---
title: 'SAFE-T: Non-Stochastic Control Under Safety Constraints'
url: https://www.emergentmind.com/topics/safe-t
type: topic
---

# SAFE-T: Non-Stochastic Control Under Safety Constraints

SAFE-T is an overloaded acronym in the arXiv literature. Most prominently, it denotes “Safe control under Time-varying constraints with adversarial, non-stochastic disturbances,” introduced in “Safe Non-Stochastic Control of Linear Dynamical Systems,” where safe control is cast as an online sequential game over linear feedback gains with per-step safety and bounded dynamic regret guarantees [2308.12395]. The same acronym has also been used for “Safe Linear Thompson Sampling with Side Information” in stochastic linear bandits [1911.02156], for a future-aware safe active learning framework for time-varying systems using Gaussian processes [2405.10581], and for several unrelated frameworks in graph algorithms, experimentation, safe reinforcement learning, and holography. In current usage, however, SAFE-T most often refers to the non-stochastic control framework centered on robust safety under time-varying convex constraints and adversarial disturbances [2308.12395].

## 1. SAFE-T as safe non-stochastic control

In its control-theoretic sense, SAFE-T studies linear dynamical systems corrupted with non-stochastic noise and seeks two simultaneous guarantees: zero constraint violation of convex time-varying constraints, and bounded dynamic regret against an optimal clairvoyant controller that knows the future disturbance sequence a priori [2308.12395]. The motivating setting is autonomy under real-world unpredictable disturbances such as wind and wake disturbances, with safety constraints that bound both the state and the control input.

The system model is linear and may be time-varying:
\[
x_{t+1} = A_t x_t + B_t u_t + w_t,\quad t=1,\ldots,T,
\]
with state \(x_t\in\mathbb{R}^{d_x}\), input \(u_t\in\mathbb{R}^{d_u}\), known matrices \(A_t,B_t\), and disturbance \(w_t\in\mathbb{R}^{d_x}\). The time-invariant case is the special case \(A_t\equiv A\), \(B_t\equiv B\). SAFE-T assumes a known disturbance radius and imposes no probabilistic model:
\[
\|A_t\|\le\kappa_A,\quad \|B_t\|\le\kappa_B,\quad w_t\in\mathcal{W}\triangleq\{w:\|w\|\le W\}.
\]

Safety is enforced through time-varying convex constraints on state and input:
\[
\begin{aligned}
x_t &\in \mathcal{X}_t \triangleq \{x : L_{x,t} x \le l_{x,t}\},\\
u_t &\in \mathcal{U}_t \triangleq \{u : L_{u,t} u \le l_{u,t}\},
\end{aligned}
\]
with polytopes as the leading special case. The per-round control objective is encoded through a convex loss \(c_t(x_{t+1},u_t)\), with bounded values and gradients on bounded domains. A common example is the quadratic cost
\[
c_t(x_{t+1},u_t)=x_{t+1}^\top Q x_{t+1} + u_t^\top R u_t.
\]

SAFE-T rewrites the problem as a sequential game. At each step \(t\), before observing \(w_t\), the controller selects a linear state-feedback gain \(K_t\) and applies
\[
u_t = -K_t x_t.
\]
After \(w_t\) is revealed through the next state, the controller incurs
\[
f_t(K_t)\triangleq c_t\big((A_t-B_t K_t) x_t + w_t,\ -K_t x_t\big),
\]
which is convex in \(K_t\) given \(A_t,B_t,x_t,w_t\). The controller therefore faces an online convex optimization problem with adversarial disturbances and time-varying safe domains.

## 2. Safety-certified gain sets and robust tightening

The central SAFE-T construction is the time-varying admissible set of feedback gains \(\mathcal{K}_t\), built so that every chosen gain is robustly safe for all disturbances satisfying \(\|w_t\|\le W\) [2308.12395]. The controller selects only gains in
\[
\boxed{
\begin{aligned}
\mathcal{K}_t \triangleq \big\{K\ \big|\ &-L_{x,t+1}\,B_t K x_t \ \le\ l_{x,t+1}-L_{x,t+1}A_t x_t - W\|L_{x,t+1}\|,\\
&-L_{u,t}\,K x_t \ \le\ l_{u,t},\\
&\|K\|\le \kappa,\quad \|A_t - B_t K\| \le 1-\gamma\big\},
\end{aligned}
}
\]
where \(\kappa>0\) and \(\gamma\in(0,1)\) are design parameters.

The first inequality is a robust tightening of the next-state constraint. Starting from
\[
x_{t+1}=A_t x_t + B_t u_t + w_t,\quad u_t=-K_t x_t,
\]
and requiring \(L_{x,t+1}x_{t+1}\le l_{x,t+1}\) for all \(\|w_t\|\le W\), one obtains
\[
L_{x,t+1} (A_t x_t - B_t K_t x_t) + L_{x,t+1} w_t \le l_{x,t+1}\quad \forall w_t\in\mathcal{W}.
\]
Using the disturbance bound and triangle inequality yields the worst-case term \(W\|L_{x,t+1}\|\), hence
\[
-L_{x,t+1} B_t K_t x_t \le l_{x,t+1}-L_{x,t+1}A_t x_t - W\|L_{x,t+1}\|.
\]
Similarly, the input constraint becomes
\[
-L_{u,t}K_t x_t\le l_{u,t}.
\]

The resulting guarantee is per-step and robust: if \(K_t\in\mathcal{K}_t\), then for any disturbance \(\|w_t\|\le W\),
\[
u_t=-K_t x_t\in\mathcal{U}_t,\qquad x_{t+1}\in\mathcal{X}_{t+1}.
\]
Thus SAFE-T enforces zero per-step violations of the time-varying convex constraints. The framework requires one-step-ahead state-constraint information \((L_{x,t+1},l_{x,t+1})\) to construct \(\mathcal{K}_t\), and it uses the technical assumption of recursive feasibility, \(\mathcal{K}_t\neq\emptyset\) for all \(t\). A plausible implication is that feasibility management is part of the safety architecture rather than a secondary implementation detail.

## 3. Safe-OGD and the online update mechanism

SAFE-T solves the sequential game with “Safe Online Gradient Descent” (Safe-OGD), which generalizes Online Gradient Descent to time-varying safe domains by projecting onto the next admissible set [2308.12395]. The per-round update is:

1. Pick \(K_t\in\mathcal{K}_t\) and apply \(u_t=-K_t x_t\).
2. Observe \(x_{t+1}\) and compute \(w_t=x_{t+1}-A_t x_t - B_t u_t\).
3. Suffer loss \(c_t(x_{t+1},u_t)\) and form \(f_t(K_t)\).
4. Compute \(\nabla_K f_t(K_t)\).
5. Obtain \(\mathcal{K}_{t+1}\).
6. Perform a gradient step and projection:
\[
K_{t+1}^\prime = K_t - \eta\,\nabla_K f_t(K_t),\quad K_{t+1} = \Pi_{\mathcal{K}_{t+1}}(K_{t+1}^\prime),
\]
where \(\Pi_{\mathcal{K}}\) denotes projection onto \(\mathcal{K}\).

The boundedness and stability constraints in \(\mathcal{K}_t\) are not merely regularizers; they are part of the regret analysis. The conditions \(\|K_t\|\le\kappa\) and \(\|A_t-B_tK_t\|\le 1-\gamma\) imply
\[
\|x_t\|\le D,\quad \|u_t\|\le D,\quad D\triangleq \max\Big\{\frac{W}{\gamma},\frac{W\kappa}{\gamma}\Big\}.
\]
This, in turn, yields a Frobenius-norm gradient bound
\[
\|\nabla_K f_t(K)\|_{\mathrm{F}}\le G_f,\quad G_f\triangleq G\,D\,d_x d_u(\kappa_B+1),
\]
and an admissible-set diameter bound
\[
\|K_1-K_2\|_{\mathrm{F}}\le D_f,\quad D_f=2\kappa\sqrt{d},\quad d=\min\{d_u,d_x\}.
\]

This structure distinguishes SAFE-T from standard online convex optimization with static domains. The projection set \(\mathcal{K}_{t+1}\) depends on the next-step safety description and also on the current state \(x_t\). As a result, even when \(\mathcal{X}_t\) and \(\mathcal{U}_t\) are time-invariant, the feasible gain domain can still vary over time.

## 4. Dynamic regret, comparator structure, and theoretical guarantees

SAFE-T evaluates performance through dynamic policy regret against a safe clairvoyant sequence of linear feedback gains \(\{K_t^\star\}\) that knows the disturbances in advance [2308.12395]. The regret is
\[
\boxed{
\operatorname{Regret}_{\text{NSC}}
\triangleq
\sum_{t=1}^{T} c_t(x_{t+1},u_t)\ -\ \sum_{t=1}^{T} c_t(x_{t+1}^\star,u_t^\star),
}
\]
where \(u_t^\star=-K_t^\star x_t\), \(x_{t+1}^\star=A_t x_t + B_t u_t^\star + w_t\), and both trajectories are evaluated under the same disturbance sequence.

The regret bound depends on two variation quantities. The first is the comparator path length,
\[
C_T \triangleq \sum_{t=2}^{T} \|K_{t-1}^\star - K_t^\star\|_{\mathrm{F}},
\]
which measures how rapidly the clairvoyant optimal controller changes. The second is the domain-variation term,
\[
S_T \triangleq \sum_{t=1}^{T} \zeta_t,\quad
\zeta_t\triangleq \|\Pi_{\mathcal{K}_t}(K_{t+1}^\prime) - \Pi_{\mathcal{K}_{t+1}}(K_{t+1}^\prime)\|_{\mathrm{F}},
\]
which quantifies the effect of changing the projection set from \(\mathcal{K}_t\) to \(\mathcal{K}_{t+1}\).

The main SAFE-T bound is
\[
\boxed{
\operatorname{Regret}_{\text{NSC}}
\le
\frac{\eta T G_f^2}{2}
+
\frac{7 D_f^2}{4\eta}
+
\frac{D_f C_T}{\eta}
+
\frac{D_f S_T}{\eta}.
}
\]
With \(\eta=\Theta(1/\sqrt{T})\), this yields
\[
\boxed{
\operatorname{Regret}_{\text{NSC}}
=
\mathcal{O}\Big(\sqrt{T}\,\big(1+C_T+S_T\big)\Big).
}
\]

This bound extends the familiar dynamic-regret scaling from online convex optimization to time-varying safe domains. When the domain is time-invariant, \(S_T=0\), and the bound reduces to \(\mathcal{O}(\sqrt{T}(1+C_T))\). This suggests that the extra term \(S_T\) is the precise analytical price of safety-domain variation, rather than a generic looseness of the proof.

A recurring theme is the safety–performance trade-off. Robust tightening subtracts \(W\|L_{x,t+1}\|\) from the state constraint, and the stability condition \(\|A_t-B_tK_t\|\le 1-\gamma\) restricts the feasible set further. These features guarantee zero violations, but may shrink \(\mathcal{K}_t\) and degrade achievable performance. Conversely, in high-rate control with slowly varying constraints, \(S_T\) can be small, improving the regret bound.

## 5. Computation, empirical validation, and limitations

The per-round computation in SAFE-T has two main parts: gradient evaluation for \(c_t\big((A_t-B_tK)x_t+w_t,-Kx_t\big)\), and projection onto \(\mathcal{K}_{t+1}\) [2308.12395]. The projection is a convex optimization problem with linear inequalities and convex norm constraints \(\|K\|\le\kappa\) and \(\|A_t-B_tK\|\le 1-\gamma\). The latter is convex in \(K\) as an operator-norm ball, and the projection can be solved via second-order cone or semidefinite relaxation depending on norm choice. Its complexity scales polynomially in the problem size and number of constraints.

The step size is set to \(\eta=\Theta(1/\sqrt{T})\) for the regret guarantee. Larger \(\eta\) accelerates adaptation but increases the \((\eta T G_f^2)/2\) term; smaller \(\eta\) decreases that term but enlarges the \((1/\eta)\)-scaled terms. The stability margin \(\gamma\) and the gain bound \(\kappa\) also affect the bound through \(D\) and \(D_f\).

The reported simulation domain is a quadrotor hovering task with LTI dynamics
\[
x_{t+1}=A x_t + B u_t + w_t,
\]
a 6D state, 3D inputs, disturbance bound \(\|w_t\|\le 0.1\), horizon \(T=500\), state constraints
\[
-\mathbf{1}_{6}\le x_t\le \mathbf{1}_{6},
\]
input constraints
\[
[-\pi,-\pi,-20]^\top\le u_t\le [\pi,\pi,20]^\top,
\]
and cost
\[
c_t(x_{t+1},u_t)=x_{t+1}^\top x_{t+1}+u_t^\top u_t.
\]

The reported outcomes are threefold. First, all methods considered, including SAFE-OGD, ensure zero violations of state and input constraints. Second, SAFE-OGD achieves competitive or superior cumulative loss relative to safe \(H_2\) and \(H_\infty\) controllers with shorter horizons such as \(N=1,5\), while \(N=10\) baselines can achieve lower loss at substantially higher computational cost. Third, SAFE-OGD is significantly faster per round, reported as up to \(9\times\) faster than \(H_2\) and \(114\times\) faster than \(H_\infty\) at \(N=10\).

The framework nonetheless rests on restrictive assumptions: known disturbance bound \(W\), known \(A_t,B_t\), convex safety sets, one-step-ahead knowledge of safety parameters, and recursive feasibility. The paper notes that robust MPC or tube-MPC ideas with a terminal invariant set and a baseline safe stabilizing controller can be used to enforce recursive feasibility when a lookahead horizon is available. Extensions proposed in the source include nonlinear systems, model uncertainty, partial observations, and stochastic disturbances treated through a high-probability or moment bound playing the role of \(W\). These are described as directions rather than established guarantees.

## 6. Other arXiv uses of the acronym

The acronym SAFE-T has been reused in several technically unrelated literatures. The following usage patterns appear in the supplied arXiv corpus.

| SAFE-T usage | Area | Core idea |
|---|---|---|
| “Safe Linear Thompson Sampling with Side Information” [1911.02156] | Linear stochastic bandits | Stage-wise linear safety constraints with side measurements \(w_t=x_t^\top\mu^\star+\zeta_t\) |
| Future-aware safe active learning [2405.10581] | Gaussian-process active learning | T-IMSPE minimizes posterior variance over current and future states under safety filtering |
| “Safety in \(s\)-\(t\) Paths, Trails and Walks” [2007.04726] | Graph algorithms | Safe subwalks common to all \(s\)-\(t\) paths, trails, or walks |
| “Safe Testing” [2310.20450] | Sequential experimentation | Anytime-valid inference via e-values and e-processes |
| Safe online RL in 1D LQR [2504.18657] | Reinforcement learning | High-probability safety and \(\tilde{O}_T(\sqrt{T})\)-regret with truncated linear controllers |
| “Safe Gauge-String Correspondence” [1912.06097] | High-energy theory | Safe gauge theories dual to safe noncritical strings on asymptotically AdS |
| Safe sets in weighted trees [1711.11412] | Combinatorics | PTAS for the connected safe number of a weighted tree |

In linear bandits, SAFE-T denotes a frequentist Thompson-sampling algorithm for linear reward maximization under an unknown linear safety constraint \(x_t^\top\mu^\star\le C\), with side information \(w_t=x_t^\top\mu^\star+\zeta_t\) observed at each played action [1911.02156]. The safe action set is the robust inner approximation
\[
D_t^s = \{ x \in D_0 : x^\top \hat{\mu}_t + \beta_t(\delta') \|x\|_{V_t^{-1}} \le C \},
\]
and the reported regret order matches that of linear Thompson sampling without safety constraints up to logarithmic factors.

In Gaussian-process active learning, SAFE-T refers to a future-aware safe acquisition rule for time-varying systems. Its central objective is T-IMSPE,
\[
\mathrm{T\text{-}IMSPE}(\mathcal{D}_t)
=
\sum_{\tau=0}^{H}
\int_{\mathcal{X}} w(x,t+\tau)\,\operatorname{Var}\bigl(f(x,t+\tau)\mid \mathcal{D}_t\bigr)\,\mathrm{d}x,
\]
optimized subject to safety filtering \(\xi(x,t)\ge \alpha\), with \(\alpha=0.977\) used in the experiments [2405.10581].

In graph theory, SAFE-T is the study of safe subwalks with respect to all \(s\)-\(t\) paths, trails, or walks of a directed graph [2007.04726]. The paper shows linear-time characterizations and algorithms in several cases, an \(O(m+n)\) compact representation for maximal safe walks, and an NP-hardness dichotomy for visible-subset variants of the path and trail problems.

In large-scale experimentation, Safe Testing uses e-values and e-processes for anytime-valid inference [2310.20450]. The defining condition is
\[
\forall P \in \mathcal{H}_0: \quad \mathbb{E}_P[E] \le 1,
\]
with stopping rule \(E_t\ge 1/\alpha\). This enables continuous monitoring without inflating Type I error under optional stopping.

In safe online reinforcement learning for one-dimensional LQR, SAFE-T denotes a high-probability safety-constrained learning framework with regret
\[
R_T \le \tilde{O}_T(\sqrt{T}),
\]
measured relative to a baseline of truncated linear controllers that clip controls at the safety boundary [2504.18657]. The safety requirement is formulated as
\[
D_L \le a^\* x_t + b^\* u_t \le D_U
\]
with high probability for all times.

The remaining uses are unrelated to control or learning. “Safe Gauge-String Correspondence” studies holographic duality for safe gauge theories with non-Gaussian ultraviolet fixed points [1912.06097]. “Approximating Connected Safe Sets in Weighted Trees” studies \(w\)-safe sets and proves a PTAS for the connected safe number of a weighted tree [1711.11412].

Across these usages, the shared lexical theme is invariance of a “safe” set under uncertainty, but the mathematical object varies widely: convex gain domains, ellipsoidal confidence-safe action sets, posterior safe regions, safe subwalks, e-processes, truncated controllers, or graph cuts. A plausible implication is that SAFE-T is better understood as a family of domain-specific safety formalisms than as a single unified framework.

Source: https://www.emergentmind.com/topics/safe-t