Stepwise Exact Methods

• Stepwise exact methods are sequential procedures that break complex problems into discrete, analyzable steps, ensuring finite-sample error control and exact inference.
• They are applied across multiple testing, joint model selection, quantum transport, and optimal control to achieve precise and computationally efficient solutions.
• By leveraging stepwise refinement and explicit analytic constraints, these methods enable scalable and robust performance in high-dimensional data indexing and various scientific domains.

Stepwise exact methods refer to a broad class of methodologies across statistical inference, model selection, control theory, computational physics, and high-dimensional indexing, in which complex problems are decomposed into a sequence of discrete, explicitly analyzable steps. The common feature is the use of “stepwise” (i.e., sequential or piecewise-constant) procedures, combined with analytic or algorithmic constructs yielding exact or provably controlled solutions at each stage. Applications range from stepwise multiple testing with error-rate control, to conditional inference in forward stepwise regression, to stepwise truncation in many-body quantum transport, to discrete-time optimal control, to stepwise refinement in high-dimensional nearest-neighbor search.

1. Control of Directional and Type I Errors in Multiple Testing

Guo and Romano (Guo et al., 2016) developed stepwise exact methods for controlling both type I (standard) and type III (directional) errors in multiple hypothesis testing under a variety of dependence structures. For two-sided nulls $H_i: \theta_i=0$ versus $H_i^A: \theta_i \ne 0$, rejections typically involve directional claims (via the sign of $T_i$); thus, control of mixed directional familywise error rate (mdFWER) and mixed directional false discovery rate (mdFDR) is crucial.

Mixed Error Rates

Let $V$ be the number of type I errors and $S$ the number of directional errors. Set $U = V + S$. Then,

• $FWER = P[V \ge 1]$, $mdFWER = P[U \ge 1]$
• $FDR = \mathbb{E}[V/\max(R,1)]$, $mdFDR = \mathbb{E}[U/\max(R,1)]$

Stepwise Procedures under Independence

The two-sided problem is reformulated by splitting each hypothesis into three one-sided nulls, focusing on the family $F_1$ of $2n$ one-sided hypotheses. Key procedures include:

• Two-stage adaptive Bonferroni (Procedure 1): Rejects hypotheses with $P_i \le \alpha/n$; further rejections among the remainder use $\alpha/(n-r)$. With correction, finite-sample FWER is controlled at level $\alpha$.
• Holm-type stepdown with enlarged critical values (Procedure 3): Orders $2n$ $P$-values and uses critical values $a_i = \alpha/(n-i+1) + \alpha$. This achieves finite-sample mdFWER control exactly at level $\alpha$.

Extensions adapt these procedures to block dependence and positive dependence (PRDS), employing blockwise splitting and Hochberg-type step-up approaches, with explicit guarantees.

Stepwise FDR Control

Procedures analogous to BH are constructed, leveraging the pairing property $P_i + P_{n+i} = 1$. Under independence and several blockwise/positive dependence regimes, these offer strong control of mdFDR at the prescribed level.

Implementation Guidance

For independence, Holm-type stepdown is preferred for mdFWER, BH-type for mdFDR. All constants are explicit. These results represent the first exact finite-sample controls for mixed directional error rates in stepwise multiple testing with a wide range of dependence models (Guo et al., 2016).

2. Exact Inference in Stepwise Joint Model Selection

Taylor et al. (Loftus et al., 2014) derived stepwise exact methods for inference in forward stepwise model selection with grouped variables. The entire process is embedded in a general quadratic-inequality selection framework, enabling the derivation of finite-sample exact p-values accounting for the selection event.

Quadratic-Inequality Selection

Selection events are characterized by polyhedral or quadratic regions:

$$\{S(y)=s\} = \bigcap_{i\in I(s)} \{ y : y^\top Q_i y + a_i^\top y \le b_i \}$$

for symmetric $Q_i$, $a_i$, and $b_i$ in $\mathbb{R}^n$.

Tχ Statistic

For the selected group $g^*$ at each step, the key contrast $T = \|X_{g^*}^\top y\|_2$ admits an exact truncated $\chi$ distribution after conditioning on both the selection event and the projected direction. The test statistic is:

$$\mathrm{T}\chi(t_{\rm obs}; r, \theta, M) = \frac{\int_{z \ge t_{\rm obs}/\theta,\; z \in M/\theta} f_{\chi_r}(z) dz} {\int_{z \in M/\theta} f_{\chi_r}(z) dz}$$

where $M$ is the slice of the selection event, and finite-sample exact uniformity holds under the global null.

Iterative Application and Practical Validity

At each stage, the response and predictors are orthogonalized with respect to selections so far, and the Tχ test is reapplied. Exactness holds at the first step; later stages are conservative unless all selection-induced constraints are tracked. This method generalizes to categorical grouping, hierarchical interaction selection, and adaptive additive models, without modification to the test machinery.

Computational Aspects

Algorithmic evaluation involves solving a set of quadratic inequalities and root-finding; the process is parallelizable at group level (Loftus et al., 2014).

3. Stepwise Exact Truncation in Quantum Many-Body Transport

Gudmundsson et al. (Gudmundsson et al., 2012) employed a stepwise introduction of model complexity and Fock-space truncation, combining exact diagonalization at each stage, to analyze electron transport in cavity-embedded nanostructures.

Layered Model Construction

Four layers are introduced stepwise:

1. Geometry: Single-electron Schrödinger problem in anisotropic confinement.
2. Magnetic field: Landau quantization and basis adjustment.
3. Coulomb interaction: Diagonalization in a truncated many-electron Slater determinant basis.
4. Photon coupling: Tensor product of Coulomb eigenstates with photon number states; full diagonalization of the combined electron-photon Hamiltonian.

Master Equation and Observable Control

The non-Markovian generalized master equation is solved with memory kernels explicitly computed in the exactly diagonalized basis. Each truncation step is rigorously controlled by convergence tests on observable spectra and occupations, ensuring accurate representation of quantum transport and transient dynamics.

Stepwise Truncation Protocol

At each layer, only levels required for physical fidelity (controlled via energy shifts and observable convergence) are retained, avoiding the exponential blow-up of the total Fock space. This methodology enables ab initio simulation of rich Coulomb–photon–transport phenomena (Gudmundsson et al., 2012).

4. Stepwise Methods in Optimal Control Problems

Afshar et al. (Afshar et al., 2015) formulated the stepwise method for finite-dimensional relaxation of continuous optimal control problems, motivated by the practical infeasibility of continuously variable controls in real-world resource allocation.

Mathematical Formulation

The control trajectory $u(t)$ is replaced by a piecewise-constant function over a partition $0=t_0<\dots<t_N=T$, reducing the continuous-time ODE-constrained optimization to:

$$J(u_1,\dots,u_N) = \sum_{i=1}^N \int_{t_{i-1}}^{t_i} L(x(t), u_i, t) dt + \Phi(x(T))$$

subject to $\dot{x}(t) = f(x(t), u_i, t)$ for $t \in [t_{i-1}, t_i]$.

Numerical Realization

Decision variables are the step values (and optionally, step endpoints). Standard finite-dimensional metaheuristics (pattern search, genetic algorithms, simulated annealing) are employed for optimization. The approximation is refined by increasing $N$ or optimizing the interval partition.

Comparison and Limitations

Stepwise methods avoid solving boundary-value adjoint equations of PMP and provide flexibility for non-smooth and nonlinear cost functions. They converge to the classical optimum as $N \to \infty$ but are susceptible to the curse of dimensionality and rely on heuristic solvers for nonconvexity.

Empirical Results

On practical problems (quadratic, chemotherapy, epidemic control), a small $N$ (3–5) yields results nearly indistinguishable from Pontryagin’s Maximum Principle. The method is advantageous when controls are naturally updated discretely or when analytical PMP is intractable (Afshar et al., 2015).

5. Stepwise Dimensionality Increasing Indexing in High-Dimensional Data

Thomasian and Zhang (Thomasian, 2024) introduced the Stepwise Dimensionality Increasing (SDI) index for similarity search in high-dimensional feature spaces, targeting efficient exact $k$-nearest neighbor ($k$-NN) queries.

Index Construction

Given a data matrix $X \in \mathbb{R}^{M \times N}$, principal components are computed via SVD. The SDI tree has $L$ levels, each corresponding to $n_\ell$ dimensions capturing increasing cumulative variance (parameter $p$). At each level, the data is projected and partitioned into hyperspheres, with increasing dimensionality per level.

Stepwise Query Algorithm

The $k$-NN search proceeds stepwise through levels of increasing dimension:

1. At each level, a $k$-NN search in $\mathbb{R}^{n_\ell}$ yields candidate neighbors.
2. A range query with the maximal candidate distance at $\mathbb{R}^{n_{\ell+1}}$ retrieves points for the next level, preventing false dismissals by the lower-bounding property ($\ell_2$ distance under projection never overestimates).
3. The process continues until the full dimension, guaranteeing exactness.

Complexity and Optimization

The primary computational burden is the initial SVD ($O(MN^2)$) and tree construction, with subsequent queries benefiting from dramatic pruning in lower dimensions. The parameter $p$ (variance step size) controls the trade-off between number of levels and per-level complexity; an optimal $p^*$ minimizes average CPU time for queries, found by profiling on sample queries.

Empirical Performance

Experiments on image-texture datasets show SDI outperforms SR-trees, VAMSR-trees, VA-Files, iDistance, and OMNI indexes in both I/O and CPU time, particularly as dimensionality grows ($N\geq50$), providing robust scalability for very high-dimensional similarity search (Thomasian, 2024).

6. Exact Solution Methods for the ASEP with Step Initial Data

In stochastic particle systems, “stepwise exact methods” refer to explicit algebraic or analytic formulas for distributions, constructed via stepwise (coordinate or moment) approaches.

Coordinate Bethe Ansatz

For the asymmetric simple exclusion process (ASEP) with step initial data, Tracy and Widom (Corwin, 2012) used a coordinate Bethe ansatz, building $k$-particle Green’s functions through stepwise contour integration. Taking limits as $k\to\infty$, they obtained exact formulas for current distributions as Fredholm determinants.

Replica Trick and Duality Approach

Borodin–Corwin–Sasamoto employed self-duality and a “replica trick,” solving the moment hierarchy via stepwise nested-contour integrals. Resummation and Laplace transforms yield full distributional formulas, which again reduce to Fredholm determinants amenable to asymptotic analysis.

Asymptotic Analysis

Both approaches, via steepest-descent analysis, recover the Tracy–Widom GUE distribution in the KPZ universality limit. The stepwise nature is reflected in either the construction of multi-integral solutions (coordinatewise) or the recursive deformation and resummation of contours and partition-indexed terms (momentwise) (Corwin, 2012).

---

Stepwise exact methods thus encompass a family of strategies which leverage sequential structure or decomposition—for exact or finite-sample controlled inference, optimization, or computation—across diverse domains. Each context exploits the granular, step-by-step buildup for analytic tractability, computational efficiency, or statistical rigor.