Can Noncommutative Maximal Ergodic Inequalities Extend to Polynomial Averages?

What the maximal inequality actually controls

The object being bounded is not a single average but the whole family of averages indexed by N. The paper proves that the maximal norm ‖(A_N f)_{N∈N}‖_{L^p(N;ℓ^∞)} is bounded by a constant times ‖f‖_{L^p(N)}, with the constant depending on p and on the polynomial P but not on f or on the averaging scale.

This distinction matters because L^p(N;ℓ^∞) is not a pointwise supremum. In the noncommutative setting there is no underlying point space, so the maximal function is replaced by a factorization norm: a family (x_λ) belongs to L^p(M;ℓ^∞) when it can be written as x_λ = a y_λ b with a,b in L^{2p}(M) and (y_λ) bounded in M. For self-adjoint families this norm coincides with the smallest a in L^p(M)_+ dominating every x_λ, which is why the notation sup^+_λ x_λ is used.

The averages themselves are built from a polynomial P(n) = Σ_{i=1}^{d} b_i n^i with integer coefficients and leading coefficient b_d > 0, and γ is a trace-preserving automorphism, meaning τ∘γ = τ. The theorem covers every such integer-coefficient polynomial, not only the monomial case P(n) = n^d.

Why the earlier result stopped at p above 1.62

Chen, Hong, and Wang had obtained a maximal ergodic inequality for averages along powers only in the range p > (1+√5)/2, approximately p > 1.62. That restriction is the benchmark this paper removes: the new theorem holds for every 1<p<∞, so the endpoint neighborhood near p = 1 is no longer excluded.

The gap was not a technicality of bookkeeping. Noncommutativity removes several tools that scalar proofs rely on: maximal functions, stopping times, the Borel–Cantelli or covering lemma, and monotonicity or convexity arguments either fail to exist or cease to hold. The paper states that its proof of the localized estimate required entirely new ideas precisely because difficulties that are trivial in the commutative case become substantial for operators.

Bourgain's major-arc strategy meets operator obstacles

The proof follows Bourgain's major-arc strategy, and a standard transference argument reduces the theorem to the corresponding maximal inequality for the shift action on Z. Writing A_Z(N) = ℓ^∞(Z)⊗N with the canonical tensor-product trace, and defining A_N f = f * K_N with K_N = (1/N)Σ_{n=1}^{N} δ_{P(n)}, a doubling argument further reduces the problem to the dyadic maximal inequality over scales 2^k.

The central localized ingredient is a noncommutative analogue of Bourgain's estimate: for 1<p<∞ and N_0 ≥ 3, the family (f * K_N) over the window N_0 < N < N_0^2 is bounded in L^p(A_Z(N);ℓ^∞) by a constant times log log N_0 times ‖f‖_{L^p(A_Z(N))}. The logarithmic loss in the window length is what makes the dyadic summation close.

Three obstacles are named explicitly. First, the moment expansion used by Bourgain for positive scalars fails because positive operators need not commute and products of positive operators need not be positive. Second, the distinct-index terms cannot be handled by the scalar positivity argument. Third, the exponentially decaying L^2 approximation in the scalar proof relies on a multi-frequency maximal inequality whose pointwise metric-entropy argument does not directly extend to L^2(A_Z(M);ℓ^∞).

Stein extrapolation rebuilt with projections

The first new tool is a noncommutative version of Stein's extrapolation, developed in Section 3. Classically, Stein showed that almost everywhere convergence of a sequence of convolution operators on L^p implies a weak type (p,p) estimate for the associated maximal operator when 1≤p≤2, an analogue of the Uniform Boundedness Principle for almost everywhere convergence rather than norm convergence.

Stein's proof cannot be transplanted because it uses a pointwise maximal function and a Borel–Cantelli argument. The paper replaces the exceptional sets E_k on which the maximal function exceeds a growing threshold λ_k by projections, and proves a Borel–Cantelli type result for those projections. A tensor-product construction first makes the relevant projections commute; applying the conditional expectation onto the commutative algebra they generate then yields scalar inequalities corresponding to the operator inequalities, and the commutative contradiction argument finishes the proof.

The result obtained is deliberately weaker than a full analogue: for 1<p≤2, Theorem 3.1 gives a form of Stein's implication sufficient for this paper, while the appendix proves an L^1 result in the spirit of Burkholder's theorem. The authors describe a general noncommutative analogue of Stein's implication as an extremely difficult problem.

Möbius inversion and a multilinear Doob inequality

The second new tool handles the repeated-index terms. Bourgain expanded a q-th power and used commutativity and nonnegativity of the factors; neither holds for positive operators. The paper instead combines Möbius inversion over the partition lattice of [q−1] with probabilistic and interpolation arguments. Writing the distinct-index contribution through linear combinations of products of martingale averages, it reduces the estimate to a maximal bound for the family (D_i)_{i∈I} in L^{q/(q−1)}(N;ℓ^∞), bounded by a constant times ‖Σ_{i∈I} a_i‖_{(q−1)/q}.

The third new tool is what the abstract calls a surprising multilinear version of Doob's maximal inequality. For an increasing filtration (A_n,E_n) in a normal semifinite tracial von Neumann algebra, and for x^{(r)} in L^{p_r}(A) with martingale averages x^{(r)}_n = E_n(x^{(r)}), the family of products x^{(1)}_n x^{(2)}_n ⋯ x^{(d)}_n is bounded in L^p(A;ℓ^∞) by the product of the individual L^{p_r} norms, where 1/p = Σ_{r=1}^{d} 1/p_r and 1<p<∞.

The proof iterates a column estimate whose constant is the product of the constants of the noncommutative Doob inequalities at the intermediate indices, and taking adjoints gives the corresponding row estimate. The authors note that a maximal inequality for products of martingales is rather unexpected in the noncommutative setting and may be of independent interest. A direct product estimate can fail here, which is why the combinatorial route is needed.

Avoiding the multi-frequency maximal inequality

The remaining difficulty is the exponentially decaying L^2 approximation needed to make the major-arc estimates summable. In the scalar proof, Bourgain obtained these estimates from the multi-frequency maximal inequality, whose bound depends only logarithmically on the number of rational points and which avoids using a common denominator. Its pointwise metric-entropy argument does not directly extend to L^2(A_Z(M);ℓ^∞).

Instead, the paper combines two of Bourgain's constructions. It uses the earlier decomposition that groups rational numbers a/q with (a,q)=1 according to common denominators rather than only the size of q, together with a noncommutative sampling principle, to estimate the contribution of each such set, and then sums using the decay of the exponential sums S(a/q). This retains stronger arithmetic input, including Weil's estimate and multiplicative properties of the exponential sums, entering through a scalar approximation whose error is O_P((log N)^{−σ}) for some σ > 1/2.

A key new step compares the truncated multiplier L^{(≤2^s)}_{2^k} with the multiplier cM_{s,k} in the noncommutative maximal L^2 norm. That comparison is what lets the decaying L^2 estimates follow from the sampling-principle estimate and the scalar approximation without any noncommutative analogue of the multi-frequency maximal inequality.

Where the theorem applies and where it stops

The scope is precise. The algebra is a semifinite von Neumann algebra with a normal faithful semifinite trace, the automorphism is trace-preserving, and the polynomial has integer coefficients with positive leading coefficient. The maximal inequality holds for every 1<p<∞, and the convergence statement is bilateral almost uniform convergence to an element of L^p(N), not norm convergence and not almost everywhere convergence in a pointwise sense.

The convergence half is not proved from scratch here. The paper adapts the argument of Hong and Wang from the monomial case P(n)=n^d to general integer-coefficient polynomials, and combines that with the maximal inequality to obtain convergence across the full range 1<p<∞. The maximal inequality (1.2) is the main task of the paper.

Two boundaries deserve attention. The constants depend on the polynomial P and on fixed cutoff functions, and are independent of the averaging and scale parameters unless stated otherwise, so the result is not uniform over all polynomials. And the noncommutative Stein extrapolation delivered here is a weaker form sufficient for this paper, not the general implication; the appendix's L^1 statement is in the spirit of Burkholder's theorem rather than a complete noncommutative counterpart.

What the full-range result changes for readers

The decision-relevant comparison is between two ranges, not between two theorems of the same strength. The earlier noncommutative result for averages along powers covered p > (1+√5)/2; this paper covers 1<p<∞ for all integer-coefficient polynomials. If a reader's application sits at small p, the earlier result simply did not apply, and the new theorem is the first noncommutative maximal inequality available there.

The second thing to weigh is which tool a reader actually needs. The multilinear noncommutative Doob maximal inequality for products of martingale averages, and the projection-based replacement for the Borel–Cantelli argument inside Stein extrapolation, are presented as results of independent interest, not merely as internal steps. For anyone working on noncommutative maximal estimates, those two tools are the transferable output; the polynomial averaging theorem is what they were built to deliver.

The third is the arithmetic trade-off. The paper gives up the multi-frequency maximal inequality and its logarithmic dependence on the number of rational points, and pays for it by returning to a decomposition by common denominators that keeps Weil's estimate and the multiplicative structure of the exponential sums. That choice is the reason the L^2 approximation still decays fast enough to sum, and it is the point where a reader judging whether to reuse this machinery should look first.

Sources used in this answer