SH02-GAM — Directional neighborhoods and a sheaf projector

Local unit: SH02-GAM. Original programme text: CC0 1.0 Universal.

We keep the advanced course convention: kk is a commutative ring with identity and finite global dimension. Modules need not be flat, finite, or free. All derived categories in this unit are bounded below. No constructibility condition occurs. Let VV be a finite-dimensional real vector space and let γ⊂V\gamma\subset V be a closed convex cone containing 00. In particular, γ\gamma may have lines, may have empty interior, and may be {0}\{0\} or VV. Write γa=−γ\gamma^a=-\gamma. A superscript aa in this unit always means the antipodal image, never a polar cone.

The purpose of the topology below is to organize information that can be continued in the directions of γ\gamma. Convexity has two distinct roles: it connects alternative continuations, and it supplies a contraction for a correspondence of points. We separate these roles in the proofs.

The directional topology and continuation argument are compared with Kashiwara and Schapira, Microlocal Study of Sheaves, Astérisque 128 (1985), Section 1.5.2, printed pp. 33–36. That section assumes a closed convex cone containing zero; it does not require a pointed cone or nonempty interior. The kernel calculation below then supplies its own explicit comparison and contraction, with the support and sign checks kept separate.

Directional open sets

SH02-GAM-TOPO — Directional open sets

Definition and elementary properties (SH02-GAM-TOPO). Put a topology on the set VV by declaring WW open precisely when it is ordinarily open and W+γ=WW+\gamma=W. Denote the resulting space by VγV_\gamma. For any subset X⊂VX\subset V, let XγX_\gamma have the subspace topology from VγV_\gamma, and let

ϕ=ϕγX:X⟶Xγ \phi=\phi_\gamma^X:X\longrightarrow X_\gamma

be the identity map of underlying sets, with the ordinary subspace topology on its domain. It is continuous. Arbitrary unions preserve both conditions defining directional openness. Finite intersections do as well: (W1∩W2)+γ⊂W1∩W2(W_1\cap W_2)+\gamma\subset W_1\cap W_2, and the opposite inclusion follows from 0∈γ0\in\gamma. Thus this is a topology.

If γ1⊂γ2\gamma_1\subset\gamma_2, the identity Xγ1→Xγ2X_{\gamma_1}\to X_{\gamma_2} is continuous. For γ={0}\gamma=\{0\} the topology is the ordinary one. The sets

Bϵ(x)+γ(ϵ>0) B_\epsilon(x)+\gamma\qquad(\epsilon>0)

form a convex directional neighborhood basis at xx in VγV_\gamma. Indeed, a directional open set containing xx contains some Bϵ(x)B_\epsilon(x) and consequently its sum with γ\gamma. Here any fixed norm can be used. For a compact convex KK and a directional open WW containing KK, compactness gives ϵ>0\epsilon>0 with

(K+Bϵ(0))+γ⊂W.(G1) (K+B_\epsilon(0))+\gamma\subset W. \qquad\text{(G1)}

To verify this, choose a positive ordinary distance from KK to the complement of WW; if the complement is empty any positive radius works. The left side is convex and directionally open.

The definition makes sense for arbitrary XX. From now on, assertions about cohomology on XγX_\gamma require XX itself to be directionally open in VV, unless explicitly stated otherwise. This ensures U+γ⊂XU+\gamma\subset X whenever U⊂XU\subset X.

The directional topology need not be Hausdorff. If v∈γv\in\gamma is nonzero, every directional neighborhood of xx contains x+vx+v, so these distinct points have no disjoint neighborhoods. We therefore use the ordinary sheaf operations ϕ−1\phi^{-1}, Rϕ*R\phi_* and RΓR\Gamma on arbitrary topological spaces. We do not apply a locally compact Hausdorff six-operation theorem to XγX_\gamma. The later uses of proper base change occur on ordinary locally compact Hausdorff spaces.

Continuing sections across a cone

SH02-GAM-SECTIONS — Continuing sections

Section lemma (SH02-GAM-SECTIONS). Suppose XX is directionally open, AA is a sheaf of kk-modules on XγX_\gamma, and H=ϕ−1AH=\phi^{-1}A. If U⊂XU\subset X is ordinarily open and convex, the pullback-and-restriction map

Γ(U+γ;A)⟶Γ(U;H)(G2) \Gamma(U+\gamma;A)\longrightarrow\Gamma(U;H) \qquad\text{(G2)}

is an isomorphism. The same statement holds for U=⌀U=\varnothing, where both groups vanish.

Proof. Inverse image does not change the stalk at a point because ϕ\phi is the identity of underlying sets. Equality of two germs of AA at xx therefore means equality on a directional neighborhood of xx. It implies equality of their germs at every point of x+γx+\gamma at which the sections in question are defined.

For injectivity, let a section on U+γU+\gamma pull back to zero on UU. For any y=u+vy=u+v with u∈Uu\in U and v∈γv\in\gamma, its zero germ at uu propagates to yy. Every germ is zero, so the section is zero.

For surjectivity, represent a section ss of HH on an ordinary open cover U=⋃iUiU=\bigcup_i U_i by sections aia_i of AA on Ui+γU_i+\gamma. Such representatives exist: inverse-image sections are locally represented on directional neighborhoods; shrink the ordinary representing open UiU_i inside that neighborhood and then restrict to Ui+γU_i+\gamma.

Fix y∈(Ui+γ)∩(Uj+γ)y\in(U_i+\gamma)\cap(U_j+\gamma). Choose ui∈Ui∩(y−γ)u_i\in U_i\cap(y-\gamma) and uj∈Uj∩(y−γ)u_j\in U_j\cap(y-\gamma). The segment joining these two points lies in the convex set U∩(y−γ)U\cap(y-\gamma). At any point uu of the segment, and for any two charts Ur,UtU_r,U_t containing uu, the germs of ar,ata_r,a_t at uu agree because both represent ss. They agree on a directional neighborhood of uu, which contains yy. Consequently their germs at yy agree. Along a small subinterval of the segment a single chart can be used, so this common germ at yy is locally constant as a function of the segment parameter. Connectedness of the interval makes it constant. In particular, (ai)y=(aj)y(a_i)_y=(a_j)_y. This argument for every yy proves equality on the overlaps of Ui+γU_i+\gamma and Uj+γU_j+\gamma. Sheaf gluing gives a section on U+γU+\gamma whose pullback is ss. The construction is inverse to (G2), hence canonical. ▫\square

Applying (G2) to convex directional opens gives an isomorphism on a basis, and therefore gives the underived unit

A→∼ϕ*ϕ−1A.(G3) A\xrightarrow{\sim}\phi_*\phi^{-1}A. \qquad\text{(G3)}

SH02-GAM-COMPACT-SECTIONS — Compact convex tests

Compact continuation (SH02-GAM-COMPACT-SECTIONS). For a compact convex subset K⊂XK\subset X, restriction induces

Γ(K+γ;H)→∼Γ(K;H).(G4) \Gamma(K+\gamma;H)\xrightarrow{\sim}\Gamma(K;H). \qquad\text{(G4)}

These are sections of the restrictions of the ordinary sheaf HH to the indicated subsets.

Proof. We use the elementary compact-neighborhood continuity of sections: for a compact subset KK of a Hausdorff space, sections of a restricted sheaf on KK are the filtered colimit of sections on its open neighborhoods. One can check it directly by representing a section near each point of KK, taking a finite subcover, and shrinking the finitely many neighborhoods so that the representatives agree on their overlaps. Equality of two representatives is checked by the same finite-neighborhood shrinking argument. For compact convex KK in a finite-dimensional vector space, convex open neighborhoods are cofinal.

Using (G2), this continuity and (G1) give

Γ(K;H)≃colim⁡U⊃K convex openΓ(U+γ;A)≃colim⁡W⊃K directionally openΓ(W;A).(G5) \Gamma(K;H) \simeq\mathop{\mathrm{colim}}_{U\supset K\text{ convex open}}\Gamma(U+\gamma;A) \simeq\mathop{\mathrm{colim}}_{W\supset K\text{ directionally open}}\Gamma(W;A). \qquad\text{(G5)}

If K⊂L⊂K+γK\subset L\subset K+\gamma and LL is compact convex, a directional open set contains KK if and only if it contains LL. Hence (G5) identifies the restriction Γ(L;H)→Γ(K;H)\Gamma(L;H)\to\Gamma(K;H) with an isomorphism.

The set C=K+γC=K+\gamma is closed: from a convergent sequence kn+vnk_n+v_n choose a convergent subsequence of kn∈Kk_n\in K, and then use closedness of γ\gamma on vnv_n. Choose increasing radii rn→∞r_n\to\infty, all large enough that K⊂B¯rn(0)K\subset\overline B_{r_n}(0), and put Ln=C∩B¯rn(0)L_n=C\cap\overline B_{r_n}(0). These compact convex sets contain KK, and their interiors int⁡C(Ln)\operatorname{int}_C(L_n) in the topological space CC cover CC. Therefore sections on CC are compatible sections on the LnL_n: these topological interiors give the required open cover for gluing; affine relative interiors are not being used. Every restriction Γ(Ln;H)→Γ(K;H)\Gamma(L_n;H)\to\Gamma(K;H) is the isomorphism just proved. Taking their inverse limit proves (G4). If KK is empty, so is CC, and the assertion is immediate. ▫\square

Derived continuation

The ordinary-space acyclicity inputs are SH02-OR-CONVEX-COMPACT and SH02-OR-CONVEX-EXHAUSTION proved in Extending sections on convex sets. Their precise contract is this: for a nonempty locally closed convex subset CC of a finite-dimensional real vector space, a sheaf HCH_C is acyclic for sections on CC if every restriction

Γ(C;HC)⟶Γ(L;HC) \Gamma(C;H_C)\longrightarrow\Gamma(L;H_C)

is onto for compact convex L⊂CL\subset C. Compact convex CC is the compact case of this criterion; the locally closed case uses a countable compact convex exhaustion with surjective transition maps. These are mathematical prerequisite proofs, not assertions that arbitrary sheaves on convex sets are acyclic. The topology in this criterion is the ordinary topology.

SH02-GAM-FLABBY-ACYCLIC — Acyclicity after inverse image

Acyclic resolution lemma (SH02-GAM-FLABBY-ACYCLIC). If AA is flabby on XγX_\gamma, then ϕ−1A\phi^{-1}A is acyclic on every ordinary locally closed convex C⊂XC\subset X.

Proof. Let L⊂CL\subset C be compact convex, and let s∈Γ(L;ϕ−1A)s\in\Gamma(L;\phi^{-1}A). Formula (G5) represents ss by a section of AA on a directional open neighborhood WW of LL in XX. Flabbiness extends it to all of XγX_\gamma. Pull back that extension and restrict to CC. It extends ss, so the cited acyclicity criterion applies. Empty sets cause no exception. Notice that we proved convex-set acyclicity; we have not claimed that ϕ−1A\phi^{-1}A is flabby in the ordinary topology. ▫\square

SH02-GAM-COH-OPEN — Open convex cohomology

Derived open continuation (SH02-GAM-COH-OPEN). For G∈D+(kXγ)G\in D^+(k_{X_\gamma}) and an ordinary convex open U⊂XU\subset X, the canonical morphism is an isomorphism:

RΓ(U+γ;G)→∼RΓ(U;ϕ−1G).(G6) R\Gamma(U+\gamma;G)\xrightarrow{\sim} R\Gamma(U;\phi^{-1}G). \qquad\text{(G6)}

SH02-GAM-COH-COMPACT — Compact convex cohomology

Derived compact continuation (SH02-GAM-COH-COMPACT). For compact convex K⊂XK\subset X and the same GG, restriction is an isomorphism:

RΓ(K+γ;ϕ−1G)→∼RΓ(K;ϕ−1G).(G7) R\Gamma(K+\gamma;\phi^{-1}G)\xrightarrow{\sim} R\Gamma(K;\phi^{-1}G). \qquad\text{(G7)}

SH02-GAM-UNIT — The derived adjunction unit

Derived full faithfulness (SH02-GAM-UNIT). The adjunction unit is an isomorphism on D+(kXγ)D^+(k_{X_\gamma}):

G→∼Rϕ*ϕ−1G.(G8) G\xrightarrow{\sim}R\phi_*\phi^{-1}G. \qquad\text{(G8)}

Proof of the three assertions. Resolve GG by a bounded-below complex I•I^\bullet of injective sheaves on XγX_\gamma. The existence of injectives and exactness of inverse image hold on arbitrary topological spaces; see SH02-IMP-INJECTIVE, SH02-IMP-INVERSE, and SH02-IMP-DERIVE in the prerequisite contracts. Injective sheaves are flabby, and flabby sheaves are acyclic on every open set, without a Hausdorff assumption.

The acyclic resolution lemma says that ϕ−1I•\phi^{-1}I^\bullet computes derived sections on UU, KK, and K+γK+\gamma. The latter is closed convex, even when it is unbounded. Applying (G2) and (G4) term by term gives (G6) and (G7). The maps are the indicated pullback or restriction maps because their underived versions are exactly the maps used in those lemmas.

For (G8), use convex directional open neighborhoods. On such a neighborhood WW, ϕ−1Im\phi^{-1}I^m has no higher ordinary cohomology, and Γ(W;Im)=Γ(W;ϕ−1Im)\Gamma(W;I^m)=\Gamma(W;\phi^{-1}I^m) by (G2). It follows that every ϕ−1Im\phi^{-1}I^m is ϕ*\phi_*-acyclic, and (G3) gives a termwise isomorphism I•→ϕ*ϕ−1I•I^\bullet\to\phi_*\phi^{-1}I^\bullet computing the derived unit. A bounded-below complex of acyclic objects computes the right derived functor, so all three arguments apply in D+D^+, rather than only to a sheaf concentrated in degree zero. ▫\square

In particular, ϕ−1:D+(kXγ)→D+(kX)\phi^{-1}:D^+(k_{X_\gamma})\to D^+(k_X) is fully faithful: apply derived adjunction and (G8) to the target of a Hom group. Its right adjoint is Rϕ*R\phi_*. Thus Pγ=ϕ−1Rϕ*P_\gamma=\phi^{-1}R\phi_* is a projector on D+(kX)D^+(k_X), with counit PγF→FP_\gamma F\to F. The isomorphism Pγ2≃PγP_\gamma^2\simeq P_\gamma comes from (G8); it is compatible with the counit by the triangle identities of this same adjunction.

A contraction with variable coefficients

The next lemma supplies the topological part of the kernel calculation. It does not identify the cohomology of an arbitrary nonproper map with the cohomology of its point fibers.

SH02-GAM-RELATIVE-CONTRACTION — A contraction with variable coefficients

Relative contraction lemma (SH02-GAM-RELATIVE-CONTRACTION). Let p:E→Bp:E\to B be a continuous map of ordinary locally compact Hausdorff spaces. Suppose there is a section s:B→Es:B\to E and a homotopy h:E×[0,1]→Eh:E\times[0,1]\to E satisfying

h0=idE,h1=sp,ph=pπ, h_0=\mathrm{id}_E,\qquad h_1=s p,\qquad p h=p\pi,

where π:E×[0,1]→E\pi:E\times[0,1]\to E is projection. Then, for every F∈D+(kB)F\in D^+(k_B), pullback gives a canonical isomorphism

RΓ(B;F)→∼RΓ(E;p−1F).(G9) R\Gamma(B;F)\xrightarrow{\sim}R\Gamma(E;p^{-1}F). \qquad\text{(G9)}

Proof. The projection π\pi is proper because [0,1][0,1] is compact. Proper base change, in the arbitrary-module form SH02-IMP-PROPER-BASECHANGE, computes the stalks of its unit

T⟶Rπ*π−1T T\longrightarrow R\pi_*\pi^{-1}T

as the maps Te→RΓ([0,1];(Te)[0,1])T_e\to R\Gamma([0,1];(T_e)_{[0,1]}). These are isomorphisms. For a module MM, the constant sheaf M[0,1]M_{[0,1]} has global sections MM, and its restrictions to nonempty compact convex subintervals are onto. The compact convex acyclicity criterion therefore gives zero higher cohomology. Applying this to the cohomology modules of a bounded-below complex, or using its truncation spectral sequence, gives the assertion for Te∈D+(k)T_e\in D^+(k). Thus RΓ(E;T)→RΓ(E×[0,1];π−1T)R\Gamma(E;T)\to R\Gamma(E\times[0,1];\pi^{-1}T) is an isomorphism.

Take T=p−1FT=p^{-1}F. The identity ph=pπph=p\pi identifies h−1Th^{-1}T with π−1T\pi^{-1}T. The two maps on cohomology induced by the endpoint inclusions into E×[0,1]E\times[0,1] are the same: both are inverses of pullback by π\pi. Composing with pullback by hh shows h0*=h1*h_0^*=h_1^*, that is, id=p*s*\mathrm{id}=p^*s^* on RΓ(E;p−1F)R\Gamma(E;p^{-1}F). On the other side s*p*=ids^*p^*=\mathrm{id} because ps=idBps=\mathrm{id}_B. These identities are in D+(k)D^+(k) and prove (G9), with inverse s*s^*. The canonical isomorphism is p*p^*; a choice of contraction only proves its invertibility. ▫\square

The correspondence projector

SH02-GAM-KERNEL — The correspondence projector

Kernel theorem (SH02-GAM-KERNEL). Set X=VX=V and let q1,q2:V×V→Vq_1,q_2:V\times V\to V be the two ordinary projections. Put

Zγ={(x,y):y−x∈γ}. Z_\gamma=\{(x,y):y-x\in\gamma\}.

For a closed subset ZZ, the notation TZT_Z means restriction to ZZ followed by its exact closed pushforward; equivalently T⊗LkZT\otimes^L k_Z. No shift occurs. For every F∈D+(kV)F\in D^+(k_V) there is a canonical isomorphism

ϕ−1Rϕ*F→∼Rq1*((q2−1F)Zγ).(G10) \phi^{-1}R\phi_*F \xrightarrow{\sim} Rq_{1*}\bigl((q_2^{-1}F)_{Z_\gamma}\bigr). \qquad\text{(G10)}

This is ordinary direct image, denoted **, not direct image with proper support.

Construction of the comparison. Write pi:Zγ→Vp_i:Z_\gamma\to V for the restricted projections. If WW is directionally open, then p1−1W⊂p2−1Wp_1^{-1}W\subset p_2^{-1}W: x∈Wx\in W and y−x∈γy-x\in\gamma imply y∈Wy\in W. Restriction of sections therefore defines

R(ϕp2)*T⟶R(ϕp1)*T R(\phi p_2)_*T\longrightarrow R(\phi p_1)_*T

on D+(kZγ)D^+(k_{Z_\gamma}). Apply it to T=p2−1FT=p_2^{-1}F and precede it by the derived unit for p2p_2. This gives

Rϕ*F⟶Rϕ*Rp2*p2−1F⟶Rϕ*Rp1*p2−1F. R\phi_*F\longrightarrow R\phi_*Rp_{2*}p_2^{-1}F \longrightarrow R\phi_*Rp_{1*}p_2^{-1}F.

The direct-image composition identifications are valid here because inverse images are exact and hence their right adjoints preserve injectives. Derived adjunction for ϕ\phi gives a map from the left side of (G10) to Rp1*p2−1FRp_{1*}p_2^{-1}F, which equals its right side. This specifies the comparison, including its direction.

Convex-neighborhood calculation. For an ordinary nonempty convex open U⊂VU\subset V, let

EU={(x,y):x∈U,y−x∈γ},B=U+γ, E_U=\{(x,y):x\in U,\ y-x\in\gamma\},\qquad B=U+\gamma,

and let p:EU→Bp:E_U\to B be projection to yy. These are ordinary locally compact Hausdorff spaces: EUE_U is closed in U×VU\times V, and BB is open in VV. The fiber over yy is U∩(y−γ)U\cap(y-\gamma), which is nonempty and convex. We need more than that assertion about fibers.

For y0∈By_0\in B, choose x0∈Ux_0\in U with y0−x0∈γy_0-x_0\in\gamma. On a sufficiently small open neighborhood B0B_0 of y0y_0, the formula

σ0(y)=x0+y−y0 \sigma_0(y)=x_0+y-y_0

lies in UU and satisfies y−σ0(y)=y0−x0∈γy-\sigma_0(y)=y_0-x_0\in\gamma. It is a local section in the first coordinate. Choose a locally finite partition of unity on the ordinary open set BB subordinate to such neighborhoods and form the weighted sum σ(y)\sigma(y) of these local sections. Local finiteness makes the sum continuous. Convexity of UU gives σ(y)∈U\sigma(y)\in U, and convexity of γ\gamma gives y−σ(y)∈γy-\sigma(y)\in\gamma. Thus s(y)=(σ(y),y)s(y)=(\sigma(y),y) is a global section.

The formula

ht(x,y)=((1−t)x+tσ(y),y) h_t(x,y)=((1-t)x+t\sigma(y),y)

stays in EUE_U and is a contraction over BB from the identity to spsp. The relative contraction lemma gives the canonical pullback isomorphism

RΓ(U+γ;F)→∼RΓ(EU;p2−1F).(G11) R\Gamma(U+\gamma;F)\xrightarrow{\sim} R\Gamma(E_U;p_2^{-1}F). \qquad\text{(G11)}

Stalk comparison. For ordinary open UU, the definition of derived direct image and open restriction give

RΓ(U;Rp1*p2−1F)≃RΓ(EU;p2−1F). R\Gamma(U;Rp_{1*}p_2^{-1}F) \simeq R\Gamma(E_U;p_2^{-1}F).

Fix x∈Vx\in V. Ordinary convex open neighborhoods UU of xx are cofinal among its ordinary neighborhoods; their directional enlargements U+γU+\gamma are cofinal among directional neighborhoods of xx. Taking the filtered colimit of cohomology in (G11), and using exactness of filtered colimits, identifies the stalks of the two sides of (G10). Under these identifications, the constructed comparison is exactly the pullback in (G11): the unit pulls back sections and the subsequent map restricts them from p2−1(U+γ)p_2^{-1}(U+\gamma) to p1−1Up_1^{-1}U. Hence it is an isomorphism at every stalk in every degree. This proves (G10). ▫\square

The proof also records why compact fibers alone would have been insufficient. The map p:EU→U+γp:E_U\to U+\gamma need not be proper; its relative contraction is what gives (G11). By contrast, for compact KK, the analogous map {(x,y):x∈K,y−x∈γ}→K+γ\{(x,y):x\in K,y-x\in\gamma\}\to K+\gamma is proper: the inverse image of a compact LL is closed in the compact set K×LK\times L. Its fibers are compact convex. That valid observation does not by itself justify base change for the different, generally nonproper, projection p1p_1 along a closed inclusion K↪VK\hookrightarrow V.

SH02-GAM-SUPPORT — A supported projector

Directional support compatibility (SH02-GAM-SUPPORT). Let XX be directionally open, let ZZ be locally closed in XγX_\gamma, and let F∈D+(kX)F\in D^+(k_X). Then

Rϕ*RΓZXF→∼RΓZXγRϕ*F.(G12) R\phi_*R\Gamma_Z^{X}F\xrightarrow{\sim} R\Gamma_Z^{X_\gamma}R\phi_*F. \qquad\text{(G12)}

Here RΓZR\Gamma_Z is the sheaf of derived sections with the indicated locally closed support convention, equivalently Rℋom(kZ,−)R\mathcal Hom(k_Z,-); it is not global cohomology with support. The statement does not cover an arbitrary ordinarily locally closed set ZZ.

Proof. First take Z=CZ=C closed in XγX_\gamma. Its complement WW is open in both topologies. The localization triangle on XX is

RΓCXF⟶F⟶Rj*F|W⟶. R\Gamma_C^X F\longrightarrow F\longrightarrow Rj_*F|_W\longrightarrow.

Apply Rϕ*R\phi_*. Direct-image composition and restriction to an open subset identify its third term with Rjγ*(Rϕ*F)|WγRj_{\gamma*}(R\phi_*F)|_{W_\gamma}. This is the third term of the localization triangle for CC on XγX_\gamma, and the middle arrow is the same restriction map. The functorial fiber comparison is therefore an isomorphism. Only open restriction is used; no properness statement is needed.

For Z=W∩CZ=W\cap C with WW directionally open and CC directionally closed, use the locally closed support identity

RΓZXF=Rj*RΓC∩WW(F|W). R\Gamma_Z^X F=Rj_*R\Gamma_{C\cap W}^{W}(F|_W).

It follows either from the definition Rℋom(kZ,F)R\mathcal Hom(k_Z,F) and the adjunction for extension by zero along jj, or by composing the open and closed support functors. Apply the closed case on WW, then compose direct images. This proves (G12) and fixes its canonical map independently of a chosen locally closed presentation. ▫\square

Worked calculations and problems

SH02-GAM-EX-SKY — A skyscraper calculation

A point becomes a reverse cone (SH02-GAM-EX-SKY). Let a∈Va\in V and let MM be any kk-module. If F=M{a}F=M_{\{a\}}, then

PγF≃Ma−γ P_\gamma F\simeq M_{a-\gamma}

in degree zero. Indeed q2−1Fq_2^{-1}F is the constant MM-sheaf on V×{a}V\times\{a\} extended by zero, and its restriction to ZγZ_\gamma is the same sheaf on {(x,a):a−x∈γ}\{(x,a):a-x\in\gamma\}. Projection to xx is a homeomorphism onto the closed subset a−γa-\gamma, so its direct image is exact. This checks the sign in (G10). For instance, take k=ℤk=\mathbb Z, M=ℤ/6ℤM=\mathbb Z/6\mathbb Z, V=ℝ2V=\mathbb R^2, and γ={(r,s):r≥|s|}\gamma=\{(r,s):r\geq|s|\}. The result is supported on the closed cone with vertex aa opening in the negative first-coordinate direction. Neither flatness of MM nor smoothness of the cone boundary is used.

SH02-GAM-EX-EXTREMES — Extreme cones

Two extreme projectors (SH02-GAM-EX-EXTREMES). For γ={0}\gamma=\{0\}, ZγZ_\gamma is the diagonal and Pγ=idP_\gamma=\mathrm{id}. For γ=V\gamma=V, the directional topology is indiscrete. Its sheaves of modules are simply modules, and

PVF≃(RΓ(V;F))V. P_VF\simeq (R\Gamma(V;F))_V.

Thus (G10) includes the pullback of global cohomology as well as the identity projector. If V=0V=0, these descriptions coincide.

SH02-GAM-EXERCISES — Exercises with solutions

Exercises with solutions (SH02-GAM-EXERCISES).

  1. Let L⊂VL\subset V be a linear subspace and take γ=L\gamma=L. Identify VγV_\gamma in terms of the quotient q:V→V/Lq:V\to V/L, and identify its category of sheaves.

    Solution. An ordinary open subset WW satisfies W+L=WW+L=W precisely when W=q−1(O)W=q^{-1}(O) for a subset O⊂V/LO\subset V/L. Since the quotient map is open, OO is open exactly when its inverse image is open. Hence the lattices of open sets of VLV_L and V/LV/L are isomorphic, including their covers. Sheaves and their restriction maps are therefore the same data on these lattices, so Sh(VL;k)≃Sh(V/L;k)\mathrm{Sh}(V_L;k)\simeq\mathrm{Sh}(V/L;k). The underlying map need not be a homeomorphism: distinct points in a coset of LL remain distinct but topologically indistinguishable in VLV_L. This example explains why a Hausdorff assumption on VγV_\gamma would exclude valid cases.

  2. Replace q1*q_{1*} in (G10) by q1!q_{1!}. Show that the resulting formula fails for V=ℝV=\mathbb R, γ=[0,∞)\gamma=[0,\infty) and F=kVF=k_V, with k≠0k\neq0.

    Solution. The sheaf kVk_V comes from the constant sheaf on VγV_\gamma, so (G8) gives PγkV≃kVP_\gamma k_V\simeq k_V. Proper-support base change for the ordinary projection identifies a stalk of the proposed replacement with RΓc([0,∞);k)R\Gamma_c([0,\infty);k). This complex is zero: compactify the ray by one endpoint to a closed interval, and compute cohomology relative to that endpoint; restriction from the interval to the endpoint is the identity on kk and both have zero higher cohomology. Thus the replacement yields zero, not kVk_V. This problem uses the ordinary proper-support base-change theorem as a further course prerequisite; it is not supplied by SH02-IMP-PROPER-BASECHANGE, which concerns Rf*R f_* for proper ff.

  3. For V=ℝ2V=\mathbb R^2 and γ=ℝ≥0(1,0)\gamma=\mathbb R_{\geq0}(1,0), describe a directional neighborhood basis at the origin. Explain why replacing γ\gamma by its ordinary interior would change the topology.

    Solution. The sets Bϵ(0)+ℝ≥0(1,0)B_\epsilon(0)+\mathbb R_{\geq0}(1,0) form a basis. They contain the entire nonnegative horizontal ray and a tubular neighborhood with a rounded left end. The ordinary interior of this cone in ℝ2\mathbb R^2 is empty. The equation W+⌀=WW+\varnothing=W admits only W=⌀W=\varnothing, so it does not even provide the same topology or a topology on nonempty VV. Passing to the interior is not allowed in the hypotheses or definitions.

  4. State the actual stalk formula obtained in the proof of (G10). Show that a single closed cone cannot replace the neighborhood colimit for arbitrary FF, even when the cone is pointed and has nonempty interior.

    Solution. For every integer jj,

    Hj(PγF)x≃colim⁡U∋x convex openHjRΓ(U+γ;F). H^j(P_\gamma F)_x \simeq\mathop{\mathrm{colim}}_{U\ni x\text{ convex open}} H^jR\Gamma(U+\gamma;F).

    For a counterexample take V=ℝ2V=\mathbb R^2, γ=ℝ≥02\gamma=\mathbb R_{\geq0}^2, k≠0k\neq0, and

    yn=(−1/n,n),F=⨁n≥1k{yn}. y_n=(-1/n,n),\qquad F=\bigoplus_{n\geq1}k_{\{y_n\}}.

    The set of yny_n is closed and discrete in VV: any compact set meets only finitely many of its points. The sheaf direct sum is therefore the closed pushforward of the constant kk-sheaf on this discrete set. Sections on any open set are a product over the points it contains; local finiteness permits arbitrary independently chosen values. The ordinary restriction F|γF|_\gamma is zero, so RΓ(γ;F)=0R\Gamma(\gamma;F)=0. On the other hand, Bϵ(0)+γB_\epsilon(0)+\gamma contains exactly the yny_n with 1/n<ϵ1/n<\epsilon. Consequently

    H0(PγF)0≃colim⁡N∏n>Nk≃(∏n≥1k)/(⨁n≥1k)≠0. H^0(P_\gamma F)_0 \simeq\mathop{\mathrm{colim}}_N\prod_{n>N}k \simeq\left(\prod_{n\geq1}k\right)\big/\left(\bigoplus_{n\geq1}k\right)\neq0.

    The last nonvanishing follows from the class of a constant sequence with nonzero value. Thus replacing the neighborhood colimit by RΓ(x+γ;F)R\Gamma(x+\gamma;F) is false without extra hypotheses. Formula (G7) does not assert such a replacement: it concerns ϕ−1G\phi^{-1}G, whereas the sheaf FF in this problem is arbitrary.

    The same construction diagnoses a closed-set base-change error for the correspondence itself. Let KK be the closed Euclidean unit disk and replace yny_n by zn=(−1−1/n,n)z_n=(-1-1/n,n). No znz_n belongs to K+γK+\gamma, so p2−1Fp_2^{-1}F restricts to zero on p1−1Kp_1^{-1}K. Yet every convex neighborhood K+Bϵ(0)K+B_\epsilon(0) has a directional enlargement containing a tail of the znz_n. Compact-neighborhood continuity of ordinary sections and (G11) therefore give

    Γ(K;p1*p2−1F)≃(∏k)/(⨁k),Γ(p1−1K;p2−1F)=0. \Gamma(K;p_{1*}p_2^{-1}F) \simeq(\prod k)/(\bigoplus k),\qquad \Gamma(p_1^{-1}K;p_2^{-1}F)=0.

    This explicitly rules out base change along K↪VK\hookrightarrow V for this nonproper p1p_1. The properness of the other projection on a compact slice cannot repair that inference.

Proof boundaries

The advanced statements have been drafted at the full cone generality above. Exercise 2 additionally imports proper-support base change and compact-support localization; their proofs are not given in this lesson. The unit makes no claim about microsupport characterization, cutoff under proper cones, Fourier–Sato inversion, specialization, or involutivity.

Compare Kashiwara and Schapira, Microlocal Study of Sheaves, Astérisque 128 (1985). In the inspected Numdam scan, printed page numbers are three less than PDF page numbers. Section 1.5.2 defines the directional topology on printed p. 33; Lemma 1.5.4 proves section continuation on pp. 34–35; Lemma 1.5.5, Theorem 1.5.3 and Corollary 1.5.6 supply the underived unit, derived unit and convex-open cohomology comparison on pp. 34–36. These are the precise antecedents for (G2), (G3), (G6) and (G8).

The section proof here follows the same mathematical mechanism: equality of germs propagates along the cone, and a segment in the convex backward slice connects two possible representatives. The presentation makes the local inverse-image representatives and their overlap gluing explicit. Compact continuation (G4) then uses the cofinality in (G5) and a compact convex exhaustion whose interiors are taken in the topological space being exhausted. This is a claim about sheaves pulled back from the directional topology. It is not the false claim, tested in Exercise 4, that one closed cone computes a directional-neighborhood colimit for an arbitrary sheaf.

The source’s Proposition 1.5.1 and Lemma 1.5.2, printed pp. 31–33, give the convex acyclicity and interval ingredients behind its derived argument. Here SH02-GAM-FLABBY-ACYCLIC invokes the exact compact and exhaustion contracts in Extending sections on convex sets, then the bounded-below acyclic-resolution argument computes the named maps term by term. No ordinary flabbiness of the inverse-image sheaf or Hausdorff property of the directional topology is asserted. The coefficient hypotheses and the zero, full, lower-dimensional and nonpointed cone cases stated at the start are retained.

The correspondence formula (G10) is justified by the additional proof in SH02-GAM-KERNEL. Its map is constructed from the derived unit and restriction on directional opens; local sections, a partition of unity and fibrewise convexity provide a contraction over the base for the ordinary projection. For the relative contraction lemma, Stacks, Tag 09V6 was compared in the native chapter at revision a04446e57ec1, label theorem-proper-base-change. That theorem applies to the proper interval projection used in (G9); it does not justify base change for the generally nonproper correspondence projection. The actual contraction and neighborhood calculation supply (G11). The skyscraper and escaping-sequence calculations check, respectively, the sign and the failure of the tempting closed-set substitution.

The support comparison (G12) is supplied by open restriction, direct-image composition and the localization triangle for a directionally locally closed support. It neither enlarges that support class nor imports a microsupport characterization. The source’s Proposition 3.2.2, printed pp. 58–60, is a later characterization involving microsupport; that separate theorem is not needed for the topology and kernel proofs here. The comparison distinguishes these arguments instead of assigning all of them to one source theorem.

Original programme expression and this source comparison are CC0 1.0 Universal. The cited Astérisque volume retains the Société mathématique de France’s 1985 copyright and archive terms. The proper-base-change input retains the Stacks attribution and component terms recorded in the prerequisite contracts. Reading access does not relicense either human source.