---
title: Signed-Knockoff Procedure for FDR Control
url: https://www.emergentmind.com/topics/signed-knockoff-procedure
type: topic
---

# Signed-Knockoff Procedure for FDR Control

The signed-knockoff procedure denotes a line of FDR-controlling methodology in which sign information is built into knockoff inference. In the classical knockoff literature, sign enters through antisymmetric statistics \(W_j\): \(|W_j|\) measures evidence against the null, while \(\operatorname{sign}(W_j)\) records whether the original variable appears more important than its knockoff. In a later and more explicit formulation, the “signed-knockoff procedure” is a direction-adaptive multiple-testing method that operates on signed \(p\)-values and their knockoff reflections, allowing asymmetric rejection boundaries while retaining finite-sample FDR control [2208.09542, 2507.15631].

## 1. Historical emergence and scope

Earlier knockoff work already treated directional information as a first-class inferential target, even when the exact phrase “signed-knockoff procedure” was not used. In high-dimensional selective inference for linear models, one framework split the observations into a screening part and an inference part, then applied the knockoff filter on the reduced model and proved control of the **directional false discovery rate**. In that formulation, the procedure “discovers” important variables as well as the directions, or signs, of their effects, and the expected proportion of wrongly chosen signs is controlled below the target level; the guarantee is non-asymptotic and holds for any distribution of the original features and any values of the unknown regression coefficients [1602.03574].

A later development made the terminology explicit. “A powerful procedure that controls the false discovery rate with directional information” introduces a method named the **signed-knockoff procedure (SK)** for large-scale multiple testing with directional information. Its stated goal is to use the signs of test statistics, such as up- or down-regulation in genomics, to improve power while controlling the ordinary FDR in finite samples [2507.15631].

The literature represented here suggests two closely related uses of the term. One use is broad: any knockoff procedure based on a signed antisymmetric comparison between an original variable and its knockoff. The other is narrow: the 2025 SK procedure based on signed \(p\)-values \(q_i\) and knockoff reflections \(\tilde q_i\) [2507.15631].

## 2. Classical signed knockoff mechanism

In fixed-\(X\) and model-\(X\) knockoffs, the central object is a feature statistic \(W_j\in\mathbb R\) constructed from the original feature and its knockoff. The standard interpretation is that \(|W_j|\) measures the overall importance of the pair, while \(\operatorname{sign}(W_j)\) records whether the original variable appears more important than its knockoff. This is the fundamental signed structure of knockoff inference [2208.09542].

A common construction begins with paired importance scores \(Z_j\) and \(\tilde Z_j\) and forms a signed contrast
\[
W_j = Z_j - \tilde Z_j.
\]
Other standard choices include the lasso coefficient-difference statistic
\[
W_j^{\mathrm{LCD}} = |\hat\beta_j^\lambda| - |\hat\beta_{j+m}^\lambda|
\]
and the lasso signed-max statistic
\[
W_j^{\mathrm{LSM}} = (\lambda_j^* \vee \lambda_{j+m}^*)\cdot(\lambda_j^*-\lambda_{j+m}^*),
\qquad
\lambda_j^*=\sup\{\lambda:\hat\beta_j^\lambda\neq 0\}.
\]
These satisfy antisymmetry: swapping the original and knockoff flips the sign of \(W_j\) [2208.09542, 1807.06214].

Under the null, this antisymmetry yields the sign-flip law
\[
\operatorname{sign}(W_j)\mid |W_j|,\mathbf W_{-j} \stackrel{H_j}{\sim} \mathrm{Unif}\{-1,+1\},
\]
which is the calibration device behind knockoff FDR control. The standard knockoff+ filter defines
\[
C(w)=\{j:W_j\ge w\},\qquad A(w)=\{j:W_j\le -w\},
\]
with FDP estimator
\[
\widehat{\mathrm{FDP}}(w)=\frac{1+|A(w)|}{|C(w)|},
\]
and threshold
\[
\hat w=\min\left\{w\ge 0:\widehat{\mathrm{FDP}}(w)\le \alpha\right\}.
\]
The rejection set is then
\[
R^{\mathrm{kn}}=\{j:W_j\ge \hat w\}.
\]
In this classical formulation, “signed” refers to original-versus-knockoff dominance, not necessarily to the sign of a regression coefficient or scientific effect [2208.09542].

## 3. The explicit signed-knockoff procedure based on signed \(p\)-values

The 2025 SK procedure begins with test statistics \(t_i\) and two-sided \(p\)-values \(p_i\), and defines the **signed \(p\)-value**
\[
q_i = \operatorname{sign}(t_i)(1-p_i).
\]
Under the null, if \(p_i\sim \mathrm{Uniform}(0,1)\) and \(\operatorname{sign}(t_i)\) is independent of \(p_i\), then \(q_i\sim \mathrm{Uniform}(-1,1)\). The associated knockoff is
\[
\tilde q_i = \operatorname{sign}(q_i)-q_i,
\]
so that \(q_i\) and \(\tilde q_i\) are reflected around \(1/2\) on the positive side and around \(-1/2\) on the negative side [2507.15631].

The procedure splits the signed \(p\)-values into positive and negative groups, orders them by closeness to \(1/2\) or \(-1/2\), and iteratively shrinks a rejection region from the middle toward the extremes. After \(k\) shrinkage steps, the rejection region is
\[
R_k= [-1,\ q^-_{(j_k)}\wedge \tilde q^-_{(j_k)})\ \cup\  (q^+_{(i_k)}\vee \tilde q^+_{(i_k)},\ 1].
\]
At each step, the algorithm removes exactly one pair from one side, with the side choice required to be measurable with respect to the masking filtration
\[
\mathcal F_k = \sigma\Big(\{\min(q_i,\tilde q_i)\}_{i=1}^n,\ \{b_i\}_{i\in\mathcal I_k}\Big),
\qquad
b_i = I(|q_i|>1/2).
\]
The estimated FDR is
\[
\widehat{\mathrm{FDR}}_k = \frac{1+\#\{i:\tilde q_i\in R_k\}}{\#\{i:q_i\in R_k\}\vee 1},
\]
and the algorithm stops when \(\widehat{\mathrm{FDR}}_k\le \alpha\), or when both sides are exhausted [2507.15631].

For power, the paper proposes a side-selection rule based on estimated local FDR under a parametric mixture
\[
f(q)=\pi_0 f_0(q)+(1-\pi_0)f_1(q),
\qquad
f_0(q)=\frac12,
\]
with alternative density
\[
f_1(q) = \lambda\frac{\alpha}{2}\left(\frac{q+1}{2}\right)^{\alpha-1} + (1-\lambda)\frac{\beta}{2}\left(\frac{1-q}{2}\right)^{\beta-1}.
\]
The next side is chosen by comparing the estimated local FDRs of the next positive-side and negative-side masked pairs [2507.15631].

## 4. Error criteria and theoretical guarantees

The 2025 SK paper controls the **ordinary** false discovery rate,
\[
\mathrm{FDP} = \frac{\#\{i: i\in \mathcal H_0,\ S_i=1\}}{\#\{i:S_i=1\}\vee1},
\qquad
\mathrm{FDR}=\mathrm E(\mathrm{FDP}),
\]
not a separate directional FDR. Its main theorem states that, under the null independence condition, if null \(p\)-values are \(\mathrm{Uniform}(0,1)\) and independent of the signs of the test statistics, then the signed-knockoff procedure controls the FDR at level \(\alpha\) [2507.15631].

This differs from earlier directional knockoff work in high-dimensional linear models, where the inferential target was explicitly the **directional false discovery rate** in a reduced model after screening. There the guarantee concerned wrongly chosen signs among selected variables, and the result was non-asymptotic [1602.03574].

A separate asymptotic strand studied sign errors for standard model-\(X\) knockoffs augmented with sign decisions. In that setting, after selecting
\[
\widehat{\mathcal S}=\{j:W_j\ge t\},
\]
one reports
\[
\widehat{\mathrm{sign}}_j=\mathrm{sign}(\widehat\beta_j),
\]
and analyzes the **false sign proportion**
\[
FSP \equiv \frac{ \left| \left\{ j\in\widehat{\mathcal S}:\mathrm{sign}(\beta_j)\neq \widehat{\mathrm{sign}}_j \right\} \right| }{ |\widehat{\mathcal S}| }.
\]
That work treats signed inference as an analytical extension of standard knockoff selection, rather than as a new finite-sample directional filter [2007.15346].

The resulting picture is technically important. Some signed-knockoff procedures control ordinary FDR while exploiting directional information in the ranking or rejection geometry; other knockoff procedures target directional or sign-error criteria directly. The distinction is substantive, not terminological [2507.15631, 1602.03574].

## 5. Variants, generalizations, and power-oriented refinements

Several later developments preserve the signed core of knockoff inference while altering its stability, power, or geometric resolution. One extension is **simultaneous multi-knockoffs**, where the binary sign of a single \(W_j\) is replaced by a winner label \(\kappa_j\in\{0,\dots,\kappa\}\) and a margin statistic \(\tau_j\). In the single-knockoff case, \(\tau_j=|W_j|\) and \(\kappa_j=0\) corresponds to \(W_j>0\); under the null, the binary sign symmetry becomes uniform label symmetry on \(\{0,\dots,\kappa\}\). This lowers the effective detection threshold and improves stability and power in sparse-signal regimes [1810.11378].

Another extension is **Aggregation of Multiple Knockoffs (AKO)**, which keeps the signed statistic \(W_j\) within each run but repeats knockoff generation, converts each run into an intermediate \(p\)-value
\[
\pi_j =
\begin{cases}
\dfrac{1 + \# \{k: W_k \leq -W_j\}}{p}, & W_j > 0,\\[1ex]
1, & W_j \leq 0,
\end{cases}
\]
and aggregates across runs. AKO therefore stabilizes signed-knockoff evidence rather than replacing it [2002.09269].

Power-improving wrappers preserve the same signed/antisymmetric core. The calibrated knockoff procedure begins from any valid fixed-\(X\) or model-\(X\) knockoff rejection set \(R^{\mathrm{kn}}\), uses the usual signed statistics \(W_j\), and augments the rejection set by conditionally calibrated fallback tests while retaining FDR control. Its purpose is to improve power without altering the underlying sign-flip mechanism [2208.09542].

The quality of signed comparisons also depends on knockoff construction. “Powerful Knockoffs via Minimizing Reconstructability” argues that minimizing mean absolute correlation can make original-versus-knockoff signs unreliable, because machine-learning procedures may reconstruct signal through the knockoffs. It proposes minimizing reconstructability instead, with the aim of producing stronger and more reliable antisymmetric comparisons such as
\[
W_j = |\hat\beta_j| - |\hat\beta_{j+p}|.
\]
This places knockoff construction, not only downstream statistics, at the center of signed-knockoff performance [2011.14625].

## 6. Interpretation, applications, and recurrent misconceptions

A persistent interpretive issue is the meaning of “sign.” In standard knockoff methodology, the sign of \(W_j\) indicates whether the original feature defeats its knockoff, not whether the underlying scientific effect is positive or negative. A finance application makes this distinction explicit: the relevant sign is induced by comparing the importance score of each original factor with that of its knockoff, and it is distinct from the sign of returns and also distinct from the sign of regression coefficients [2103.05921].

By contrast, the 2025 SK procedure uses the sign of the original test statistic \(t_i\) to define
\[
q_i=\operatorname{sign}(t_i)(1-p_i),
\]
so its signed structure is directly directional in the scientific sense. This is why the method can use asymmetric rejection regions of the form
\[
[-1,a)\cup (b,1]
\]
and adapt differently to the positive and negative sides [2507.15631].

Applications reflect both meanings. Earlier high-dimensional knockoff work applied directional sign control to genome-wide association analysis with a continuous phenotype [1602.03574]. The explicit SK procedure is motivated by genetics and illustrated on gene-expression-style settings where the sign of the test statistic corresponds to up- or down-regulation [2507.15631]. Other knockoff-based applications, including finance and multi-resolution localization, rely on the signed antisymmetry of original-versus-knockoff comparisons even when no effect-direction claim is made [2103.05921, 2306.09976].

Construction validity is another recurring theme. In case-control studies, knockoff variables can be built using controls only, cases only, or arbitrary mixtures, and the resulting exchangeability validates the construction step for any downstream signed-knockoff pipeline that depends on the usual swap symmetry [1812.11433]. Robustness work on approximate knockoffs shows, in a different way, that the standard signed/antisymmetric logic can survive misspecified feature models asymptotically when approximate statistics can be coupled closely to ideal model-\(X\) statistics [2307.04400].

Taken together, these developments establish the signed-knockoff procedure as both a specific direction-adaptive testing algorithm and a broader inferential principle. In the narrow sense, it is the signed-\(p\)-value procedure with knockoff reflections and finite-sample ordinary FDR control [2507.15631]. In the broader sense, it is the use of signed antisymmetric knockoff statistics—together with null sign symmetry, positive-versus-negative thresholding, and knockoff construction—to perform controlled variable selection or multiple testing with sign-aware evidence [2208.09542, 1602.03574].

Source: https://www.emergentmind.com/topics/signed-knockoff-procedure