Original paper

Statistical Gains from Looped Estimation under Parameter Budgets

Xinyu Tian · Xiaotong Shen

arXiv preprint · v2 · 2026Read the paper ↗

A reading note on the work cited above. Explanations and reading questions are identified separately from the paper’s reported findings.

Based on the public abstract

This note reads the public abstract of arXiv v2. Its scope is the stated question and claims; the full proofs and experiments have not been reviewed here.

The claim and its conditions

Under a common parameter budget, the paper compares looped and untied estimation. For specified architectures and known Hölder smoothness, the abstract reports minimax polynomial rates up to logarithmic factors with a fixed count of bounded real parameters. More looping raises computation and fitted-class complexity.

Editorial explanation: reading a budget

A useful general habit is to name the resource being counted. Parameter count, runtime, numerical precision, and training effort describe different constraints. A comparison becomes easier to interpret once the reader identifies which constraint matches the intended application.

Editorial reading questions

Questions for a fuller reading: What changes if runtime and numerical precision are also constrained? How can a statistical risk bound be connected to a feasible training procedure? Which experiment would make that connection convincing?

Sources & further reading

Public abstract · arXiv v2

Continue reading

Reusing a Transformer block