1: \begin{abstract}

2: Suppose that we first apply the Lasso to a  design matrix, 

3: and then update one of its columns.

4: In general, the signs of the Lasso coefficients may change,

5: and there is no closed-form expression for updating the Lasso solution exactly.

6: In this work, we propose an approximate formula for updating a debiased Lasso coefficient.

7: We provide general nonasymptotic error bounds 

8: in terms of the norms and correlations of a given design matrix's columns,

9: and then prove asymptotic convergence results for the case of 

10: a random design matrix with i.i.d.\

11: sub-Gaussian row vectors and i.i.d.\ Gaussian noise. 

12: Notably, the approximate formula is asymptotically correct for most coordinates in the proportional growth regime, 

13: under the mild assumption that each row of the design matrix is sub-Gaussian with a covariance matrix having a bounded condition number.

14: Our proof only requires certain concentration and anti-concentration properties to control various error terms and the number of sign changes.

15: In contrast, rigorously establishing distributional limit properties (e.g.\ Gaussian limits for the debiased Lasso) 

16: under similarly general assumptions has been considered open problem in the universality theory. 

17: As applications, we show that the approximate formula allows us to reduce the computation complexity of variable selection algorithms that require solving multiple Lasso problems,

18: such as the conditional randomization test and a variant of the knockoff filter.

19: \end{abstract}

20: