Turnstile Streaming Algorithms Might (Still) as Well Be Linear Sketches, for Polynomial-Length Streams

Abstract: A fundamental question in streaming complexity is whether every space-efficient turnstile algorithm is implicitly a linear sketch. The landmark work of Li, Nguyen, and Woodruff [LNW14] established an equivalence between the two, but their reduction requires a stream length that is at least doubly exponential in the dimension $n$. In the opposite direction, results by Kallaugher and Price [KP20] demonstrate a separation for streams of linear length, showing that the equivalence does not hold in general. The most natural and practically relevant regime -- polynomial-length streams -- has therefore remained open. We show that polynomial-length turnstile algorithms admit linear-sketch simulations. More precisely, if a turnstile algorithm uses $S$ bits of space and succeeds on all streams of length $\mathrm{poly}(D, n)$, then on final vectors $x$ with $|x|_2 \le D$, its output can be recovered from $O(S)$ linear measurements of $x$, using $O(S \log S)$ bits overall. For smooth problems under appropriate input distributions, a mollified version of the reduction yields a bounded-entry sketch with $O(S / \log D)$ measurements and optimal $O(S)$ total space. Our results extend to strict turnstile streams and non-uniform Read-Once Branching Programs (ROBPs). Our proof departs from prior transition-graph based machinery, relying instead on a Fourier-analytic framework and tools from additive combinatorics to extract discrete linear measurements. Our analysis shows that any $S$-bit algorithm can only be sensitive to a low-dimensional lattice of heavy Fourier frequencies, which we then use to construct the rows of the sketching matrix. Consequently, we obtain new lower bounds for polynomial-length streams via existing real sketching and communication lower bounds.

• The paper establishes that any space-efficient turnstile streaming algorithm for polynomial-length streams can be simulated by an equivalent linear sketch using O(S) measurements with minimal extra space.
• It introduces a novel Fourier-analytic and additive-combinatorial approach to isolate heavy frequencies that capture the streaming state.
• The work bridges theoretical gaps by enabling direct transfers of sketching lower bounds to practical turnstile streaming scenarios.

Structural Equivalence of Turnstile Streaming Algorithms and Linear Sketches for Polynomial-Length Streams

Introduction and Context

The structural relationship between streaming algorithms in the turnstile model and linear sketching has been a foundational topic in data stream complexity. While linear sketching is the de facto paradigm for numerous sublinear turnstile streaming algorithms (e.g., estimating frequency moments, $\ell_p$ norms, heavy hitters), the fundamental question of whether all space-efficient turnstile algorithms are essentially linear sketches was resolved for extremely long streams and refuted for very short ones, but left open for polynomial-length streams—the regime most relevant in practice.

Previous work by Li, Nguyen, and Woodruff established that, given a sufficiently long stream (of doubly-exponential length in input dimension), every turnstile streaming algorithm could be simulated by a linear sketch of comparable space [li2014turnstile]. Conversely, Kallaugher and Price demonstrated a separation for streams of linear length, constructing problems solvable exponentially more compactly by nonlinear algorithms. This left a significant gap with respect to polynomial-length streams, the common case for modern data-intensive applications.

Main Contributions

This paper provides a definitive structural characterization for polynomial-length streams, showing that any space-efficient turnstile streaming algorithm is, up to minor overhead, equivalent to a linear sketch. More specifically:

• Exact Route: For any $S$-bit turnstile streaming algorithm that succeeds on all streams of length $\textsf{poly}(n, D)$, and for any final vector $x$ with $\|x\|_2 \le D$, there exists a linear sketching algorithm reconstructing the output from $O(S)$ integer linear measurements of $x$ with $O(S \log S)$ total bits of space.
• Mollified/Smooth Route: For "smooth" approximation problems under appropriate (e.g., Gaussian) input distributions, a refined reduction produces sketches with $O(S/\log D)$ measurements and $O(S)$ total bits, achieving optimal spatial efficiency relative to the streaming algorithm.

The equivalence extends to strict turnstile models and non-uniform descriptions such as Read-Once Branching Programs (ROBP).

Technical Approach

The proofs deviate sharply from previous graph-theoretic arguments and use a primarily Fourier-analytic and additive-combinatorial framework. The approach is summarized as follows:

Key Steps

1. Streaming-to-Noisy Block Experiment: The core reduction represents any input vector as the sum of many random "noisy" blocks (i.e., high-dimensional discrete Gaussians), plus a carefully chosen final update to land at the target. This effectively decorrelates nonlinear dependencies and exposes only the linear statistical distinguishers preserved by the streaming algorithm.
2. Conditioning on State Sequences: For a fixed state sequence (sequence of memory-states traversed by the streaming algorithm), the conditional law of the prefixes yields a convolution of independent, dense (spread-out) measures. The actual content that remains detectable by the streaming algorithm over these polynomial-length streams is then captured in the Fourier spectrum of these convolved measures.
3. Fourier-Analytic Decomposition: The paper shows that the only directions in which the final output distribution is sensitive (under translation) correspond to a small collection of "heavy" frequencies in the spectrum. These can be organized into rows of an integer matrix, forming the sketching operator.
4. Additive Combinatorics Constraint: Through a version of Chang's lemma, the number of independent heavy frequencies is shown to be $S$0, where $S$1 is the algorithm's space budget in bits. This establishes that the whole "non-diffuse" part of the spectrum is describable by a sketch of $S$2 measurements.
5. Translation-Invariance Argument: For any two vectors matching on the sketch, the conditional output distributions of the streaming algorithm are statistically indistinguishable, leading to fiberwise decoding of the final answer from the sketch value.

Smoothness and "Mollified" Cases

For input distributions and functions that are "smooth" (e.g., small additive noise does not alter answers), the convolution with one more Gaussian block further localizes the heavy spectrum near the origin. Then, with a tailored dimension reduction, the number of required measurements is improved to $S$3 while maintaining optimal bit complexity.

Implications for Lower Bounds

The structural equivalence established in this work has significant theoretical and practical implications:

• Transferring Communication and Sketching Lower Bounds: For approximation problems where lower bounds are known for real-valued sketches (e.g., $S$4 norm estimation, operator norm approximation, compressed sensing), the lifting theorems confirm that these directly imply tight lower bounds for turnstile streaming algorithms in the polynomial-length regime [GangulyW18, LiW16, NeedellSW22, SwartworthW23, gribelyuk2025lifting].
• Tight SMP-to-Streaming Reductions: For discontinuous problems (e.g., $S$5 estimation, subgraph/hypergraph counting, maximum matching), lower bounds previously shown for public-coin simultaneous message passing (SMP) protocols are now directly valid for polynomial-length turnstile streams, even in the presence of non-uniformity (ROBP), matching upper and lower bounds in parameters such as $S$6 (stream length).
• Resolution of Structural Uncertainty: The work closes the gap left by previous equivalence and separation results, demonstrating that the separation between linear and nonlinear algorithms for short streams does not extend to the practically relevant polynomial regime.

Table: Summary of Theoretical Implications

Problem/Setting	Lower Bound Scaling	Key Papers Referenced
$S$7 norm estimation ($S$8)	$S$9	[GangulyW18, gribelyuk2025lifting]
$\textsf{poly}(n, D)$0 estimation	$\textsf{poly}(n, D)$1	[DuMWY19, KNW10]
Approximate maximum matching	$\textsf{poly}(n, D)$2	[assadi2016maximum]
Operator norm, Ky Fan norm, eigenvalue	See detailed formulas (all scaling as $\textsf{poly}(n, D)$3 or $\textsf{poly}(n, D)$4)	[LiW16, SwartworthW23, gribelyuk2025lifting]
Subgraph/hypergraph counting	$\textsf{poly}(n, D)$5	[KallaugherKP18]
Compressed sensing ($\textsf{poly}(n, D)$6)	$\textsf{poly}(n, D)$7	[PriceW11, gribelyuk2025lifting]

Contrasts and Refinements

This work rigorously establishes that the only known separations between general turnstile streaming and linear sketching arise due to severe constraints on stream length or intermediate state, not in the conventional polynomial regime. It also shows that, in contrast to earlier reductions, the conversion from streaming algorithms to linear sketches can be made with negligible dimension and space overhead for all "natural" streaming problems of interest.

For "smooth" problems and natural input laws, the sketch can be made dimension-optimal and with bounded integer entries, and existing real-valued sketching lower bounds "lift" (without loss in polynomial factors) to the streaming model. The techniques apply to deterministic and randomized models, strictness constraints, and non-uniform read-once branching programs with only minor modifications.

Conclusion

This paper resolves a fundamental question in streaming complexity theory for polynomial-length streams, unifying the landscape between combinatorial, algebraic, and analytic approaches. The stated equivalence means that the design of efficient turnstile algorithms in the polynomial regime reduces to the design of efficient linear sketches, and that lower bounds for sketches precisely capture the landscape of streaming space complexity for all known natural problems in this model. On the theoretical side, this closes a major open gap; practically, it simplifies the pathway for both lower and upper bound research in sublinear streaming.

Key references:

• "Turnstile Streaming Algorithms Might as Well Be Linear Sketches" [li2014turnstile]
• "Separations and Equivalences between Turnstile Streaming and Linear Sketching" [KP20]
• "Lifting Linear Sketches: Optimal Bounds and Adversarial Robustness" [gribelyuk2025lifting]
• Lower bounds for specific problems [GangulyW18, LiW16, SwartworthW23, PriceW11, NeedellSW22]

---

arXiv reference: (2604.22052)