Finite first-order recognition of F(X)F(X) and Z(X)\mathbb Z^{(X)} from subsemigroup lattices

2.81(a,b)

Problem

For nonempty semigroups SS, let L(S)\mathcal L(S) be the lattice of all subsemigroups, including ∅\varnothing. Can free groups and free abelian groups be recognized by first-order lattice sentences, uniformly in their rank?

L(S)⊨Φfr  ⟺  S≅F(X),L(S)⊨Φab  ⟺  S≅Z(X).\begin{aligned} \mathcal L(S)\models\Phi_{\mathrm{fr}}&\iff S\cong F(X),\\ \mathcal L(S)\models\Phi_{\mathrm{ab}}&\iff S\cong\mathbb Z^{(X)}. \end{aligned}

The language is {∧,∨}\{\wedge,\vee\}; XX may be empty or infinite. The result below gives one finite sentence for each class.

Setup and result

For a semigroup SS, write

L(S)={U⊆S:UU⊆U},U∧V=U∩V,U∨V=⟨U∪V⟩+.\mathcal L(S)=\{U\subseteq S: UU\subseteq U\},\qquad U\wedge V=U\cap V,\qquad U\vee V=\langle U\cup V\rangle_+.

The empty subsemigroup is included. We consider first-order sentences in the two-function language

Llat={∧,∨}.\mathscr L_{\rm lat}=\{\wedge,\vee\}.

An elementary class of lattices is a class defined by first-order sentences in this language. The lattice identities can always be adjoined to an axiom set; they impose no additional restriction on the lattices L(S)\mathcal L(S).

Shevrin proved finite lattice recognition results for torsion-free groups and for torsion-free abelian groups among semigroups (4; 5). At the end of (5, p. 684), he asked whether free groups and free abelian groups admit analogous first-order axiomatizations. These are the two questions in Kourovka Problem 2.81.

Theorem 1. There are Llat\mathscr L_{\rm lat}-sentences Φfr\Phi_{\rm fr} and Φab\Phi_{\rm ab} such that, for every nonempty semigroup SS,

L(S)⊨Φfr⟺S≅F(X) as a semigroup for some set X,L(S)⊨Φab⟺S≅Z(X) as a semigroup for some set X.\begin{align*} \mathcal L(S)\models\Phi_{\rm fr} &\quad\Longleftrightarrow\quad S\cong F(X)\text{ as a semigroup for some set }X, \tag{3a}\\ \mathcal L(S)\models\Phi_{\rm ab} &\quad\Longleftrightarrow\quad S\cong\mathbb Z^{(X)}\text{ as a semigroup for some set }X. \tag{3b} \end{align*}

Here F(X)F(X) is the free group on XX, and Z(X)\mathbb Z^{(X)} is the group of finitely supported integer-valued functions on XX. The set XX may be empty or have any infinite cardinality.

Elements and bases will be represented inside the lattice by

a⟷Xa=⟨a⟩+,basis⟷I(M),a\longleftrightarrow X_a=\langle a\rangle_+, \qquad \text{basis}\longleftrightarrow I(M),

where I(M)I(M) is the set of irreducibles of a submonoid MM. Generation is expressed through containment in subsemigroups and definable subgroups. The free-group argument reconstructs an oriented multiplication graph by Beth’s theorem. These constructions use finitely many variables independently of the rank.

Sections “Recognizing torsion-free groups”–“A finite sentence for free abelian groups” prove (3b), including the initial recognition results. Sections “Recovering unordered multiplication”–“The finite sentence for free groups” prove (3a). We write ⟨A⟩+\langle A\rangle_+, ⟨A⟩mon\langle A\rangle_{\rm mon} and ⟨A⟩grp\langle A\rangle_{\rm grp} for semigroup, monoid and group generation, respectively. The cyclic group {1}\{1\} is permitted throughout.

Recognizing torsion-free groups

Order, strict order and covering are definable by

X≤Y  ⟺  X∧Y=X,X<Y  ⟺  X≤Y∧X≠Y,X≺Y  ⟺  X<Y∧¬∃Z (X<Z<Y).\begin{align*} X\le Y&\iff X\wedge Y=X,\\ X<Y&\iff X\le Y\wedge X\ne Y,\\ X\prec Y&\iff X<Y\wedge\neg\exists Z\,(X<Z<Y). \end{align*}

The symbols ⊥,⊤\bot,\top denote the uniquely definable bounds. Put

C(X)  ⟺  ∃D(D<X ∧ ∀U (U<X⇒U≤D)),C∞(X)  ⟺  C(X) ∧ ¬∃E (Atom(E)∧E≤X).\begin{align*} \mathsf C(X) &\;\Longleftrightarrow\; \exists D\bigl(D<X\ \wedge\ \forall U\,(U<X\Rightarrow U\leq D)\bigr),\\ \mathsf C_\infty(X) &\;\Longleftrightarrow\; \mathsf C(X)\ \wedge\ \neg\exists E\,(\mathsf{Atom}(E)\wedge E\leq X). \end{align*}

Every C\mathsf C-code is monogenic. Indeed, if DD is its greatest proper subsemigroup and a∈X∖Da\in X\setminus D, then ⟨a⟩+=X\langle a\rangle_+=X. Here

⟨a⟩+={an:n≥1}.\langle a\rangle_+=\{a^n:n\geq1\}.

A monogenic semigroup with a repeated power is finite and contains an idempotent. Conversely, if all positive powers of aa are distinct, deleting aa gives the greatest proper subsemigroup of ⟨a⟩+\langle a\rangle_+. Thus C∞\mathsf C_\infty codes are exactly the infinite monogenic subsemigroups. Atoms are exactly singleton idempotents.

For lattice elements E≤SE\leq S, define

P(E,S)  ⟺  ∀H(E<H≤S⇒∃A (C∞(A)∧A≤H)).\mathsf P(E,S)\;\Longleftrightarrow\; \forall H\bigl(E<H\leq S\Rightarrow \exists A\,(\mathsf C_\infty(A)\wedge A\leq H)\bigr).

Define Z(E,H)\mathsf Z(E,H) by E≤HE\leq H and the existence of C∞\mathsf C_\infty-codes A,B,VA,B,V such that

A∧B=⊥,A∨B=H,V<B,A∨V=H,∀W(C∞(W)∧W≤H⇒W≤A ∨ W≤B).\begin{gather*} A\wedge B=\bot,\qquad A\vee B=H,\qquad V<B,\qquad A\vee V=H, \tag{9a}\\ \forall W\bigl(\mathsf C_\infty(W)\wedge W\leq H \Rightarrow W\leq A\ \vee\ W\leq B\bigr). \tag{9b} \end{gather*}

Our local group guard T(S)\mathsf T(S) is

∃E  [  Atom(E)∧E≤S∧∀F (Atom(F)∧F≤S⇒F=E)∧P(E,S)∧∀A (C∞(A)∧A≤S⇒∃H (A≤H≤S∧Z(E,H)))  ].\begin{align*} \exists E\;\bigl[\;&\mathsf{Atom}(E)\wedge E\leq S \wedge\forall F\,(\mathsf{Atom}(F)\wedge F\leq S\Rightarrow F=E)\\ &\wedge\mathsf P(E,S)\\ &\wedge\forall A\,(\mathsf C_\infty(A)\wedge A\leq S\Rightarrow \exists H\,(A\leq H\leq S\wedge\mathsf Z(E,H)))\;\bigr]. \end{align*}

This is a finite first-order formula. It is a variant of the cyclic-interval recognition criteria of Shevrin (3; 5).

Lemma 2. A subsemigroup SS of a monoid satisfies T(S)\mathsf T(S) if and only if its multiplication makes it a torsion-free group, with an identity that need not equal the identity of the ambient monoid.

Proof. Suppose T(S)\mathsf T(S) holds and write E={e}E=\{e\}. If x∈Sx\in S has repeated powers, then ⟨x⟩+\langle x\rangle_+ is finite and contains an idempotent, necessarily ee. Unless x=ex=e, the interval

E<⟨x⟩+≤SE<\langle x\rangle_+\leq S

contradicts P(E,S)\mathsf P(E,S). Hence every x≠ex\neq e has distinct powers.

Fix an interval satisfying Z(E,H)\mathsf Z(E,H) and write A=⟨a⟩+A=\langle a\rangle_+, B=⟨b⟩+B=\langle b\rangle_+. The preceding observation and (9b) imply

H=A∪˙B∪˙{e}.H=A\mathbin{\dot\cup}B\mathbin{\dot\cup}\{e\}.

We first show ab=ba=eab=ba=e. If neither cross product equals ee, both belong to A∪BA\cup B. Whenever ab=akab=a^k, induction gives

ambn=am+n(k−1)(m,n≥1),a^m b^n=a^{m+n(k-1)}\qquad(m,n\geq1),

and whenever ab=bkab=b^k it gives ambn=bn+m(k−1)a^m b^n=b^{n+m(k-1)}. The corresponding formulas for baba show that A∪BA\cup B is a subsemigroup. This contradicts e∈A∨Be\in A\vee B. By symmetry, take ab=eab=e. If ba=akba=a^k, then ea=ak+1ea=a^{k+1}, and e2=ee^2=e gives

a2k+1=ak+1,a^{2k+1}=a^{k+1},

contrary to injectivity of powers. The case ba=bkba=b^k is identical with the roles reversed. Thus ba=eba=e.

Associativity now gives ea=aeea=ae and eb=beeb=be. If ea=akea=a^k, then idempotence implies ak2=aka^{k^2}=a^k, hence k=1k=1. If ea=bkea=b^k, multiplying by bb gives bk+1=eb^{k+1}=e, which is impossible. Therefore eaea is either aa or ee, and similarly ebeb is either bb or ee. The identities ab=ba=eab=ba=e exclude mixed choices. If ee absorbs both rays, then

A∪V∪{e}A\cup V\cup\{e\}

is a proper subsemigroup of HH containing AA and VV, contrary to (9a). Consequently ee is the identity on HH, and a,ba,b are mutual inverses. Every element of HH is a positive power of one of them or is ee.

Every x≠ex\neq e lies in such an interval. All these cyclic groups have the same identity ee. It follows that ee is an identity on SS and every element of SS has an inverse in SS. The injectivity of powers already proved gives torsion-freeness.

Conversely, in a torsion-free group the only idempotent is 11. Every subsemigroup properly containing {1}\{1\} contains a nonidentity element and hence an infinite monogenic subsemigroup. Set

H=⟨a⟩,A=⟨a⟩+,B=⟨a−1⟩+,V=⟨a−2⟩+(a≠1).H=\langle a\rangle,\qquad A=\langle a\rangle_+, \qquad B=\langle a^{-1}\rangle_+,\qquad V=\langle a^{-2}\rangle_+\qquad(a\neq1).

The two rays are disjoint, V<BV<B, and a(a−2)=a−1a(a^{-2})=a^{-1} gives A∨V=HA\vee V=H. Every nonidentity element of HH lies on one ray, so every infinite monogenic subsemigroup of HH lies on one ray. Thus all the required conditions hold. The trivial group satisfies the same formula, with the conditions involving infinite codes vacuous. ◻

The ambient-monoid convention in Lemma 2 can be removed within the finite lattice language. Adjoin a new identity to an arbitrary semigroup SS, obtaining S1S^1. There is an order isomorphism

L(S1)≅L(S)×{0,1},T⟼(T∩S,1{1new∈T}).\mathcal L(S^1)\cong\mathcal L(S)\times\{0,1\},\qquad T\longmapsto\bigl(T\cap S,\boldsymbol1_{\{1_{\rm new}\in T\}}\bigr).

Terms in the product are evaluated coordinatewise. A quantifier over its second coordinate is a conjunction or disjunction of two cases. Induction on formulas therefore eliminates the finite factor. Applying this translation to T\mathsf T at (S,0)(S,0) gives a finite sentence TF\mathsf{TF} such that

L(S)⊨TF⟺S is isomorphic as a semigroup to a torsion-free group.(18) \mathcal L(S)\models\mathsf{TF} \quad\Longleftrightarrow\quad S\text{ is isomorphic as a semigroup to a torsion-free group}. \tag{18}

The multiplication in the reconstructed group is the original multiplication of SS.

Commutativity in the subsemigroup lattice

For the remainder of this section let GG be a torsion-free group. Its C\mathsf C-codes are exactly the subsemigroups

Xa=⟨a⟩+(a∈G),X_a=\langle a\rangle_+\qquad(a\in G),

and a↦Xaa\mapsto X_a is a bijection. Indeed, Xa=XbX_a=X_b gives a=bma=b^m, b=anb=a^n with m,n>0m,n>0. If a≠1a\neq1, injectivity of its powers gives mn=1mn=1. The identity is represented by the unique atomic code E=X1E=X_1. In formulas below, quantification over elements means quantification over C\mathsf C-codes; membership a∈Ha\in H means Xa≤HX_a\leq H.

Subgroups are definable by

Q(H):1∈H ∧ ∀a∈H∖{1} ∃b∈H∖{1}1∈Xa∨Xb.(20) \mathsf Q(H):\quad 1\in H\ \wedge\ \forall a\in H\setminus\{1\}\ \exists b\in H\setminus\{1\}\quad 1\in X_a\vee X_b. \tag{20}

To verify soundness, a positive word in a,ba,b equal to 11 must contain aa, since bb has infinite order. Writing it as uav=1uav=1 gives a−1=vu∈Ha^{-1}=vu\in H. Completeness follows by taking b=a−1b=a^{-1}. Hence the least subgroup containing XaX_a is definable and equals ⟨a⟩\langle a\rangle.

The inverse code is definable as well. For a≠1a\neq1, Xa−1X_{a^{-1}} is the greatest nonidentity element code in ⟨a⟩\langle a\rangle disjoint from XaX_a. A nonidentity power ana^n has this disjointness precisely when n<0n<0. The identity code is excluded here; it is also disjoint from XaX_a. At a=1a=1, inversion fixes EE.

Lemma 3. Let a≠1a\neq1 and let DD be the greatest proper subsemigroup of XaX_a. Then Xb=Xa2X_b=X_{a^2} if and only if XbX_b is an element code and there are element codes Z,UZ,U such that

Xb∨Z=D,Xb∨U≺D.X_b\vee Z=D,\qquad X_b\vee U\prec D.

Proof. Here D={an:n≥2}=Xa2∨Xa3D=\{a^n:n\geq2\}=X_{a^2}\vee X_{a^3}. If Xb∨Z=DX_b\vee Z=D, write b=amb=a^m and let ana^n generate ZZ, where m,n≥2m,n\geq2. Representing a2,a3a^2,a^3 in the join forces

{m,n}={2,3}.\{m,n\}=\{2,3\}.

For m=2m=2, choose U=Xa5U=X_{a^5}. The join has exponent set {2}∪{n:n≥4}\{2\}\cup\{n:n\geq4\}, so adding its only missing element a3a^3 gives DD.

Suppose m=3m=3 and Xb∨U≺DX_b\vee U\prec D. The generator of UU is aka^k with k≥3k\geq3, since k=2k=2 would make the join equal to DD. This join lies in T={an:n≥3}T=\{a^n:n\geq3\}, and T<DT<D. The covering relation forces the join to equal TT. But representing both a4a^4 and a5a^5 with the exponents 3,k3,k forces respectively k=4k=4 and k=5k=5. This is impossible. ◻

The identity-code case completes a finite formula for b=a2b=a^2. This is the square-code construction of (4, Lemma 4).

Write H(a,b)=Xa∨XbH(a,b)=X_a\vee X_b. Define the special-product relation by

SP(a,b;p):p∈H(a,b),∣H(a,b)∣≥3,∀z∈H(a,b)∖{p},(p∈H(a,z)⇒z=b) ∧ (p∈H(z,b)⇒z=a).\begin{align*} \mathsf{SP}(a,b;p):\quad&p\in H(a,b),\qquad |H(a,b)|\geq3,\\ &\forall z\in H(a,b)\setminus\{p\},\quad \bigl(p\in H(a,z)\Rightarrow z=b\bigr) \ \wedge\ \bigl(p\in H(z,b)\Rightarrow z=a\bigr). \end{align*}

The cardinality condition means the existence of three distinct element codes below H(a,b)H(a,b). All parts are finite lattice formulas. A shortest positive word for pp proves

SP(a,b;p)⟹p=ab or p=ba.(24) \mathsf{SP}(a,b;p)\quad\Longrightarrow\quad p=ab\ \text{or}\ p=ba. \tag{24}

Indeed, the three-element condition supplies an element different from both generators; the defining implications then exclude p=ap=a and p=bp=b. A shortest word for pp therefore has length at least two. Removing its first letter gives a different element z∈H(a,b)z\in H(a,b), and the defining implication identifies zz with the other generator.

We use the following local predicate, corresponding to (4, Lemma 6). Let A(a,b)\mathsf A(a,b) assert:

  1. 1∉H(a,b)1\notin H(a,b), a∉Xba\notin X_b, and b∉Xab\notin X_a.

  2. There is exactly one pp with SP(a,b;p)\mathsf{SP}(a,b;p).

  3. Every q∈H(a,b)∖(Xa∪Xb)q\in H(a,b)\setminus(X_a\cup X_b) satisfies SP(u,v;q)\mathsf{SP}(u,v;q) for some u∈Xau\in X_a, v∈Xbv\in X_b.

  4. If SP(a,b;p)\mathsf{SP}(a,b;p), u∈Xa∖{a}u\in X_a\setminus\{a\}, v∈Xb∖{b}v\in X_b\setminus\{b\}, and

    p2∈H(u,v)∖(Xa∪Xb),p^2\in H(u,v)\setminus(X_a\cup X_b),

    then u=a2u=a^2 and v=b2v=b^2.

  5. If SP(a,b;p)\mathsf{SP}(a,b;p), q∈H(a,b)q\in H(a,b), q≠pq\neq p, and q2=p2q^2=p^2, then q∈Xa∪Xbq\in X_a\cup X_b.

Lemma 3 makes the fourth and fifth conditions finite lattice formulas.

Lemma 4. In a torsion-free group, A(a,b)\mathsf A(a,b) implies ab=baab=ba. In a torsion-free abelian group, its first condition implies all five conditions.

Proof. For soundness, symmetry and (24) allow the unique special product to be p=abp=ab. Suppose ab≠baab\neq ba. Then ba∉Xa∪Xbba\notin X_a\cup X_b, since either membership would imply commutativity. The third condition and (24) give

ba=akblorba=blak,k,l≥1.ba=a^k b^l\quad\text{or}\quad ba=b^l a^k, \qquad k,l\geq1.

The case k=l=1k=l=1 contradicts uniqueness of pp. The second equality gives

1=bl−1ak−1∈H(a,b),1=b^{l-1}a^{k-1}\in H(a,b),

contrary to the first condition. In the first case,

(ab)2=ak+1bl+1.(ab)^2=a^{k+1}b^{l+1}.

If this square lies in neither original ray, condition 4 gives ak+1=a2a^{k+1}=a^2, bl+1=b2b^{l+1}=b^2, and hence k=l=1k=l=1, another contradiction. The square therefore lies in XaX_a or XbX_b. Conjugation gives

(ba)2=a−1(ab)2a=b(ab)2b−1.(ba)^2=a^{-1}(ab)^2a=b(ab)^2b^{-1}.

Using the expression appropriate to that ray gives (ba)2=(ab)2(ba)^2=(ab)^2. Condition 5 now puts baba in an original ray, the final contradiction.

For necessity, work in an abelian group. The first condition gives

ambn=1 (m,n≥0)⟹m=n=0.a^m b^n=1\ (m,n\geq0)\quad\Longrightarrow\quad m=n=0.

It also gives

ambn=ab⟹m=n=1.(31) a^m b^n=ab\quad\Longrightarrow\quad m=n=1. \tag{31}

A zero exponent in (31) would put one generator in the other positive ray, or make it the identity. With both exponents positive, cancellation reduces the equality to the preceding positive identity relation.

To check SP(a,b;ab)\mathsf{SP}(a,b;ab), write z=ambnz=a^m b^n and ab=arzsab=a^r z^s. Equation (31) gives

r+ms=1,ns=1.r+ms=1,\qquad ns=1.

Thus n=s=1n=s=1. If z≠abz\neq ab, then m=0m=0, so z=bz=b. The other defining implication is symmetric. Uniqueness follows from (24).

For a mixed element q=ambnq=a^m b^n, both exponents are positive. Neither ama^m nor bnb^n is a positive power of the other, since that would put qq in an original ray. The same argument gives the required special product.

For condition 4, write u=amu=a^m, v=bnv=b^n with m,n≥2m,n\geq2. A representation

(ab)2=amrbns(ab)^2=a^{mr}b^{ns}

has r,s>0r,s>0 because the square lies in neither ray. Cancellation gives

amr−2bns−2=1,a^{mr-2}b^{ns-2}=1,

so mr=ns=2mr=ns=2 and m=n=2m=n=2. Finally, squaring is injective in a torsion-free abelian group: q2=p2q^2=p^2 gives (qp−1)2=1(qp^{-1})^2=1, hence q=pq=p. Thus condition 5 holds. ◻

Define AB\mathsf{AB} by requiring, for every a,b≠1a,b\neq1 such that

b∉⟨a⟩,a∉Xb,a−1∉Xb,b\notin\langle a\rangle,\qquad a\notin X_b,\qquad a^{-1}\notin X_b,

that A(a,b)\mathsf A(a,b) or A(a−1,b)\mathsf A(a^{-1},b) holds. All excluded pairs already commute. Lemma 4 therefore proves that AB\mathsf{AB} implies commutativity. Conversely, in a torsion-free abelian group, both H(a,b)H(a,b) and H(a−1,b)H(a^{-1},b) cannot contain 11: equations

ambn=1=a−rbs(m,n,r,s>0)a^m b^n=1=a^{-r}b^s\qquad(m,n,r,s>0)

would give ams+rn=1a^{ms+rn}=1. Hence one orientation satisfies the first condition of A\mathsf A, and the lemma applies. We have proved

G torsion-free⟹(L(G)⊨AB  ⟺  G is abelian).(37) G\text{ torsion-free}\quad\Longrightarrow\quad \bigl(\mathcal L(G)\models\mathsf{AB}\iff G\text{ is abelian}\bigr). \tag{37}

A finite sentence for free abelian groups

Put

Π(a,b;p)  ⟺  (a=1∧p=b) ∨ (b=1∧p=a) ∨SP(a,b;p),\Pi(a,b;p)\;\Longleftrightarrow\; (a=1\wedge p=b)\ \vee\ (b=1\wedge p=a)\ \vee\mathsf{SP}(a,b;p),

and define

D(a,b;p)  ⟺  Π(a,b;p) ∨ ∃z,u,v(Π(b,z;u)∧Π(a,u;v)∧Π(p,z;v)).\begin{align*} \mathsf D(a,b;p)\;\Longleftrightarrow\;&\Pi(a,b;p)\ \vee\ \exists z,u,v\bigl(\Pi(b,z;u)\wedge\Pi(a,u;v)\wedge\Pi(p,z;v)\bigr). \tag{39} \end{align*}

In an abelian group, (24) gives

D(a,b;p)⟹p=ab.(40) \mathsf D(a,b;p)\quad\Longrightarrow\quad p=ab. \tag{40}

For the second disjunct, u=bzu=bz, v=au=pzv=au=pz, and cancellation gives the conclusion. Let Tot\mathsf{Tot} be the finite sentence ∀a,b ∃p D(a,b;p)\forall a,b\,\exists p\,\mathsf D(a,b;p).

Lemma 5. Every free abelian group of rank at least two satisfies Tot\mathsf{Tot}.

Proof. Use additive notation. If a,ba,b are independent over Z\mathbb Z, independent integer coordinates prove SP(a,b;a+b)\mathsf{SP}(a,b;a+b) by counting the two letters in a positive word. The identity cases are immediate.

Suppose a,b≠0a,b\neq0 are dependent. Choose zz independent of aa; such a zz exists because the rank is at least two. A nonzero relation ma+nb=0ma+nb=0 has n≠0n\neq0, and it shows that

(b,z),(a,b+z),(a+b,z)(b,z),\qquad (a,b+z),\qquad (a+b,z)

are independent whenever their first entries are nonzero. For example, from ka+l(b+z)=0ka+l(b+z)=0 one obtains

(nk−lm)a+nlz=0,(nk-lm)a+nlz=0,

so l=0l=0 and then k=0k=0. Taking u=b+zu=b+z and v=a+b+zv=a+b+z proves the second disjunct of (39). If a+b=0a+b=0, its last partial product is an identity case. ◻

Assume now that GG is a torsion-free abelian group satisfying Tot\mathsf{Tot}. By (40), D\mathsf D defines its multiplication on element codes. For a submonoid M≤GM\leq G, let

I(M)={a∈M∖{1}: a=xy, x,y∈M⇒x=1 or y=1}.I(M)=\{a\in M\setminus\{1\}:\ a=xy,\ x,y\in M\Rightarrow x=1\text{ or }y=1\}.

The following properties form a finite lattice formula Cone(M)\mathsf{Cone}(M):

1∈M,M∩M−1={1},M=⟨I(M)⟩mon,G=⟨M⟩grp,∀a∈I(M) ∀x,y∈M,a∣Mxy⇒a∣Mx or a∣My.\begin{align*} &1\in M,\qquad M\cap M^{-1}=\{1\}, \tag{44a}\\ &M=\langle I(M)\rangle_{\rm mon},\qquad G=\langle M\rangle_{\rm grp}, \tag{44b}\\ &\forall a\in I(M)\ \forall x,y\in M,\quad a\mid_M xy\Rightarrow a\mid_M x\text{ or }a\mid_M y. \tag{44c} \end{align*}

Here a∣Mxa\mid_M x means x=arx=ar for some r∈Mr\in M. Products and divisibility use D\mathsf D. The first equality in (44b) is expressed by

∀T∈L(G),(1∈T ∧ I(M)⊆T)⇒M≤T.\forall T\in\mathcal L(G),\quad \bigl(1\in T\ \wedge\ I(M)\subseteq T\bigr)\Rightarrow M\leq T.

The second uses the subgroup formula (20): every subgroup containing MM is GG. Thus no quantification over arbitrary subsets or over lists is introduced into the sentence.

Lemma 6. If an abelian group GG contains a submonoid satisfying (44a)–(44c), then GG is free abelian on I(M)I(M). Conversely, the nonnegative cone of a free abelian group has these properties.

Proof. A divisor in I(M)I(M) of another element of I(M)I(M) equals that element: the remaining factor must be a unit, hence 11 by (44a). Primality therefore implies that an irreducible dividing a finite product of irreducibles occurs among its factors. Cancelling that factor and inducting on the number of factors proves that the product map

N(I(M))⟶M\mathbb N^{(I(M))}\longrightarrow M

is injective. It is surjective by (44b). The induced group homomorphism

Z(I(M))⟶G\mathbb Z^{(I(M))}\longrightarrow G

is injective: split a finitely supported integer vector into its positive and negative parts, and apply injectivity of the positive product map. Its image contains MM, so it is surjective by the second part of (44b).

For G=Z(X)G=\mathbb Z^{(X)}, take M=N(X)M=\mathbb N^{(X)}. Its only unit is 00, its irreducibles are the standard basis vectors, and each is prime by the corresponding nonnegative coordinate. These vectors generate MM and generate GG as a group. ◻

Let Cyc\mathsf{Cyc} assert that the greatest lattice element is the least subgroup containing some element code. In a torsion-free group this means that the group is trivial or infinite cyclic. Define

Φab=TF ∧ AB ∧ (Cyc ∨ (Tot ∧ ∃M Cone(M))).(48) \Phi_{\rm ab} =\mathsf{TF}\ \wedge\ \mathsf{AB}\ \wedge\ \left(\mathsf{Cyc}\ \vee\ \left(\mathsf{Tot}\ \wedge\ \exists M\,\mathsf{Cone}(M)\right)\right). \tag{48}

Theorem 7. For every nonempty semigroup SS,

L(S)⊨Φab⟺S≅Z(X) as a semigroup for some set X.\mathcal L(S)\models\Phi_{\rm ab} \quad\Longleftrightarrow\quad S\cong\mathbb Z^{(X)}\text{ as a semigroup for some set }X.

In particular, the class in part (b) of Kourovka Problem 2.81 is recognized by a single finite sentence in the pure lattice language.

Proof. If Φab\Phi_{\rm ab} holds, (18) reconstructs a torsion-free group with the given multiplication, and (37) makes it abelian. The cyclic alternative is free abelian of rank zero or one. Otherwise (40) and Tot\mathsf{Tot} identify the multiplication used in Cone\mathsf{Cone} with the group multiplication, and Lemma 6 applies.

Conversely, a free abelian group satisfies TF\mathsf{TF} and AB\mathsf{AB}. In rank zero or one it satisfies Cyc\mathsf{Cyc}. In every other rank, Lemma 5 supplies Tot\mathsf{Tot} and the nonnegative cone supplies Cone\mathsf{Cone}. A semigroup isomorphism induces an isomorphism of subsemigroup lattices and preserves the sentence. ◻

Recovering unordered multiplication

The special-product relation of Section “Commutativity in the subsemigroup lattice” is also useful without commutativity. Let

CP(x,y)  ⟺  ∃cx,y∈⟨c⟩.\mathsf{CP}(x,y)\;\Longleftrightarrow\; \exists c\quad x,y\in\langle c\rangle.

This is a finite lattice formula: the least subgroup containing an element code is definable by (20). Its truth implies xy=yxxy=yx.

We impose two finite conditions on a torsion-free group GG:

¬CP(x,y) ⟹ ∃p≠qSP(x,y;p)∧SP(x,y;q),x≠1 ⟹ ∃HCyc(H)∧∀y (y∈H⟺CP(x,y)).\begin{align*} \neg\mathsf{CP}(x,y)&\ \Longrightarrow\ \exists p\neq q\quad\mathsf{SP}(x,y;p)\wedge\mathsf{SP}(x,y;q), \tag{51a}\\ x\neq1&\ \Longrightarrow\ \exists H\quad \mathsf{Cyc}(H)\wedge \forall y\,(y\in H\Longleftrightarrow\mathsf{CP}(x,y)). \tag{51b} \end{align*}

Here Cyc(H)\mathsf{Cyc}(H) means that HH is a cyclic subgroup, rather than the global sentence Cyc\mathsf{Cyc}. Denote their conjunction by CC\mathsf{CC}.

Lemma 8. If a torsion-free group satisfies CC\mathsf{CC}, every nonidentity centralizer is cyclic, and

CP(x,y)⟺xy=yx.\mathsf{CP}(x,y)\quad\Longleftrightarrow\quad xy=yx.

Every free group satisfies CC\mathsf{CC}.

Proof. If x,yx,y commute, (24) gives at most one special product. Thus (51a) implies CP(x,y)\mathsf{CP}(x,y). The reverse implication always holds. Consequently the subgroup HH in (51b) is exactly CG(x)C_G(x).

In a free group, CG(x)≅ZC_G(x)\cong\mathbb Z for x≠1x\ne1, so commuting pairs satisfy CP\mathsf{CP}. If [x,y]≠1[x,y]\ne1, the subgroup theorem for free groups gives

K=⟨x,y⟩≅F2,K/[K,K]≅Z2.K=\langle x,y\rangle\cong F_2, \qquad K/[K,K]\cong\mathbb Z^2.

Indeed, the rank is at most two by generation and is not at most one by noncommutativity. The images of x,yx,y generate Z2\mathbb Z^2 and form an integer basis. Every positive word equal to xyxy or yxyx has exactly one occurrence of each generator. The defining implications for SP\mathsf{SP} follow by substituting a positive word for the intermediate element and counting the two letters. Hence

SP(x,y;xy),SP(x,y;yx),xy≠yx,\mathsf{SP}(x,y;xy),\qquad \mathsf{SP}(x,y;yx),\qquad xy\neq yx,

which establish (51a). The centralizer itself supplies (51b). ◻

Lemma 9. Let GG be torsion-free, with cyclic centralizers of nonidentity elements. If x,u≠1x,u\neq1, u∈CG(x)u\in C_G(x) and gug−1∈CG(x)gug^{-1}\in C_G(x), then g∈CG(x)g\in C_G(x). In particular, conjugate elements of the same nonidentity centralizer are equal.

Proof. Elements commuting with the same nonidentity element lie in a cyclic group and commute with each other. Thus

CG(u)=CG(x)=CG(gug−1).C_G(u)=C_G(x)=C_G(gug^{-1}).

Conjugation by gg preserves CG(x)=⟨r⟩C_G(x)=\langle r\rangle. Its effect on rr is therefore r↦rr\mapsto r or r↦r−1r\mapsto r^{-1}. In the first case gg commutes with rr and hence with xx.

In the second case g≠1g\neq1, and g2≠1g^2\neq1 by torsion-freeness. Both gg and rr commute with g2g^2, so they belong to its cyclic centralizer and commute. This contradicts grg−1=r−1≠rgrg^{-1}=r^{-1}\neq r. Finally, if a,gag−1∈CG(x)a,gag^{-1}\in C_G(x) and a≠1a\neq1, the first assertion puts gg in the cyclic group CG(x)C_G(x), giving gag−1=agag^{-1}=a. The case a=1a=1 is immediate. ◻

Continue to write Π\Pi for the partial product defined before (39). Define J(x,y;p)\mathsf J(x,y;p) as the disjunction

(x=1∧p=y) ∨ (y=1∧p=x),x,y≠1 ∧ ¬CP(x,y) ∧ SP(x,y;p),x,y≠1 ∧ CP(x,y) ∧ CP(x,p) ∧ ∃z,u,v  [¬CP(x,z)∧Π(y,z;u)∧Π(x,u;v)∧Π(p,z;v)].\begin{align*} &(x=1\wedge p=y)\ \vee\ (y=1\wedge p=x), \tag{56a}\\ &x,y\neq1\ \wedge\ \neg\mathsf{CP}(x,y)\ \wedge\ \mathsf{SP}(x,y;p), \tag{56b}\\ &x,y\neq1\ \wedge\ \mathsf{CP}(x,y)\ \wedge\ \mathsf{CP}(x,p) \ \wedge\ \exists z,u,v\;\bigl[ \neg\mathsf{CP}(x,z)\nonumber\\[-2pt] &\hspace{37mm}\wedge\Pi(y,z;u)\wedge\Pi(x,u;v)\wedge\Pi(p,z;v)\bigr]. \tag{56c} \end{align*}

All variables range over element codes. Thus J\mathsf J is a fixed finite lattice formula.

Proposition 10. For a torsion-free group satisfying CC\mathsf{CC},

J(x,y;p)⟹p=xy or p=yx.\mathsf J(x,y;p)\quad\Longrightarrow\quad p=xy\text{ or }p=yx.

For every noncyclic free group, the converse also holds.

Proof. The first two alternatives follow from (24). In (56c), let C=CG(x)C=C_G(x). Then x,y,p∈Cx,y,p\in C, x,y≠1x,y\neq1, and z∉Cz\notin C. The partial-product witnesses give

v∈{xyz,yzx,xzy,zyx}∩{pz,zp}.v\in\{xyz,yzx,xzy,zyx\}\cap\{pz,zp\}.

Since xy=yxxy=yx, the first and last possibilities are xyzxyz and zxyzxy. Equality xyz=pzxyz=pz or zxy=zpzxy=zp gives p=xyp=xy by cancellation. Equality xyz=zpxyz=zp or zxy=pzzxy=pz makes pp conjugate to xyxy inside CC, so Lemma 9 gives the same conclusion, including xy=1xy=1.

Each remaining equality contradicts malnormality. More explicitly,

xzy=pz ⟹ zyz−1=x−1p∈C,xzy=zp ⟹ z−1xz=py−1∈C,yzx=pz ⟹ zxz−1=y−1p∈C,yzx=zp ⟹ z−1yz=px−1∈C.\begin{align*} xzy=pz&\ \Longrightarrow\ zyz^{-1}=x^{-1}p\in C,& xzy=zp&\ \Longrightarrow\ z^{-1}xz=py^{-1}\in C,\\ yzx=pz&\ \Longrightarrow\ zxz^{-1}=y^{-1}p\in C,& yzx=zp&\ \Longrightarrow\ z^{-1}yz=px^{-1}\in C. \end{align*}

In all four cases a nonidentity element of CC is conjugated into CC by zz or z−1z^{-1}, contrary to z∉Cz\notin C.

For completeness in a noncyclic free group, identity pairs and noncommuting pairs were treated in Lemma 8. Suppose x,y≠1x,y\neq1 commute and p=xy=yxp=xy=yx. Choose z∉CG(x)z\notin C_G(x); otherwise G=CG(x)G=C_G(x) would be cyclic. Cyclic centralizers imply that yy does not commute with zz, and that xx does not commute with yzyz. If xy≠1xy\neq1, then xyxy also does not commute with zz. Therefore

u=yz,v=xyzu=yz,\qquad v=xyz

satisfy all three partial products in (56c), using the special products for these noncommuting pairs. When xy=1xy=1, the final partial product is an identity case. This proves completeness. ◻

Subsemigroup cuts and free bases

The next criterion uses no cyclic-centralizer assumption in its sufficient direction. Let t∈Gt\in G have infinite order and let (bi)i∈I(b_i)_{i\in I} be a family of elements. For each i∈Ii\in I, suppose that a set Di⊆GD_i\subseteq G satisfies

1∈Di,bi∉Di,htn∈Di⟺h∈Di(h∈G, n∈Z),h∈Di⟺hbj∈Di(h≠1 or j≠i).\begin{align*} &1\in D_i,\qquad b_i\notin D_i, \tag{61a}\\ &ht^n\in D_i\quad\Longleftrightarrow\quad h\in D_i &&(h\in G,\ n\in\mathbb Z), \tag{61b}\\ &h\in D_i\quad\Longleftrightarrow\quad hb_j\in D_i &&(h\neq1\text{ or }j\neq i). \tag{61c} \end{align*}

The indices matter in (61c); different labels are not identified even if their values in GG agree.

Lemma 11. If G=⟨t,bi (i∈I)⟩grpG=\langle t,b_i\ (i\in I)\rangle_{\rm grp} and (61a)–(61c) hold, the homomorphism

F({∗}⊔I)⟶G,∗=t,i=bi,F(\{*\}\sqcup I)\longrightarrow G,\qquad *=t,\quad i=b_i,

is an isomorphism.

Proof. Let t^,b^i\widehat t,\widehat b_i be the free generators. We prove the stronger statement: whenever a finite signed word ww evaluates in GG to tnt^n, its value in the free group is t^n\widehat t^{n}. Induct on the number of letters of ww. For the empty word, the infinite order of tt gives n=0n=0. If the first letter is tεt^\varepsilon, ε∈{1,−1}\varepsilon\in\{1,-1\}, the remaining word evaluates to tn−εt^{n-\varepsilon}, so induction applies.

Suppose the first letter is bib_i. The path described by the remaining letters starts at bi∉Dib_i\notin D_i and ends at tn∈Dit^n\in D_i. Choose its first crossing into DiD_i. By (61b), the crossing is not a tt-letter. By (61c), the only possible crossing from outside to inside is the step bi→1b_i\to1 labelled bi−1b_i^{-1}. Thus

w=bi u bi−1 v,u=1 in G,v=tn in G.w=b_i\,u\,b_i^{-1}\,v, \qquad u=1\text{ in }G,\qquad v=t^n\text{ in }G.

Both uu and vv are shorter than ww. Their free values are respectively 11 and t^n\widehat t^{n}, so the asserted equality follows. If the first letter is bi−1b_i^{-1}, the remaining word runs from 1∈Di1\in D_i to bitn∉Dib_it^n\notin D_i. Its first crossing out of DiD_i is 1→bi1\to b_i, giving

w=bi−1 u bi vw=b_i^{-1}\,u\,b_i\,v

with the same values for u,vu,v, and the same argument applies.

Taking n=0n=0 proves that every word in the kernel is the identity in the free group. Generation gives surjectivity. ◻

These cuts can be required using actual subsemigroups. For each i∈Ii\in I, require Ui∈L(G)U_i\in\mathcal L(G) such that

t∈Ui,bitbi−1∉Ui,hth−1∈Ui⟺(hbj)t(hbj)−1∈Ui(h≠1 or j≠i).\begin{align*} &t\in U_i,\qquad b_itb_i^{-1}\notin U_i, \tag{65a}\\ &hth^{-1}\in U_i\quad\Longleftrightarrow\quad (hb_j)t(hb_j)^{-1}\in U_i &&(h\neq1\text{ or }j\neq i). \tag{65b} \end{align*}

Indeed, Di={h:hth−1∈Ui}D_i=\{h:hth^{-1}\in U_i\} satisfies (61a)–(61c); the power invariance follows from

(htn)t(htn)−1=hth−1.(ht^n)t(ht^n)^{-1}=hth^{-1}.

Lemma 12. A free basis {t}⊔{bi:i∈I}\{t\}\sqcup\{b_i:i\in I\} admits subsemigroups satisfying (65a)–(65b).

Proof. In the reduced-word model, let DiD_i consist of the elements whose reduced word does not begin with the positive letter bib_i. Right multiplication by a basis generator changes this property only on the edge 1→bi1\to b_i. To see this, cancellation at the right end cannot change the first letter unless the whole word cancels; in that case the two endpoints are precisely 1,bi1,b_i. This proves (61c), and applying it repeatedly to t,t−1t,t^{-1} proves (61b).

Let χ:F({t}⊔I)→Z\chi:F(\{t\}\sqcup I)\to\mathbb Z count the exponent sum of tt, and put

Ui=⟨hth−1:h∈Di⟩+.U_i=\bigl\langle hth^{-1}:h\in D_i\bigr\rangle_+.

Every displayed generator has χ\chi-value 11. Therefore a conjugate hth−1hth^{-1} belonging to UiU_i must be a single generator, rather than a product of two or more. If

hth−1=ktk−1,hth^{-1}=ktk^{-1},

then h−1k∈CG(t)=⟨t⟩h^{-1}k\in C_G(t)=\langle t\rangle. Consequently k=htnk=ht^n for some n∈Zn\in\mathbb Z, and (61b) gives h∈Dih\in D_i if and only if k∈Dik\in D_i. We conclude that

hth−1∈Ui⟺h∈Di.hth^{-1}\in U_i\quad\Longleftrightarrow\quad h\in D_i.

The required subsemigroup cuts now follow from the set cuts. ◻

For a submonoid M≤GM\leq G, use the same irreducible set I(M)I(M) as in Section “A finite sentence for free abelian groups”. Let FC(M,t)\mathsf{FC}(M,t) assert

1∈M,M∩M−1={1},t∈I(M),M=⟨I(M)⟩mon,G=⟨M⟩grp,\begin{align*} &1\in M,\qquad M\cap M^{-1}=\{1\},\qquad t\in I(M), \tag{70a}\\ &M=\langle I(M)\rangle_{\rm mon},\qquad G=\langle M\rangle_{\rm grp}, \tag{70b} \end{align*}

and, for every c∈I(M)∖{t}c\in I(M)\setminus\{t\}, the existence of a subsemigroup UU such that

t∈U,ctc−1∉U,hth−1∈U⟺(hd)t(hd)−1∈U(h≠1 or d≠c),\begin{align*} &t\in U,\qquad ctc^{-1}\notin U, \tag{71a}\\ &hth^{-1}\in U\quad\Longleftrightarrow\quad (hd)t(hd)^{-1}\in U \quad(h\neq1\text{ or }d\neq c), \tag{71b} \end{align*}

where h∈Gh\in G and d∈I(M)∖{t}d\in I(M)\setminus\{t\}. As before, generation by irreducibles is expressed by quantifying over subsemigroups containing 11 and every irreducible. Hence FC\mathsf{FC} is a finite formula in the lattice language together with a relation for multiplication on codes.

Proposition 13. A nontrivial torsion-free group GG is free if and only if FC(M,t)\mathsf{FC}(M,t) holds for some M∈L(G)M\in\mathcal L(G) and t∈Gt\in G.

Proof. Assume FC(M,t)\mathsf{FC}(M,t). Set I=I(M)∖{t}I=I(M)\setminus\{t\} and bc=cb_c=c. The two generation clauses imply

G=⟨{t}∪I⟩grp.G=\langle\{t\}\cup I\rangle_{\rm grp}.

Since t≠1t\neq1, it has infinite order. The cut clauses and Lemma 11 give a free basis.

Conversely, choose a generator tt of a nonempty free basis and let MM be the positive-word monoid on that basis. Word length shows that MM has no unit other than 11, and that its irreducibles are exactly the basis letters. They generate MM as a monoid and GG as a group. Lemma 12 supplies the cut clauses. Isomorphisms transport all the conditions. ◻

Eliminating the multiplication relation

We use Scott’s theorem in the following form (2): an injective map f:G→Hf:G\to H between groups satisfying

f(xy)∈{f(x)f(y),f(y)f(x)}(x,y∈G)f(xy)\in\{f(x)f(y),f(y)f(x)\}\qquad(x,y\in G)

is either a homomorphism or an antihomomorphism. There is no cardinality restriction. The noncommuting-pair proof in (6) gives the same global orientation and is the proof used here.

Let LL be any first-order structure, let D⊆LD\subseteq L be defined by a formula, and let JJ be a definable ternary relation. Name four elements

c=(e,a,b,p)\boldsymbol c=(e,a,b,p)

of LL, and add a ternary relation symbol RR. Consider the following finite theory Grp(R;c)\mathsf{Grp}(R;\boldsymbol c):

  1. R(x,y,z)R(x,y,z) implies x,y,z∈Dx,y,z\in D, and e∈De\in D.

  2. For x,y∈Dx,y\in D there is a unique zz with R(x,y,z)R(x,y,z).

  3. The operation on DD described by RR is associative, has identity ee, and every element has a left inverse.

  4. For x,y,z∈Dx,y,z\in D,

    J(x,y,z)⟺R(x,y,z)∨R(y,x,z).(75) J(x,y,z)\quad\Longleftrightarrow\quad R(x,y,z)\vee R(y,x,z). \tag{75}
  5. R(a,b,p)R(a,b,p) and ¬R(b,a,p)\neg R(b,a,p).

Each group condition is first-order in RR. For example, associativity is

R(x,y,u)∧R(y,z,v) ⟹ (R(u,z,w)⟺R(x,v,w)).R(x,y,u)\wedge R(y,z,v) \ \Longrightarrow\ \bigl(R(u,z,w)\Longleftrightarrow R(x,v,w)\bigr).

The identity conditions are R(e,x,x)R(e,x,x) and R(x,e,x)R(x,e,x) for x∈Dx\in D, and the inverse condition is ∀x∈D ∃y R(y,x,e)\forall x\in D\,\exists y\,R(y,x,e). Together with totality and uniqueness they define an ordinary group on the nonempty set DD.

Lemma 14. For a fixed LL and fixed c\boldsymbol c, at most one relation RR satisfies Grp(R;c)\mathsf{Grp}(R;\boldsymbol c).

Proof. Suppose R1,R2R_1,R_2 describe operations ∗1,∗2*_1,*_2. Equation (75) gives

{x∗1y,y∗1x}={x∗2y,y∗2x}(x,y∈D).\{x*_1y,y*_1x\}=\{x*_2y,y*_2x\} \qquad(x,y\in D).

The identity map is a bijective half-homomorphism between the two groups. By Scott’s theorem, x∗2y=x∗1yx*_2y=x*_1y for all pairs, or x∗2y=y∗1xx*_2y=y*_1x for all pairs. The latter alternative contradicts the common orientation:

a∗1b=p=a∗2b,b∗1a≠p.a*_1b=p=a*_2b,\qquad b*_1a\neq p.

Thus the operations agree on DD. Both relations are false outside D3D^3, so R1=R2R_1=R_2. ◻

Lemma 15. Let TT be a first-order theory in L∪{R}\mathscr L\cup\{R\}, where RR is a relation symbol. If each L\mathscr L-structure has at most one expansion satisfying TT, then there is an L\mathscr L-formula ρ\rho such that

T⊨∀x (R(x)⟺ρ(x)).T\models\forall\boldsymbol x\, \bigl(R(\boldsymbol x)\Longleftrightarrow\rho(\boldsymbol x)\bigr).

Proof. This is the implicit-to-explicit definability theorem of Beth (1), in its at-most-one-expansion form. One standard proof uses compactness and the joint-consistency theorem: two theories with no common expansion admit a sentence of their common language separating them. Take disjoint copies R0,R1R_0,R_1 of the added symbol and the corresponding renamed theories T0,T1T_0,T_1. Name a tuple d\boldsymbol d in the common base language. Implicit uniqueness makes T0∪{R0(d)}T_0\cup\{R_0(\boldsymbol d)\} and T1∪{¬R1(d)}T_1\cup\{\neg R_1(\boldsymbol d)\} jointly inconsistent. Separation yields a formula of the common base language; removing the names d\boldsymbol d gives the defining formula.

If TT is finite, its reducts are defined by the finite conjunction T[R←ρ]T[R\leftarrow\rho]. Indeed, a model of TT satisfies this conjunction because R=ρR=\rho there. Conversely, expand a reduct satisfying T[R←ρ]T[R\leftarrow\rho] by declaring R(x)R(\boldsymbol x) to mean ρ(x)\rho(\boldsymbol x). Induction on formulas shows that this expansion satisfies every sentence of TT. ◻

Apply the lemma to Grp\mathsf{Grp}, with the four named elements included in the base language. Removing their constant names gives a fixed formula

ρ(x,y,z;e,a,b,p)\rho(x,y,z;e,a,b,p)

in the original language, with seven free variables. Put

B(c)=Grp(R;c)[R←ρ].(81) \mathsf B(\boldsymbol c)=\mathsf{Grp}(R;\boldsymbol c)[R\leftarrow\rho]. \tag{81}

This is a finite formula in the base language. For any interpretation of c\boldsymbol c, its truth means precisely that interpreting RR by ρ(⋅;c)\rho(\cdot;\boldsymbol c) produces the required relational group.

Lemma 14 applies to every base structure LL. Thus Beth’s hypothesis holds before the specialization L=L(G)L=\mathcal L(G).

The finite sentence for free groups

Take L=Llat\mathscr L=\mathscr L_{\rm lat}, let DD be the element-code predicate C\mathsf C, and let J=JJ=\mathsf J from Section “Recovering unordered multiplication”. Let ρ\rho and B\mathsf B be the resulting formulas. In FC(M,t)\mathsf{FC}(M,t), replace every multiplication relation by ρ\rho, and denote the resulting pure lattice formula by

FCρ(M,t;c).\mathsf{FC}_{\rho}(M,t;\boldsymbol c).

Inversion and conjugation in the cut clauses are expressed through existentially quantified intermediate products and the identity parameter ee. All intermediate elements are required to be C\mathsf C-codes. The quantifiers over MM and the cut subsemigroups remain ordinary lattice quantifiers.

Define

Φfr=TF ∧ [Cyc ∨ (CC ∧ ∃c (B(c) ∧ ∃M,t  FCρ(M,t;c)))].(83) \begin{aligned} \Phi_{\rm fr}=\mathsf{TF}\ \wedge\ \bigl[\mathsf{Cyc}\ \vee\ \bigl(\mathsf{CC}\ \wedge\ \exists\boldsymbol c\, (\mathsf B(\boldsymbol c)\ \wedge\ \exists M,t\;\mathsf{FC}_{\rho}(M,t;\boldsymbol c))\bigr)\bigr]. \end{aligned} \tag{83}

The variables e,a,b,p,te,a,b,p,t are lattice variables representing elements by their codes. Each displayed abbreviation expands to a finite first-order formula. In particular, (83) contains no quantifier over relations.

Lemma 16. Suppose a torsion-free group GG satisfies CC\mathsf{CC} and its subsemigroup lattice has an expansion satisfying Grp\mathsf{Grp}. Under the identification g↔Xgg\leftrightarrow X_g, the multiplication defined by RR is either the multiplication of GG or its opposite.

Proof. Let ∗* be the relational group operation on the element codes. By (75) and Proposition 10,

Xx∗Xy=Xz⟹z=xy or z=yx.X_x*X_y=X_z\quad\Longrightarrow\quad z=xy\text{ or }z=yx.

The decoding bijection Xx↦xX_x\mapsto x is therefore a half-isomorphism from the relational group to GG. Scott’s theorem makes it an isomorphism or an anti-isomorphism. Only soundness of J\mathsf J was used here; its completeness is not assumed for this GG. ◻

Proof of (3a). Assume L(S)⊨Φfr\mathcal L(S)\models\Phi_{\rm fr}. By (18), the multiplication of SS makes it a torsion-free group GG, up to semigroup isomorphism. If the cyclic alternative holds, GG is free of rank zero or one.

In the other alternative, interpret RR using the formula ρ\rho at the chosen parameters. Lemma 15 gives the relational group axioms, and Lemma 16 identifies its multiplication with that of GG or GopG^{\rm op}. Reversing multiplication preserves every subsemigroup as a set, and hence induces

L(G)≅L(Gop),Xg⟼Xgop.\mathcal L(G)\cong\mathcal L(G^{\rm op}),\qquad X_g\longmapsto X_{g^{\rm op}}.

Consequently the chosen M,tM,t satisfy FC\mathsf{FC} in GG or GopG^{\rm op}. Proposition 13 proves that group free. Inversion g↦(g−1)opg\mapsto(g^{-1})^{\rm op} is an isomorphism G≅GopG\cong G^{\rm op}, so GG is free in either case.

Conversely, let GG be free. In ranks zero and one it satisfies Cyc\mathsf{Cyc}. Otherwise GG is noncyclic and nonabelian. Lemma 8 gives CC\mathsf{CC} and Proposition 10 gives

J(x,y;p)⟺p=xy or p=yx.\mathsf J(x,y;p)\quad\Longleftrightarrow\quad p=xy\text{ or }p=yx.

Choose a,ba,b with ab≠baab\neq ba and put e=1e=1, p=abp=ab, all represented by their element codes. The actual multiplication graph satisfies Grp\mathsf{Grp}. By Beth definability, ρ\rho agrees with this graph. A nonempty free basis gives M,tM,t satisfying FC\mathsf{FC} by Proposition 13. Hence (83) holds. Transport along a semigroup isomorphism finishes the proof for arbitrary SS. ◻

Theorem 7 proves (3b), so both assertions of Theorem 1 follow. Adjoining the finitely many lattice identities to each of the two sentences gives the required finitely axiomatized classes of lattices.

References

Preprint · Lean (GitHub)

  1. E. W. Beth, On Padoa’s method in the theory of definition, Indagationes Mathematicae (Proceedings) 56 (1953), 330–339. doi:10.1016/S1385-7258(53)50042-3.
  1. W. R. Scott, Half-homomorphisms of groups, Proc. Amer. Math. Soc. 8, no. 6 (1957), 1141–1144. doi:10.1090/S0002-9939-1957-0095890-3.
  1. L. N. Shevrin, On lattice properties of semigroups, Sibirsk. Mat. Zh. 3, no. 3 (1962), 446–470 (Russian).
  1. L. N. Shevrin, The lattice-subsemigroup characterization of commutative nonperiodic groups, Sibirsk. Mat. Zh. 5, no. 3 (1964), 671–678 (Russian).
  1. L. N. Shevrin, Elementary lattice properties of semigroups, Sibirsk. Mat. Zh. 7, no. 3 (1966), 664–684 (Russian).
  1. W. Vipismakul, Another proof of half homomorphisms, Burapha Science Journal 25, no. 1 (2020), 187–193.