Bourbaki

Appendix

Characterization of terms and relations

Metamathematics, when it goes beyond the very elementary level of the present chapter, makes considerable use of the results of mathematics, as we remarked in the Introduction. The purpose of this Appendix is to give a simple example of this type of reasoning [^1]. We shall begin by establishing certain results which belong to the mathematical theory of free semigroups; we shall then use these results in a metamathematical " application " to obtain a characterization of the terms and relations in a theory.

1. Signs and words

\* Let S be a non-empty set, whose elements will be called signs (this terminology being appropriate to the metamathematical application we have in mind). Let L0(S)\mathrm{L}_0(\mathrm{S}) be the free semigroup generated by S; the elements of L0(S)\mathrm{L}_0(\mathrm{S}) are called words and are identified with finite sequences A=(si)0⩽i⩽nA = (s_i)_{0 \leqslant i \leqslant n} of elements of S. The law of composition in L0(S)\mathrm{L}_0(\mathrm{S}) will be written multiplicatively, so that ABAB is the sequence obtained by juxtaposition of AA and BB. The empty word ∅\emptyset is the identity element of L0(S)\mathrm{L}_0(\mathrm{S}). We recall that the length l(A)l(A) of a word A∈L0(S)A \in \mathrm{L}_0(\mathrm{S}) is the number of elements in the sequence AA; thus l(AB)=l(A)+l(B)l(AB) = l(A) + l(B), and the words of length 1 are the signs. Let L(S)\mathrm{L}(\mathrm{S}) denote the set of non-empty words in L0(S)\mathrm{L}_0(\mathrm{S}). ¶ Suppose, moreover, that we are given a mapping s→n(s)s \rightarrow n(s) of S into the set N\mathbf{N} of integers ⩾0\geqslant 0. For each non-empty word A=(si)0⩽i⩽kA = (s_i)_{0 \leqslant i \leqslant k} of L(S), we put

n(A)  =  ∑i=0k n(si)n(A) \;=\; \sum_{i=0}^{k} \, n(s_i)

and n(∅)=0n(\emptyset) = 0. n(A)n(A) is called the weight of AA. Clearly n(AB)=n(A)+n(B)n(AB) = n(A) + n(B). ¶ If A=A′BA′′A = A'BA'', the word BB is said to be a segment of AA (a proper segment if also B≠AB \neq A). If A′A' (resp. A′′A'') is empty, BB is said to be an initial (resp. final) segment of AA. If l(A′)=kl(A') = k, then BB is said to begin at the (k+1)(k+1)th place.

¶ If A=BCDEFA = BCDEF (where the words BB, CC, DD, EE, FF may be empty) the segments CC and EE of AA are said to be disjoint.

2. Significant words

A significant sequence is any sequence (Aj)1⩽j⩽n(A_j)_{1 \leqslant j \leqslant n} of words of L0(S)\mathrm{L}_0(\mathrm{S}) with the following property : for each word AiA_i of the sequence, one of the following two conditions is satisfied :

(1) AiA_i is a sign of weight 0.

(2) There exist pp words Ai1A_{i_1}, Ai2A_{i_2}, ..., AipA_{i_p} in the sequence, with indices less than ii, and a sign ff of weight pp such that

Ai=f Ai1Ai2 … Aip.A_i = f\, A_{i_1} A_{i_2} \, \ldots \, A_{i_p}.

Words which appear in significant sequences are called significant words. Then we have :

Proposition 1 03P6

If A1A_1, A2A_2, ..., ApA_p are pp significant words and if ff is a sign of weight pp, then the word fA1A2…Apf A_1 A_2 \ldots A_p is significant.

3. Characterization of significant words

A word A∈L0(S)A \in \mathrm{L}_0(\mathrm{S}) is said to be balanced if it has the following two properties :

(1) l(A)=n(A)+1l(A) = n(A) + 1 (which implies that AA is not empty).

(2) For every proper initial segment BB of AA, l(B)⩽n(B)l(B) \leqslant n(B).

Proposition 2 03P7

A word is significant if and only if it is balanced.

Let AA be a significant word belonging to a significant sequence

A1,  A2,  …,  An.A_1, \; A_2, \; \ldots, \; A_n.

We shall show by induction on kk that each AkA_k is balanced. Suppose that this has been established for the AjA_j with index j<kj < k, and let us show that it is true for AkA_k. If AkA_k is a sign of weight 0 (which is the only possibility when k=1k = 1), AkA_k is balanced because

l(Ak)=1andn(Ak)=0.l(A_k) = 1 \qquad \text{and} \qquad n(A_k) = 0.

If AkA_k is not a sign of weight 0, then Ak=fB1B2…BpA_k = fB_1B_2 \ldots B_p, where ff is a sign of weight pp, and the BjB_j are of the form AijA_{i_j}, where ij<ki_j < k, and are therefore balanced words by the inductive hypothesis. We have

l(Ak)=1+l(B1)+l(B2)+⋯+l(Bp)=1+(n(B1)+1)+(n(B2)+1)+⋯+(n(Bp)+1)=1+p+n(B1)+n(B2)+⋯+n(Bp)=1+n(Ak).\begin{aligned} l(A_k) &= 1 + l(B_1) + l(B_2) + \cdots + l(B_p) \\ &= 1 + (n(B_1) + 1) + (n(B_2) + 1) + \cdots + (n(B_p) + 1) \\ &= 1 + p + n(B_1) + n(B_2) + \cdots + n(B_p) \\ &= 1 + n(A_k). \end{aligned}

Let CC be a proper initial segment of AA, and let qq be the largest of the integers m<pm < p such that BmB_m is a segment of CC, so that

C=fB1B2…BqD,C = f B_1 B_2 \ldots B_q D,

where DD is a proper initial segment of Bq+1B_{q+1}. Then

l(C)=1+l(B1)+l(B2)+⋯+l(Bq)+l(D)⩽1+(n(B1)+1)+(n(B2)+1)+⋯+(n(Bq)+1)+n(D)⩽p+n(B1)+⋯+n(Bq)+n(D)=n(C).\begin{aligned} l(C) &= 1 + l(B_1) + l(B_2) + \cdots + l(B_q) + l(D) \\ &\leqslant 1 + (n(B_1) + 1) + (n(B_2) + 1) + \cdots + (n(B_q) + 1) + n(D) \\ &\leqslant p + n(B_1) + \cdots + n(B_q) + n(D) = n(C). \end{aligned}

Hence AkA_k is balanced.

¶ To prove that, conversely, every balanced word is significant, we need the following two lemmas :

Lemma 1 03P8

Let A be a balanced word. Then for each integer k such that 0⩽k<l(A)0 \leqslant k < l(A) there exists exactly one balanced segment S of A which begins at the (k+1)(k + 1)th place.

The uniqueness of SS is an immediate consequence of the following remark : if TT is a balanced word, then by definition no proper initial segment of TT is balanced. Let us prove the existence of SS. Write A=BCA = BC where l(B)=kl(B) = k. For each ii such that 0⩽i⩽q=l(C)0 \leqslant i \leqslant q = l(C), let CiC_i be the initial segment of CC of length ii. Since BB is a proper initial segment of AA, we have

l(Cq)=l(A)−l(B)⩾n(A)+1−n(B)=n(Cq)+1.l(C_q) = l(A) - l(B) \geqslant n(A) + 1 - n(B) = n(C_q) + 1.

On the other hand, we have 0=l(C0)⩽n(C0)=00 = l(C_0) \leqslant n(C_0) = 0. Let ii be the largest of the integers j<qj < q such that l(Ch)⩽n(Ch)l(C_h) \leqslant n(C_h) for 0⩽h⩽j0 \leqslant h \leqslant j; then we have l(Ci)≤n(Ci)l(C_i)\leq n(C_i) and l(Ci+1)≥n(Ci+1)+1l(C_{i+1})\geq n(C_{i+1})+1. We show that Ci+1C_{i+1} is balanced: the condition relating to proper initial segments is satisfied by reason of the definition of ii; on the other hand, we have

n(Ci+1)+1≤l(Ci+1)=l(Ci)+1≤n(Ci)+1≤n(Ci+1)+1, n(C_{i+1})+1\leq l(C_{i+1})=l(C_i)+1\leq n(C_i)+1\leq n(C_{i+1})+1,

so that l(Ci+1)=n(Ci+1)+1l(C_{i+1})=n(C_{i+1})+1, and the proof of Lemma 1 is complete.

Lemma 2 03GT

Every balanced word AA can be put in the form

A=fA1A2…Ap, A=fA_1A_2\ldots A_p,

where the AiA_i are balanced and n(f)=pn(f)=p.

Let ff be the initial sign of AA. By Lemma 1, AA can be written as

fA1A2…Ap, fA_1A_2\ldots A_p,

where the AiA_i are balanced : we need only define AiA_i inductively as the balanced segment of AA which begins at the k(i)k(i)th place, where

k(i)=2+∑j<il(Aj). k(i)=2+\sum_{j<i}l(A_j).

Moreover, we have

1+l(A1)+…+l(Ap)=l(A)=n(A)+1 1+l(A_1)+\ldots+l(A_p)=l(A)=n(A)+1
=n(f)+n(A1)+…+n(Ap)+1 =n(f)+n(A_1)+\ldots+n(A_p)+1
=n(f)+(l(A1)−1)+…+(l(Ap)−1)+1, =n(f)+(l(A_1)-1)+\ldots+(l(A_p)-1)+1,

from which it follows that n(f)=pn(f)=p.

¶ Now that the two lemmas have been proved, it is obvious by induction on the length of AA that every balanced word AA is significant, by reason of Lemma 2 and Proposition 1.

Corollary 1 03GU

Let AA be a significant word. For each integer kk such that 0≤k<l(A)0\leq k<l(A) there is exactly one significant segment of AA which begins at the (k+1)(k+1)th place.

Corollary 2 03GV

Every significant word may be written in exactly one way in the form fA1A2…ApfA_1A_2\ldots A_p, where the AiA_i are significant and n(f)=p.n(f)=p.

4. Application to assemblies in a mathematical theory

Suppose that the set S is the set of signs of a mathematical theory T\mathscr{T}. We put n(□)=0n(\square)=0, n(τ)=n(η)=1n(\tau)=n(\eta)=1, n(∨)=2n(\vee)=2, n(x)=0n(x)=0 for every letter xx;

and finally, for every specific sign ss of T\mathscr{T}, n(s)n(s) is the weight of ss, which is fixed when T\mathscr{T} is given.

¶ Let AA be an assembly in T\mathscr{T}. We denote by A∗A^* the word obtained by deleting the links in AA, and we shall say that AA is balanced if A∗A^* is balanced [in L0(S)L_0(S)]. A segment of AA is any assembly obtained by replacing, in a segment SS of A∗A^*, the links which, in AA, join pairs of signs in SS.

CRITERION 1. If AA is a term or a relation in T\mathscr{T}, then AA is balanced.

Let A1A_1, A2A_2, …\ldots, AnA_n be a formative construction in T\mathscr{T}, in which AA appears. Let us argue by induction and suppose that we have proved that the AjA_j of indices j<ij<i are balanced; then we have to prove that AiA_i is balanced. The proof goes just as in the first part of the proof of Proposition 2, except when AiA_i is of the form τx(B)\tau_x(B) with B=AjB=A_j and j<ij<i. In this case, let CC be the assembly obtained by replacing xx, wherever it occurs in BB, by □\square. The word A∗A^* is then identical with τC∗\tau C^*; now B∗B^* is balanced, and therefore C∗C^* is balanced (because n(□)=n(x)=0n(\square)=n(x)=0). Consequently Ai∗A_i^* is balanced.

¶ We have thus obtained a necessary condition for an assembly in T\mathscr{T} to be a term or a relation. But this condition is not sufficient, as we shall see.

¶ Let AA be a balanced assembly in T\mathscr{T}. If AA begins with a letter or a □\square, then AA must consist of this sign alone (Proposition 2, Corollary 2). For all other cases, we shall now define the assembly or assemblies antecedent to AA.

(1) If AA begins with a ∧\land, or a ∨\vee, or a specific sign, then A∗A^* can be put in exactly one way in the form fB1B2…BpfB_1B_2\ldots B_p, where ff is a sign of weight p≥1p\geq 1 and the BiB_i are balanced (Proposition 2, Corollary 2). The segments A1A_1, A2A_2, …\ldots, ApA_p of AA which correspond to the segments B1B_1, B2B_2, …\ldots, BpB_p of A∗A^* are called the assemblies antecedent to AA. Moreover, we shall say that AA is perfectly balanced if AA is identical with

fA1A2…Ap, fA_1A_2\ldots A_p,

in other words if no link in AA joins ff to one of the BiB_i, or joins two distinct BiB_i.

(2) If AA begins with a τ\tau, then A∗A^* is of the form τB\tau B, where BB is balanced (Proposition 2, Corollary 2). In this case an assembly antecedent to AA is any of the assemblies A1A_1 defined as follows : replace the signs □\square in BB which are linked in AA to the initial τ\tau by a letter xx distinct from the other letters which appear in BB, and replace the links which join two signs of BB in AA. (If, instead of xx, we substitute a letter yy which also does not appear in BB, we obtain an assembly which is just (y∣x)A1(y|x)A_1.) Moreover, we shall say that AA is perfectly balanced if AA is identical with τx(A1)\tau_x(A_1), in other words if no link joins the initial τ\tau to any sign of BB other than a □\square.

¶ We can now state the following criterion :

C2. Let AA be a balanced assembly in T\mathscr{T}.

For AA to be a term it is necessary and sufficient that one of the following conditions be satisfied : (1) AA consists of a single letter; (2) AA begins with a τ\tau, is perfectly balanced, and its antecedent assemblies are relations (by CF8, it is enough to verify that one antecedent assembly is a relation); (3) AA begins with a substantific sign, is perfectly balanced, and its antecedent assemblies are terms.For AA to be a relation it is necessary and sufficient that one of the following conditions be satisfied : (1) AA begins with a ∨\vee or a ¬\neg, is perfectly balanced, and its antecedent assemblies are relations; (2) AA begins with a relational sign, is perfectly balanced, and its antecedent assemblies are terms.

The criteria CF1 to CF4 (§ 1, no. 4) show that the conditions are sufficient. Let us show that they are necessary. We have already seen (§ 1, no. 3) that if AA is a relation, then AA begins with a ∨\vee, or a ¬\neg, or a relational sign. The reasoning is similar in each of the three cases. If, for example, AA begins with a ∨\vee, then AA is of the form ∨BC\vee BC, where BB and CC are relations, so that BB and CC are the assemblies antecedent to AA; hence AA is perfectly balanced. If AA is a term, then there are two cases : it consists of a single letter, or it begins with either a substantific sign or a τ\tau. In the second case, we argue as above. If AA begins with a τ\tau, the definition of a formative construction shows that AA is of the form τx(B)\tau_x(B), where BB is a relation and xx a letter, so that we may take BB as assembly antecedent to AA, and AA is perfectly balanced.

When we wish to test whether a given assembly AA (not consisting of a single letter) is a relation (resp. a term) in T\mathscr{T}, we first verify that AA is balanced and that it begins with a ∨\vee, a ¬\neg, or a relational sign (resp. with a τ\tau or a substantific sign). Then we form the antecedent assembly or assemblies, and verify (if appropriate) that AA is perfectly balanced. Having done this, we are left with an analogous problem relating to shorter assemblies. Thus, step by step, we come down to assemblies each of which consists of a single sign, for which the solution is immediate.

Remark 03GW

Except in certain mathematical theories which are particularly weak in axioms (cf. Exercise 7) we have no general procedure of this type to enable us to test whether or not a given relation RR in a theory T\mathscr{T} is a theorem in T\mathscr{T}.

Exercises

See the exercises for Appendix 0.

[^1]: The results established in this Appendix will not be used anywhere else in this series.

Transcribed from the printed edition, 2004, Springer. Not translated.