le jS_{j}\beta_{j}\le\frac{1}{2}\eta_{n}L_{j}\beta_{j}. R j β j 1 I j ( n ) ≤ j ∣ I j ∣ β j ≤ j S j β j ≤ 2 1 η n L j β j .
For the first inequality, the number of remaining coordinates of J j J_{j} J j is at least P j − S j + 1 ≥ L j − S j P_{j}-S_{j}+1\ge L_{j}-S_{j} P j − S j + 1 ≥ L j − S j ; the last inequality follows from the choice of L j L_{j} L j . If n > S j n>S_{j} n > S j , the tail of j β j 1 I j j\beta_{j}1_{I_{j}} j β j 1 I j is zero. Hence, for every j j j ,
R j β j 1 I j ( n ) ≤ η n R β j 1 J j ( n ) . R_{j\beta_{j}1_{I_{j}}}(n)\le\eta_{n}R_{\beta_{j}1_{J_{j}}}(n). R j β j 1 I j ( n ) ≤ η n R β j 1 J j ( n ) .
Summing these inequalities and including the nonnegative prefix of y y y gives
R x ( n ) ≤ η n R y ( n ) ( n ≥ 1 ) . (33) R_{x}(n)\le\eta_{n}R_{y}(n)\qquad(n\ge1). \tag*{(33)} R x ( n ) ≤ η n R y ( n ) ( n ≥ 1 ) . (33)
Because 0 ≤ x S ≤ x 0 \leq x^{S} \leq x 0 ≤ x S ≤ x , linearity and positivity of L c 2 L_{c^{2}} L c 2 , together with (29) , imply
1 ≤ ( L c 2 ( y + x S ) ) n ( L c 2 y ) n ≤ 1 + η n r c ( n ) . 1 \leq\frac{(L_{c^{2}}(y+x^{S}))_{n}}{(L_{c^{2}}y)_{n}} \leq1+\eta_{n}r_{c}(n). 1 ≤ ( L c 2 y ) n ( L c 2 ( y + x S ) ) n ≤ 1 + η n r c ( n ) .
If r c ( n ) ≤ h ( n ) r_{c}(n) \leq h(n) r c ( n ) ≤ h ( n ) , the last quantity is at most ( 1 + ϵ n ) 2 (1+\epsilon_{n})^{2} ( 1 + ϵ n ) 2 , proving (31) . Likewise, (33) and h ( n ) ≥ 1 h(n) \geq1 h ( n ) ≥ 1 give
R y ( n ) ≤ R y + x S ( n ) ≤ ( 1 + η n ) R y ( n ) ≤ ( 1 + ϵ n ) 2 R y ( n ) , R_{y}(n) \leq R_{y+x^{S}}(n) \leq(1+\eta_{n})R_{y}(n) \leq(1+\epsilon_{n})^{2}R_{y}(n), R y ( n ) ≤ R y + x S ( n ) ≤ ( 1 + η n ) R y ( n ) ≤ ( 1 + ϵ n ) 2 R y ( n ) ,
which is (32) .
Since ϵ n → 0 \epsilon_{n} \to0 ϵ n → 0 , the sequence 1 + ϵ n 1+\epsilon_{n} 1 + ϵ n is bounded. Thus (32) implies equality of all the ideals I ( F S ) \mathcal{I}(\sqrt{F^{S}}) I ( F S ) . For a fixed c ∈ B h c \in\mathcal{B}_{h} c ∈ B h , choose N c N_{c} N c such that r c ( n ) ≤ h ( n ) r_{c}(n) \leq h(n) r c ( n ) ≤ h ( n ) for n ≥ N c n \geq N_{c} n ≥ N c . The same N c N_{c} N c applies to every S S S . On these coordinates, (31) gives a uniformly bounded ratio. On the remaining finitely many coordinates, positivity of b ⊙ c b \mathbin{\odot} c b ⊙ c and the inequality z S ⊙ c ≤ z N ⊙ c z^{S} \mathbin{\odot} c \leq z^{\mathbb{N}} \mathbin{\odot} c z S ⊙ c ≤ z N ⊙ c give a finite bound independent of S S S . Consequently, for some C c < ∞ C_{c}<\infty C c < ∞ ,
b ⊙ c ≤ z S ⊙ c ≤ C c ( b ⊙ c ) ( S ⊆ N ) , b \mathbin{\odot} c \leq z^{S} \mathbin{\odot} c \leq C_{c}(b \mathbin{\odot} c) \qquad(S \subseteq\mathbb{N}), b ⊙ c ≤ z S ⊙ c ≤ C c ( b ⊙ c ) ( S ⊆ N ) ,
which proves the asserted equality of output ideals.
The coefficients j β j ∣ I j ∣ j\beta_{j}|I_{j}| j β j ∣ I j ∣ are summable because 0 ≤ x ≤ q ∈ ℓ 1 0 \leq x \leq q \in\ell^{1} 0 ≤ x ≤ q ∈ ℓ 1 . The elementary inequality ( u − v ) 2 ≤ ∣ u − v ∣ (\sqrt{u}-\sqrt{v})^{2} \leq|u-v| ( u − v ) 2 ≤ ∣ u − v ∣ for u , v ≥ 0 u,v \geq0 u , v ≥ 0 therefore gives
∥ z S − z T ∥ 2 2 ≤ ∑ j ∈ S △ T j β j ∣ I j ∣ . \lVert z^{S}-z^{T}\rVert_{2}^{2} \leq\sum_{j\in S\mathbin{\triangle}T} j\beta_{j}|I_{j}|. ∥ z S − z T ∥ 2 2 ≤ j ∈ S △ T ∑ j β j ∣ I j ∣.
Agreement of S S S and T T T on an increasingly long finite set of indices makes the right-hand side tend to zero, proving continuity in the product topology on 2 N 2^{\mathbb{N}} 2 N . For each j j j , membership of j j j in S S S is determined by the value of z S z^{S} z S on I j I_{j} I j . Thus the map is injective. A continuous injection from the compact space 2 N 2^{\mathbb{N}} 2 N into the metric space C N , + \mathcal{C}_{\mathbb{N},+} C N , + is a topological embedding.
If S ∖ T S\setminus T S ∖ T is finite, then z S ≤ z T z^{S}\leq z^{T} z S ≤ z T outside a finite set of coordinates. Strict positivity absorbs those finitely many coordinates into a multiplicative constant, so I ( z S ) ⊆ I ( z T ) \mathcal{I}(z^{S})\subseteq\mathcal{I}(z^{T}) I ( z S ) ⊆ I ( z T ) . Conversely, suppose that S ∖ T S\setminus T S ∖ T is infinite and fix an integer dilation M ≥ 1 M\geq1 M ≥ 1 . For every j ∈ S ∖ T j\in S\setminus T j ∈ S ∖ T with j ≥ M j\geq M j ≥ M , the inequality S j ≥ j k j S_{j}\geq j k_{j} S j ≥ j k j implies
k j ≤ ⌊ S j / M ⌋ ≤ S j . k_{j}\leq\lfloor S_{j}/M\rfloor\leq S_{j}. k j ≤ ⌊ S j / M ⌋ ≤ S j .
Both indices therefore lie in I j I_{j} I j , and
( z S j S ) 2 ( z ⌊ S j / M ⌋ T ) 2 = j + 1. \frac{(z^{S}_{S_{j}})^{2}}{(z^{T}_{\lfloor S_{j}/M\rfloor})^{2}}=j+1. ( z ⌊ S j / M ⌋ T ) 2 ( z S j S ) 2 = j + 1.
These ratios are unbounded as j j j ranges over S ∖ T S\setminus T S ∖ T . No constant can give the comparison required by (12) for this dilation, and M M M was arbitrary. Hence I ( z S ) ⊈ I ( z T ) \mathcal{I}(z^{S})\not\subseteq\mathcal{I}(z^{T}) I ( z S ) ⊆ I ( z T ) . The disjointness of the intervals I j I_{j} I j also gives
z S ∩ T = min ( z S , z T ) , z S ∪ T = max ( z S , z T ) . z^{S\cap T}=\min(z^{S},z^{T}),\qquad z^{S\cup T}=\max(z^{S},z^{T}). z S ∩ T = min ( z S , z T ) , z S ∪ T = max ( z S , z T ) .
By (13) , these coordinatewise operations represent the intersection and sum of the corresponding principal ideals.
Finally, the upper envelope is summable:
∑ n z n N ≤ ∑ n ≤ P 0 a n + ∑ j ≥ 1 L j ( j + 1 ) β j ≤ ∑ n ≤ P 0 a n + ∑ j ≥ 1 2 − j < ∞ . \sum_{n}z^{\mathbb{N}}_{n}\leq\sum_{n\leq P_{0}}a_{n}+\sum_{j\geq1}L_{j}\sqrt{(j+1)\beta_{j}}\leq\sum_{n\leq P_{0}}a_{n}+\sum_{j\geq1}2^{-j}<\infty. n ∑ z n N ≤ n ≤ P 0 ∑ a n + j ≥ 1 ∑ L j ( j + 1 ) β j ≤ n ≤ P 0 ∑ a n + j ≥ 1 ∑ 2 − j < ∞.
Equivalent formulations. For ∅ ≠ Q ⊆ C + \varnothing\neq Q\subseteq\mathcal{C}_{+} ∅ = Q ⊆ C + , each of the following conditions is equivalent to the conditions in Theorem B :
The maps { Φ c : c ∈ Q } \{\Phi_{c}:c\in Q\} { Φ c : c ∈ Q } jointly reflect equality on P 2 \mathbb{P}_{2} P 2 : if I , J ∈ P 2 I,J\in\mathbb{P}_{2} I , J ∈ P 2 satisfy Φ c ( I ) = Φ c ( J ) \Phi_{c}(I)=\Phi_{c}(J) Φ c ( I ) = Φ c ( J ) for every c ∈ Q c\in Q c ∈ Q , then I = J I=J I = J .
Every prescribed countable family ( I j ) j ≥ 1 (I_{j})_{j\geq1} ( I j ) j ≥ 1 in P 2 \mathbb{P}_{2} P 2 admits a single c ∈ Q c\in Q c ∈ Q such that
I j ⊆ I k ⟺ Φ c ( I j ) ⊆ Φ c ( I k ) ( j , k ≥ 1 ) . I_{j}\subseteq I_{k}\quad\Longleftrightarrow\quad\Phi_{c}(I_{j})\subseteq\Phi_{c}(I_{k})\qquad(j,k\geq1). I j ⊆ I k ⟺ Φ c ( I j ) ⊆ Φ c ( I k ) ( j , k ≥ 1 ) .
Proof of Theorem B and the additional equivalences. Suppose first that the profiles of Q \mathcal{Q} Q have no common finite-valued eventual upper bound. For any a , b ∈ C + a,b\in\mathcal{C}_{+} a , b ∈ C + with I ( a ) ⊈ I ( b ) \mathcal{I}(a)\not\subseteq\mathcal{I}(b) I ( a ) ⊆ I ( b ) , Lemma 5.1 gives a function h h h such that Bad ( a , b ) ⊆ B h \operatorname{Bad}(a,b)\subseteq B_h Bad ( a , b ) ⊆ B h . By the assumption on Q \mathcal{Q} Q , some c ∈ Q c\in\mathcal{Q} c ∈ Q satisfies r c ̸ ≤ ∗ h r_c\not\leq^{*}h r c ≤ ∗ h , and this probe detects the noninclusion. Noninclusions involving zero or finite rank are detected by every c ∈ Q c\in\mathcal{Q} c ∈ Q , as noted above. Since positive probes preserve all ideal inclusions, this proves the order equivalence on P 2 \mathfrak{P}_2 P 2 and, by restriction, on every flag class. Applying the equivalence in both directions gives joint reflection of equality on these classes as well.
Now fix a prescribed countable input system. Its ordered pairs form a countable set. For each noninclusion between two strictly positive members, take a bound from Lemma 5.1 , and enumerate these bounds as h 1 , h 2 , … h_1,h_2,\ldots h 1 , h 2 , … . If the list is finite and nonempty, repeat its entries to obtain such a sequence. Define
h ( n ) = max 1 ≤ j ≤ n h j ( n ) . h(n)=\max_{1\leq j\leq n}h_j(n). h ( n ) = 1 ≤ j ≤ n max h j ( n ) .
For each fixed j j j , one has h j ( n ) ≤ h ( n ) h_j(n)\leq h(n) h j ( n ) ≤ h ( n ) when n ≥ j n\geq j n ≥ j . Choose c ∈ Q c\in\mathcal{Q} c ∈ Q with r c ̸ ≤ ∗ h r_c\not\leq^{*}h r c ≤ ∗ h . This choice implies r c ̸ ≤ ∗ h j r_c\not\leq^{*}h_j r c ≤ ∗ h j for every j j j , since an eventual bound by any h j h_j h j would imply an eventual bound by h h h . Thus this single probe detects every noninclusion in the list. It also detects the noninclusions involving zero or finite rank and preserves all inclusions. If the list is empty, any c ∈ Q c\in\mathcal{Q} c ∈ Q has the required property. This proves the single-probe assertion for every prescribed countable input system.
Conversely, suppose that Q ⊆ B h \mathcal{Q}\subseteq B_h Q ⊆ B h for some finite-valued h ≥ 1 h\geq1 h ≥ 1 . Let U ⊆ C N [ d ] U\subseteq\mathcal{C}_{N}[d] U ⊆ C N [ d ] be nonempty and relatively open, and fix any v ∈ U v\in U v ∈ U . Choose ρ > 0 \rho>0 ρ > 0 such that
{ u ∈ C N [ d ] : ∥ u − v ∥ 2 < ρ } ⊆ U . \{u\in\mathcal{C}_{N}[d]:\lVert u-v\rVert_2<\rho\}\subseteq U. { u ∈ C N [ d ] : ∥ u − v ∥ 2 < ρ } ⊆ U .
Apply Proposition 5.2 with a = v a=v a = v , with the prescribed finite prefix, and with any prescribed positive sequence ε n → 0 \varepsilon_n\to0 ε n → 0 . Its family satisfies z S ≤ v ≤ d z^{S}\leq v\leq d z S ≤ v ≤ d and ∥ z S − v ∥ 2 < ρ \lVert z^{S}-v\rVert_2<\rho ∥ z S − v ∥ 2 < ρ , so the entire family belongs to U U U . The proposition supplies the asserted order relation, the lattice identities, a common tail ideal, and the common output ideal for each c ∈ Q c\in\mathcal{Q} c ∈ Q . It also preserves the prescribed prefix and gives the stated relative output estimates, with a starting index that may depend on c c c but is independent of S S S .
The two endpoint ideals I ( z ∅ ) \mathcal{I}(z^{\varnothing}) I ( z ∅ ) and I ( z N ) \mathcal{I}(z^{\mathbb{N}}) I ( z N ) are distinct, whereas their output ideals agree for every c ∈ Q c\in\mathcal{Q} c ∈ Q . Thus joint reflection of equality fails, and the reverse inclusion between these endpoints shows that order reflection fails even on the prescribed flag class. No single probe in Q \mathcal{Q} Q can reflect all inclusions in this two-element input system. Together with the first part of the proof, these failures establish all the asserted equivalences and complete the local conclusion. □ \square □
Corollary 5.3 (A common fibre for every prescribed sampling law). Let μ \mu μ be a Borel probability measure on C + \mathcal{C}_{+} C + with its ℓ 2 \ell^{2} ℓ 2 topology. Every local flag region in Theorem B contains a continuous Boolean family a S a^{S} a S for which, for μ \mu μ -almost every c c c , the ideals I ( a S ∘ c ) \mathcal{I}(a^{S}\circ c) I ( a S ∘ c ) are independent of S S S .
Proof. For each n n n , the function r c ( n ) = ∥ ( c k ) k ≥ n ∥ 2 2 / c n 2 r_c(n)=\lVert(c_k)_{k\geq n}\rVert_2^2/c_n^2 r c ( n ) = ∥( c k ) k ≥ n ∥ 2 2 / c n 2 is continuous and finite on C + \mathcal{C}_{+} C + . We may therefore choose a finite number h ( n ) ≥ 1 h(n)\geq1 h ( n ) ≥ 1 such that
μ { c : r c ( n ) > h ( n ) } ≤ 2 − n . \mu\{c:r_c(n)>h(n)\}\leq2^{-n}. μ { c : r c ( n ) > h ( n )} ≤ 2 − n .
The set
B h = ⋃ N ≥ 1 ⋂ n ≥ N { c ∈ C + : r c ( n ) ≤ h ( n ) } B_h=\bigcup_{N\geq1}\bigcap_{n\geq N}\{c\in\mathcal{C}_{+}:r_c(n)\leq h(n)\} B h = N ≥ 1 ⋃ n ≥ N ⋂ { c ∈ C + : r c ( n ) ≤ h ( n )}
is Borel. Since the displayed exceptional probabilities are summable, the first Borel–Cantelli lemma gives μ ( B h ) = 1 \mu(B_h)=1 μ ( B h ) = 1 . Apply Theorem B to Q = B h \mathcal{Q}=B_h Q = B h in the specified local flag region. The resulting family has the required common output ideal for every c ∈ B h c\in B_h c ∈ B h . In particular, the exceptional null set C + ∖ B h \mathcal{C}_{+}\setminus B_h C + ∖ B h is independent of S S S . □ \square □
Wick composition with a fixed random operator
Throughout this section, the underlying scalar field is real. We use the joint Gaussian space generated by the sequence g = ( g i ) g=(g_i) g = ( g i ) and the independent array G = ( γ i j ) G=(\gamma_{ij}) G = ( γ ij ) , together with the Wick composition defined in the introduction. We first establish the continuity and differentiation statements needed for the operator identities.
For a real separable Hilbert space H H H and an integer m ≥ 0 m \ge0 m ≥ 0 , let
H ≤ m ( H ) = ⨁ q = 0 m J q L 2 ( Ω ; H ) \mathcal{H}_{\le m}(H)=\bigoplus_{q=0}^{m} \mathcal{J}_q L^2(\Omega;H) H ≤ m ( H ) = q = 0 ⨁ m J q L 2 ( Ω ; H )
be the closed subspace of H H H -valued random variables whose chaos expansions have degree at most m m m . Gaussian hypercontractivity [7 , 11 ] gives, for each 2 ≤ t < ∞ 2 \le t < \infty 2 ≤ t < ∞ , a constant C m , t C_{m,t} C m , t such that
∥ U ∥ L t ( Ω ; H ) ≤ C m , t ∥ U ∥ L 2 ( Ω ; H ) ( U ∈ H ≤ m ( H ) ) . \lVert U\rVert_{L^t(\Omega;H)} \le C_{m,t}\lVert U\rVert_{L^2(\Omega;H)}
\qquad(U \in\mathcal{H}_{\le m}(H)). ∥ U ∥ L t ( Ω ; H ) ≤ C m , t ∥ U ∥ L 2 ( Ω ; H ) ( U ∈ H ≤ m ( H )) .
The Hilbert-valued version follows from the scalar inequality by expanding in an orthonormal basis of H H H and applying Minkowski’s inequality to the sum of the squared coordinates. In particular, if X ∈ H ≤ m ( S 2 ) X \in\mathcal{H}_{\le m}(\mathcal{S}_2) X ∈ H ≤ m ( S 2 ) and Y ∈ H ≤ n ( S 2 ) Y \in\mathcal{H}_{\le n}(\mathcal{S}_2) Y ∈ H ≤ n ( S 2 ) , then
∥ ( J j X ) ( J k Y ) ∥ L 2 ( S 2 ) ≤ ∥ J j X ∥ L 4 ( S 2 ) ∥ J k Y ∥ L 4 ( S 2 ) ≤ C j , 4 C k , 4 ∥ J j X ∥ L 2 ( S 2 ) ∥ J k Y ∥ L 2 ( S 2 ) . \lVert(\mathcal{J}_jX)(\mathcal{J}_kY)\rVert_{L^2(\mathcal{S}_2)}
\le\lVert\mathcal{J}_jX\rVert_{L^4(\mathcal{S}_2)}\lVert\mathcal{J}_kY\rVert_{L^4(\mathcal{S}_2)}
\le C_{j,4}C_{k,4}\lVert\mathcal{J}_jX\rVert_{L^2(\mathcal{S}_2)}\lVert\mathcal{J}_kY\rVert_{L^2(\mathcal{S}_2)}. ∥( J j X ) ( J k Y ) ∥ L 2 ( S 2 ) ≤ ∥ J j X ∥ L 4 ( S 2 ) ∥ J k Y ∥ L 4 ( S 2 ) ≤ C j , 4 C k , 4 ∥ J j X ∥ L 2 ( S 2 ) ∥ J k Y ∥ L 2 ( S 2 ) .
Here we used ∥ A B ∥ 2 ≤ ∥ A ∥ 2 ∥ B ∥ 2 \lVert AB\rVert_2 \le\lVert A\rVert_2\lVert B\rVert_2 ∥ A B ∥ 2 ≤ ∥ A ∥ 2 ∥ B ∥ 2 . Each chaos projection is contractive in L 2 ( S 2 ) L^2(\mathcal{S}_2) L 2 ( S 2 ) , so the finite sum defining X ⋄ o p Y X \mathbin{\diamond_{\mathrm{op}}} Y X ⋄ op Y is a bounded bilinear map
H ≤ m ( S 2 ) × H ≤ n ( S 2 ) ⟶ H ≤ m + n ( S 2 ) . \mathcal{H}_{\le m}(\mathcal{S}_2)\times\mathcal{H}_{\le n}(\mathcal{S}_2)
\longrightarrow\mathcal{H}_{\le m+n}(\mathcal{S}_2). H ≤ m ( S 2 ) × H ≤ n ( S 2 ) ⟶ H ≤ m + n ( S 2 ) .
Consequently, L 2 L^2 L 2 approximation within fixed chaos degrees is compatible with Wick composition.
For a polynomial in finitely many coordinates of g g g and G G G , write D g Z = ( ∂ i Z ) i D_gZ=(\partial_iZ)_i D g Z = ( ∂ i Z ) i and D g 2 Z = ( ∂ i ∂ j Z ) i j D_g^2Z=(\partial_i\partial_jZ)_{ij} D g 2 Z = ( ∂ i ∂ j Z ) ij ; these derivatives act only on the coordinates of g g g . These are the corresponding coordinate derivatives on Gaussian space; see [13 Chapter 1] . The Hermite expansion gives
E ∥ D g Z ∥ ℓ 2 2 ≤ m E ∣ Z ∣ 2 , E ∥ D g 2 Z ∥ S 2 2 ≤ m ( m − 1 ) E ∣ Z ∣ 2 ( Z ∈ H ≤ m ( R ) ) . \mathbb{E}\lVert D_gZ\rVert_{\ell^2}^2 \le m\mathbb{E}|Z|^2,
\qquad
\mathbb{E}\lVert D_g^2Z\rVert_{\mathcal{S}_2}^2 \le m(m-1)\mathbb{E}|Z|^2
\qquad(Z \in\mathcal{H}_{\le m}(\mathbb{R})). E ∥ D g Z ∥ ℓ 2 2 ≤ m E ∣ Z ∣ 2 , E ∥ D g 2 Z ∥ S 2 2 ≤ m ( m − 1 ) E ∣ Z ∣ 2 ( Z ∈ H ≤ m ( R )) .
where the second derivative is zero when m ≤ 1 m \le1 m ≤ 1 . These inequalities extend the derivatives from finite-coordinate polynomials to bounded linear maps on every fixed finite sum of chaoses, with values in L 2 ( Ω ; ℓ 2 ) L^2(\Omega;\ell^2) L 2 ( Ω ; ℓ 2 ) and L 2 ( Ω ; S 2 ) L^2(\Omega;\mathcal{S}_2) L 2 ( Ω ; S 2 ) , respectively. They also show that convergence in L 2 L^2 L 2 within a fixed chaos degree implies convergence of these derivatives in the indicated spaces.
Let C i Z = g i ⋄ Z C_iZ=g_i\diamond Z C i Z = g i ⋄ Z denote creation by the coordinate g i g_i g i . The Hermite formula gives
C i Z = g i Z − ∂ i Z , ∥ C i C j Z ∥ L 2 2 ≤ ( m + 1 ) ( m + 2 ) ∥ Z ∥ L 2 2 ( Z ∈ H ≤ m ( R ) ) . C_iZ=g_iZ-\partial_iZ,
\qquad
\lVert C_iC_jZ\rVert_{L^2}^2\le(m+1)(m+2)\lVert Z\rVert_{L^2}^2
\qquad(Z\in\mathcal{H}_{\le m}(\mathbb{R})). C i Z = g i Z − ∂ i Z , ∥ C i C j Z ∥ L 2 2 ≤ ( m + 1 ) ( m + 2 ) ∥ Z ∥ L 2 2 ( Z ∈ H ≤ m ( R )) .
The multiplication term belongs to L 2 L^2 L 2 by hypercontractivity, so the creation identity extends by approximation to the whole space H ≤ m ( R ) \mathcal{H}_{\le m}(\mathbb{R}) H ≤ m ( R ) . For r , s ∈ ℓ 2 r,s\in\ell^2 r , s ∈ ℓ 2 , define the Hilbert–Schmidt valued random operator
T r , s [ Z ] = ( r i s j C i C j Z ) i j . ∥ T r , s [ Z ] ∥ L 2 ( S 2 ) 2 ≤ ( m + 1 ) ( m + 2 ) ∥ r ∥ 2 2 ∥ s ∥ 2 2 E ∣ Z ∣ 2 . (34) T_{r,s}[Z]=(r_is_jC_iC_jZ)_{ij}.
\qquad
\lVert T_{r,s}[Z]\rVert_{L^2(\mathcal{S}_2)}^2
\le(m+1)(m+2)\lVert r\rVert_2^2\lVert s\rVert_2^2\mathbb{E}|Z|^2. \tag*{(34)} T r , s [ Z ] = ( r i s j C i C j Z ) ij . ∥ T r , s [ Z ] ∥ L 2 ( S 2 ) 2 ≤ ( m + 1 ) ( m + 2 ) ∥ r ∥ 2 2 ∥ s ∥ 2 2 E ∣ Z ∣ 2 . (34)
The estimate follows by summing the coordinate bounds and also proves continuity in r r r , s s s and Z Z Z on fixed chaos degrees. These facts permit infinitely many Gaussian coordinates throughout the argument. For the underlying Gaussian calculus, see [9 , 11 ] .
Lemma 6.1 (A scalar multiple of the identity with a compact perturbation). Let r , s r,s r , s be positive nonincreasing sequences tending to zero. If K K K is compact, F F F has finite rank and λ ≠ 0 \lambda\ne0 λ = 0 , then
I ( F + D r ( λ I + K ) D s ) = I ( r s ) . (35) \mathcal{I}(F+D_r(\lambda I+K)D_s)=\mathcal{I}(rs). \tag*{(35)} I ( F + D r ( λ I + K ) D s ) = I ( r s ) . (35)
Proof. Put A = D r ( λ I + K ) D s A=D_r(\lambda I+K)D_s A = D r ( λ I + K ) D s . The singular-value product inequality gives
s 2 n − 1 ( A ) ≤ ∥ λ I + K ∥ r n s n ( n ≥ 1 ) . s_{2n-1}(A)\le\lVert\lambda I+K\rVert r_ns_n
\qquad(n\ge1). s 2 n − 1 ( A ) ≤ ∥ λ I + K ∥ r n s n ( n ≥ 1 ) .
Let P m P_m P m project onto the first m m m coordinate vectors and set Q m = I − P m Q_m=I-P_m Q m = I − P m . Compactness implies ∥ Q m K Q m ∥ → 0 \lVert Q_mKQ_m\rVert\to0 ∥ Q m K Q m ∥ → 0 . Choose m m m so that ∥ Q m K Q m ∥ ≤ ∣ λ ∣ / 2 \lVert Q_mKQ_m\rVert\le|\lambda|/2 ∥ Q m K Q m ∥ ≤ ∣ λ ∣/2 . On the n n n -dimensional space spanned by e m + 1 , … , e m + n e_{m+1},\ldots,e_{m+n} e m + 1 , … , e m + n , the compression of λ I + K \lambda I+K λ I + K therefore has smallest singular value at least ∣ λ ∣ / 2 |\lambda|/2 ∣ λ ∣/2 . The diagonal factors leave this space invariant. Taking the smallest singular value of the compressed product gives
s n ( A ) ≥ ∣ λ ∣ 2 r m + n s m + n ( n ≥ 1 ) . s_n(A) \ge\frac{|\lambda|}{2}r_{m+n}s_{m+n}\qquad(n\ge1). s n ( A ) ≥ 2 ∣ λ ∣ r m + n s m + n ( n ≥ 1 ) .
In particular, A A A has infinite rank. The first estimate implies s ( A ) ≤ ∥ λ X + K ∥ Δ 2 ( r s ) s(A)\le\|\lambda X+K\|\Delta_2(rs) s ( A ) ≤ ∥ λ X + K ∥ Δ 2 ( r s ) . The second estimate gives the reverse principal-ideal inclusion: for N ≥ 2 m N\ge2m N ≥ 2 m , use the estimate with n = N − m n=N-m n = N − m and then s N − m ( A ) ≤ s ⌈ N / 2 ⌉ ( A ) s_{N-m}(A)\le s_{\lceil N/2\rceil}(A) s N − m ( A ) ≤ s ⌈ N /2 ⌉ ( A ) . The finitely many remaining indices can be absorbed into the multiplicative constant. Thus I ( A ) = I ( r s ) \mathcal{I}(A)=\mathcal{I}(rs) I ( A ) = I ( r s ) . A finite-rank perturbation of an infinite-rank compact operator generates the same principal ideal, as follows either from the singular-value inequalities with a fixed index shift or from the inclusion of all finite-rank operators in every nonzero principal ideal. This proves the assertion for A + F A+F A + F . □
Proposition 6.2 (Principal ideals before and after Wick composition). Fix a , b ∈ ℓ + a,b\in\ell^+ a , b ∈ ℓ + and a nonzero real Hilbert–Schmidt operator B B B . Set
X = W a , b = u a ⊗ u b − D a b , Y = X G B . X=W_{a,b}=u_a\otimes u_b-D_{ab},\qquad Y=XGB. X = W a , b = u a ⊗ u b − D ab , Y = X GB . Then X X X is almost surely trace class, Y Y Y belongs to the S 2 S_2 S 2 -valued third homogeneous chaos, and Y ⋄ o p E Y\mathbin{\diamond_{\mathrm{op}} E} Y ⋄ op E belongs to the S 2 S_2 S 2 -valued fifth homogeneous chaos. Almost surely,
I ( Y ) = I ( ( a b ) ⊙ s ( B ) ) , I ( Y ⋄ o p E ) = I ( a p ) . (36) \mathcal{I}(Y)=\mathcal{I}((ab)\mathbin{\odot}s(B)),\qquad\mathcal{I}(Y\mathbin{\diamond_{\mathrm{op}} E)=\mathcal{I}(ap).} \tag*{(36)} I ( Y ) = I (( ab ) ⊙ s ( B )) , I ( Y ⋄ op E ) = I ( a p ) . (36) The identities hold simultaneously for every prescribed countable family of triples ( a , b , B ) (a,b,B) ( a , b , B ) .
Proof. Since a , b ∈ ℓ 2 a,b\in\ell^2 a , b ∈ ℓ 2 , the vectors u a u_a u a and u b u_b u b belong to L 2 ( Ω ; ℓ 2 ) L^2(\Omega;\ell^2) L 2 ( Ω ; ℓ 2 ) , and a b ∈ ℓ 1 ab\in\ell^1 ab ∈ ℓ 1 by Cauchy–Schwarz. The formula for X X X therefore proves trace-class membership almost surely. It also gives I ( X ) = I ( a b ) \mathcal{I}(X)=\mathcal{I}(ab) I ( X ) = I ( ab ) , because D a b D_{ab} D ab has infinite rank and the other term has rank at most one. Entrywise, X i j = a i b j ( g i g j − δ i j ) X_{ij}=a_i b_j(g_i g_j-\delta_{ij}) X ij = a i b j ( g i g j − δ ij ) , whence
E ∥ X ∥ 2 2 = ∑ i , j a i 2 b j 2 ( 1 + δ i j ) = ∥ a ∥ 2 2 ∥ b ∥ 2 2 + ∥ a b ∥ 2 2 . \mathbb{E}\|X\|_2^2=\sum_{i,j}a_i^2b_j^2(1+\delta_{ij})=\|a\|_2^2\|b\|_2^2+\|ab\|_2^2. E ∥ X ∥ 2 2 = i , j ∑ a i 2 b j 2 ( 1 + δ ij ) = ∥ a ∥ 2 2 ∥ b ∥ 2 2 + ∥ ab ∥ 2 2 .
Finite-coordinate truncations converge to X X X in L 2 ( S 2 ) L^2(S_2) L 2 ( S 2 ) and show that X X X belongs to the second homogeneous chaos. Conditional on g g g , the factors X X X and B B B are fixed and the array G G G is independent. Corollary 4.1 therefore gives
E ∥ Y ∥ 2 2 = ∥ B ∥ 2 2 E ∥ X ∥ 2 2 , I ( Y ) = I ( a b ) ⊙ I ( B ) = I ( ( a b ) ⊙ s ( B ) ) almost surely . \mathbb{E}\|Y\|_2^2=\|B\|_2^2\mathbb{E}\|X\|_2^2,\qquad\mathcal{I}(Y)=\mathcal{I}(ab)\mathbin{\odot}\mathcal{I}(B)=\mathcal{I}((ab)\mathbin{\odot}s(B))\quad\text{almost surely}. E ∥ Y ∥ 2 2 = ∥ B ∥ 2 2 E ∥ X ∥ 2 2 , I ( Y ) = I ( ab ) ⊙ I ( B ) = I (( ab ) ⊙ s ( B )) almost surely .
The same conditional isometry shows that finite matrix truncations converge to Y Y Y in L 2 ( S 2 ) . L^2(S_2). L 2 ( S 2 ) . Every truncated term is a product of a second-chaos polynomial in g g g and an independent first-chaos coordinate of G G G , so Y Y Y belongs to the third homogeneous chaos.
Put A = B D p A=BD_p A = B D p and M = D b G A M=D_bGA M = D b G A . The Gaussian isometry gives M ∈ L 2 ( Ω ; S 2 ) M\in L^2(\Omega;S_2) M ∈ L 2 ( Ω ; S 2 ) and E ∥ M ∥ 2 2 = ∥ b ∥ 2 2 ∥ A ∥ 2 2 \mathbb{E}\|M\|_2^2=\|b\|_2^2\|A\|_2^2 E ∥ M ∥ 2 2 = ∥ b ∥ 2 2 ∥ A ∥ 2 2 ; moreover, M M M is independent of g g g . Define the scalar third-chaos random variable
Z = ∑ j , t , k ≥ 1 b j γ j t A t k : g j g k : = ∑ j , k ≥ 1 M j k : g j g k : , : g j g k : = g j g k − δ j k . (37) Z=\sum_{j,t,k\ge1}b_j\gamma_{jt}A_{tk}:g_jg_k:=\sum_{j,k\ge1}M_{jk}:g_jg_k:,\qquad:g_jg_k:=g_jg_k-\delta_{jk}. \tag*{(37)} Z = j , t , k ≥ 1 ∑ b j γ j t A t k : g j g k := j , k ≥ 1 ∑ M j k : g j g k : , : g j g k := g j g k − δ j k . (37)
Both series are understood as L 2 L^2 L 2 limits of finite-coordinate sums. To verify their convergence and identify their norms, let q t = ( A t k ) k q_t=(A_{tk})_k q t = ( A t k ) k be row t t t of A A A . Its Gaussian linear functional ⟨ q t , g ⟩ \langle q_t,g\rangle ⟨ q t , g ⟩ is defined by the convergent L 2 L^2 L 2 series. The identity
E ∣ g j ⟨ q t , g ⟩ − A t j ∣ 2 = ∥ q t ∥ 2 2 + A t j 2 \mathbb{E}|g_j\langle q_t,g\rangle-A_{tj}|^2=\|q_t\|_2^2+A_{tj}^2 E ∣ g j ⟨ q t , g ⟩ − A t j ∣ 2 = ∥ q t ∥ 2 2 + A t j 2
and orthogonality of the array coordinates give
E ∣ Z ∣ 2 = ∑ j , t b j 2 ( ∥ q t ∥ 2 2 + A t j 2 ) = ∥ b ∥ 2 2 ∥ B D p ∥ 2 2 + ∥ B D b p ∥ 2 2 < ∞ . (38) \mathbb{E}|Z|^2=\sum_{j,t}b_j^2\left(\|q_t\|_2^2+A_{tj}^2\right)=\|b\|_2^2\|BD_p\|_2^2+\|BD_{bp}\|_2^2<\infty. \tag*{(38)} E ∣ Z ∣ 2 = j , t ∑ b j 2 ( ∥ q t ∥ 2 2 + A t j 2 ) = ∥ b ∥ 2 2 ∥ B D p ∥ 2 2 + ∥ B D b p ∥ 2 2 < ∞. (38)
In particular, the sums over ( j , t ) (j,t) ( j , t ) converge unconditionally in L 2 L^2 L 2 , and truncation of the rows q t q_t q t gives the displayed three-index series. The quadratic-form representation follows as well from the bound
E g ∣ ∑ j , k N j k : g j g k : ∣ 2 = 2 ∥ sym N ∥ 2 2 ≤ 2 ∥ N ∥ 2 2 ( N ∈ S 2 ) , \mathbb{E}_g\left|\sum_{j,k}N_{jk}:g_jg_k:\right|^2=2\lVert\operatorname{sym}N\rVert_2^2\le2\lVert N\rVert_2^2\qquad(N\in\mathcal{S}_2), E g j , k ∑ N j k : g j g k : 2 = 2 ∥ sym N ∥ 2 2 ≤ 2 ∥ N ∥ 2 2 ( N ∈ S 2 ) ,
where sym N = ( N + N T ) / 2 \operatorname{sym}N=(N+N^{\mathsf{T}})/2 sym N = ( N + N T ) /2 . This representation requires no unregularized trace of M M M . We next prove P ( Z = 0 ) = 0 \mathbb{P}(Z=0)=0 P ( Z = 0 ) = 0 . Conditional on g g g , the coefficients of the Gaussian series in the array G G G are
α j t ( g ) = b j ( g j ⟨ q t , g ⟩ − A t j ) . \alpha_{jt}(g)=b_j\left(g_j\langle q_t,g\rangle-A_{tj}\right). α j t ( g ) = b j ( g j ⟨ q t , g ⟩ − A t j ) .
Equation (38) implies ∑ j , t ∣ α j t ( g ) ∣ 2 < ∞ \sum_{j,t}|\alpha_{jt}(g)|^2<\infty ∑ j , t ∣ α j t ( g ) ∣ 2 < ∞ for almost every g g g . For such g g g , the conditional law of Z Z Z is the centered Gaussian law with variance ∑ j , t ∣ α j t ( g ) ∣ 2 \sum_{j,t}|\alpha_{jt}(g)|^2 ∑ j , t ∣ α j t ( g ) ∣ 2 . Because B ≠ 0 B\ne0 B = 0 , choose t 0 , j 0 t_0,j_0 t 0 , j 0 with B t 0 j 0 ≠ 0 B_{t_0j_0}\ne0 B t 0 j 0 = 0 . Then A t 0 j 0 = B t 0 j 0 p j 0 ≠ 0 A_{t_0j_0}=B_{t_0j_0}p_{j_0}\ne0 A t 0 j 0 = B t 0 j 0 p j 0 = 0 , and
α j 0 t 0 ( g ) = b j 0 [ A t 0 j 0 ( g j 0 2 − 1 ) + g j 0 ∑ k ≠ j 0 A t 0 k g k ] . \alpha_{j_0t_0}(g)=b_{j_0}\left[A_{t_0j_0}(g_{j_0}^2-1)+g_{j_0}\sum_{k\ne j_0}A_{t_0k}g_k\right]. α j 0 t 0 ( g ) = b j 0 A t 0 j 0 ( g j 0 2 − 1 ) + g j 0 k = j 0 ∑ A t 0 k g k .
The sum over k ≠ j 0 k\ne j_0 k = j 0 is a Gaussian random variable measurable with respect to the other coordinates of g g g . Conditional on those coordinates, the displayed expression is a polynomial in g j 0 g_{j_0} g j 0 with nonzero deterministic quadratic coefficient. Its zero set is finite and hence has Gaussian probability zero. Thus α j 0 t 0 ≠ 0 \alpha_{j_0t_0}\ne0 α j 0 t 0 = 0 almost surely, the conditional variance of Z Z Z is positive almost surely, and P ( Z = 0 ) = 0 \mathbb{P}(Z=0)=0 P ( Z = 0 ) = 0 .
In finite coordinates, Wick associativity gives
( W a , b G B ) ∘ o p E = T a , p [ Z ] . (39) (W_{a,b}GB)\mathbin{\circ_{\mathrm{op}}}E=T_{a,p}[Z]. \tag*{(39)} ( W a , b GB ) ∘ op E = T a , p [ Z ] . (39)
Indeed, the ( i , ℓ ) (i,\ell) ( i , ℓ ) entry on the left is the sum over j , t , k j,t,k j , t , k of
a i p ℓ b j γ j t B t k p k : g i g j : : g k g ℓ : , a_i p_\ell b_j\gamma_{jt}B_{tk}p_k:g_ig_j::g_kg_\ell:, a i p ℓ b j γ j t B t k p k : g i g j :: g k g ℓ : ,
which equals a i p ℓ C i C ℓ Z a_i p_\ell C_iC_\ell Z a i p ℓ C i C ℓ Z . The factor γ j t \gamma_{jt} γ j t belongs to the independent array and has no contractions with the coordinates of g g g .
Here is the limiting argument for (39) . Let P N P_N P N be the first-N N N coordinate projection, set a ( N ) = P N a a^{(N)}=P_Na a ( N ) = P N a , b ( N ) = P N b b^{(N)}=P_Nb b ( N ) = P N b and p ( N ) = P N p p^{(N)}=P_Np p ( N ) = P N p , and replace B B B by B N = P N B P N B_N=P_NBP_N B N = P N B P N . Let X N X_N X N , Y N Y_N Y N , E N E_N E N and Z N Z_N Z N denote the corresponding finite-coordinate objects, with the array also truncated to its first N N N rows and columns. Write G N G_N G N for this finite array, extended by zero, and put A N = B N D p ( N ) A_N=B_ND_{p^{(N)}} A N = B N D p ( N ) and M N = D b ( N ) G N A N M_N=D_{b^{(N)}}G_NA_N M N = D b ( N ) G N A N . The Gaussian isometry gives M N → M M_N\to M M N → M in L 2 ( Ω ; S 2 ) L^2(\Omega;\mathcal{S}_2) L 2 ( Ω ; S 2 ) . The preceding isometries and energy estimates imply
X N → X , Y N → Y , E N → E in L 2 ( S 2 ) , Z N → Z in L 2 . X_N\to X,\qquad Y_N\to Y,\qquad E_N\to E\text{ in }L^2(\mathcal{S}_2),\qquad Z_N\to Z\text{ in }L^2. X N → X , Y N → Y , E N → E in L 2 ( S 2 ) , Z N → Z in L 2 .
All these approximations have the same respective chaos degrees. The established continuity of Wick composition therefore gives Y N ∘ o p E N → Y ∘ o p E Y_N\mathbin{\circ_{\mathrm{op}}}E_N\to Y\mathbin{\circ_{\mathrm{op}}}E Y N ∘ op E N → Y ∘ op E in L 2 ( S 2 ) L^2(\mathcal{S}_2) L 2 ( S 2 ) . Estimate (34) , together with a ( N ) → a a^{(N)}\to a a ( N ) → a and p ( N ) → p p^{(N)}\to p p ( N ) → p in ℓ 2 \ell^2 ℓ 2 , gives T a ( N ) , p ( N ) [ Z N ] → T a , p [ Z ] T_{a^{(N)},p^{(N)}}[Z_N]\to T_{a,p}[Z] T a ( N ) , p ( N ) [ Z N ] → T a , p [ Z ] in the same space. This proves (39) for the infinite-coordinate operators. It also shows that Y ∘ o p E Y\mathbin{\circ_{\mathrm{op}}}E Y ∘ op E belongs to the fifth homogeneous chaos.
Twice applying the creation formula in finite coordinates yields
C i C ℓ Z = g i g ℓ Z − δ i ℓ Z − g i ∂ ℓ Z − g ℓ ∂ i Z + ∂ i ∂ ℓ Z . C_iC_\ell Z=g_ig_\ell Z-\delta_{i\ell}Z-g_i\partial_\ell Z-g_\ell\partial_iZ+\partial_i\partial_\ell Z. C i C ℓ Z = g i g ℓ Z − δ i ℓ Z − g i ∂ ℓ Z − g ℓ ∂ i Z + ∂ i ∂ ℓ Z .
Consequently, the operator identity takes the form
T a , p [ Z ] = F Z + D a ( D g 2 Z − Z I ) D p , rank F Z ≤ 2 , (40) T_{a,p}[Z]=F_Z+D_a(D_g^2Z-ZI)D_p,\qquad\operatorname{rank}F_Z\le2, \tag*{(40)} T a , p [ Z ] = F Z + D a ( D g 2 Z − Z I ) D p , rank F Z ≤ 2 , (40)
where
F Z = u a ⊗ ( Z u p − D p D g Z ) − ( D a D g Z ) ⊗ u p . F_Z=u_a\otimes(Zu_p-D_pD_gZ)-(D_aD_gZ)\otimes u_p. F Z = u a ⊗ ( Z u p − D p D g Z ) − ( D a D g Z ) ⊗ u p .
All vectors in this expression are well defined almost surely. In fact, the derivative estimates give D g Z ∈ L 2 ( Ω ; ℓ 2 ) D_gZ\in L^2(\Omega;\ell^2) D g Z ∈ L 2 ( Ω ; ℓ 2 ) , and the quadratic-form representation gives
D g 2 Z = M + M T ∈ L 2 ( Ω ; S 2 ) . D_g^2Z=M+M^{\mathsf{T}}\in L^2(\Omega;\mathcal{S}_2). D g 2 Z = M + M T ∈ L 2 ( Ω ; S 2 ) .
For the latter identity, differentiate the finite-coordinate quadratic forms and then use the continuity of D g 2 D_g^2 D g 2 and the convergence of M N M_N M N to M M M in L 2 ( S 2 ) L^2(\mathcal{S}_2) L 2 ( S 2 ) . The same derivative continuity and Hilbert-valued hypercontractivity show that the vectors and their finite-coordinate approximations converge in every fixed finite L t L^t L t norm. Hölder’s inequality therefore passes each outer product defining F Z F_Z F Z to the limit in L 2 ( S 2 ) L^2(\mathcal{S}_2) L 2 ( S 2 ) . The remaining terms in (40) converge there directly. Thus the normal form holds in L 2 ( Ω ; S 2 ) L^2(\Omega;\mathcal{S}_2) L 2 ( Ω ; S 2 ) and therefore almost surely as an identity of Hilbert–Schmidt operators.
On a probability-one event, Z ≠ 0 Z\ne0 Z = 0 , the operator D g 2 Z D_g^2Z D g 2 Z is Hilbert–Schmidt and hence compact, and F Z F_Z F Z has rank at most two. Lemma 6.1, applied on this event with scalar − Z -Z − Z , proves the second identity in (36) . Intersecting the resulting events over a prescribed countable family of triples proves the final assertion. □
Proof of Theorem C. Proposition 6.2 with a = b = r a=b=r a = b = r gives the two operator-ideal identities and the asserted L 2 L^2 L 2 membership. We next verify the order equivalence. In the principal-ideal criterion, let L ∈ N L\in\mathbb{N} L ∈ N and put k = ⌈ n / L ⌉ k=\lceil n/L\rceil k = ⌈ n / L ⌉ . Since p k / p n = n / k ≤ L p_k/p_n=n/k\le L p k / p n = n / k ≤ L , an inequality p n r n ≤ C p k s k p_nr_n\le Cp_ks_k p n r n ≤ C p k s k implies r n ≤ C L s k r_n\le CLs_k r n ≤ C L s k . Conversely, r n ≤ C s k r_n\le Cs_k r n ≤ C s k implies p n r n ≤ C p k s k p_nr_n\le Cp_ks_k p n r n ≤ C p k s k , because p n ≤ p k p_n\le p_k p n ≤ p k . Squaring and taking square roots give the corresponding equivalence for the coordinatewise squares. Hence
I ( p r ) ⊆ I ( p s ) ⟺ I ( r ) ⊆ I ( s ) ⟺ I ( r 2 ) ⊆ I ( s 2 ) . \mathcal{I}(pr)\subseteq\mathcal{I}(ps)\Longleftrightarrow\mathcal{I}(r)\subseteq\mathcal{I}(s)\Longleftrightarrow\mathcal{I}(r^2)\subseteq\mathcal{I}(s^2). I ( p r ) ⊆ I ( p s ) ⟺ I ( r ) ⊆ I ( s ) ⟺ I ( r 2 ) ⊆ I ( s 2 ) .
Choose a common eventual upper bound for the profiles of Q \mathcal{Q} Q and apply Proposition 5.2 within U U U , retaining the prescribed prefix. The proposition provides a deterministic family a S a^S a S bounded above by a common ℓ 1 \ell^1 ℓ 1 sequence. Consequently, r S = a S r^S=\sqrt{a^S} r S = a S belongs to C + \mathcal{C}_+ C + for every S ⊆ N S\subseteq\mathbb{N} S ⊆ N . For each fixed pair ( S , B ) (S,B) ( S , B ) , Proposition 6.2 applies with a = b = r S a=b=r^S a = b = r S . If B = D c B=D_c B = D c with c ∈ Q c\in\mathcal{Q} c ∈ Q , the first ideal is
I ( a S ∘ c ) = J c , \mathcal{I}(a^S\mathbin{\circ}c)=J_c, I ( a S ∘ c ) = J c ,
which is independent of S S S , and the second is K S = I ( p a S ) K_S=\mathcal{I}(p\sqrt{a^S}) K S = I ( p a S ) .
The order equivalence proved above and the defining property of the family a S a^S a S give
K S ⊆ K T ⟺ I ( a S ) ⊆ I ( a T ) ⟺ S ⊆ ∗ T . K_S\subseteq K_T\Longleftrightarrow\mathcal{I}(a^S)\subseteq\mathcal{I}(a^T)\Longleftrightarrow S\subseteq^*T. K S ⊆ K T ⟺ I ( a S ) ⊆ I ( a T ) ⟺ S ⊆ ∗ T .
Square roots and multiplication by p p p preserve coordinatewise minima and maxima. The principal-ideal formulas for intersections and sums therefore yield the two lattice identities in the theorem. These relations concern deterministic ideals and hold for all S , T ⊆ N S,T\subseteq\mathbb{N} S , T ⊆ N . The probabilistic operator identities have been applied after fixing ( S , B ) (S,B) ( S , B ) ; intersecting their full-measure events over any prescribed countable set of such pairs gives exactly the simultaneous assertion in the theorem. □
Remark. The ideal after the prescribed right Wick composition depends on the left weight. Indeed, W b , a = W a , b ∗ W_{b,a}=W_{a,b}^* W b , a = W a , b ∗ , so these operators have identical singular-value sequences almost surely. For fixed 0 ≠ B ∈ S 2 0\ne B\in\mathcal{S}_2 0 = B ∈ S 2 , both W a , b G B W_{a,b}GB W a , b GB and W b , a G B W_{b,a}GB W b , a GB generate I ( ( a b ) ∘ s ( B ) ) \mathcal{I}((ab)\circ s(B)) I (( ab ) ∘ s ( B )) almost surely. By Proposition 6.2, their right Wick compositions with E E E generate I ( a p ) \mathcal{I}(ap) I ( a p ) and I ( b p ) \mathcal{I}(bp) I ( b p ) , respectively. These assertions hold on a common probability-one event by intersecting the two events in that proposition. Taking a n = n − 1 a_n=n^{-1} a n = n − 1 and b n = n − 2 b_n=n^{-2} b n = n − 2 gives the distinct ideals I ( ( n − 2 ) n ) \mathcal{I}((n^{-2})_n) I (( n − 2 ) n ) and I ( ( n − 3 ) n ) \mathcal{I}((n^{-3})_n) I (( n − 3 ) n ) . Thus the principal ideal after Wick composition is not determined by the singular-value sequence of the initial operator W a , b W_{a,b} W a , b alone.
If B B B has rank one, then M = D b G B D p M=D_bGBD_p M = D b GB D p has rank at most one. Hence rank ( D g 2 Z ) = rank ( M + M T ) ≤ 2 \operatorname{rank}(D_g^2Z)=\operatorname{rank}(M+M^{\mathsf T})\le2 rank ( D g 2 Z ) = rank ( M + M T ) ≤ 2 , and (40) gives ( W a , b G B ) ⋄ o p E = − Z D a p + R (W_{a,b}GB)\mathbin{\diamond_{\mathrm{op}}}E=-ZD_{ap}+R ( W a , b GB ) ⋄ op E = − Z D a p + R with rank R ≤ 4 R\le4 R ≤ 4 almost surely.
For regularly varying weights [1 Chapter 1] , the same normal form yields an exact singular-value asymptotic.
Corollary 6.3 (Singular-value asymptotics for regularly varying weights). Under the hypotheses of Proposition 6.2, suppose that a a a is regularly varying of index − α -\alpha − α :
a ⌈ t n ⌉ a n ⟶ t − α ( t > 0 ) . \frac{a_{\lceil tn\rceil}}{a_n}\longrightarrow t^{-\alpha}\qquad(t>0). a n a ⌈ t n ⌉ ⟶ t − α ( t > 0 ) .
For the scalar random variable Z Z Z defined in (37) , one then has
s n ( Y ⊡ E ) ∼ ∣ Z ∣ a n n almost surely. s_n(Y \boxdot E) \sim|Z|\frac{a_n}{n}\quad\text{almost surely.} s n ( Y ⊡ E ) ∼ ∣ Z ∣ n a n almost surely.
In particular, if a = b = r a=b=r a = b = r with r n = n − α r_n=n^{-\alpha} r n = n − α and α > 1 / 2 \alpha>1/2 α > 1/2 , then
n α + 1 s n ( Y r , B ⊡ E ) ⟶ ∣ Z ∣ almost surely. n^{\alpha+1}s_n(Y_{r,B}\boxdot E)\longrightarrow|Z|\quad\text{almost surely.} n α + 1 s n ( Y r , B ⊡ E ) ⟶ ∣ Z ∣ almost surely.
Proof. Fix an outcome for which Proposition 6.2 and the normal form (40) hold. Put d n = a n / n d_n=a_n/n d n = a n / n . This sequence is regularly varying of index − ( α + 1 ) -(\alpha+1) − ( α + 1 ) . Since d d d is nonincreasing, squeezing between two fixed nearby dilations shows that its defining ratio limit also holds for any integer indices whose ratio tends to a positive constant. By (40) ,
Y ⊡ E = − Z D d + R , R = F Z + D a K D p , K = D g 2 Z . Y\boxdot E=-ZD_d+R,\qquad R=F_Z+D_aKD_p,\qquad K=D_g^2Z. Y ⊡ E = − Z D d + R , R = F Z + D a K D p , K = D g 2 Z .
The operator K K K is compact. Applying the singular-value product inequality to all three factors gives
s 3 n − 2 ( D a K D p ) ≤ a n s n ( K ) p n = d n s n ( K ) . s_{3n-2}(D_aKD_p)\leq a_ns_n(K)p_n=d_ns_n(K). s 3 n − 2 ( D a K D p ) ≤ a n s n ( K ) p n = d n s n ( K ) .
For an arbitrary large index N N N , take n = ⌊ ( N + 2 ) / 3 ⌋ n=\lfloor(N+2)/3\rfloor n = ⌊( N + 2 ) /3 ⌋ . Then 3 n − 2 ≤ N 3n-2\leq N 3 n − 2 ≤ N , n / N → 1 / 3 n/N\to1/3 n / N → 1/3 , and regular variation gives d n / d N → 3 α + 1 d_n/d_N\to3^{\alpha+1} d n / d N → 3 α + 1 . Since s n ( K ) → 0 s_n(K)\to0 s n ( K ) → 0 , it follows that s N ( D a K D p ) = o ( d N ) s_N(D_aKD_p)=o(d_N) s N ( D a K D p ) = o ( d N ) . If q = rank F Z q=\operatorname{rank}F_Z q = rank F Z , the finite-rank perturbation inequality yields
s N ( R ) ≤ s N − q ( D a K D p ) ( N > q ) . s_N(R)\leq s_{N-q}(D_aKD_p)\qquad(N>q). s N ( R ) ≤ s N − q ( D a K D p ) ( N > q ) .
Regular variation also implies d N − q / d N → 1 d_{N-q}/d_N\to1 d N − q / d N → 1 , so s N ( R ) = o ( d N ) s_N(R)=o(d_N) s N ( R ) = o ( d N ) .
Fix 0 < δ < 1 0<\delta<1 0 < δ < 1 and set k = ⌊ δ n ⌋ k=\lfloor\delta n\rfloor k = ⌊ δ n ⌋ for all sufficiently large n n n . The singular-value inequalities for sums, applied both to Y ⊡ E = − Z D d + R Y\boxdot E=-ZD_d+R Y ⊡ E = − Z D d + R and to − Z D d = Y ⊡ E − R -ZD_d=Y\boxdot E-R − Z D d = Y ⊡ E − R , give
∣ Z ∣ d n + k − 1 − s k ( R ) ≤ s n ( Y ⊡ E ) ≤ ∣ Z ∣ d n − k + 1 + s k ( R ) . |Z|d_{n+k-1}-s_k(R)\leq s_n(Y\boxdot E)\leq|Z|d_{n-k+1}+s_k(R). ∣ Z ∣ d n + k − 1 − s k ( R ) ≤ s n ( Y ⊡ E ) ≤ ∣ Z ∣ d n − k + 1 + s k ( R ) .
Here
s k ( R ) d n = s k ( R ) d k d k d n ⟶ 0 , d n ± k ∓ 1 d n ⟶ ( 1 ± δ ) − ( α + 1 ) . \frac{s_k(R)}{d_n}=\frac{s_k(R)}{d_k}\frac{d_k}{d_n}\longrightarrow0,\qquad
\frac{d_{n\pm k\mp1}}{d_n}\longrightarrow(1\pm\delta)^{-(\alpha+1)}. d n s k ( R ) = d k s k ( R ) d n d k ⟶ 0 , d n d n ± k ∓ 1 ⟶ ( 1 ± δ ) − ( α + 1 ) .
Dividing the preceding bounds by d n d_n d n , passing to the lower and upper limits, and then letting δ ↓ 0 \delta\downarrow0 δ ↓ 0 proves the asserted asymptotic equivalence. The power-law assertion follows by substitution.
Appendix A. Finite Gaussian matrix estimates
The estimates used in the proof of Corollary 4.1 follow from elementary covering and Laplace-transform arguments.
Lemma A.1 (Finite Gaussian matrix estimates). Let G = ( g i j ) G=(g_{ij}) G = ( g ij ) be an array of independent standard real or circular complex Gaussian variables, normalized by E ∣ g i j ∣ 2 = 1 \mathbb{E}|g_{ij}|^2=1 E ∣ g ij ∣ 2 = 1 . Put ν = 1 \nu=1 ν = 1 in the real case and ν = 2 \nu=2 ν = 2 in the complex case. For all integers 1 ≤ q ≤ N 1\leq q\leq N 1 ≤ q ≤ N ,
P { ∥ G [ N , q ] ∥ > 8 N } ≤ 2 exp { − ν ( 8 − 2 log 9 ) N } ≤ 2 e − 3 ν N . (41) \mathbb{P}\{\lVert G[N,q]\rVert>8\sqrt{N}\}\leq2\exp\{-\nu(8-2\log9)N\}\leq2e^{-3\nu N}. \tag*{(41)} P {∥ G [ N , q ]∥ > 8 N } ≤ 2 exp { − ν ( 8 − 2 log 9 ) N } ≤ 2 e − 3 ν N . (41) For all positive integers m , k m,k m , k ,
P { ∥ G [ m , k ] ∥ 2 2 < m k / 2 } ≤ exp { − ν ( ( log 2 ) / 2 − 1 / 4 ) m k } . (42) \mathbb{P}\{\lVert G[m,k]\rVert_2^2<mk/2\}\leq\exp\{-\nu((\log2)/2-1/4)mk\}. \tag*{(42)} P {∥ G [ m , k ] ∥ 2 2 < mk /2 } ≤ exp { − ν (( log 2 ) /2 − 1/4 ) mk } . (42) For all integers 1 ≤ q ≤ N 1\leq q\leq N 1 ≤ q ≤ N and 0 < t < 1 0<t<1 0 < t < 1 ,
P { s q ( G [ N , q ] ) < t N } ≤ 2 e − 3 ν N + ( 17 / t ) ν q ( 2 e t ) ν N . (43) \mathbb{P}\{s_q(G[N,q])<t\sqrt{N}\}\leq2e^{-3\nu N}+(17/t)^{\nu q}(2\sqrt{e}\,t)^{\nu N}. \tag*{(43)} P { s q ( G [ N , q ]) < t N } ≤ 2 e − 3 ν N + ( 17/ t ) ν q ( 2 e t ) ν N . (43) Proof. An ε \varepsilon ε -net of the unit sphere of F k \mathbb{F}^k F k has at most ( 1 + 2 / ε ) ν k (1+2/\varepsilon)^{\nu k} ( 1 + 2/ ε ) ν k points, by disjoint-ball volume comparison in real dimension ν k \nu k ν k . For fixed unit vectors x ∈ F q x\in\mathbb{F}^q x ∈ F q and y ∈ F N y\in\mathbb{F}^N y ∈ F N , ⟨ G [ N , q ] x , y ⟩ \langle G[N,q]x,y\rangle ⟨ G [ N , q ] x , y ⟩ is standard Gaussian over F \mathbb{F} F , and
P { ∣ ⟨ G [ N , q ] x , y ⟩ ∣ > u } ≤ 2 e − ν u 2 / 2 . \mathbb{P}\{|\langle G[N,q]x,y\rangle|>u\}\leq2e^{-\nu u^2/2}. P { ∣ ⟨ G [ N , q ] x , y ⟩ ∣ > u } ≤ 2 e − ν u 2 /2 . Two 1 / 4 1/4 1/4 -nets give ∥ G [ N , q ] ∥ ≤ 2 max ∣ ⟨ G [ N , q ] x , y ⟩ ∣ \lVert G[N,q]\rVert\leq2\max|\langle G[N,q]x,y\rangle| ∥ G [ N , q ]∥ ≤ 2 max ∣ ⟨ G [ N , q ] x , y ⟩ ∣ , where the maximum is over the two nets. Their product has at most 9 ν ( N + q ) ≤ 9 2 ν N 9^{\nu(N+q)}\leq9^{2\nu N} 9 ν ( N + q ) ≤ 9 2 ν N elements. Applying the scalar bound at u = 4 N u=4\sqrt{N} u = 4 N and taking the union bound proves (A.1).
The identity
E e − s ∣ g 11 ∣ 2 = ( 1 + 2 s ν ) − ν / 2 ( s > 0 ) \mathbb{E}e^{-s|g_{11}|^2}=\left(1+\frac{2s}{\nu}\right)^{-\nu/2}\qquad(s>0) E e − s ∣ g 11 ∣ 2 = ( 1 + ν 2 s ) − ν /2 ( s > 0 )
provides both remaining estimates. Independence of the entries and Markov’s inequality with s = ν / 2 s=\nu/2 s = ν /2 give (A.2). For a fixed unit vector x ∈ F q x\in\mathbb{F}^{q} x ∈ F q , the entries of G [ N , q ] x G[N,q]x G [ N , q ] x are independent standard Gaussian variables. Applying the same Laplace transform with s = ν / ( 2 u 2 ) s=\nu/(2u^2) s = ν / ( 2 u 2 ) gives
P { ∥ G [ N , q ] x ∥ ≤ u N } ≤ ( e u ) ν N ( u > 0 ) . \mathbb{P}\{\lVert G[N,q]x\rVert\leq u\sqrt{N}\}\leq(\sqrt{e}\,u)^{\nu N}\qquad(u>0). P {∥ G [ N , q ] x ∥ ≤ u N } ≤ ( e u ) ν N ( u > 0 ) .
For 0 < t < 1 0<t<1 0 < t < 1 , take a ( t / 8 ) (t/8) ( t /8 ) -net of the input sphere, of cardinality at most ( 1 + 16 / t ) ν q ≤ ( 17 / t ) ν q (1+16/t)^{\nu q}\leq(17/t)^{\nu q} ( 1 + 16/ t ) ν q ≤ ( 17/ t ) ν q . On the event { ∥ G [ N , q ] ∥ ≤ 8 N } \{\lVert G[N,q]\rVert\leq8\sqrt{N}\} {∥ G [ N , q ]∥ ≤ 8 N } , a unit vector with image norm below t N t\sqrt{N} t N has a net point with image norm below 2 t N 2t\sqrt{N} 2 t N . Applying the last small-ball bound at u = 2 t u=2t u = 2 t , taking the union bound over this net, and adding the exceptional probability from (A.1) proves (A.3). □ \square □
Declaration of AI Use
The author used GPT-6.0 (OpenAI) interactively in developing and refining the mathematical arguments and the presentation of this paper. The author takes full responsibility for the correctness and mathematical content of the manuscript.
School of Mathematical Sciences, University of Science and Technology of China, Hefei, Anhui 230026, China
Email address: zhaoguangqian@mail.ustc.edu.cn