AN-04 proof edition · CC0; exact source credit and separately licensed prerequisites
Resolved corner kernels and their exact inverse
These connected components retain AN03-U032, Totally characteristic operators on the half space, Sections 2–5, Section 6 through Example 6.7, and Sections 7–13. Original author: Claude Opus 5.5 (Anthropic), September 2026; editorial additions: Codex, September 2026. Both were dedicated to the public domain (CC0). Current prerequisite connections and proof clarifications: AN-04 course-writing task and OpenAI Codex, 5 October 2026, also CC0. The selected components retain every mathematical display, the full scalar and finite-matrix hypotheses, and all five original solved exercises.
The approved mathematical antecedent is Hörmander III, 2007 eBook, ISBN 978-3-540-49938-1, Section 18.3. Its use and ordinary citation are valid. Complete proofs are supplied in the components and the exact earlier programme proofs. The earlier linked components retain their individual licences.
The four components, in proof order, are Boundary tests, lacunary symbols and all normal jets, Resolved corner kernels and their exact inverse, Boundary adjoints, complete composition and distributional action, Boundary operator bounds, conormal action and the residual obstruction. Original section and equation numbers are retained across them. Sections 6.8 (polyhomogeneous corner characterization) and 14 (arbitrary positive-order Sobolev loss) are separate unadopted obligations; the theorems below do not substitute for those results or for the global compressed wave-front calculus.
The symbol class, Fourier-kernel construction, conventions and exact prerequisites are in Sections 1–5. This component proves all uniform estimates at the resolved corner, their converse and the finite negative-order Schur kernel bound.
6. Kernels near the corner
Coordinates at the corner
Kernels of totally characteristic operators live on
Q={(x,y)∈R2n:xn≥0, yn≥0},∂2Q={(x,y):xn=yn=0},
the quarter space and its distinguished boundary. Near ∂2Q we use
t=2xn+yn,r=txn−yn(t>0);xn=t(1+2r),yn=t(1−2r).(6.1)
Proposition 6.1 (Blow-up coordinates). Write Φ(t,r)=(t(1+r/2),t(1−r/2)).
- Φ maps (0,∞)×R diffeomorphically onto {xn+yn>0}, with ∣detΦ′∣=t; hence dxndyn=tdtdr.
- Q∖{xn=yn=0} corresponds to t>0, ∣r∣≤2. The face {xn=0<yn} is r=−2, the face {yn=0<xn} is r=2, the diagonal xn=yn is r=0, and yn/xn=(2−r)/(2+r).
- Φ extends smoothly to [0,∞)×R and maps the whole line t=0 to the corner: the corner is blown up into the front face t=0, of which the segment ∣r∣≤2 lies over Q.
- Normal dilations (xn,yn)↦λ(xn,yn) are (t,r)↦(λt,r), and the radial field is xn∂xn+yn∂yn=t∂t.
- On Q, ∣w∣/2≤t≤∣w∣ for w=(xn,yn); r is homogeneous of degree 0, so ∣∂wβr∣≤Cβ∣w∣−∣β∣ on {xn+yn>0}∩{∣r∣≤3}.
- The rescaled normal variable in (4.4) is a function of r alone: (xn−yn)/xn=2r/(2+r).
Proof. The Jacobian matrix of Φ has rows (1+2r,2t) and (1−2r,−2t), with determinant −t; the inverse is (6.1). Parts 2–4 and 6 are direct substitutions. For part 5, xn+yn≥∣w∣ when both are nonnegative, and xn+yn≤2∣w∣. The set {∣r∣≤3}={∣xn−yn∣≤23(xn+yn)} is a closed cone that meets the line xn+yn=0 only at the origin; its intersection with the unit circle is a compact subset of the open set where r is smooth. The derivatives of order k of r are homogeneous of degree −k, so they are bounded by Ck∣w∣−k on that cone. □
The point of these coordinates is part 6. The kernel formula (4.4) involves xn−1 and a function of (xn−yn)/xn, and neither is smooth at the corner. But tK becomes a smooth function of (t,r), as Theorem 6.2 shows.
Residual kernels
Theorem 6.2 (Residual kernels).
(a) Let a∈Sla−∞ and A(x,z)=(2π)−n∫eiz⋅ξa(x,ξ)dξ. Then A∈C∞(R+n×Rn), A=0 for zn≥1, and for all α,β,N
∣∂xα∂zβA(x,z)∣≤CαβN(1+∣z∣)−N(1+xn)−N,(6.2)
with CαβN bounded by seminorms of a. Conversely every A∈C∞(R+n×Rn) satisfying (6.2) and vanishing for zn>1 comes in this way from exactly one a∈Sla−∞, namely a(x,ξ)=∫e−iz⋅ξA(x,z)dz.
(b) For a∈Sla−∞ the kernel of Ta is the locally integrable function
K(x,y)=xn−1A(x,x′−y′,xnxn−yn)(xn>0),K(x,y)=0(xn<0).(6.3)
For xn>0 and u∈S, Tau(x)=∫K(x,y)u(y)dy, and ∫∣K(x,y)∣dy=∫∣A(x,z)∣dz≤C(1+xn)−N. Moreover suppK⊂Q and K∈C∞(R2n∖∂2Q).
(c) The function F(x′,y′,t,r)=tK(x′,t(1+2r),y′,t(1−2r)), t>0, extends to a C∞ function on {t≥0}×Rx′n−1×Ry′n−1×Rr, which vanishes for ∣r∣≥2. For all α,β,τ,ρ,ν and r>−2,
∣Dx′αDy′βDtτDrρF∣≤C(1+∣x′−y′∣+t)−ν(2+r)ν,(6.4)
and in particular
∣Dx′αDy′βDtτDrρF∣≤C′(1+∣x′−y′∣+t)−ν.(6.5)
On the front face,
F(x′,y′,0,r)=2+r2A(x′,0,x′−y′,2+r2r)(r>−2).(6.6)
(d) Conversely, let K∈Lloc1(R2n) with suppK⊂Q, and suppose that the function F of (c) agrees almost everywhere on t>0 with a function in C∞({t≥0}) that vanishes for ∣r∣≥2 and satisfies (6.5). Then K is, almost everywhere, the kernel of Ta for exactly one a∈Sla−∞.
Proof. (a) For a∈S+−∞, integration by parts gives
zγ∂zβ∂xαA=(2π)−n∫eiz⋅ξ(−Dξ)γ[(iξ)β∂xαa]dξ, with an integrand bounded by C(1+∣ξ∣)−n−1(1+xn)−N. This proves smoothness and (6.2). For fixed (x,ξ′), a(x,ξ′,⋅)∈S(R), so Fna is a continuous function; by (4.5) it vanishes for t<−1. Since A(x,z)=(2π)−n∫eiz′⋅ξ′(Fna)(x,ξ′,−zn)dξ′, A=0 for zn>1, and by continuity for zn≥1. Conversely, if A satisfies (6.2), then ξγ∂ξβ∂xαa=∫e−iz⋅ξDzγ[(−iz)β∂xαA]dz is bounded by C(1+xn)−N, so a∈S+−∞; Fourier inversion recovers A from a; and Fna(x,ξ′,t)=2π∫e−iz′⋅ξ′A(x,z′,−t)dz′ vanishes for t<−1. Uniqueness is Fourier inversion.
(b) For xn>0 fixed, a♭(x,⋅)∈S(Rn), because a decreases rapidly in (ξ′,xnξn). So K(x,y)=(2π)−n∫ei(x−y)⋅ξa♭(x,ξ)dξ converges absolutely, and the substitution ηn=xnξn gives (6.3). Fubini gives Tau(x)=∫K(x,y)u(y)dy. The change of variables z=(x′−y′,(xn−yn)/xn), dy=xndz, gives ∫∣K(x,y)∣dy=∫∣A(x,z)∣dz. Hence (Tau,v)=∬K(x,y)u(y)v(x)dydx for u,v∈S, with absolute convergence; so K, which is locally integrable, is the Schwartz kernel. If xn<0, K=0. If xn>0>yn, then (xn−yn)/xn>1 and A=0. So K vanishes outside Q, up to the null set xn=0.
Smoothness off ∂2Q. Near a point with xn>0, (6.3) is smooth. Near a point with xn<0, or with xn=0 and yn<0, K=0. Let xn0=0<yn0. For xn>0 small and yn near yn0, the normal argument zn=(xn−yn)/xn satisfies ∣zn∣≥yn0/(2xn). By the chain rule, a derivative of order ∣γ∣ of K is a finite sum of terms xn−kynl(∂A)(x,z) with k≤1+2∣γ∣, l≤∣γ∣, and ∣∂A∣≤CM(1+∣zn∣)−M≤CM(2xn/yn0)M. So K and all its derivatives tend to 0 as xn→0+, uniformly near the point. Since K=0 for xn≤0, K is smooth there.
(c) For t>0 and r>−2 we have xn=t(1+r/2)>0, (xn−yn)/xn=2r/(2+r) and t/xn=2/(2+r). So by (6.3)
F=2+r2A(x′,t(1+2r),x′−y′,2+r2r)=:G.(6.7)
The right side is smooth on {t≥0, r>−2}, because A is smooth up to xn=0. It vanishes for r≥2, where 2r/(2+r)≥1. For t>0, r<−2 we have xn<0 and F=0.
Now let −2<r<2 and write zn=2r/(2+r), z′=x′−y′, xn=t(2+r)/2. Then
2+r1≤21+∣zn∣,t=2+r2xn≤xn(1+∣zn∣).(6.8)
The first inequality holds because 1+∣zn∣≥1≥2/(2+r) when r≥0, and 1+∣zn∣=(2−r)/(2+r)≥2/(2+r) when r<0. By the chain rule, Dx′αDy′βDtτDrρG is a finite sum of terms ctk(2+r)−l(∂xμ∂zκA)(x′,xn,z′,zn) with 0≤k≤ρ and l≤1+2ρ: an r-derivative may hit (2+r)−l, the argument t(1+r/2) (factor t/2) or the argument 2r/(2+r) (factor 4/(2+r)2); a t-derivative brings the factor (2+r)/2. By (6.8) and (6.2), each term is at most
Cxnk(1+∣zn∣)k+l+(1+∣z∣)−M(1+xn)−M.
On the other hand 1+∣x′−y′∣+t≤(1+∣z′∣)(1+xn)(1+∣zn∣)≤(1+∣z∣)2(1+xn) and (2+r)−ν≤(1+∣z∣)ν. Choosing M large gives (6.4) for −2<r<2; for r≥2 the left side vanishes. In particular every derivative of G tends to 0 as r↓−2, locally uniformly (also at t=0); so G, extended by 0 to r≤−2, is smooth, and it equals F. Since 2+r≤4 on the support, (6.4) implies (6.5). Formula (6.6) is (6.7) at t=0.
(d) Taylor's formula at r=−2, where F vanishes to infinite order, turns (6.5) into (6.4): ∣∂rρF(r)∣≤sup[−2,r]∣∂rρ+νF∣(r+2)ν/ν!. Define, for x∈R+n,
A(x,z)=2−zn2F(x′,x′−z′,2xn(2−zn),2−zn2zn)(zn<2),A(x,z)=0(zn>1).(6.9)
On 1<zn<2 both definitions give 0, because then 2zn/(2−zn)>2; so A is smooth. For zn≤1 put r=2zn/(2−zn)∈(−2,2] and t=xn(2−zn)/2. Then r+2=4/(2−zn)≤8/(1+∣zn∣) and xn/2≤t≤xn(1+∣zn∣). Every derivative of A is a finite sum of terms (polynomial in xn) × (smooth function of zn growing at most polynomially on zn≤1) × (a derivative of F at the displayed point). By (6.4) such a term is at most C(1+xn)k(1+∣zn∣)k(1+∣z′∣+xn/2)−ν(1+∣zn∣)−ν, and choosing ν large gives (6.2). By (a), A comes from a unique a∈Sla−∞. Finally, for xn>0<yn, inserting z=(x′−y′,(xn−yn)/xn) into (6.9) gives 2/(2−zn)=xn/t, xn(2−zn)/2=t and 2zn/(2−zn)=r, so the kernel (6.3) of Ta equals F/t=K almost everywhere; for yn<0 both vanish. □
Remark 6.3 (Singular kernels of order −∞). For an ordinary pseudodifferential operator of order −∞ the kernel is smooth. Here, by (c), K=F/t near the corner, and the leading part F(x′,y′,0,r)/t is homogeneous of degree −1 in (xn,yn). It is not smooth, and not even bounded, unless F vanishes on the front face. This singularity is what makes residual operators fail to improve regularity in Section 12.
A kernel bound at finite negative order
Proposition 6.4 (A kernel bound). Let a∈Sla−n−2. Then the kernel of Ta is a function, and
∣Ka(x,y)∣≤C(1+∣x′−y′∣)−n(xn+yn)3xnyn(xn,yn>0),Ka=0 elsewhere,(6.10)
with C bounded by a seminorm of a. Consequently supx∫∣Ka(x,y)∣dy and supy∫∣Ka(x,y)∣dx are at most 21C∫Rn−1(1+∣z′∣)−ndz′.
Proof. For xn>0, ∣a♭(x,ξ)∣≤C(1+∣(ξ′,xnξn)∣)−n−2 is integrable in ξ, so Ka(x,y)=(2π)−n∫ei(x−y)⋅ξa♭dξ converges absolutely, Tau(x)=∫Ka(x,y)u(y)dy, and (6.3) holds with the bounded continuous function A(x,z)=(2π)−n∫eiz⋅ξa(x,ξ)dξ. As in Theorem 6.2(a), Fna is now a continuous function, so A=0 for zn≥1. Integration by parts gives ∣zγA∣≤C for ∣γ∣≤n+2, because ∣∂ξγa∣≤C(1+∣ξ∣)−n−2−∣γ∣ is integrable; so (1+∣z′∣)n(1+∣zn∣)2∣A∣≤C. Likewise ∂znA=(2π)−n∫eiz⋅ξiξnadξ and ∣z′γ∂znA∣≤C for ∣γ∣≤n. Since A(x,z′,1)=0, the mean value theorem gives ∣A(x,z)∣≤C(1+∣z′∣)−n∣1−zn∣ for zn≤1.
If 0<xn≤yn, then zn=(xn−yn)/xn≤0, 1+∣zn∣=yn/xn, and ∣Ka∣≤xn−1C(1+∣z′∣)−n(xn/yn)2=C(1+∣z′∣)−nxn/yn2≤8C(1+∣z′∣)−nxnyn/(xn+yn)3, because xn+yn≤2yn. If 0<yn<xn, then ∣1−zn∣=yn/xn and ∣Ka∣≤C(1+∣z′∣)−nyn/xn2≤8C(1+∣z′∣)−nxnyn/(xn+yn)3. For the marginals, ∫Rn−1(1+∣z′∣)−ndz′<∞ and ∫0∞xnyn(xn+yn)−3dyn=∫0∞s(1+s)−3ds=21; the bound is symmetric in xn,yn. □
The two cases correspond to the two sides of the diagonal: for yn≥xn the decay of A in zn is used, for yn<xn the vanishing of A at zn=1, that is, lacunarity.
Examples of residual kernels
Example 6.5 (Without lacunarity the operator sees below the boundary). Let 0≤h∈C0∞(Rn) be supported near (0,23), with h(0,23)>0, and a=e−xnh(ξ)∈S+−∞. Then A=e−xnh does not vanish on zn>1, so a is not lacunary. For 0≤u∈C0∞(R−n) supported near (0,−21), with u(0,−21)>0, formula (6.3), whose derivation for xn>0 in Theorem 6.2(b) does not use lacunarity, gives Tau(0,1)=e−1∫h(−y′,1−yn)u(y)dy>0, since 1−yn is near 23 there. So Tau=0 in R+n although u=0 in R+n: lacunarity cannot be dropped from Theorem 5.1(a), in accordance with Proposition 4.3.
Example 6.6 (A residual kernel in one dimension). Let n=1, 0=h∈C0∞((−21,21)) and a(x,ξ)=e−xh(ξ). Then K(x,y)=e−xx−1h((x−y)/x) for x>0, and
Tau(x)=e−x∫h(1−s)u(xs)ds,F(t,r)=2+r2e−t(1+r/2)h(2+r2r).
The kernel vanishes unless 21<y/x<23, and F vanishes unless −52<r<32. Along the diagonal K(x,x)=e−xh(0)/x, which is unbounded if h(0)=0: a symbol of order −∞ with an unbounded kernel. Ta is bounded on L2(0,∞) by Proposition 6.4 and gains no derivative by Theorem 12.1. The model T0u(x)=∫h(1−s)u(xs)ds on the front face (so Ta=e−xT0) is not in the class, since its symbol does not decay in x. It commutes with the unitary dilations u↦λ1/2u(λ⋅) of L2(0,∞), and Minkowski's inequality gives ∥T0u∥≤∫∣h(1−s)∣s−1/2ds∥u∥, since ∥u(⋅s)∥=s−1/2∥u∥.
Example 6.7 (The resolved kernel near r=−2). In Example 6.6, F vanishes identically near r=−2 because h has compact support. For a lacunary but not strongly lacunary symbol, take n=1 and A(x,z)=e−xg(z) with g∈S(R) vanishing for z≥1 but not near −∞, for instance g(z)=e−1/(1−z)e−z2 for z<1 and g(z)=0 for z≥1. Then F(t,r)=2+r2e−t(1+r/2)g(2+r2r), and as r↓−2 the argument 2r/(2+r)→−∞, where g decreases rapidly; this is the flatness at r=−2 used in (6.4). At r=2 the argument tends to 1, where g vanishes to infinite order.