Coherence keeps the modes of the carrier and changes how they evolve. The choice of symbol controls the phase rates, the resonances and the stability of the resulting budget.
6. The Coherent / Reversible Face
The coherent face uses a bounded real symbol h of K.
6.1 Admissible coherence symbols
Definition 6.1.Admissible coherence symbol.
An admissible coherence symbol is a Borel h:[0,∞)→R with: h real-valued; h bounded; h(0)=0; h nondecreasing; if K=0, the coherent response is required to be nontrivial, h(K)=0. The zero carrier is admitted as a degenerate case.
Example 6.2.Default symbol.
For λ0>0, hλ0(λ)=π2arctan(λ/λ0) is real, bounded by 1, satisfies h(0)=0, and is strictly increasing.
6.2 The coherence generator and unitary flow
Definition 6.3.Coherence generator.
C:=h(K). Bounded selfadjoint with ∥C∥≤∥h∥∞ and C∈W∗(K).
Definition 6.4.Reversible flow.
Ut:=e−itC=e−ith(K), t∈R. A strongly continuous one-parameter unitary group.
Proposition 6.5.Mode preservation.
If Kej=λjej, then Cej=h(λj)ej and Utej=e−ith(λj)ej. Every spectral subspace of K is Ut-invariant.
6.3 The reversible Fejér compression
Definition 6.6.Coherence Fejér operator.
ΓT(C):=(1−e−iTC)/(iTC)=T1∫0Te−itCdt, identity on kerC.
Definition 6.7.Coherent compression.
CT:=ΓT∗ΓT=T2C22(1−cos(TC)), with eigenvalues sinc2(Tμ/2) on (kerC)⊥. Here sincx:=sinx/x, with sinc0:=1.
6.4 The coherence budget
Definition 6.8.Coherence compression budget: full convention.
Let
qT(μ):=sinc2(Tμ/2),sincx:=xsinx,sinc0:=1.
For C=h(K), define the resonant zero-mode subspace
RT(C):=EC({μ∈σ(C)∖{0}:2Tμ∈πZ∖{0}})H.
Equivalently, RT(C) is the part of (kerC)⊥ annihilated by CT=ΓT(C)∗ΓT(C).
Under (D3), the full coherence budget convention used in this paper is
Thus any nonzero coherence mode annihilated by CT produces infinite full compression loss. If (kerC)⊥={0}, we set BT(C)=0.
In finite rank, with nonzero coherence eigenvalues μj, this reads
BT(C)=⎩⎨⎧−j∑logsinc2(Tμj/2),+∞,sinc(Tμj/2)=0 for every j,sinc(Tμj/2)=0 for at least one j.
When no confusion is possible we write BT(K) for BT(h(K)).
Remark 6.9.Full versus reduced resonant convention.
The coherent Fejér compression differs from the dissipative Fejér transmission in one essential respect. The dissipative eigenvalues
fT(λ)=Tλ1−e−Tλ
are strictly positive for λ>0, whereas the coherent compression eigenvalues
qT(μ)=sinc2(Tμ/2)
vanish at the resonant values Tμ/2∈πZ∖{0}. There are therefore two distinct conventions.
Full determinant convention. The convention used in the paper unless explicitly stated otherwise. The kernel of C is ignored because ΓT(C) is the identity there and contributes no loss. Any resonant zero mode inside (kerC)⊥ is not ignored: it makes the full budget equal to +∞.
Reduced resonant pseudodeterminant convention. One may instead remove both kerC and the resonant zero modes:
HT,red:=(kerC⊕RT(C))⊥,
and define
BT,red(C):=−logdet∗(CTHT,red),
when the determinant on the reduced subspace is defined. In finite rank this is the product over the remaining strictly positive compression eigenvalues only:
On nonresonant scale windows, RT(C)={0}, so the full and reduced conventions agree up to the harmless omission of kerC, where the compression is the identity. At resonant scales they do not agree: the reduced budget is a pseudodeterminant after deleting total-loss modes, and is therefore not the full compression-loss budget.
Proposition 6.10.Properties of BT.
For T>0 the full budget satisfies BT(K)≥0, with equality if and only if C=0 on the retained space. The compression CT is a Borel function of K, and BT(UKU∗)=BT(K) whenever the determinant convention is transported by U. Hence the budget is basis-independent.
Proof. For real μ, 0≤qT(μ)≤1, with qT(μ)=1 exactly when μ=0, since ∣sinx∣<∣x∣ for x=0. Each nonzero retained mode therefore contributes a strictly positive loss, or infinite loss at resonance. In the Fredholm case, the positive compact operator I−CT has a spectral product with the same conclusion: zero loss forces every factor to be one. Functional calculus and unitary invariance of the determinant prove the remaining assertions. ◻
6.5 Symbol-class stability
Proposition 6.11.Borel-symbol stability.
Let K≥0 be selfadjoint, and let h1,h2:[0,∞)→R be bounded Borel symbols. Then
∥h1(K)−h2(K)∥≤λ∈σ(K)sup∣h1(λ)−h2(λ)∣.
For a spectral window [0,R],
(h1(K)−h2(K))EK([0,R])≤∥h1−h2∥∞,[0,R],
where the right-hand side denotes the supremum over [0,R]∩σ(K), or over [0,R] if the symbols are declared on the whole interval.
Proof. Let φ=h1−h2. By the bounded Borel functional calculus, h1(K)−h2(K)=φ(K) and ∥φ(K)∥≤∥φ∥L∞(σ(K)). Moreover, (h1(K)−h2(K))EK([0,R])=(φ1[0,R])(K), so the same functional-calculus norm bound gives the windowed estimate. ◻
Let PR:=EK([0,R]) have finite rank r<∞. Fix T>0 and define qT(u):=sinc2(Tu/2). For a real bounded Borel symbol h, define the finite-window coherence budget
BT,R(h):=−λ∈σ(K∣PRH)∑m(λ)logqT(h(λ)),
whenever all terms are positive. Here eigenvalues are counted with multiplicity, and modes with h(λ)=0 contribute −logqT(0)=0.
Assume there is δ>0 such that qT(hℓ(λ))∈[δ,1] for every λ∈σ(K∣PRH) and for ℓ=1,2. Then both finite-window budgets are full-convention nonresonant budgets on PRH, and
Proof. For each eigenvalue λ in the finite window, put aℓ(λ):=qT(hℓ(λ)) for ℓ=1,2. By hypothesis aℓ(λ)∈[δ,1]. Since x↦logx has derivative bounded by δ−1 on [δ,1],
∣loga1(λ)−loga2(λ)∣≤δ−1∣a1(λ)−a2(λ)∣.
The function qT is smooth, hence Lipschitz on [−M,M] with Lipschitz constant LT,M. Therefore
In the setting of Lemma 6.12, suppose h satisfies qT(h(λ))≥δ for every λ∈σ(K∣PRH), and let k be another real bounded Borel symbol with ε:=∥h−k∥∞,[0,R]≤1. Put Mh:=∥h∥∞,[0,R] and LT,Mh+1:=sup∣u∣≤Mh+1∣qT′(u)∣. If LT,Mh+1ε≤δ/2, then
qT(k(λ))≥2δon σ(K∣PRH),
and
BT,R(k)−BT,R(h)≤δ2rLT,Mh+1ε.
Remark 6.14.Scope of the stability statement.
Proposition 6.11 is a general bounded-functional-calculus estimate. Lemma 6.12 and Corollary 6.13 are deliberately finite-rank, nonresonant determinant statements. They do not assert an infinite-dimensional determinant-continuity theorem without additional Fredholm, heat-kernel, zeta, or finite-volume hypotheses.
6.6 Properties shared by the two faces
Lemma 6.15.
Cdom(K)⊆dom(K) and KC=CK on dom(K); likewise KUt=UtK on dom(K), [ΛT(K),Us]=0 for all s,t,T. Both {e−tK}t≥0 and {Us}s∈R lie in W∗(K) and commute with each other.
Proof. All bounded members are Borel functions of K; their commutation with K includes preservation of its operator domain. The algebra W∗(K) is abelian. ◻
The carrier is canonical given its closed form and Hilbert metric. The coherent face is uniquely determined once the specific symbol h is fixed; the admissible symbol class permits multiple choices.