12 · Regularized maxima, approximation and barriers

6 NAG Tags

12.1 The normalized regularized maximum

A regularized maximum must do more than approximate a nonsmooth function. In a gluing argument it must agree exactly with a branch which dominates the others. For logarithmic branches it must also ignore a branch near its zero set. We prove these properties with one normalization.

Definition 12.1.1. Fix an even nonnegative function θ∈Cc∞((−1,1))\theta\in C_c^\infty((-1,1)) with integral one. For m≥1m\geq 1 and ε>0\varepsilon>0 set Mε,m(s)=∫Rmmax⁡1≤i≤m(si+εti)∏i=1mθ(ti) dt1⋯dtm.M_{\varepsilon,m}(s)=\int_{\mathbb R^m}\max_{1\leq i\leq m}(s_i+\varepsilon t_i) \prod_{i=1}^m\theta(t_i)\,dt_1\cdots dt_m. The subscript mm is retained when branches are removed. For m=1m=1, the zero mean gives Mε,1(s)=sM_{\varepsilon,1}(s)=s.

Theorem 12.1.2. The function M=Mε,mM=M_{\varepsilon,m} is smooth, convex, symmetric in its coordinates, and nondecreasing in each coordinate. It satisfies M(s+a1)=M(s)+a,max⁡isi≤M(s)≤max⁡isi+ε,∣M(s)−M(r)∣≤max⁡i∣si−ri∣.\begin{align} M(s+a\mathbf1)&=M(s)+a,\\ \max_i s_i\leq M(s)&\leq\max_i s_i+\varepsilon,\\ |M(s)-M(r)|&\leq\max_i|s_i-r_i|. \end{align} If si<max⁡j≠isj−2εs_i<\max_{j\ne i}s_j-2\varepsilon, then on a neighbourhood of ss the function is independent of sis_i and equals the regularized maximum of the remaining coordinates. Furthermore, ∂iM≥0,∑i∂iM=1,(∂i∂jM)ij≥0,D2M 1=0.\partial_iM\geq0,\qquad \sum_i\partial_iM=1,\qquad (\partial_i\partial_jM)_{ij}\geq0,\qquad D^2M\,\mathbf1=0.

Proof. Put F(s)=max⁡isiF(s)=\max_i s_i. After changing variables, MM is the convolution of the locally integrable function FF with the smooth compactly supported probability kernel ε−m∏iθ(ti/ε)\varepsilon^{-m}\prod_i\theta(t_i/\varepsilon). Differentiating that kernel proves smoothness. For 0≤λ≤10\leq\lambda\leq 1, apply the convexity inequality for FF to λ(s+εt)+(1−λ)(r+εt)\lambda(s+\varepsilon t)+(1-\lambda)(r+\varepsilon t) and integrate. This proves convexity. Monotonicity and permutation symmetry follow directly from the integral.

Translation by a1a\mathbf 1 adds aa inside the maximum, and the kernel has mass one. Jensen’s inequality gives the lower bound in (12.1.2), since ∫tiθ(ti)dti=0\int t_i\theta(t_i)dt_i=0. The upper bound follows from ∣ti∣<1|t_i|<1. The pointwise estimate ∣F(s+εt)−F(r+εt)∣≤∥s−r∥∞|F(s+\varepsilon t)-F(r+\varepsilon t)|\leq\|s-r\|_\infty proves (12.1.3).

For the deletion assertion choose j≠ij\ne i realizing the other maximum. Then si+εti<sj+εtjs_i+\varepsilon t_i<s_j+\varepsilon t_j throughout the kernel support. The maximum does not use the iith coordinate; integrating out tit_i contributes the factor one. The strict gap persists near ss. Finally, monotonicity and convexity give the derivative inequalities. Differentiate (12.1.1) once in aa, then once in ss, to obtain the two remaining identities. ◻

The factor ε−m\varepsilon^{-m} is essential if the kernel is written as ∏iθ(ti/ε)\prod_i\theta(t_i/\varepsilon). Without it the mass is εm\varepsilon^m, and translation adds aεma\varepsilon^m. This is the normalization error in the displayed regularization in CLD25, Proposition 16.9.1, p. 466. Formula (12.1.1) supplies a direct check independent of any positivity argument.

12.2 Composition and gluing

The smoothness just proved is an ordinary real statement. To use it in CLD geometry, all functions involved must first be represented on a common refinement of their charts. The following chain rule then descends to actual forms.

Proposition 12.2.1. Let u1,…,umu_1,\ldots,u_m be original smooth functions on an analytic open set, and let FF be smooth near their joint image. Then d′d′′F(u)=∑iFi(u)d′d′′ui+∑i,jFij(u)d′ui∧d′′uj.\mathrm d'\mathrm d''F(u)=\sum_i F_i(u)\mathrm d'\mathrm d''u_i+ \sum_{i,j}F_{ij}(u)d'u_i\wedge d''u_j. If every d′d′′ui\mathrm d'\mathrm d''u_i is positive, FF is convex, and Fi≥0F_i\geq 0, then d′d′′F(u)\mathrm d'\mathrm d''F(u) is positive. In particular Mε,m(u)M_{\varepsilon,m}(u) is smooth psh.

Proof. Apply d′d' to d′′F(u)=∑iFi(u)d′′uid''F(u)=\sum_iF_i(u)d''u_i and use the product rule. At a point, factor the positive semidefinite Hessian as D2F=BtBD^2F=B^{\mathsf t}B. The second sum becomes ∑kαk∧Jαk\sum_k\alpha_k\wedge J\alpha_k, where αk=∑iBkid′ui\alpha_k=\sum_i B_{ki}d'u_i. This is a positive (1,1)(1,1)-form. The first sum is positive because its scalar coefficients are nonnegative. These identities commute with the pullbacks defining the sheaf. ◻

Corollary 12.2.2. (Gluing with a collar). Let V⊂UV\subset U be open, let uu be smooth psh on UU, and let vv be smooth psh on VV. Suppose an open neighbourhood OO of the relative boundary of VV in UU satisfies u>v+2εu>v+2\varepsilon on O∩VO\cap V. Then w={Mε,2(u,v)on V,uon U∖Vw=\begin{cases}M_{\varepsilon,2}(u,v)&\text{on }V,\\u&\text{on }U\setminus V\end{cases} is smooth psh on UU, and equals vv wherever v>u+2εv>u+2\varepsilon.

Proof. Deletion and Mε,1(u)=uM_{\varepsilon,1}(u)=u show equality on the collar, so the definitions agree on an open neighbourhood of every interface point. Composition proves positivity on VV, and the old function supplies it elsewhere. The last assertion follows by deleting the other branch. ◻

A partition of unity does not replace this argument. Although it glues unrestricted smooth functions, derivatives of its factors introduce mixed terms of no fixed sign. The collar inequality is the geometric input that makes positive gluing possible.

12.3 Logarithmic branches and finite barriers

Proposition 12.3.1. Let f1,…,fmf_1,\ldots,f_m be analytic functions on an analytic open set UU, let aia_i and cc be real constants, and put q=max⁡{c,log⁡∣f1∣+a1,…,log⁡∣fm∣+am}.q=\max\{c,\log|f_1|+a_1,\ldots,\log|f_m|+a_m\}. Interpret log⁡0=−∞\log 0=-\infty. Applying MεM_{\varepsilon} to the finite branches and deleting the branches equal to −∞-\infty defines a global original smooth psh function qεq_\varepsilon satisfying q≤qε≤q+εq\leq q_\varepsilon\leq q+\varepsilon.

Proof. Fix xx. If fi(x)=0f_i(x)=0, continuity of ∣fi∣|f_i| gives a neighbourhood on which log⁡∣fi∣+ai<c−2ε\log|f_i|+a_i<c-2\varepsilon. Delete that branch throughout this neighbourhood. For every remaining branch, shrink the neighbourhood so that fif_i is a unit. Its logarithmic absolute value is then an original smooth function with zero d′d′′\mathrm d'\mathrm d''. The constant branch remains finite everywhere. Proposition 12.2.1 therefore proves smoothness and positivity locally. On overlaps the deletion identity shows that the local formulas coincide. The estimates follow from (12.1.2). ◻

Theorem 12.3.2. (A barrier from nested rational domains). Let VV be an affinoid domain in a good Hausdorff analytic space XX. Let f1,…,fm,g∈O(V)f_1,\ldots,f_m,g\in\mathcal O(V) have no common zero and let λi>0\lambda_i>0. For real a<ba<b, define Wt={y∈V:∣fi(y)∣≤etλi∣g(y)∣ for all i}.W_t=\{y\in V:|f_i(y)|\leq e^t\lambda_i|g(y)|\text{ for all }i\}. Suppose Wb⊂Int⁡X(V)W_b\subset\operatorname{Int}_X(V), and let xx satisfy all the inequalities for WaW_a strictly. There exist an open neighbourhood UU of WaW_a, an original smooth psh function ρ\rho on UU, and an open neighbourhood U0U_0 of xx such that ρ<0 on U0,{y∈U:ρ(y)≤0} is compact in U.\rho<0\text{ on }U_0,\qquad \{y\in U:\rho(y)\leq0\}\text{ is compact in }U.

Proof. The no-common-zero assumption implies g≠0g\ne 0 on every WtW_t: otherwise every fif_i would vanish there. These sets are compact rational domains. Choose c<ac<a and set q=max⁡{c,log⁡∣fi/g∣−log⁡λi:1≤i≤m}q=\max\{c,\log|f_i/g|-\log\lambda_i:1\leq i\leq m\} on the locus D(g)D(g) in VV. It is finite and continuous. Choose cc so that q(x)<aq(x)<a, and put U={q<b}⊂D(g)U=\{q<b\}\subset D(g). This subset is contained in WbW_b, hence in the ordinary interior of VV; it is therefore open in XX. It contains WaW_a. For c≤t<bc\leq t<b, its sublevel {q≤t}\{q\leq t\} is exactly WtW_t.

Choose ε>0\varepsilon>0 and a real ss with q(x)+ε<s<aq(x)+\varepsilon<s<a. The preceding proposition applies to the analytic functions fi/gf_i/g on UU. Put ρ=qε−s\rho=q_\varepsilon-s. Then ρ(x)<0\rho(x)<0. Its nonpositive sublevel is closed in UU and contained in Ws⊂WaW_s\subset W_a. As a closed subset of the compact set WaW_a, it is compact. Continuity gives the required U0U_0. ◻

The theorem is deliberately stated with the nested-domain input visible. Establishing such rational data near every point uses the local domain theorems of analytic geometry. The analytic construction above proves what follows once those data have been supplied. It also distinguishes the larger domain UU from the smaller set on which ρ\rho is negative; these two sets must not be identified.

12.4 Uniform approximation and decreasing sequences

The domain of approximation is part of the assertion. The useful elementary conversion below does not change that domain with the accuracy parameter.

Lemma 12.4.1. Let uu be continuous on UU and let F(U)⊂C0(U)\mathcal F(U)\subset C^0(U) be stable under addition of real constants. Suppose uu is a locally uniform limit of functions in F(U)\mathcal F(U). For every relatively compact V⋐UV\Subset U, there is a sequence in F(U)∣V\mathcal F(U)|_V decreasing to u∣Vu|_V.

Proof. Choose vj∈F(U)v_j\in\mathcal F(U) with sup⁡V‾∣vj−u∣≤4−j\sup_{\overline V}|v_j-u|\leq 4^{-j}. Set wj=vj+2⋅4−jw_j=v_j+2\cdot 4^{-j}. Then u+4−j≤wj≤u+3⋅4−j,wj+1≤u+34 4−j<wj.u+4^{-j}\leq w_j\leq u+3\cdot4^{-j},\qquad w_{j+1}\leq u+\tfrac34\,4^{-j}<w_j. The squeeze proves the limit. Every function remains defined on the common set VV. ◻

Conversely, decreasing continuous functions with a continuous limit converge uniformly on every compact subset, by Dini’s theorem. Neither direction proves that a continuous section at an arbitrary transfinite decreasing-limit stage admits continuous approximants in the preceding stage. That is the missing replacement step discussed in the source concordance, not a consequence of Dini’s theorem.