Wiki
Wiki

Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.

Updated

Main theorem (Section 2)


Source. Section 2, printed p. 60 (PDF p. 2).

Statement. Let λ\lambda be Lebesgue measure on R\mathbb R, and let λ2\lambda^2 denote two-dimensional Lebesgue measure on R2\mathbb R^2, equivalently the completion of the product measure λ×λ\lambda\times\lambda. Suppose f:R→Rf:\mathbb R\to\mathbb R and there is a Lebesgue-measurable set N⊆R2N\subseteq\mathbb R^2 with λ2(N)=0\lambda^2(N)=0 such that

f(x+y)=f(x)+f(y)f(x+y)=f(x)+f(y)

whenever (x,y)∉N(x,y)\notin N. Then there is a function h:R→Rh:\mathbb R\to\mathbb R such that

h(x+y)=h(x)+h(y)h(x+y)=h(x)+h(y)

for every x,y∈Rx,y\in\mathbb R, and f(x)=h(x)f(x)=h(x) for λ\lambda-almost every xx.

Proof. For x∈Rx\in\mathbb R, write

Nx={y∈R:(x,y)∈N}.N_x=\{y\in\mathbb R:(x,y)\in N\}.

Fubini's theorem gives a null set M⊆RM\subseteq\mathbb R such that NxN_x is null for every x∉Mx\notin M.

Fix x∈Rx\in\mathbb R. Both MM and x−Mx-M are null, so their union cannot be all of R\mathbb R. Choose x1x_1 outside that union. Thus x1∉Mx_1\notin M and x−x1∉Mx-x_1\notin M. The corresponding vertical sections are null, so

f(x1+y)−f(y)=f(x1)f(x_1+y)-f(y)=f(x_1)

for almost every yy, and

f(x−x1+z)−f(z)=f(x−x1)f(x-x_1+z)-f(z)=f(x-x_1)

for almost every zz. Translation invariance of Lebesgue measure permits the substitution z=x1+yz=x_1+y in the second almost-everywhere identity. Adding the resulting two identities gives

f(x+y)−f(y)=f(x1)+f(x−x1)f(x+y)-f(y)=f(x_1)+f(x-x_1)

for almost every yy. Hence, for each xx, the function y↦f(x+y)−f(y)y\mapsto f(x+y)-f(y) is almost everywhere equal to a constant. That constant is unique, since two conull subsets of R\mathbb R have nonempty intersection. Define it to be h(x)h(x). We have therefore proved that, for every xx,

f(x+y)−f(y)=h(x)(2)f(x+y)-f(y)=h(x) \tag{2}

for almost every yy. Notice that the auxiliary x1x_1 above was chosen after xx was fixed and may depend on xx.

If x∉Mx\notin M, the original equation also gives f(x+y)−f(y)=f(x)f(x+y)-f(y)=f(x) for almost every yy. Uniqueness of the almost-everywhere constant in (2) yields h(x)=f(x)h(x)=f(x). Thus h=fh=f almost everywhere.

It remains to prove that hh is additive. For each t∈Rt\in\mathbb R, choose a null set KtK_t outside which (2) holds with x=tx=t. Fix a,b∈Ra,b\in\mathbb R. Consider the following five exceptional subsets of the (w,z)(w,z)-plane:

E1=Ka×R,E2=R×Kb,E3={(w,z):w+z∈Ka+b},E4=N,E5={(w,z):(a+w,b+z)∈N}.\begin{aligned} E_1&=K_a\times\mathbb R,\\ E_2&=\mathbb R\times K_b,\\ E_3&=\{(w,z):w+z\in K_{a+b}\},\\ E_4&=N,\\ E_5&=\{(w,z):(a+w,b+z)\in N\}. \end{aligned}

Each is a two-dimensional null set. For the coordinate cylinders this follows by first intersecting the unrestricted coordinate with [−n,n][-n,n] and then taking a countable union. The set E3E_3 is null because the shear (w,z)↦(w,w+z)(w,z)\mapsto(w,w+z) preserves Lebesgue measure and sends it to R×Ka+b\mathbb R\times K_{a+b}. Finally, E5=(−a,−b)+NE_5=(-a,-b)+N, so it is null by translation invariance.

A finite union of null sets cannot cover R2\mathbb R^2. Choose (w,z)(w,z) outside E1∪⋯∪E5E_1\cup\cdots\cup E_5. The five corresponding identities are

f(a+w)−f(w)=h(a),f(b+z)−f(z)=h(b),f(a+b+w+z)−f(w+z)=h(a+b),f(w+z)=f(w)+f(z),f(a+b+w+z)=f(a+w)+f(b+z).\begin{aligned} f(a+w)-f(w)&=h(a),\\ f(b+z)-f(z)&=h(b),\\ f(a+b+w+z)-f(w+z)&=h(a+b),\\ f(w+z)&=f(w)+f(z),\\ f(a+b+w+z)&=f(a+w)+f(b+z). \end{aligned}

Substituting the last two identities into the third and regrouping with the first two gives

h(a+b)=h(a)+h(b).h(a+b)=h(a)+h(b).

Since aa and bb were arbitrary, hh is additive. □\square

Dependencies. Fubini's theorem for Lebesgue measure, translation invariance of null sets, and invariance of two-dimensional Lebesgue measure under determinant-one linear transformations.

Formalization. A Lean 4 proof formalizes this real-valued theorem in mathlib v4.29.1. The proof was located during compilation but was not built here.

Bears on. #1126