Stiefel Optimizer Overview

• Stiefel optimizer is an optimization technique for matrices with orthonormal columns, using geometric constraints on the Stiefel manifold.
• It incorporates methods such as retraction-based gradient updates, mixed-direction line searches, and intrinsic momentum schemes to ensure feasibility and convergence.
• Applied in areas like quantum channel control, neural network training, and system identification, it enhances computational efficiency and model performance.

A Stiefel optimizer is an optimization method for problems whose variables are matrices with orthonormal columns, typically posed on a Stiefel manifold such as $$\mathrm{St}(n,p)=\{X\in\mathbb{R}^{n\times p}:X^\top X=I_p\}$$, on a generalized Stiefel manifold with $X^\top M X=I_p$, or on related quotient constructions in which orthogonality appears together with gauge symmetries. In the cited literature, the term spans exact Riemannian methods based on tangent projections and retractions, second-order and proximal composite solvers, multiplier-correction and penalty formulations, continuous-time flows that asymptotically land on the manifold, and application-specific training rules that couple standard optimizers with QR or polar retractions (Oviedo et al., 2017, Kong et al., 2022, Jiang et al., 5 Feb 2026).

1. Manifold setting and geometric primitives

The standard Stiefel manifold is the set of orthonormal $p$-frames in $\mathbb{R}^n$, written in several papers as $\mathrm{St}(n,p)$ or $\mathrm{St}_p^n$. Its tangent space at $X$ is characterized by

$$T_X\mathrm{St}(n,p)=\{\Delta\in\mathbb{R}^{n\times p}:X^\top \Delta+\Delta^\top X=0\},$$

or, equivalently, by the decomposition

$$T_U\mathrm{St}_p^n=\{UA+(I_n-UU^\top)C: A=-A^\top,\ C\in M_{n\times p}(\mathbb{R})\}.$$

These formulas are the basic constraints that any Stiefel optimizer must respect, either exactly at every iterate or asymptotically along a flow (Oviedo et al., 2017, Birtea et al., 2018).

Generalized variants enlarge the class of admissible constraints. For symmetric positive definite $B$, the generalized Stiefel manifold is

$X^\top M X=I_p$0

with tangent space

$X^\top M X=I_p$1

For positive semidefinite, possibly singular, $X^\top M X=I_p$2, the feasible set is

$X^\top M X=I_p$3

and the singular-$X^\top M X=I_p$4 case requires a careful separation of range and null-space components (Shustin et al., 2019, Jiang et al., 5 Feb 2026).

The same geometric pattern extends to complex and quotient settings. In quantum-channel optimization, stacking Kraus operators into a block matrix $X^\top M X=I_p$5 yields the isometry constraint $X^\top M X=I_p$6, so channels are parameterized by the complex Stiefel manifold $X^\top M X=I_p$7. Because Kraus representations are non-unique up to unitary mixing, the physically meaningful space is the quotient $X^\top M X=I_p$8, and the paper establishes that this quotient is homeomorphic to the space of quantum channels (Russkikh et al., 2024).

2. Feasibility-preserving update mechanisms

A large part of the Stiefel-optimizer literature is organized around how orthogonality is preserved. Some methods move in tangent directions and retract back to the manifold; some use ambient-space descent plus a feasibility correction; some preserve feasibility through exact group actions; and some deliberately leave the manifold during intermediate iterates but are designed to land on it asymptotically.

Family	Core update idea	Feasibility mechanism
Retraction-based first order	Tangent projection of gradients, then descent step	SVD, QR, polar, or Cayley retraction
Mixed-direction line search	$X^\top M X=I_p$9 with BB steps	Projection $p$0 onto $p$1
Intrinsic momentum methods	Evolve position and momentum on the tangent bundle	Polar retraction with exact tangent-bundle preservation
Euclidean optimizer plus correction	Apply SGD, Adam, or AdamW to factors, then orthonormalize	QR or polar retraction after each step
Group-action methods	Optimize an orthogonal transform $p$2 acting on Stiefel data	Matrix exponential on $p$3

The mixed-direction method “Grad-Retrac” is a representative feasible first-order scheme. It uses

$p$4

combines them as $p$5 with $p$6, and globalizes the update through a non-monotone Barzilai–Borwein line search. Feasibility is maintained either by a second-order approximation or by explicit SVD projection, and the paper proves tangency, descent, and retraction properties (Oviedo et al., 2017).

A different line is the Momentum Stiefel Optimizer. It derives a dissipative second-order system on the tangent bundle $p$7, decomposes the tangent variable as $p$8, and discretizes the resulting dynamics so that both $p$9 and tangent compatibility of the momentum are preserved exactly. Its discrete update uses a polar step

$\mathbb{R}^n$0

and the paper emphasizes that no additional transport or projection is needed to keep momentum in the changing tangent spaces (Kong et al., 2022).

Several recent deep-learning implementations are deliberately hybrid rather than fully intrinsic. StelLA constrains the adapter factors $\mathbb{R}^n$1 and $\mathbb{R}^n$2, projects Euclidean gradients onto the tangent spaces, runs any Euclidean optimizer on the projected gradients, reprojects the perturbation to the tangent space, and retracts by the polar map $\mathbb{R}^n$3. The method is explicitly described as a flexible and modular design that converts any Euclidean optimizer to a Riemannian one (Li et al., 2 Oct 2025). A related LoRA method constrains only the $\mathbb{R}^n$4 factor to $\mathbb{R}^n$5, uses Adam-style preconditioning in ambient space, projects the resulting direction onto the tangent space, and retracts by the $\mathbb{R}^n$6-factor of a QR decomposition (Park et al., 25 Aug 2025). Spectral Compact Training uses standard AdamW on spectral factors $\mathbb{R}^n$7 and then retracts $\mathbb{R}^n$8 and $\mathbb{R}^n$9 to the Stiefel manifold by QR after every optimizer step (Kohlberger, 1 Apr 2026).

In system identification on Stiefel data, the optimization variable can be moved from the manifold itself to the orthogonal group. For observations $\mathrm{St}(n,p)$0, the paper on autoregressive processes estimates $\mathrm{St}(n,p)$1 by minimizing

$\mathrm{St}(n,p)$2

and updates $\mathrm{St}(n,p)$3 through conjugate gradient descent with line search along $\mathrm{St}(n,p)$4, which preserves orthogonality automatically because $\mathrm{St}(n,p)$5 (Figueras et al., 29 Sep 2025).

3. Second-order, proximal, and penalty formulations

Beyond first-order retraction methods, the literature contains an extensive second-order and composite-optimization theory. A central embedded-manifold result is the Hessian formula

$\mathrm{St}(n,p)$6

together with an explicit local tangent-frame construction and a QR-retraction Newton algorithm. This framework is used to recover second-order conditions for the Procrustes and Penrose regression problems and to derive necessary and sufficient local-minimum conditions for the Brockett problem (Birtea et al., 2018).

Composite objectives of the form

$\mathrm{St}(n,p)$7

with $\mathrm{St}(n,p)$8 smooth and $\mathrm{St}(n,p)$9 convex but possibly nonsmooth, motivate proximal quasi-Newton and proximal Newton methods on the Stiefel manifold. ManPQN solves a tangent-space proximal subproblem with a variable positive definite operator $\mathrm{St}_p^n$0, updates by a retraction, and uses a nonmonotone line search. The paper proves global convergence, an $\mathrm{St}_p^n$1 iteration bound for obtaining an $\mathrm{St}_p^n$2-stationary point, and local linear convergence under a pullback-Hessian positivity condition (Wang et al., 16 Jan 2025).

ARPQN and ARPN add adaptive quadratic regularization. ARPQN solves

$\mathrm{St}_p^n$3

while ARPN uses a more geometric model with $\mathrm{St}_p^n$4. The paper establishes global convergence and local linear convergence for ARPQN, and local q-superlinear convergence for ARPN under stronger assumptions (Wang et al., 2024).

Multiplier-correction methods represent another exact-feasibility strategy. Starting from a feasible iterate $\mathrm{St}_p^n$5, they first perform a function-value reduction step along an arbitrary Euclidean descent direction and then apply a proximal orthogonal correction inside the range space of the intermediate point, $\mathrm{St}_p^n$6 with $\mathrm{St}_p^n$7. The correction is obtained from the SVD of a $\mathrm{St}_p^n$8 matrix and is designed to reduce the asymmetry of the Lagrange multipliers. The paper proves monotone decrease of the objective, convergence of every accumulation point to a first-order stationary point, and an $\mathrm{St}_p^n$9 complexity estimate (Wang et al., 2020).

Penalty formulations are prominent for generalized constraints. SLEP reformulates

$X$0

as the smooth unconstrained problem

$X$1

For sufficiently large $X$2, the paper proves local equivalence of first-order stationary points, second-order stationary points, and local minimizers between the penalty model and the original generalized-Stiefel problem, including the singular-$X$3 case. The stated motivation is that SLEP eliminates the need for retractions and vector transports and permits direct application of unconstrained methods (Jiang et al., 5 Feb 2026).

Preconditioning changes the geometry rather than the objective. On $X$4, a metric defined by an SPD matrix $X$5 yields a Riemannian gradient

$X$6

and corresponding projection, retraction, vector-transport, and Hessian formulas. The paper’s central claim is that metric choice functions as a preconditioner: it can reduce geometric cost and improve conditioning of the Riemannian Hessian near the solution (Shustin et al., 2019).

4. Convergence guarantees, landscape results, and computational limits

The convergence theory of Stiefel optimizers is strong at the level of stationarity but limited at the level of worst-case global optimality. Retraction-based first-order methods prove descent and feasibility; second-order and proximal methods prove convergence of accumulation points to stationary points; multiplier-correction methods prove monotone decrease; and adaptive regularization yields local linear or q-superlinear rates under suitable curvature conditions (Oviedo et al., 2017, Wang et al., 16 Jan 2025, Wang et al., 2024, Wang et al., 2020).

Continuous-time analysis adds a different perspective. The landing-flow system

$X$7

is designed so that trajectories may leave the manifold at finite time but satisfy $X$8 asymptotically. The paper proves global existence, monotone decrease of the orthogonality defect, convergence to the set of critical points on the Stiefel manifold, and asymptotic stability of isolated local minima (Gao et al., 2022).

In special application domains, the geometry can remove suboptimal local extrema. For continuous functionals on the space of quantum channels, the quotient map from the complex Stiefel manifold preserves non-strict local extrema, and if the functional is continuous and convex or concave then there are no local but non-global minima or maxima on either the channel space or the Stiefel manifold. The paper explicitly applies this “no traps” result to mean-value, entropy-related, channel-generation, and gate-generation objectives (Russkikh et al., 2024).

Worst-case complexity is much harsher. Linear programming with linear constraints over a Stiefel manifold, and unconstrained quadratic programming over a Stiefel manifold, are both NP-hard in general. The same paper proves nonexistence of FPTAS for these problem classes unless $X$9, and extends related hardness results to Grassmann and flag manifolds. Its practical lesson is explicit: local methods and heuristics are essential, and tractable relaxations such as semidefinite programming may be indispensable for certification or approximation (Lai et al., 3 Jul 2025).

Taken together, these results delimit the scope of any Stiefel optimizer. Exact feasibility, descent, and high-quality local behavior are well supported theoretically; uniform polynomial-time guarantees for global optimization are not.

5. Application-specific realizations

In quantum control and open quantum systems, the Stiefel manifold provides the geometric arena for optimization over Kraus representations of quantum channels. Stacking Kraus operators produces the isometry constraint $$T_X\mathrm{St}(n,p)=\{\Delta\in\mathbb{R}^{n\times p}:X^\top \Delta+\Delta^\top X=0\},$$0, the quotient $$T_X\mathrm{St}(n,p)=\{\Delta\in\mathbb{R}^{n\times p}:X^\top \Delta+\Delta^\top X=0\},$$1 removes Kraus non-uniqueness, and the induced quotient metric defines a distance on channel space. The framework is then used to analyze control landscapes for mean values, entropy-related objectives, channel generation, and gate generation, for both Markovian and non-Markovian dynamics (Russkikh et al., 2024).

In parameter-efficient fine-tuning, StelLA treats LoRA as a subspace-learning problem with

$$T_X\mathrm{St}(n,p)=\{\Delta\in\mathbb{R}^{n\times p}:X^\top \Delta+\Delta^\top X=0\},$$2

The paper reports that product Stiefel geometry $$T_X\mathrm{St}(n,p)=\{\Delta\in\mathbb{R}^{n\times p}:X^\top \Delta+\Delta^\top X=0\},$$3 works better than a plain Euclidean three-factor model, outperforms a quotient geometry formulation, and that polar retraction performs about as well as the exponential map but is cheaper (Li et al., 2 Oct 2025). A separate LoRA paper constrains only $$T_X\mathrm{St}(n,p)=\{\Delta\in\mathbb{R}^{n\times p}:X^\top \Delta+\Delta^\top X=0\},$$4 in $$T_X\mathrm{St}(n,p)=\{\Delta\in\mathbb{R}^{n\times p}:X^\top \Delta+\Delta^\top X=0\},$$5 to $$T_X\mathrm{St}(n,p)=\{\Delta\in\mathbb{R}^{n\times p}:X^\top \Delta+\Delta^\top X=0\},$$6, uses Adam-preconditioned Riemannian updates for $$T_X\mathrm{St}(n,p)=\{\Delta\in\mathbb{R}^{n\times p}:X^\top \Delta+\Delta^\top X=0\},$$7, and measures effective rank by SVD entropy. A representative result reported there is that with nominal rank $$T_X\mathrm{St}(n,p)=\{\Delta\in\mathbb{R}^{n\times p}:X^\top \Delta+\Delta^\top X=0\},$$8 on LLaMA-3.2-1B, the Stiefel method uses all $$T_X\mathrm{St}(n,p)=\{\Delta\in\mathbb{R}^{n\times p}:X^\top \Delta+\Delta^\top X=0\},$$9 dimensions, while AdamW averages about $$T_U\mathrm{St}_p^n=\{UA+(I_n-UU^\top)C: A=-A^\top,\ C\in M_{n\times p}(\mathbb{R})\}.$$0 effective dimensions (Park et al., 25 Aug 2025).

In compact neural training, Spectral Compact Training permanently stores each dense layer as

$$T_U\mathrm{St}_p^n=\{UA+(I_n-UU^\top)C: A=-A^\top,\ C\in M_{n\times p}(\mathbb{R})\}.$$1

runs standard backpropagation and AdamW on $$T_U\mathrm{St}_p^n=\{UA+(I_n-UU^\top)C: A=-A^\top,\ C\in M_{n\times p}(\mathbb{R})\}.$$2, and then retracts $$T_U\mathrm{St}_p^n=\{UA+(I_n-UU^\top)C: A=-A^\top,\ C\in M_{n\times p}(\mathbb{R})\}.$$3 and $$T_U\mathrm{St}_p^n=\{UA+(I_n-UU^\top)C: A=-A^\top,\ C\in M_{n\times p}(\mathbb{R})\}.$$4 by QR after each step. The paper reports up to $$T_U\mathrm{St}_p^n=\{UA+(I_n-UU^\top)C: A=-A^\top,\ C\in M_{n\times p}(\mathbb{R})\}.$$5 memory reduction per MLP layer at rank $$T_U\mathrm{St}_p^n=\{UA+(I_n-UU^\top)C: A=-A^\top,\ C\in M_{n\times p}(\mathbb{R})\}.$$6, a full training step of a $$T_U\mathrm{St}_p^n=\{UA+(I_n-UU^\top)C: A=-A^\top,\ C\in M_{n\times p}(\mathbb{R})\}.$$7B-parameter architecture on a Steam Deck handheld with $$T_U\mathrm{St}_p^n=\{UA+(I_n-UU^\top)C: A=-A^\top,\ C\in M_{n\times p}(\mathbb{R})\}.$$8 GB peak memory versus $$T_U\mathrm{St}_p^n=\{UA+(I_n-UU^\top)C: A=-A^\top,\ C\in M_{n\times p}(\mathbb{R})\}.$$9 GB for dense FP32 training with Adam, and orthogonality error below $B$0 in the $B$1B validation benchmark (Kohlberger, 1 Apr 2026).

In probabilistic partial least squares, exact Stiefel optimization enforces the identifiable constraints $B$2 and $B$3 over the product manifold

$B$4

The solver uses projected Riemannian gradients, QR retractions, Armijo backtracking, and a block-coordinate variant with closed-form scalar updates. The reported empirical outcomes include near-nominal predictive coverage without post-hoc recalibration and improved stability of parameter recovery (Hu et al., 12 May 2026).

In autoregressive modeling on manifolds, the Stiefel constraint appears in system identification rather than direct state evolution. The unknown parameter is an orthogonal transform $B$5 acting on Stiefel observations, estimated by minimizing $B$6 using conjugate gradient descent and matrix-exponential updates on $B$7 (Figueras et al., 29 Sep 2025).

In spectrum-controlled neural layers, ManifoldFlow keeps a Stiefel basis $B$8 but relaxes the fixed unit spectrum through

$B$9

Because $X^\top M X=I_p$00, the eigenvalues of $X^\top M X=I_p$01 are exactly the squared singular values of the realized weight, so eigenvalue clipping on $X^\top M X=I_p$02 gives direct singular-value control. The $X^\top M X=I_p$03 factor uses the same Stiefel optimizer as the fixed-Stiefel baseline, while $X^\top M X=I_p$04 is updated on the SPD cone by an affine-invariant rule (Yi et al., 5 Jul 2026).

6. Conceptual distinctions and common misconceptions

A frequent misconception is that a Stiefel optimizer is necessarily a fully intrinsic Riemannian algorithm. The literature is broader. Some methods are intrinsic from the outset, such as retraction-based gradient, conjugate-gradient, Newton, and proximal quasi-Newton schemes (Oviedo et al., 2017, Birtea et al., 2018, Wang et al., 16 Jan 2025). Others are explicitly hybrid: standard AdamW or Adam is applied in ambient space and orthogonality is restored afterward by QR or polar retraction, as in Spectral Compact Training and Stiefel-constrained LoRA (Kohlberger, 1 Apr 2026, Park et al., 25 Aug 2025). SLEP goes further: it is presented not as a Riemannian method, but as a smooth unconstrained reformulation whose local stationary structure matches the original generalized-Stiefel problem for sufficiently large penalty parameters (Jiang et al., 5 Feb 2026).

A second misconception is that “orthogonal” and “spectrally constrained” are interchangeable. Standard fixed-spectrum Stiefel layers force all represented singular values to equal one, whereas ManifoldFlow keeps the basis on the Stiefel manifold but learns a bounded positive spectrum through an SPD factor. The paper’s point is precisely that an orthonormal basis need not imply a frozen singular spectrum (Yi et al., 5 Jul 2026).

A third misconception is that the full Stiefel manifold is always the correct search space. In quantum channels, physically equivalent Kraus families are related by a unitary action, so the meaningful parameter space is the quotient $X^\top M X=I_p$05, not the unreduced complex Stiefel manifold (Russkikh et al., 2024).

The practical implication is that “Stiefel optimizer” is best understood as a family of orthogonality-aware optimization schemes rather than a single algorithmic template. What unifies them is the treatment of orthogonality as a geometric constraint to be preserved, exploited, or encoded exactly; what separates them is whether the constraint is handled by intrinsic Riemannian motion, exact retraction after Euclidean updates, quotient reduction, penalty equivalence, or coupled product-manifold dynamics.