We give a sufficient condition in terms of sectional BMO for almost everywhere Pringsheim convergence of rectangular trigonometric Fourier sums. For continuous functions, the condition includes Oskolkov's criterion, and an example shows that the inclusion is strict.
Introduction
Let T=R/Z, with normalized Lebesgue measure m, and write m2=m⊗m for the product measure on T2. For f∈L1(T2), put
f(k,ℓ)=∫T2f(x,y)e−2πi(kx+ℓy)dxdy,(1)
SM,Nf(x,y)=∣k∣≤M∑∣ℓ∣≤N∑f(k,ℓ)e2πi(kx+ℓy).(2)
The Pringsheim limit means that min(M,N)→∞, with the two indices otherwise independent. For f∈C(T2), define
Fefferman [3] constructed a continuous function whose rectangular Fourier sums diverge in the Pringsheim sense. Bakhbukh and Nikishin [1] obtained a divergence example with
ω(f,δ)C=O(log(1/δ)1)(δ↓0).(3)
Oskolkov [9 Theorem 5] proved almost-everywhere Pringsheim convergence under the condition
ω(f,δ)C=o(log(1/δ)logloglog(1/δ)1).(4)
Masyutin [8] later obtained criteria for Lp functions, 1<p<∞, in terms of their Lp modulus of continuity.
where a summand with aq=0 is interpreted as zero, then SM,Nf(x,y)→f(x,y) for almost every (x,y) as min(M,N)→∞.
Corollary 1. Let f∈sBMO(T2). If
Ω(f,δ)=o(log(1/δ)logloglog(1/δ)1),
then SM,Nf→f almost everywhere in the Pringsheim sense.
Proof. For δq=(22q+1)−1, we have log(1/δq)≍2q and logloglog(1/δq)∼logq. Consequently aq=o(1/logq). Given η>0, eventually η/aq≥2logq whenever aq>0; thus the summands in (1.10) are at most q−2. Apply Theorem 1. □
For f∈C(T2), the bound
Ω(f,δ)≤2ω(f,δ)C(11)
shows that Corollary 1 includes Oskolkov’s criterion. The next example shows that (1.10) can hold for a continuous function that does not satisfy Oskolkov’s hypothesis.
Theorem 2. There is F∈C(T2) such that, for every η>0,
q=1∑∞exp(−2qω(F,(nq+1)−1)Cη)<∞,(12)
with the usual zero-summand convention, but
rloglogrω(F,2−r)C→0(r→∞).(13)
Thus Oskolkov’s hypothesis need not hold when Theorem 1 applies.
One-dimensional estimates
Write Sn(1) and Sn(2) for Fourier partial sums in x and y, respectively. Explicitly,
Sn(1)f(x,y)=∣k∣≤n∑(∫Tf(s,y)e−2πiksds)e2πikx,
Sn(2)f(x,y)=∣ℓ∣≤n∑(∫Tf(x,t)e−2πiℓtdt)e2πiℓy.
By Fubini,
SM,N=SM(1)SN(2)=SN(2)SM(1).(14)
A polynomial in one coordinate may have measurable coefficients depending on the other coordinate.
Lemma 1. There is an absolute CJ>0 such that, for f∈sBMO(T2), n≥1, and ν=1,2, there exists a polynomial Pn(ν) of degree at most n in coordinate ν. The residual gn(ν)=f−Pn(ν) has mean zero in the active coordinate, and
Then ∣h∣≤δ and N(t)≤1+∣t∣/δ. Write uy=f(⋅,y) and (τsu)(x)=u(x+s). The identity
uy−τ−tuy=j=0∑N(t)−1τ−jh(uy−τ−huy)
is a telescoping sum. For each fixed t, the definition of Ω1, together with ∣h∣≤δ, gives [uy−τ−huy]BMO≤Ω1(f,δ) for almost every y. The triangle inequality and translation invariance of the BMO seminorm therefore imply
The exceptional set of y’s may depend on t; Fubini’s theorem shows that this estimate holds for almost every t for almost every y. We may therefore integrate it against Jn(t). Since Jn≥0, ∫TJn=1, and (2.4) gives δ−1∫T∣t∣Jn(t)dt≤C, we obtain
For u=0, the left-hand side is zero. The estimate for Cg=supn∣Sng∣ in [7 Remark 6.1] is an L∞→BMO statement.
Lemma 2.There are absolute A,c>0 such that, for every mean-zero g∈BMO(T) and u>0,
m{x:n≥0sup∣g(x)−Sng(x)∣>u[g]BMO(T)}≤Ae−cu.(20)
If [g]BMO=0, the maximal remainder vanishes almost everywhere.
Proof. Let σ=[g]BMO(T). If σ=0, the mean-zero function g vanishes almost everywhere. Otherwise choose a,b in (2.5). Since Sn commutes with H,
nsup∣Sng∣≤nsup∣Sna∣+nsup∣Snb∣.
Apply (2.6) to a and b, and use ∥a∥∞+∥b∥∞≤Cσ. This yields
m{x:nsup∣Sng(x)∣>uσ}≤A1e−c1u.
John–Nirenberg [6], applied to the whole circle and using gT=0, likewise yields m{∣g∣>uσ}≤A2e−c2u. Finally supn∣g−Sng∣≤∣g∣+supn∣Sng∣; a union bound proves (2.7). □
Let M(1) and M(2) be the Hardy–Littlewood maximal operators along the indicated coordinates, applied to the absolute value of their arguments.
Lemma 3.For h∈L1(T2), m≥0, and ν=1,2,
∣Sm(ν)h∣≤Clog(m+2)M(ν)halmost everywhere.(21)
Proof. In the active coordinate the Dirichlet kernel obeys ∣Dm(t)∣≤Cmin{m+1,∣t∣−1}. Split the convolution into ∣t∣≤(m+1)−1 and dyadic annuli. Each piece is bounded by a fixed multiple of the one-dimensional maximal function, and there are O(log(m+2)) pieces. □
We also use, for 1<s<∞,
∥M(ν)H∥Ls(T2)≤Cs∥H∥Ls(T2).(22)
Proof of Theorem 1
Put δq=(nq+1)−1, and for ν=1,2 set εq(ν)=CJΩν(f,δq). Choose Pq(ν)=Pnq(ν) from Lemma 1 and write gq(ν)=f−Pq(ν). Each active-coordinate section of gq(ν) has zero mean and BMO seminorm at most εq(ν), outside a null set of passive coordinates. Define
Bq(ν)=nq≤n<nq+1supf−Sn(ν)f.
Since Sn(ν)Pq(ν)=Pq(ν) for n≥nq, we have
Bq(ν)=nq≤n<nq+1supgq(ν)−Sn(ν)gq(ν).
Write ε=εq(1) and, for a passive coordinate y, put gy=gq(1)(⋅,y) and σy=[gy]BMO(T). For almost every y, gy has mean zero and σy≤ε. Also,
Bq(1)(x,y)≤n≥0sup∣gy(x)−Sngy(x)∣.
Suppose ε>0. If σy>0, then uε≥uσy, so Lemma 2 yields
If σy=0, the mean-zero section gy vanishes almost everywhere and the set on the left is empty. Thus the same bound holds for almost every y. Since m(T)=1, Fubini gives
For ν=2, fix x instead and integrate over x. We have therefore proved, for both ν=1,2 whenever εq(ν)>0, that
m2{Bq(ν)>uεq(ν)}≤Ae−cu,u>0.(23)
Finally, if εq(ν)=0, almost every active-coordinate section of gq(ν) has BMO seminorm zero and mean zero. Such a section is zero almost everywhere, so Bq(ν)=0 almost everywhere by Fubini. We construct λq≥1 with
q≥1∑e−λq<∞,aqλq⟶0.(24)
Indeed, (1.10) implies aq→0. For k≥1, choose 1≤Q1<Q2<⋯ so that, for q≥Qk, aq≤1/k and
q≥Qk∑exp[−1/(kaq)]≤2−k.
For Qk≤q<Qk+1, set λq=(kaq)−1 when aq>0, and λq=q otherwise. For the finitely many q<Q1, put λq=q. The sums over successive blocks are at most 2−k, while the zero cases contribute at most ∑qe−q. This proves the first part of (3.2); on a nonzero block aqλq=1/k, and in zero cases the product is zero. Thus the second part holds as well. In particular,
The last expression is summable in k, since Rk grows doubly exponentially. The finitely many initial terms cause no difficulty. This proves (12).
Now set hk=(4Nk)−1<dRk. For ℓ>k, Nℓ/Nk is divisible by 4, so sin(2πNℓhk)=0. For ℓ<k, 0<2πNℓhk<π/2, so those sine terms are nonnegative. It follows that
ω(F,dRk)C≥∣F(hk,0)−F(0,0)∣≥Ak.
With rk=2Rk, we have dRk=2−rk and
rkloglogrkω(F,2−rk)C≥logRklog(Rklog2)⟶1.
This proves (13). In fact, the specific F has uniformly convergent rectangular Fourier sums, since its Fourier coefficients are absolutely summable.
Department of Mathematical Sciences, United Arab Emirates University, P.O. Box 15551, Al Ain, Abu Dhabi, United Arab Emirates
[1]M. Bakhbukh and E. M. Nikishin, The convergence of the double Fourier series of continuous functions, Siberian Math. J. 14 (1973), 832–839. doi:10.1007/BF00975888.DOI
[2]L. Carleson, On convergence and growth of partial sums of Fourier series, Acta Math. 116 (1966), 135–157.DOI
[3]C. Fefferman, On the divergence of multiple Fourier series, Bull. Amer. Math. Soc. 77 (1971), 191–195.DOI
[4]C. Fefferman and E. M. Stein, Hᵖ spaces of several variables, Acta Math. 129 (1972), 137–193.DOI
[5]R. A. Hunt, On the convergence of Fourier series, in Orthogonal Expansions and Their Continuous Analogues, Southern Illinois University Press, 1968, pp. 235–255.
[6]F. John and L. Nirenberg, On functions of bounded mean oscillation, Comm. Pure Appl. Math. 14 (1961), 415–426.
[7]G. A. Karagulyan, Bounded oscillation operators on BMO spaces, Math. Z. 312 (2026), article 12. doi:10.1007/s00209-025-03901-9.DOI
[8]D. I. Masyutin, Estimate of the rate of convergence of Fourier sums for functions from Lebesgue classes on a set of full measure, arXiv:2606.28786 (2026).arxiv.org/abs/2606.28786
[9]K. I. Oskolkov, An estimate of the rate of approximation of a continuous function and its conjugate by Fourier sums on a set of total measure, Math. USSR-Izv. 8 (1974), no. 6, 1372–1386.DOI