Square-Root LASSO Pivotal Second-Order Cone Sparse Regression
Algorithms AIMech:SquareRootLassoPivotalConicSparseRegression DRAFT
Aliases: Square-Root Lasso, Sqrt-LASSO Pivotal Sparse Recovery, Belloni-Chernozhukov-Wang Second-Order Cone L1-Penalized Root-MSE Estimator, Self-Normalized Score Pivotal L1-Regularized Regression
Square-Root LASSO Pivotal Second-Order Cone Sparse Regression is a high-dimensional sparse linear regression estimator introduced by Belloni, Chernozhukov, and Wang (2011, Biometrika 98(4):791-806, DOI:10.1093/biomet/asr043, arXiv:1009.5689v5) for the model $y_i = x_i'\beta_0 + \sigma\epsilon_i$ ($i=1,\dots,n$, $\mathbb{E}_{F_0}(\epsilon_i)=0$, $\mathbb{E}_{F_0}(\epsilon_i^2)=1$, unknown noise scale $\sigma > 0$, support $T = \operatorname{supp}(\beta_0)$ of size $s < n$, and normalized covariates $\mathbb{E}_n(x_j^2) = 1$) that minimizes the square-root of the empirical mean-squared error plus an $\ell_1$ penalty, $\widehat{\beta} \in \arg\min_{\beta \in \mathbb{R}^p} \{\widehat{Q}(\beta)\}^{1/2} + \frac{\lambda}{n}\|\beta\|_1$ where $\widehat{Q}(\beta) = \frac{1}{n}\sum_{i=1}^n (y_i - x_i'\beta)^2$ (Eq. 8). Because the criterion gradient at $\beta_0$ is the self-normalized score $\tilde{S} = \nabla \widehat{Q}^{1/2}(\beta_0) = \frac{\mathbb{E}_n(x\epsilon)}{\{\mathbb{E}_n(\epsilon^2)\}^{1/2}}$ which is invariant to $(\beta_0, \sigma)$, setting the pivotal penalty $\lambda = c\Lambda(1 - \alpha) = c n^{1/2}\Phi^{-1}(1 - \alpha/(2p)) \le c\{2n\log(2p/\alpha)\}^{1/2}$ ($c > 1$) dominates $\Lambda = n\|\tilde{S}\|_\infty$ without knowing or pre-estimating $\sigma$, achieves the near-oracle prediction and $\ell_2$ convergence rate $\|\widehat{\beta} - \beta_0\|_{2,n} \lesssim \sigma\{s\log(2p/\alpha)/n\}^{1/2}$ across Gaussian and non-Gaussian errors (Theorem 1, Corollaries 1-4), and admits an exact second-order cone programming (SOCP) primal-dual formulation and closed-form coordinatewise update (Theorem 2, Eqs. 19-20, Appendix B.3).
Identity Notes & Concept Scope
Distinct from standard LASSO (`AIO:LassoRegression`), whose score $S = \nabla \widehat{Q}(\beta_0) = 2\sigma\mathbb{E}_n(x\epsilon)$ scales with the unknown noise standard deviation $\sigma$ and therefore requires an oracle $\sigma$, conservative pilot estimates, or cross-validation to set $\lambda = \sigma c 2 n^{1/2}\Phi^{-1}(1 - \alpha/(2p))$. Also distinct from Adaptive LASSO (`AIMech:AdaptiveLassoOracleWeightedL1PenalizedEstimation`), which reweights $\ell_1$ penalties using a root-$n$-consistent initial estimator in fixed or low dimensions, and from tangent-cone projected LASSO high-SNR risk analysis (`AIMech:WorstCaseProximalDenoisingNMSEAndSubdifferentialMeanSquaredDistance`). Belloni, Chernozhukov, & Wang (2011) replace $\widehat{Q}(\beta)$ with $\{\widehat{Q}(\beta)\}^{1/2}$ to make the score self-normalized and pivotal with respect to $\sigma$, prove finite-sample and moderate-deviation non-Gaussian near-oracle bounds under restricted eigenvalues, and establish strong conic duality over the second-order (ice-cream) cone $Q^{n+1}$.
Claims & Grounded Evidence
3 Claims
Mechanism
#self-normalized-pivotal-score-and-gaussian-and-moderate-deviation-penalty-domination
At the true parameter $\beta_0$, the gradient of $\widehat{Q}^{1/2}(\beta) = \{\frac{1}{n}\sum_{i=1}^n (y_i - x_i'\beta)^2\}^{1/2}$ is the self-normalized score $\tilde{S} = \nabla \widehat{Q}^{1/2}(\beta_0) = \frac{\nabla \widehat{Q}(\beta_0)}{2\{\widehat{Q}(\beta_0)\}^{1/2}} = \frac{\mathbb{E}_n(x\epsilon)}{\{\mathbb{E}_n(\epsilon^2)\}^{1/2}}$, whose scaled supremum $\Lambda = n\|\tilde{S}\|_\infty$ is completely independent of $(\beta_0, \sigma)$ (Section 2.1, p. 3). For $c > 1$ and confidence level $1 - \alpha$, setting $\lambda$ via the exact conditional quantile $\lambda = c\Lambda_F(1 - \alpha \mid X)$, the semi-exact family maximum $\lambda = c\max_{F \in \mathcal{F}} \Lambda_F(1 - \alpha \mid X)$, or the closed-form asymptotic option $\lambda = c\Lambda(1 - \alpha) = c n^{1/2}\Phi^{-1}(1 - \alpha/(2p)) \le c\{2n\log(2p/\alpha)\}^{1/2}$ (Eqs. 12-15) guarantees the regularization event $\lambda \ge c\Lambda$ with probability at least $1 - \alpha$ (exact) or $1 - \alpha\tau = 1 - \alpha(1 + o(1))$ under Gaussian errors $F_0 = \Phi$ with $\Lambda_\Phi(1 - \alpha \mid X) \le \nu\Lambda(1 - \alpha)$ where $\nu = \frac{\{1 + 2/\log(2p/\alpha)\}^{1/2}}{1 - 2\{\log(2/\alpha)/n\}^{1/2}}$ (Lemma 1, p. 5), and with probability $1 - \alpha - o(\alpha)$ for non-Gaussian errors with finite $q$-th moments ($q > 2$ under Slastnikov moderate deviations, Lemma 2, p. 6; or $q > 4$ under Jing-Shao-Wang self-normalized moderate deviations, Lemma 3, p. 19).
Mechanism
#restricted-eigenvalue-cone-membership-and-non-asymptotic-near-oracle-error-bounds
On the regularization event $\lambda \ge c\Lambda$ ($c > 1$), convexity of $\widehat{Q}^{1/2}$ and optimality of $\widehat{\beta}$ force the estimation error $\widehat{\delta} = \widehat{\beta} - \beta_0$ into the restricted cone $\Delta_{\bar{c}} = \{\delta \in \mathbb{R}^p : \|\delta_{T^c}\|_1 \le \bar{c}\|\delta_T\|_1,\, \delta \ne 0\}$ with $\bar{c} = \frac{c+1}{c-1}$ (Appendix 1, Step 1, Eqs. 21-25, p. 13). Under the restricted eigenvalues $\kappa_{\bar{c}} = \min_{\delta \in \Delta_{\bar{c}}} \frac{s^{1/2}\|\delta\|_{2,n}}{\|\delta_T\|_1}$ and $\tilde{\kappa}_{\bar{c}} = \min_{\delta \in \Delta_{\bar{c}}} \frac{\|\delta\|_{2,n}}{\|\delta\|_2}$ (Eq. 17, p. 7) and the growth condition $\lambda s^{1/2} \le n\kappa_{\bar{c}}\rho$ for some $\rho < 1$, the quadratic-root factorization $\widehat{Q}(\widehat{\beta}) - \widehat{Q}(\beta_0) = [\{\widehat{Q}(\widehat{\beta})\}^{1/2} + \{\widehat{Q}(\beta_0)\}^{1/2}][\{\widehat{Q}(\widehat{\beta})\}^{1/2} - \{\widehat{Q}(\beta_0)\}^{1/2}]$ yields the deterministic finite-sample bound $\tilde{\kappa}_{\bar{c}}\|\widehat{\beta} - \beta_0\|_2 \le \|\widehat{\beta} - \beta_0\|_{2,n} \le A_n \sigma\{\mathbb{E}_n(\epsilon^2)\}^{1/2} \frac{\lambda s^{1/2}}{n}$ where $A_n = \frac{2(1 + 1/c)}{\kappa_{\bar{c}}(1 - \rho^2)}$ (Theorem 1, p. 8; Appendix 1, Step 2, Eqs. 26-31, pp. 13-14), achieving the near-oracle rate $\sigma\{2s\log(2p/\alpha)/n\}^{1/2}$ without knowledge of $\sigma$ (Corollaries 1-4, pp. 8-9, 19).
Mechanism
#second-order-cone-primal-dual-equivalence-and-closed-form-coordinatewise-minimization
The Square-Root LASSO problem (8) is equivalent to the primal second-order cone program $\min_{t,v,\beta^+,\beta^-} \frac{t}{n^{1/2}} + \frac{\lambda}{n}\sum_{j=1}^p (\beta_j^+ + \beta_j^-)$ subject to $v = Y - X\beta^+ + X\beta^-$, $(v, t) \in Q^{n+1} = \{(v, t) \in \mathbb{R}^n \times \mathbb{R} : t \ge \|v\|_2\}$, and $\beta^+, \beta^- \in \mathbb{R}_+^p$ (Eq. 19), which admits with zero duality gap the strongly dual problem $\max_{a \in \mathbb{R}^n} \frac{1}{n}\sum_{i=1}^n y_i a_i$ subject to $|\frac{1}{n}\sum_{i=1}^n x_{ij} a_i| \le \frac{\lambda}{n}$ ($j=1,\dots,p$) and $\|a\|_2 \le n^{1/2}$ (Eq. 20), where whenever $Y \ne X\widehat{\beta}$ the optimal dual vector is the normalized residual $\widehat{a} = n^{1/2}\widehat{v}/\|\widehat{v}\|_2 = (Y - X\widehat{\beta})/\{\widehat{Q}(\widehat{\beta})\}^{1/2}$ (Theorem 2, pp. 9-10; Proof of Theorem 2, p. 14). Moreover, single-coordinate minimization over $\beta_j$ given $\beta_{-j}$ admits an exact closed-form threshold and shrinkage update setting $\beta_j = 0$ iff $|\mathbb{E}_n\{x_j(y - x'\beta_{-j})\}| \le (\lambda/n)\gamma_j\{\widehat{Q}(\beta_{-j})\}^{1/2}$ (Appendix B.3, pp. 21-23).
⚠️ Curated Limitations & Negative Results
- Theorem 1 requires the additional sparsity-penalty growth restriction $\lambda s^{1/2} \le n\kappa_{\bar{c}}\rho$ ($\rho < 1$) to absorb the quadratic term $(\lambda s^{1/2} / (n\kappa_{\bar{c}}))^2\|\widehat{\delta}\|_{2,n}^2$ arising from expanding $\{\widehat{Q}(\widehat{\beta})\}^{1/2}$ around $\{\widehat{Q}(\beta_0)\}^{1/2}$, and induces a slightly larger finite-sample shrinkage bias than infeasible oracle-$\sigma$ LASSO because $\widehat{Q}(\widehat{\beta})^{1/2}/\sigma$ exceeds 1 (typically 1.18-1.22 in Section 5, pp. 11-12), motivating post-Square-Root-LASSO ordinary least squares refitting.
- Non-Gaussian asymptotic validity of the closed-form Gaussian quantile penalty $\lambda = c n^{1/2}\Phi^{-1}(1 - \alpha/(2p))$ requires bounded $q$-th moments ($q > 2$ in Condition M or $q > 4$ in Condition SN) and polynomial restrictions on $p$ relative to $n$ ($p \le \alpha n^{\eta(q-2)/2}/2$ in Condition R or $\log(p/\alpha)\alpha^{-2/q}n^{-1/2}\log^{1/2}(n \vee p/\alpha) = o(1)$ in Condition SN), and assumes homoscedastic i.i.d. errors $\mathbb{E}(\epsilon_i^2) = 1$.
- Primary-source typographical audit of `arXiv:1009.5689v5`: (1) Equation (10) (p. 2) prints $\|\widehat{\beta} - \beta\|_2$ instead of $\|\widehat{\beta} - \beta_0\|_2$; (2) Section 2.2 (p. 5, lines 30-31), Section 2.2 Lemma 2 (p. 6), and Section 3.2 Corollary 3 (p. 9) cite '(36)' when referring to the non-normal penalty options in Equation (15); (3) Equation (19) (p. 9) writes $\sum_{i=1}^p (\beta_j^+ + \beta_j^-)$ with summation index $i$ and summand index $j$; (4) Section 5 (p. 10, lines 36-37, 42; p. 12, lines 4, 15, 21) repeatedly cites 'Figure 5' in place of 'Figure 1'; (5) Appendix B.3 (p. 23, lines 14, 16, 28, 31) prints a leading minus sign $-\mathbb{E}_n\{x_j(y - x'\beta_{-j})\}/\mathbb{E}_n(x_j^2)$ instead of $+\mathbb{E}_n\{x_j(y - x'\beta_{-j})\}/\mathbb{E}_n(x_j^2)$ in the componentwise update formulas for both LASSO and Square-Root LASSO (and an unmatched closing bracket `]` in the numerator of the first Square-Root LASSO display on line 28); and (6) Appendix C.2 (p. 24, line 19) writes 'Figures 3 and 3 illustrates the theoretical results indicated obtained in the paepr'.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:LassoRegression
Lasso Regression
|
broad_match | Square-Root LASSO is a pivotal conic-programming modification of LASSO regression that replaces the squared-error loss Q_hat(beta) with Q_hat(beta)^{1/2} so the penalty parameter lambda is independent of the noise standard deviation sigma. |
AIO:LinearRegression
Linear Regression
|
broad_match | Estimates high-dimensional sparse linear regression models y_i = x_i' beta_0 + sigma epsilon_i when p >> n. |
Evidence Sources & Primary Literature
3 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
bcw2011-sqrt-lasso-pdf
|
Alexandre Belloni, Victor Chernozhukov, Lie Wang
|
primary paper
full_text
|
ID: https://doi.org/10.1093/biomet/asr043
sha256:b86004a8edc4…
|
bcw2011-sqrt-lasso-abs
|
Alexandre Belloni, Victor Chernozhukov, Lie Wang
|
primary paper
abstract_only
|
ID: https://arxiv.org/abs/1009.5689
sha256:de8ea376abef…
|
bcw2011-sqrt-lasso-crossref
|
A. Belloni, V. Chernozhukov, L. Wang
|
primary paper
abstract_only
|
ID: https://doi.org/10.1093/biomet/asr043
sha256:95e3377fdb0f…
|
Ontology & Grounding Topology
Open in interactive network graph →Curation Provenance
Curator: Antigravity
Method: ai_assisted
Model: Gemini
Created: 2026-10-19T02:35:00Z
Updated: 2026-10-19T02:35:00Z
Source File:
data/curated/SquareRootLassoPivotalConicSparseRegression.yaml
· Schema:
curated_concept.yaml
Explore across X-Mech
Related computational & biological Mechs:
Embedding Map
Network Graph
ProteinTraitsMech (ESM-2 pLM)
DUFMech (Domain pLM)
TraitMech
CultureMech
All 14 X-Mech Repositories