Write g=fβΟ, a real-valued function on Ωδ. By Quadratic Increment Characterisation of Semiconvexity it suffices to prove that for all x1β,x2ββΩδ and every tβR with 0β€t and tβ€1,
g(tx1β+(1βt)x2β)β€tg(x1β)+(1βt)g(x2β)+2ΞΌβt(1βt)β₯x1ββx2ββ₯2,
where 2sβ has the meaning fixed in Quadratic Increment Characterisation of Semiconvexity.
Fix such x1β, x2β and t, put z=tx1β+(1βt)x2β, which lies in Ωδ because that set is convex by The Ξ΄-Interior of a Convex Subset of Rn is Convex, and put c=2ΞΌβt(1βt)β₯x1ββx2ββ₯2.
For xβΩδ let hxβ:RnβR be the integrand of Convolution of a Continuous Function with a Compactly Supported Continuous Kernel, that is, hxβ(w)=f(xβw)Ο(w) for those wβRn with xβwβΞ© and hxβ(w)=0 for all other w; thus g(x)=β«RnβhxβdΞ»nβ, and hxβ is integrable with respect to Ξ»nβ by claim 1 of The Convolution Integrand is Continuous, Compactly Supported and Integrable. Here Rn together with its Borel Ο-algebra and Ξ»nβ is a measure space, by Lebesgue Measure on Rn.
Step 1 (a pointwise inequality). We claim that
hzβ(w)β€thx1ββ(w)+(1βt)hx2ββ(w)+cΟ(w)forΒ everyΒ wβRn.
Suppose first that Ξ΄<β₯wβ₯. Then Ο(w)=0 by clause 3 of Mollifier Kernel of Radius Ξ΄ on Rn, and consequently hxβ(w)=0 for every xβΩδ: either xβwβ/Ξ©, in which case hxβ(w)=0 by definition, or xβwβΞ©, in which case hxβ(w)=f(xβw)β
0=0 by claim 1 of Zero Products and Elementary Identities in a Field. By the same annihilation claim both sides of the asserted inequality equal 0, so it holds.
Suppose now that β₯wβ₯β€Ξ΄. For any xβΩδ we have xβw=x+(βw) by claim 3 of Euclidean Space Rn is a Real Vector Space and βw=(β1)w by claim 2 of the same result, so claims 2 and 5 of Elementary Properties of the Euclidean Norm on Rn give
d(x,xβw)=β₯βwβ₯=β£β1β£β₯wβ₯=β₯wβ₯β€Ξ΄,
whence xβwβBΛ(x,Ξ΄)βΞ©. In particular x1ββw, x2ββw and zβw all lie in Ξ©, so
hx1ββ(w)=f(x1ββw)Ο(w),hx2ββ(w)=f(x2ββw)Ο(w),hzβ(w)=f(zβw)Ο(w).
In the real vector space Rn of Euclidean Space Rn is a Real Vector Space, distributivity of scalar multiplication over vector addition and over addition of scalars, together with associativity and commutativity of addition, give
t(x1ββw)+(1βt)(x2ββw)=(tx1β+(1βt)x2β)β(t+(1βt))w=zβw,
and likewise (x1ββw)β(x2ββw)=x1ββx2β.
Since Ξ© is convex and f is semiconvex on Ξ© with constant ΞΌ, Quadratic Increment Characterisation of Semiconvexity applied to the points x1ββw and x2ββw of Ξ© and to the scalar t yields
f(zβw)β€tf(x1ββw)+(1βt)f(x2ββw)+2ΞΌβt(1βt)β₯x1ββx2ββ₯2=tf(x1ββw)+(1βt)f(x2ββw)+c.
By clause 2 of Mollifier Kernel of Radius Ξ΄ on Rn we have 0β€Ο(w), so multiplying this non-strict inequality by Ο(w), which is permitted by claim 5 of Elementary Arithmetic in an Ordered Field, and expanding by distributivity in the field of real numbers gives precisely the asserted inequality at w. This proves the claim.
Step 2 (integration). By clause 4 of Mollifier Kernel of Radius Ξ΄ on Rn the kernel Ο is integrable with respect to Ξ»nβ and β«RnβΟdΞ»nβ=1. Let u:RnβR be given by u(w)=thx1ββ(w)+(1βt)hx2ββ(w)+cΟ(w). Applying claim 2 of Linearity and Monotonicity of the Lebesgue Integral first to the integrable functions hx1ββ and hx2ββ with the scalars t and 1βt, and then to the resulting integrable function and Ο with the scalars 1 and c, shows that u is integrable with
β«RnβudΞ»nβ=tg(x1β)+(1βt)g(x2β)+c.
By Step 1, hzβ(w)β€u(w) for every wβRn, so the monotonicity assertion of claim 2 of Linearity and Monotonicity of the Lebesgue Integral, applied to the integrable functions hzβ and u, gives
g(z)=β«RnβhzβdΞ»nββ€tg(x1β)+(1βt)g(x2β)+c.
This is the required inequality. Since x1β,x2ββΩδ and t with 0β€tβ€1 were arbitrary, Quadratic Increment Characterisation of Semiconvexity shows that fβΟ is semiconvex on Ωδ with constant ΞΌ.