Double Machine Learning Now Has Bootstrap Theory

If you use double/debiased machine learning to estimate an effect, you also need to know how uncertain that estimate is. The bootstrap—the familiar trick of resampling data to see how much answers move around—has been widely used for that job, but until now it lacked a general proof for these estimators. This paper closes that gap. The authors show that bootstrap inference is valid for a broad class of double/debiased machine learning estimators under exchangeably weighted resampling schemes, with Efron’s bootstrap included as a special case. The key point is striking: the bootstrap works under exactly the same conditions needed for double machine learning itself to be valid. In the paper’s words, the bootstrap law converges conditionally weakly to the sampling law of the original estimator. That matters because double machine learning is designed for modern settings with high-dimensional or otherwise complex nuisance components, where classical assumptions are often too restrictive. With this result, the resampling methods already suggested and used in practice now have theoretical backing.

A resampling trick can feel like a magic mirror. You take your data, shuffle it into new fake samples, and watch how much an answer moves. That is the bootstrap. It is handy when you want error bars, but modern machine learning makes those error bars tricky. Double/debiased machine learning, or DML, is built for hard cases with complex nuisance parts. Nuisance parts are the side pieces of a model that help you adjust for confounding or hidden structure. Until now, many people used bootstrap tools on DML anyway, even though the proof was missing. That gap matters when a small formula feeds a big policy or medicine choice.

The surprise: the bootstrap works under the same rules as DML

The paper closes a gap that sat in plain sight. DML already gives root-n consistent and asymptotically normal estimators. Root-n consistent means the error shrinks at the same pace as the square root of sample size. Asymptotically normal means the estimate settles into a bell-shaped pattern as data grow. Even so, bootstrap methods can fail for estimators with those traits. Here, the main result says that bootstrap inference is valid for a broad class of DML estimators. It holds under exchangeably weighted resampling schemes. Efron’s bootstrap is one special case. The key point is neat and strong. The bootstrap law converges conditionally weakly to the sampling law of the original estimator. In plain English, the resampled answers move like the real answer would move.

How the proof is built

The engine of DML has two parts. One part is a Neyman-orthogonal score. A score is a rule that turns data into an estimate. Orthogonal means small errors in the nuisance parts have only a weak effect on the target. The second part is cross-fitting. Cross-fitting means splitting the sample, fitting the nuisance model on one piece, and scoring on another. That keeps the target estimate from leaning too hard on one fitted model. The proof shows that these same safeguards also protect the bootstrap. So the resampling step does not need a new, tighter cage of assumptions. It lives inside the cage DML already uses.

Why this matters for real analysis

This result helps because DML often meets messy, modern data head-on. The nuisance models can be high-dimensional or otherwise complex. That is the sort of setting where black-box learners feel natural, but theory can lag behind practice. Bootstrap methods are attractive there because they give an intuitive path to uncertainty bands. They also fit many inferential styles, including percentile, basic, and studentized confidence intervals. Now those tools do not sit on guesswork alone. They have a general proof for a broad DML setup. That gives analysts a cleaner bridge between flexible prediction tools and the need to say how sure they are about a final effect.

What still needs to hold up

The paper’s claim is broad, but it is not a blank check. Its result stays tied to the conditions DML itself needs. That is the important boundary. The good news is that the bootstrap does not ask for extra assumptions beyond that. The harder test now shifts from theory to use. Any real DML pipeline still has to build the orthogonal score, cross-fit the nuisance parts, and keep the original DML conditions in place. If those pieces hold, the bootstrap now has a sound home. If they do not, no resampling trick can rescue the analysis. That is a useful kind of discipline.

The next check for the bootstrap

The most important consequence is simple. A bootstrap interval for DML no longer has to be treated as a polite guess. It now has a proof that matches the rules behind DML itself. That could make uncertainty checks easier to trust in papers that use complex nuisance learning. The next test is whether common DML workflows keep the same behavior across different exchangeably weighted resampling schemes. The paper already covers Efron’s bootstrap as one special case. So the key practical question is not whether bootstrap ideas belong with DML. It is how widely this same guarantee travels inside the modern resampling family.