Self-similar measures and the Rajchman property

For Bernoulli convolutions, the convergence to zero of the Fourier transform at inﬁnity was characterized by successive works of Erd¨os [4] and Salem [17]. We provide a quasi-complete extension of these results to general self-similar measures on the real line


Introduction
Rajchman measures.In the present article we consider the question of extending some classical results on Bernoulli convolutions to a more general context of self-similar measures.For a Borel probability measure µ on R, define its Fourier transform as : We say that µ is Rajchman, whenever μ(t) → 0, as t → +∞.When µ is a Borel probability measure on the torus T = R\Z, we introduce its Fourier coefficients, defined as : In this study, starting from a Borel probability measure µ on R, Borel probability measures on T will naturally appear, quantifying the non-Rajchman character of µ.
For a Borel probability measure µ on R, the Rajchman property holds for example if µ is absolutely continuous with respect to Lebesgue measure L R , by the Riemann-Lebesgue lemma.The situation can be more subtle and for instance there exist Cantor sets of zero Lebesgue measure and even of zero-Hausdorff dimension which support a Rajchman measure; cf Menshov [13], Bluhm [2].Questions on the Rajchman property of a measure naturally arise in Harmonic Analysis, for example when studying sets of multiplicity for trigonometric series; cf Lyons [12] or Zygmund [29].We shall say a word on this topic at the end of the article.A classical counter-example is the uniform measure µ on the standard middle-third Cantor set, which is a continuous singular measure, not Setting ν = L(X), we obtain that L(X n ) weakly converges to ν.By construction, we have L(X n+1 ) = 0≤j≤N p j L(X n ) • ϕ −1 j .Taking the limit as n → +∞, the measure ν verifies : The previous convergence implies that the solution of this "stable fixed point equation" is unique among Borel probability measures.Also, ν has to be of pure type, i.e. either purely atomic or absolutely continuous with respect to L R or else singular continuous, since each term in its Radon-Nikodym decomposition with respect to L R verifies (1).A few remarks are in order : i) If p ∈ C N , the measure ν is purely atomic if and only if the ϕ j have a common fixed point c, in which case ν is the Dirac mass at c. Indeed, consider the necessity and suppose that ν has an atom.Let a > 0 be the maximal mass of an atom and E the finite set of points having mass a. Fixing any c ∈ E, the relation ν({c}) = j p j ν({ϕ −1 j (c)}) furnishes ϕ −1 j (c) ∈ E, 0 ≤ j ≤ N .Hence ϕ −n j (c) ∈ E, n ≥ 0, for all j.If ϕ j = id, then ϕ −1 j (c) = c, the set {ϕ −n j (c), n ≥ 0} being infinite otherwise.If ϕ j = id, it fixes all points.
ii) The equation for a hypothetical density f of ν with respect to L R , coming from (1), is : This "unstable fixed point equation" is difficult to solve directly.It is equivalently reformulated into the fact that ((r ) n≥0 is a non-negative martingale (for its natural filtration), for Lebesgue a.-e.x ∈ R. Notice that when f exists and is bounded, then p j ≤ r j for all j, because p j r −1 iii) Let f (x) = ax + b be an affine map, with a = 0.With the same p ∈ D N (r), consider the conjugate system (ψ j ) 0≤j≤N , with ψ j (x) = f • ϕ j • f −1 (x) = r j x + b(1 − r j ) + ab j .It has an invariant measure w = L(aX + b) verifying the relation ŵ(t) = ν(at)e 2iπtb , t ∈ R. In particular, ν is Rajchman if and only if w is Rajchman.iv) When supposing that the (ϕ k ) 0≤k≤N are strict contractions, some self-similar set F can be introduced, where F ⊂ R is the unique non-empty compact set verifying the self-similarity relation F = ∪ 0≤k≤N ϕ k (F ).See for example Huchinson [7] for general properties of such sets.Introducing N = {0, 1, • • • } and the compact S = {0, • • • , N } N , the hypothesis that the (ϕ k ) 0≤k≤N are strict contractions implies that F is a continuous (and even hölderian) image of S, in other words we have the following description : Whereas in the general case a self-similar invariant measure can have R as topological support, when the (ϕ k ) 0≤k≤N are strict contractions the compact self-similar set F exists and supports any self-similar measure.
Background and content of the article.Coming back to the general case, we assume in the sequel that the (ϕ j ) 0≤j≤N do not have a common fixed point (in particular N ≥ 1), so that µ is a continuous measure.A difficult problem is to characterize the absolute continuity of ν with respect to L R in terms of the parameters r, b and p.An example with a long and well-known history is that of Bernoulli convolutions, corresponding to N = 1, the affine contractions ϕ 0 (x) = λx − 1, ϕ 1 (x) = λx + 1, 0 < λ < 1, and the probability vector p = (1/2, 1/2).Notice that when the r i are all equal (to some real in (0, 1)), the situation is a little simplified, as ν is an infinite convolution (this is not true in general).Although we discuss below some works in this context, we will not present here the vast subject of Bernoulli convolutions, addressing the reader to detailed surveys, such as Peres-Schlag-Solomyak [15] or Solomyak [21].
For general self-similar measures, an important aspect of the problem, that we shall not enter, and an active line of research, concerns the Hausdorff dimension of the measure ν; cf the recent fundamental work of Hochman [6] for example.In a large generality, see Falconer [5] and more recently Jaroszewska and Rams [9], there is an "entropy/Lyapunov exponent" upper-bound : The quantity s(p, r) is called the singularity dimension of the measure and can be > 1.The equality Dim H (ν) = 1 does not mean that ν is absolutely continuous, but the inequality s(p, r) < 1 surely implies that ν is singular.The interesting domain of parameters for the question of the absolute continuity of the invariant measure therefore corresponds to s(p, r) ≥ 1.
We focus in this work on another fundamental tool, the Fourier transform ν.If ν is not Rajchman, the Riemann-Lebesgue lemma implies that ν is singular.This property was used by Erdös [4] in the context of Bernoulli convolutions.He proved that if 1/2 < λ < 1 is such that 1/λ is a Pisot number, then ν is not Rajchman.The reciprocal statement (for 1/2 < λ < 1) was next shown by Salem [17].As a result, for Bernoulli convolutions the Rajchman property always holds, except for a very particular set of parameters.Some works have next focused on the decay on average of the Fourier transform for general self-similar measures associated to strict contractions; cf Strichartz [24,25], Tsuji [26].In the same context, the non-Rajchman character was recently shown to hold for only a very small set of parameters by Li and Sahlsten [11], who showed that ν is Rajchman when log r i / log r j is irrational for some (i, j), with moreover some logarithmic decay of ν at infinity, under a Diophantine condition.Next, Solomyak [22] proved that outside a set of r of zero Hausdorff dimension, ν even has a power decay at infinity.The aim of the present article is to study for general self-similar measures the exceptional set of parameters where the Rajchman property is not true, trying to follow the line of [4] and [17].We essentially show that r and b have to be closely related to some fixed Pisot number, as for Bernoulli convolutions.We first prove a general extension of the result of Salem [17], reducing to a small island the set of parameters where the Rajchman property may not hold.Focusing then on this island of parameters, we provide a general characterization of the Rajchman character, appearing in this particular case as equivalent to absolute continuity with respect to L R .Next, supposing that the (ϕ k ) 0≤k≤N are strict contractions, we prove a partial extension of the theorem of Erdös [4], showing that for most parameters in the small island the Rajchman property is not true, with in general a few exceptions.We finally give some complements, first rather surprising numerical simulations involving the Plastic number, then an application to sets of uniqueness for trigonometric series.

Statement of the results
Let us place in the general context considered in the Introduction.Pisot numbers will play a central role in the analysis.Let us introduce a few definitions concerning Algebraic Number Theory; cf for example Samuel [19] for more details.

Definition 2.1
A Pisot number is a real algebraic integer θ > 1, with conjugates (the other roots of its minimal unitary polynomial) of modulus strictly less than 1.Fixing such a θ > 1, denote its minimal polynomial as We write D(θ) for its Z-dual (as a Z-lattice), i.e. : As a classical fact, T r θ (θ n α) ∈ Z, for all n ≥ 0, if this holds for 0 ≤ n ≤ s.Define : Remark.-In the context of the previous definition, introduce the integer-valued (s + 1) × (s + 1)companion matrix M of Q : One may show that for any µ ∈ Q[θ], setting V = (T r θ (θ 0 µ), • • • , T r θ (θ s µ)), then µ ∈ T (θ) if and only if there exists n ≥ 0 such that V M n has integral entries.
We introduce special families of affine maps, that will play the role of canonical models for the analysis of the Rajchman property.
As a first result, extending [17], the analysis of the non-Rajchman character of the invariant measure requires to consider families in Pisot form.
with no common fixed point, and 0≤j≤N p j log r j < 0. The invariant measure ν is not Rajchman if and only if there exists f (x) = ax + b, a = 0, such that the conjugate system (f • ϕ j • f −1 ) 0≤j≤N is in Pisot form, for some Pisot number 1/λ > 1, with invariant measure w verifying ŵ(λ −k ) → k+∞ 0.
In particular, one gets that r j = λ nj , for all j, for some Pisot number 1/λ > 1 et relatively prime integers (n k ) 0≤k≤N .Hence, up to an affine change of variables, the non-Rajchman character of the invariant measure ν can be read on the sequence (λ −k ) k≥0 , as in [4].In a second step, we provide a general analysis of families in Pisot form.
In the sequel we use standard inner products and Euclidean norms on all spaces R n .
that 0≤j≤N p j n j > 0 and i.i.d.random variables (ε n ) n∈Z , with law p.Let (S l ) l∈Z be the cocycle notations associated to the (n εi ) i∈Z .The real random variable X = l≥0 µ ε l λ S l has law ν.
i) Let the T-valued random variables Z k = l∈Z µ ε l λ k+S l , k ∈ Z. Then λ −n X mod 1 converges, as n → +∞, to a probability measure m on T, verifying, for all f ∈ C(T, R) and all k ∈ Z : where as n → +∞, to a probability measure M on T s+1 , with one-dimensional marginals m, verifying : for all f ∈ C(T s+1 , R) and all k ∈ Z.
ii) If the (ϕ k ) 0≤k≤N do not have a common fixed point (i.e. if ν is continuous), denoting by Z a T s+1 -valued random variable with law M, then for any 0 = n = (n 0 , • • • , n s ) t ∈ Z s+1 , Z, n has a continuous law; in particular, m and M are continuous measures.If the (ϕ k ) 0≤k≤N have a common fixed point, there exists a rational number p/q such that m = δ p/q and M = (δ p/q ) ⊗(s+1) . iii In the context of the previous theorem, ν and M are always of the same nature, with respect to the uniform measure of the space they live on.In particular, M is also of pure type.We finally consider families in Pisot form, when supposing that the (ϕ k ) 0≤k≤N are strict contractions.
i) For any p ∈ C N , if the invariant measure ν is Rajchman, then it is absolutely continuous with respect to L R , with a density bounded and with compact support.
ii) There exists 0 = a ∈ Z such that for any k = 0, for any p ∈ C N outside finitely many realanalytic graphs of dimension ≤ N − 1 (points if N = 1), we have m(ak) = 0.In this case, m = L T and ν is not Rajchman.
Remark.-In Theorem 2.5 ii), observe that when making k vary, we obtain that for all p ∈ C N outside a countable number of real-analytic graphs of dimension less than or equal to N − 1 (points if N = 1), then m(ak) = 0, for all k ∈ Z. Part ii) of Theorem 2.5 relies on an indirect argument, based on the analysis of the regularity of m(n), for some fixed n ∈ Z, as a function of p ∈ C N .
Remark.-On the existence of singular measures in the non-homogeneous case, we are essentially aware of the non-explicit examples, using algebraic curves, of Neunhäuserer [14].As suggested by the referee, Theorem 2.5 allows to give in the non-homogeneous case an explicit example of a continuous singular and not Rajchman invariant measure ν with singularity dimension > 1.Indeed, take for 1/λ > 1 the Plastic number, i.e. the real root of X 3 − X − 1.This is the smallest Pisot number; cf Siegel [20].We have 1/λ = 1.3247.... Let N = 1 and ϕ 0 (x) = λx, ϕ 1 (x) = λ 2 x + 1.For p = (p 0 , p 1 ) ∈ C 1 , if ν is absolutely continuous with respect to L R , then, by Theorem 2.5 i), the density has to be bounded.By remark ii) in the Introduction, this implies that p 0 ≤ λ = 0.7548... and p 1 ≤ λ 2 .Now, as detailed in the last section, the similarity dimension in this case is > 1 if and only if 0, 203... < p 0 < 0, 907...For example we can conclude that for p 0 ∈ [0.76, 0.90], the measure ν is continuous, singular with respect to L R , not Rajchman and with similarity dimension > 1.Still for the system ϕ 0 (x) = λx, ϕ 1 (x) = λ 2 x + 1, we will give in the last section a strong numerical support for the fact that ν is in fact continuous singular and not Rajchman for all p ∈ C 1 .
It would be interesting to find more developed examples, where ν is absolutely continuous with respect to L R .A difficulty is that a priori the probability vector p has to be chosen in accordance with the polynomial equations verified by λ.

Proof of Theorem 2.3
Let N ≥ 1 and (ϕ k ) 0≤k≤N , with ϕ k (x) = r k x + b k , where r k > 0, and having no common fixed point.Fixing p ∈ C N , introduce i.i.d.random variables (ε n ) n≥0 with law p, to which P and E refer.By hypothesis, E(log r ε0 ) < 0. Recall that the invariant measure ν is the law of the random variable that ν is supposed to be non Rajchman.Without loss of generality, we assume that 0 The proof has three parts.First we show that log r i / log r j ∈ Q, for all 0 ≤ i = j ≤ N .From this, we will get that r j = λ nj , for some 0 < λ < 1 and integers (n j ).We then show that the non Rajchman character of ν can be seen on a subsequence of the form (αλ −k ) k≥0 .We finally prove that 1/λ is a Pisot number and the family (ϕ k ) 0≤k≤N is affinely conjugated with one in Pisot form.
Step 1.Let us show that if ever log r i / log r j ∈ Q, for some 0 ≤ i = j ≤ N , then ν is Rajchman.This is established in [11] for strict contractions.We simplify their proof.
In the expectation, the first exponential term is T s -measurable.Also, the conditional expectation of the second exponential term with respect to T s is just ν(αe −Sτ s +s ), as a consequence of the strong Markov property.It follows that : This gives |ν(αe s )| ≤ E |ν(αe −Sτ s +s )| , so by the Cauchy-Schwarz inequality and the Fubini theorem, which directly applies, consecutively : E e 2πiαe −Sτ s +s (x−y) dν(x)dν(y).
Let Y := − log r ε0 .The law of Y is non-lattice, since some log r i / log r j ∈ Q and p k > 0 for all 0 ≤ k ≤ N .As Y is integrable, with 0 < E(Y ) < ∞, it is a well-known consequence of the Blackwell theorem on the law of the overshoot that (see for instance Woodroofe [28], chap.2, thm 2.3), that : for any Riemann-integrable g defined on R + .Here, all S τs −s, for s ≥ 0, (and in particular S τ0 ) have support in some [0, A].Therefore, P(S τ0 > x) = 0 for large x > 0. For any α > 0, by dominated convergence (letting s → +∞) : +∞ 0 e 2πiαe −u (x−y) P(S τ0 > u)du dν(x)dν(y).
The inside term (in the modulus) is uniformly bounded with respect to (x, y) ∈ R 2 .We shall use dominated convergence once more, this time with α → +∞.It is sufficient to show that for ν ⊗2 -almost every (x, y), the inside term goes to zero.Since the measure ν is non-atomic, ν ⊗2almost-surely, x = y.If for example x > y : +∞ 0 e 2πiαe −u (x−y) P(S τ0 > u)du = x−y 0 e 2πiαt P(S τ0 > log((x − y)/t) dt t , making the change of variable t = e −u (x − y).The last integral now converges to 0, as α → +∞, by the Riemann-Lebesgue lemma.Hence, lim t→+∞ ν(t) = 0.This ends the proof of this step.
Step 2. As ν is not Rajchman, from Step 1, log r i / log r j ∈ Q, for all (i, j).Hence r j = r pj /qj 0 , with integers p j ∈ Z, q j ≥ 1, for 1 ≤ j ≤ N .Let : Recall that 0 < r 0 < 1. Setting λ = r 1/n0 0 ∈ (0, 1), one has r j = λ nj , 0 ≤ j ≤ N .Up to taking some positive integral power of λ, one can assume that gcd(n 0 , • • • , n N ) = 1.Recall in passing that the set of Pisot numbers is stable under positive integral powers.The condition E(log r ε0 ) < 0 rewrites into E(n ε0 ) > 0 and we have Using some sub-harmonicity, we shall now show that one can reinforce the assumption that ν(t) is not converging to 0, as t → +∞.
the support of the law of S τ0 generates Z as an additive group (cf for example Woodroofe [28], thm 2.3, second part).For an integer u ≥ 1 large enough, we can fix integers r ≥ 1 and s ≥ 1 such that the support of the law of S r contains u and that of S s contains u + 1, both supports being included in some Doing the same thing with S s and taking modulus gives : In particular, |ν(t)| ≤ max 1≤l≤M |ν(λ l t)|.We now set : The previous remarks imply that as soon as l is large enough.By continuity, letting l → +∞, we get c If this were not true, there would exist ε > 0 and (m k ) → +∞, with |ν(αλ (2) with r and t = αλ −m k −u and next with s and t = αλ −m k −u−1 .Since u is in the support of the law of S r and u + 1 is in the support of the law of S s , we obtain the existence of some c 1 < c such that for k large enough : Again via (2), with successively r and t = αλ −m k −2u , next r and t = αλ −m k −2u−1 and finally s and t = αλ −m k −2u−2 , still using that u is in the support of the law of S r and u + 1 in the support of the law of S s , we get some c 2 < c such that for k large enough : Etc, for some c M −1 < c and k large enough : This contradicts the fact that V α (−k) → c, as k → ∞.We conclude that |ν(αλ −k )| → c, as k → ∞, and this ends the proof of the lemma.
Step 3. We complete the proof of Theorem 2.3.In this part, introduce the notation x = dist(x, Z), for x ∈ R. Let us consider any 1 ≤ α ≤ 1/λ, with ν(αλ −k ) = c k e 2iπθ k , verifying c k → c > 0, as k → +∞.The existence of such a α was shown in Step 2. We start from the relation : obtained when conditioning with respect to the value of ε 0 .This furnishes for k ≥ 0 : We rewrite this as : Let K > 0 be such that c k−nj ≥ c/2 > 0, for k ≥ K and all 0 ≤ j ≤ N .For L > n * , where n * = max 0≤j≤N |n j |, we sum the previous equality on Observe that the right-hand side involves a telescopic sum and is bounded by 2n * (using that |c k | ≤ 1), uniformly in K and L. In the left hand-hand side, we take the real part and use that 1 − cos(2πx) = 2(sin πx) 2 , which, as is well-known, has the same order as x 2 .We obtain, for some constant C, that for K and L large enough : Introducing the constants p * = min 0≤j≤N p j > 0 and C = 2C/(cp * ), we get that for all 0 ≤ j ≤ N and K, L large enough : In the sequel, we distinguish two cases : there exists a non-zero translation among the (ϕ k ) 0≤k≤N (case 1) or not (case 2).
-Case 1.For any non-zero-translation ϕ j (x) = x + b j , we have n j = 0 and b j = 0. Then (3) gives that for K, L large enough : This implies that ( αb j λ −k ) k≥0 ∈ l 2 (N).By a classical theorem of Pisot, cf Cassels [3], chap.8, Theorems I and II, we obtain that 1/λ is a Pisot number and b j = (1/α)µ j , with µ j ∈ T (1/λ).Consider now the non-translations ϕ j (x) = λ nj x + b j , n j = 0.By (3), for any r ≥ 0 and K, L large enough (depending on r) : Fixing l j ≥ 1 and summing over 0 ≤ r ≤ l j − 1, making use of the triangular inequality and of (x , we obtain, for K, L large enough (depending on l j ) : Changing k into k + l j n j , we obtain, for K, L large enough (depending on l j ) : Let 1 = 0≤j≤N l j n j be a Bezout relation and J ⊂ {0, • • • , N } be the subset of j where l j n j = 0, equipped with its natural order.Using successively for j ∈ J either (4) or ( 5), according to the sign of l j , we obtain with : the following relation, for a new constant C and all K, L large enough : Now, for any n j = 0, whatever the sign of n j is, we arrive at, for some constant C and all K, L large enough : Hence, for any 0 ≤ j ≤ N with n j = 0, for some new constant C and all K, L large enough, using (3) : Let 0 ≤ j ≤ N , with n j = 0.If b j = b (1 − λ nj ), then we deduce again (still by Cassels [3], chap.8, Theorems I and II) that 1/λ is a Pisot number and b j = b (1 − λ nj ) + (1/α)µ j , with µ j ∈ T (1/λ).The other case is b j = b (1 − λ nj ).In any case, we obtain that for all 0 ≤ j ≤ N : for some µ j ∈ T (1/λ).Finally, remark that (7) says that the (ϕ j ) 0≤j≤N are conjugated with the (ψ j ) 0≤j≤N , where -Case 2. Any ϕ j with n j = 0 is the identity.The conclusion is the same, because there now necessarily exists some 0 ≤ j ≤ N with n j = 0 and b j = b (1 − λ nj ), otherwise b is a common fixed point for all (ϕ j ) 0≤j≤N .
This ends the proof of Theorem 2.3.

Proof of Theorem 2.4
Let N ≥ 1 and affine maps ϕ k (x) = λ n k x + µ k , for 0 ≤ k ≤ N , with 1/λ > 1 a Pisot number, relatively prime integers (n k ) 0≤k≤N and µ k ∈ T (1/λ), for 0 ≤ k ≤ N .Let p ∈ C N and denote by (ε n ) n∈Z a two-sided family of i.i.d.random variables with law p, to which again the probability P and the expectation E refer.We suppose that E(n ε0 ) > 0. Without loss of generality, n N ≤ • • • ≤ n 0 and in particular n 0 ≥ 1.For general background on Markov chains, cf Spitzer [23].
Recall the cocycle notations for the (n εi ) i∈Z introduced before the statement of the theorem and denote by θ the formal shift such that θε l = ε l+1 , l ∈ Z.We have for all k and l in Z : Then ν is the law of X = l≥0 µ ε l λ S l .We write Q ∈ Z[X] for the minimal polynomial of 1/λ, of degree s + 1, with roots The case s = 0 corresponds to 1/λ an integer ≥ 2 (using then usual conventions regarding sums or products).Recall that for any k ∈ Z, l∈Z µ ε l λ k+S l mod 1 is a well-defined T-valued random variable.
For each 0 ≤ r < n 0 , we now observe that we can move the sum q∈Z inside the expectation, using the theorem of Fubini, if we first show the finiteness of : This is true, since, as soon as n is larger than some constant (because of the missing term for q = 0 in the second sum), this equals G − (0, −n − r) + G + (0, −n − r) < +∞, where G − (x, y) and G + (x, y) are the Green functions, finite for every integers x and y, respectively associated to the i.i.d.transient random walks (S −q ) q≥0 and (S q ) q≥0 .Let σ + k , for k ∈ Z, be the first time ≥ 0 when (S q ) q≥0 touches k.We have G + (x, y) = P 0 (σ + y−x < ∞)G + (0, 0).With some symmetric quantities, one has G − (x, y) = P 0 (σ − y−x < ∞)G − (0, 0).We therefore obtain : Let us now fix 0 ≤ r < n 0 and consider the corresponding term of the right-hand side.First of all, for n > 0 larger than some constant (so that S 0 = −n − r) : as n → +∞, since (S q ) q≥0 is transient to the right.We thus only need to consider : where N (−k − r) := q≥0 1 S−q=−n−r .Consider an integer M 0 , that will tend to +∞ at the end.The difference of T (−n) with the following expression : is bounded by A + B, where, first : because N (−n − r) is stochastically dominated by N (0).Notice that N (0) is square integrable, as it has exponential tail.The first term on the right-hand side also goes to 0, as M 0 → +∞, by dominated convergence.The other term B is : as before.The first term on the right-hand side goes to 0, as M 0 → +∞, since (S −v ) is transient to −∞, as v → +∞.As a result : where o M0 (1) goes to 0, as M 0 → +∞, uniformly in n.Now, when n > 0 is large enough, Taking inside the expectation the conditional expectation with respect to the σ-algebra generated by the (ε l ) l≥−M0 , we obtain : Now, things are simpler because G − (S −M0 , −n − r) is bounded by the constant G − (0, 0).Hence, for some new o M0 (1), with the same properties : ), as n → ∞, by renewal theory (since the (n j ) are relatively prime and p j > 0, for all 0 ≤ j ≤ N ; cf Woodroofe [28], chap.2, thm 2.1), staying bounded by G − (0, 0), we get by dominated convergence and next M 0 → +∞ : ) E e 2iπ l∈Z µε l λ k+r+S l 1 S−u<−r,u≥1 .
From the initial expression, the limit, if existing, had to be independent on the parameter k.So this gives the announced convergence and invariance, hence proving item i) in Theorem 2.4.
Step 2. We now consider the proof of Theorem 2.4 ii) and suppose that ν is continuous.We first show that m is a continuous measure.For a continuous f : T → R + and any k ∈ R, we have : on T is continuous.Since l<0 µ ε l λ k+S l mod 1 and λ k X mod 1 are independent random variables, the law of Z k on T is continuous.Thus m is a continuous measure (hence M).
More generally, if 0 = n = (n 0 , • • • , n s ) t ∈ Z s+1 and if Z is random variable with law M, then the law of Z, n on T is m α , measure corresponding to m when replacing the (µ j ) by (αµ j ), thus the (ϕ j ) by the (ψ j ), with ψ j (x) = λ nj x + αµ j , where α = 0≤u≤s n u λ −u .Since α = 0, because (λ −u ) 0≤u≤s is a basis of Q[λ] over Q, the (ψ j ) do not have a common fixed point and thus m α is continuous, by the previous reasoning.
Suppose now that the (ϕ j ) have a common fixed point c.Hence µ j = c(1 − λ nj ), 0 ≤ j ≤ N , and ν = δ c .Necessarily c ∈ Q[λ], since the n j are not all zero.We shall show that λ −n c mod 1 converges to a rational number in T, as n → +∞.First of all, for n large enough, for all 0 ≤ j ≤ N : Hence, for any fixed sequence (k j ) 0≤j≤N , for n large enough, for all 0 ≤ j ≤ N : Supposing that 0≤j≤N k j n j = 1, using the previous expression successively n replaced by n, n − , respectively with j = 0, j = 1, • • • , j = N , and finally adding the results, we obtain that for some large K > 0, for all n > K : Let T r 1/λ (cλ −K ) = p/q.For n > K, there exists an integer l n such that T r 1/λ (cλ −n ) = p/q + l n .As a result, denoting by c = c 0 , c 1 , • • • , c s the conjugates of c corresponding to Q[λ] (reminding that (α j ) 0≤j≤s are that of 1/λ = α 0 ), we get : Consequently λ −n c mod 1 converges to p/q in T, as n → +∞, as announced.
To complete the proof of iii), we show that with n * = max 0≤j≤N n j .We now fix k ≥ n * so that T r 1/λ (λ −l µ j ) ∈ Z, for 0 For 0 ≤ j ≤ N , denote by (µ Taking any 0 ≤ u ≤ s and l < 0, we have : The role of the indicator function is now fundamental.On the event {S −v < −r, v ≥ 1}, we have T r 1/λ (µ ε l λ −u−k+r+S l ) ∈ Z, by our choice of k, since l ≤ −1.As a result, introducing the real random variables : together with s ), we obtain that for any f ∈ C(T s+1 , R) : 1 Hence, for any f ∈ C(T s+1 , R + ) : Fix any 0 ≤ r < n * and let X 0 = l≥0 µ ε l λ −k+r+S l and for 1 ≤ j ≤ s, X j = − l<0 µ (j) where V is the Vandermonde matrix : .
The matrix V is invertible (since the roots of the minimal polynomial Q of 1/λ are simple).By Cramer's formula : with γ i = det(V (i) )/det(V ), where V (i) is obtained from V by replacing the first column by e i , denoting by (e i ) 0≤i≤s the canonical basis of R s+1 .
Notice now that each γ i is real (first of all, 1/λ is a real root of Q; next, regrouping the other roots in conjugate pairs, when conjugating γ i one gets permutations in the numerator det(V (i) ) and the denominator det(V ), the same ones, so γi = γ i ).As V is invertible, γ := (γ i ) 0≤i≤s = 0.
We have X 0 = Y (r) , γ .Since ν is singular with respect to L R , we also have L(X 0 ) ⊥ L R , as , for all 0 ≤ r < n * .Finally, (11) implies that M ⊥ L T s+1 , as announced.
This ends the proof of Theorem 2.4.

Proof of Theorem 2.5
The context is the same as that of Theorem 2.4, but now the (ϕ k ) 0≤k≤N are strict contractions.Precisely, let N ≥ 1 and ϕ k (x) = λ n k x + µ k , for 0 ≤ k ≤ N , with 1/λ > 1 a fixed Pisot number, relatively prime integers (n k ) 0≤k≤N , with now n 0 ≥ • • • ≥ n N ≥ 1, without loss of generality, and Step 1.We first show Theorem 2.5 i), using again the arguments appearing in the previous section.If ν is absolutely continuous with respect to L R , then M = L T s+1 .The event {S −v < 0, v ≥ 1} has this time probability one.Looking at (10) with r = 0, we get that the law of Y (0) mod Z s+1 is absolutely continuous with respect to L T s+1 , with a density bounded by E(n ε0 ).Hence the law of Y (0) on R s+1 is absolutely continuous with respect to L R s+1 , with a density also bounded by E(n ε0 ).Since the (ϕ k ) 0≤k≤N are strict contractions, the n j are ≥ 1, so the random variable Y (0) is evidently bounded, cf (9).As a result the density of the law of Y (0) with respect to L R s+1 is bounded and with compact support in R s+1 .Hence this is also the case of X 0 = Y (0) , γ , where γ = (γ i ) 0≤i≤s = 0 is the first line of the inverse of the Vandermonde matrix V .Therefore this is also verified for X = λ k−r X 0 .This ends the proof of Theorem 2.5 i).
We turn to the proof of Theorem 2.5 ii).We shall focus on some Fourier coefficient m(n), thus for some fixed n ∈ Z, of the measure m appearing in Theorem 2.4 i).We study its regularity as a function of p ∈ C N , showing its real-analytic character.We then conclude the proof of Theorem 2.5 ii) using a theorem on the structure of the set of zeros of a non constant real-analytic function.
Step 2. Considering p ∈ C N , denote by (ε n ) n∈Z a sequence of i.i.d.random variables with law p.Let us fix an integer n = 0, whose exact value will be precised at the end of the proof.We focus on the Fourier coefficient m(n) of the measure m introduced in Theorem 2.4 i).Let us write m p in place of m to mark the dependence in p ∈ C N .As n j ≥ 1, for 0 ≤ j ≤ N , we have the simplified expression for this Fourier coefficient : where this last quantity is independent on k ∈ Z, by Theorem 2.4 i).The expectation E(n ε0 ) also depends on p, but to study the zeros of p −→ mp (n) we just need to focus on ∆ p .We now consider the regularity of p −→ ∆ p on the domain C N .
For any k ∈ Z, observe first that ∆ p (k) is well-defined, with the same formula as above, on the closure CN .Fixing k ∈ Z, the map p −→ ∆ p (k) is continuous on CN , as this function is the uniform limit on CN , as L → +∞, of the continuous maps : It follows that p −→ ∆ p (k) = ∆ p is well-defined on CN , continuous and independent on k.We shall now prove using standard methods that it is in fact real-analytic in a classical sense, precised below.Let us take k = 0 and fix 0 ≤ r < n 0 .Using independence, write : = E e 2iπn l≥0 µε l λ r+S l E e 2iπn l≤−1 µε l λ r+S l 1 nε −1 >r .
Call F (p) and G(p) respectively the terms appearing in the right-hand side.We shall show that both functions are real-analytic functions of p.This property will be inheritated by p −→ ∆ p .We treat the case of p −→ F (p), the case of G(p) needing only to rewrite first the µ ε l λ r+S l , appearing in the definition of G(p), as soon as l < 0 is large enough (depending only the (µ j ) 0≤j≤N , since n k ≥ 1, for all k), as − 1≤j≤s α −r−S l j µ (j) ε l , quantity equal to µ ε l λ r+S l in T, where the (µ Introducing the product measure µ p = ( 0≤j≤N p j δ j ) ⊗N on S, we can write : Denote by C(S) the space of continuous functions f : S → C and introduce the operator P p : C(S) → C(S) defined by : where (j, x) ∈ S is the word obtained by the left concatenation of the symbol j to x.The operator P p is Markovian, i.e. f ≥ 0 ⇒ P p (f ) ≥ 0 and verifies P p 1 = 1, where 1(x) = 1, x ∈ S. The measure µ p has the invariance property S P p (f ) dµ p = S f dµ p , f ∈ C(S).For f ∈ C(S) and k ≥ 0, introduce the variation : For any 0 < α < 1, let |f | α = sup{α −k Var k (f ), k ≥ 0}, as well as f α = |f | α + f ∞ .We denote by F α the complex Banach space of fonctions f on S such that f α < ∞.Any F α is preserved by P p .Observe now that g ∈ F α for λ ≤ α < 1.We fix α = λ.
As a classical fact from Spectral Theory, cf for example Baladi [1], the operator P p : F λ → F λ satisfies a Perron-Frobenius theorem.Let us show this elementarily.For f ∈ F λ , we have : This furnishes Var k (P n p f − 1 S f dµ p ) = Var k (P n p f ) ≤ Var k+n (f ).Therefore : In a similar way, we can write : Consequently, Putting things together, finally : This shows that 1 is a simple eigenvalue and that the rest of the spectrum of P p is contained in the closed disk of radius λ < 1. Remark that this holds uniformly on p ∈ CN .
Fix some circle Γ centered at 1 and with radius 0 < r < 1 − λ.By standard functional holomorphic calculus, cf Kato [10], for any p ∈ CN , the following operator, involving the resolvent, is a continuous (Riesz) projector on Vect(1) : Moreover Π p (F λ ) and (I − Π p )(F λ ) are closed P p -invariant subspaces, with : Also, in restriction to (I − Π p )(F λ ), the spectral radius of P p is less than λ.
Recall that N ≥ 1.We view a function of p ∈ CN in terms of the first N variables (p 0 , ).For any p ∈ CN and any η (even when p + η ∈ CN ), we can define the continuous operator P p+η : F λ → F λ by (12).It always verifies the relation : where Q j (f )(x) = f (j, x) − f (N, x).Denote by B N (0, δ) the open Euclidean ball in R N of radius δ.Let λ < λ < 1 − r.For any p in CN , there exists δ > 0 such that when η ∈ B N (0, δ), then 1 is still a simple eigenfunction of P p+η , with P p+η 1 = 1, the rest of the spectrum of P p+η being contained in the disk of radius λ and Π p+η , also defined by (13), is a continuous projector on Vect(1); this follows from the implicit function theorem, cf Rosenbloom [16], Kato [10].By compacity of CN , we can choose δ > 0 uniformly on p ∈ CN .This defines some open δ-neighborhood C δ N of CN .When p ∈ CN , we have S f dµ p = 0, for f ∈ (I − Π p )(F λ ).Thus for any f ∈ F λ : Applying this to the function g of interest to us, we obtain that when p ∈ CN : union of real-analytic graphs of dimension ≤ N − 1 (points if N = 1).By compacity of CN , the set {p ∈ CN | h(p) = 0} is included in a finite union of real-analytic graphs of dimension ≤ N − 1.
For the sequel, let us write x ≡ y for equality of x and y in T. Lemma 5.2 Let d ≥ 1 and µ ∈ T (1/λ).The series l∈Z µλ ld mod 1, well-defined as an element of T, equals a rational number modulo 1.
Notice in passing that the invariance with respect to k is now obvious, as we sum over r on a full period of length n j .Now, taking k = 0, we have : ∆ p j = 0≤r<nj e 2iπn(Aj,r/Bj,r) , for rational numbers A j,r /B j,r , making use of the previous lemma, since λ r µ j ∈ T (1/λ), for any r.If for example n is a multiple of B j,r for any 0 ≤ r < n j , we get ∆ p j = n j ≥ 1, which gives what was desired.This ends the proof of the theorem.
Remark.-Lojasiewicz's stratification theorem, giving the local structure of {p ∈ C δ N | h(p) = 0}, is a difficult theorem.In an elementary way, using the implicit function theorem, one can show that the set of zeros of a real-valued real analytic non constant function is locally included in a countable union of connected real-analytic graphs of codimension one.
Remark.-In the general case, when the (ϕ k ) 0≤k≤N are not all strict contractions, the method seems to reach some limit.Using the notation D N (r) of the Introduction, with r = (λ n k ) 0≤k≤N , and considering as in Step 2 the regularity of p −→ F (p) on D N (r), it is not difficult to show continuity, using some standard coupling argument.The real-analytic character, if ever true, a priori requires more work.Still setting S = {0, • • • , N } N and µ p = ( 0≤j≤N p j δ j ) ⊗N on S, we again have : , but this function is only defined µ p -almost-everywhere.
To study an interesting example, we take into account the similarity dimension s(p, r), rewritten here as s(p, λ) : As a function of p 1 , the left-hand side has a minimum value − ln(λ + λ 2 ), attained at p 1 = λ/(1 + λ).

Applications to sets of uniqueness for trigonometric series
Let N ≥ 1 and for 0 ≤ k ≤ N affine contractions ϕ k (x) = r k x + b k , with reals (r k ) and (b k ) such that 0 < r k < 1 for all k.As a general fact, Theorem 2.3 has some consequences in terms of sets of multiplicity for trigonometric series, cf for example Salem [18] or Zygmund [29]  Let us place on the torus T and consider trigonometric series.Recall that a subset E of T is a set of uniqueness (U -set), if whenever a trigonometric series n≥0 (a n cos(2πx) + b n sin(2πx)), with complex numbers (a n ) and (b n ), converges to 0 for all x ∈ E, then a n = b n = 0 for all n ≥ 0. Otherwise E is said of multiplicity (M -set).Theorem 6.1 Let N ≥ 1 and for 0 ≤ k ≤ N affine contractions ϕ k (x) = r k x + b k , where 0 < r k < 1, with no common fixed point.Suppose that the system (ϕ k ) 0≤k≤N is not affinely conjugated to a family in Pisot form.Then F mod 1 ⊂ T is a M -set.
Proof of the theorem : Any p ∈ C N gives a Rajchman invariant probability measure ν supported by F ⊂ R. Hence F mod (1) ⊂ T supports the probability ν, image of ν under the projection x −→ x mod 1, from R to T. Then ν is a Rajchman measure on T, so, cf Salem [18] (chap.V), F mod 1 is a M -set.
In the other direction, in general more delicate, we shall simply apply existing results.For the following statement, fixing 0 < λ < 1 and integers n k ≥ 1, for 0 ≤ k ≤ N , notice that for any (x 0 , x 1 , • • • ) ∈ S, we have l≥0 λ nx 0 +•••+nx l−1 (1 − λ nx l ) = 1.Theorem 6.2 Let N ≥ 1 and suppose that the (ϕ k ) are affine contractions of the form ϕ k (x) = λ n k x + b k , with b k = ba k + c(1 − λ n k ), for some 0 < λ < 1 with 1/λ a Pisot number > N + 2, relatively prime positive integers n k ≥ 1, 0 ≤ a k ∈ Q[λ] and real numbers b ≥ 0 and c.Then the non-empty compact self-similar set F = ∪ 0≤k≤N ϕ k (F ) ⊂ R can be written as F = bG + c, where G is the compact set : Assume that bG ⊂ [0, 1), so that bG and F can be seen as subsets of T. Then F is U -set.
Proof of the theorem : Up to replacing b and the (a k ) respectively by br and (a k /r), for some r > 1 in Q, we may assume that 0 ≤ a k < 1/(1 − λ), for all 0 ≤ k ≤ N .Then : Since 1/λ > N + 2 is a Pisot number and all a 0 , • • • , a N are in Q[λ], it follows from the Salem-Zygmund theorem, cf Salem [18], chap.VII, paragraph 3, on perfect homogeneous sets, that H is a perfect U -set.Mention that in this theorem, one also assumes that max 0≤k≤N a k = 1/(1 − λ) and that successive a u < a v in [0, 1) verify a v − a u ≥ λ.These conditions serve to give a geometrical description of the perfect homogeneous set H in terms of dissection, without overlaps.They are in fact not used in the proof, where only the above description of H is important (one can indeed start reading Salem [18], chap.VII, paragraph 3, directly from line 9 of the proof).
As a subset of a U -set, G is also a U -set.This is also the case of bG, by hypothesis a subset of [0, 1), using Zygmund, Vol.I, chap.IX, Theorem 6.18 (the proof, not obvious, is in Vol.II, chap.XVI, 10.25, and relies on Fourier integrals).Hence, F = bG + c is also a U -set, as any translate on T of a U -set is a U -set.This ends the proof of the theorem.
Remark.-As a general fact, the hypothesis 1/λ > N + 2 ensures that H and F have zero Lebesgue measure, which is a necessary condition for a set to be a U -set.