--- title: 'Doubly-Uniform Regret: Robustness Across Two Axes' url: https://www.emergentmind.com/topics/doubly-uniform-regret type: topic --- # Doubly-Uniform Regret: Robustness Across Two Axes Doubly-uniform regret is a family of regret guarantees in which a single bound is required to hold simultaneously along two distinct axes of uncertainty or performance measurement. The term does not have a single universal definition across the literature. In discounted online convex optimization, it refers to uniformity over both the discount factor $\lambda$ and the horizon $T$ [2505.19491]. In online linear regression, it denotes regret that is simultaneously uniform in the comparator $w$ and invariant to scaling of the covariates [2605.01628]. In episodic reinforcement learning, the closely related Uniform-PAC framework yields guarantees that are uniform over both accuracy $\varepsilon$ and time $T$ [1703.07710]. In dynamic contextual pricing, the term is used for regret bounds that are uniform over two nonparametric model classes, namely the mean utility function and the noise distribution [2405.06866]. Earlier adjacent usages include “twice uniform regret” in sequential linear regression, meaning uniformity over comparator vectors and worst-case sequences [1805.11386], and uniformly bounded regret in both the number of candidates $n$ and the budget $k$ in the multi-secretary problem [1710.07719]. This suggests that “doubly-uniform” is best understood as a structural property: a regret guarantee that removes dependence on two problem parameters or two specification choices at once. ## 1. Terminological scope and common structure Across the cited works, doubly-uniform regret is not a single theorem but a recurring design objective. The shared pattern is that the learner is asked to compete without knowing in advance which member of a family of objectives, scales, smoothness classes, or horizons will be the relevant one. In each case, the regret guarantee is required to hold simultaneously, rather than after tuning to a single parameter. In discounted online convex optimization, the target is a bound that holds “simultaneously for all $\lambda$ in a continuous interval” and is “essentially uniform over $T$” [2505.19491]. In online linear regression with square loss, the bound must hold “for all $w \in \mathbb{R}^d$ without dependence on $\|w\|$” and also be “scale-invariant” under $x_t \mapsto c x_t$ [2605.01628]. In Uniform-PAC reinforcement learning, one high-probability event controls both the number of $\varepsilon$-suboptimal episodes for all $\varepsilon>0$ and the regret for all $T \ge 1$ [1703.07710]. In dynamic pricing, the uniformity is over two unknown nonparametric objects, $\mu$ and $F$, with regret statements of the form \[ \sup_{\mu \in \mathcal{M}_\ell,\; F \in \mathcal{F}_m} \mathbb{E}[R_T] \le \cdots \] for appropriate smoothness classes [2405.06866]. A plausible implication is that the phrase identifies a shift from single-parameter adaptivity to simultaneous robustness. Rather than optimizing for one comparator norm, one discount factor, one accuracy threshold, or one model class, the algorithm is analyzed against an entire continuum or product class. ## 2. Discounted online convex optimization In discounted online convex optimization, the core performance metric is the $\lambda$-discounted regret \[ \DReg(T,\lambda)= \sum_{t=1}^T \lambda^{T-t} f_t(w_t) - \min_{w\in W} \sum_{t=1}^T \lambda^{T-t} f_t(w), \] where recent losses are weighted more heavily than distant ones [2505.19491]. Under convexity, bounded gradients, bounded domain, and range normalization, Online Gradient Descent with update \[ w_{t+1} = \Pi_W\big[w_t - \eta\, \nabla f_t(w_t)\big] \] achieves $O(1/\sqrt{1-\lambda})$ discounted regret when $\lambda$ is known, using \[ \eta = \frac{D\sqrt{2(1-\lambda)}}{G}, \] with the explicit bound \[ \sum_{t=1}^T \lambda^{T-t} f_t(w_t) - \sum_{t=1}^T \lambda^{T-t} f_t(w) \le \frac{DG\sqrt{2}}{\sqrt{1-\lambda}}. \] The step size and bound are independent of $T$ [2505.19491]. The doubly-uniform question in this setting is whether one can adapt to an unknown discount factor. The interval considered is \[ \Lambda := \left[1-\frac{1}{\tau},\, 1-\frac{1}{T}\right], \] and Smoothed OGD (SOGD) is shown to satisfy, for every comparator $w \in W$ and all $\lambda \in \Lambda$, \[ \sum_{t=1}^T \lambda^{T-t} f_t(w_t) - \sum_{t=1}^T \lambda^{T-t} f_t(w) \le \frac{2GD}{\sqrt{1-\lambda}} \Big(4\sqrt{\log \tfrac{1}{Z}} + \sqrt{2}\Big) +\frac{GD(N+1)Z}{1-\lambda}+2GD, \] where $N = \lceil \log_2(T/\tau) \rceil$ and $Z=1/T$ in the theorem statement, yielding dominant order \[ O\Big(\frac{\sqrt{\log T}}{\sqrt{1-\lambda}}\Big) \] uniformly across all $\lambda \in \Lambda$ [2505.19491]. The algorithmic construction uses a geometric grid \[ \mathcal{S} = \left\{ \lambda_i = 1 - \frac{2^{i-1}}{T} : i = 1,\ldots,N+1 \right\}, \qquad N= \left\lceil \log_2 \tfrac{T}{\tau} \right\rceil, \] with one OGD expert per grid point, and sequentially aggregates them by Discounted-Normal-Predictor with conservative updating (DNP-cu). The combiner forms \[ w_t = (1-\omega_t)\, w_{t,1} + \omega_t\, w_{t,2}, \qquad \ell_t := \frac{f_t(w_{t,1}) - f_t(w_{t,2})}{GD} \in [-1,1], \] and the crucial technical fact is that DNP-cu can aggregate experts even when they optimize discounted regret with different discount factors [2505.19491]. Here, “doubly-uniform” has a precise local meaning: the bound is uniform in $\lambda$ over a continuous interval and uniform in $T$ up to the explicit $\sqrt{\log T}$ adaptivity overhead. Relative to known-$\lambda$ OGD, the price of adaptivity is that factor $\sqrt{\log T}$ [2505.19491]. ## 3. Online linear regression, self-normalization, and scale invariance In online linear regression with square loss, doubly-uniform regret is defined differently. The protocol is \[ \hat y_t = w_t^\top x_t,\qquad R_T(w) = \sum_{t=1}^T (y_t - w_t^\top x_t)^2 - \sum_{t=1}^T (y_t - w^\top x_t)^2. \] A regret bound is uniform over $\mathbb{R}^d$ if it holds simultaneously for all $w \in \mathbb{R}^d$ without explicit dependence on $\|w\|$, and it is doubly-uniform if it is additionally scale-invariant under $x_t \mapsto c x_t$ for any $c>0$ [2605.01628]. The analytical object underlying this definition is the self-normalized quantity \[ V_t = \sum_{s=1}^t X_s X_s^\top,\qquad S_t = \sum_{s=1}^t Y_s X_s,\qquad \|S_t\|_{V_t^\dagger}^2 = S_t^\top V_t^\dagger S_t. \] Under scaling $X_s \mapsto c X_s$, one has $V_t \mapsto c^2 V_t$ and $S_t \mapsto c S_t$, so \[ (c S_t)^\top (c^2 V_t)^\dagger (c S_t) = S_t^\top V_t^\dagger S_t, \] showing that the self-normalized ratio is intrinsically scale-invariant [2605.01628]. The main structural result is dimension-dependent. In dimension $d=1$, nontrivial scale-invariant self-normalized bounds exist without boundedness or moment assumptions on the covariates beyond predictability. Specifically, for any dyadic martingale and any $c \in (0,1/4]$, \[ E[\exp(c R_T)] \le T \exp(c/(1-2c)), \qquad R_T = \frac{S_T^2}{V_T}, \] and therefore \[ E[R_T] \le \frac{\log T}{c} + \frac{1}{1-2c} = O(\log T). \] This leads to an explicit algorithm with deterministic regret \[ \mathrm{Reg}(T) \le C m^2 \log T \] for $d=1$ and $|y_t| \le m$, uniform in $w \in \mathbb{R}$ and scale-invariant in $x_t$ [2605.01628]. For $d>1$, the paper proves impossibility in full generality. For any $\varepsilon \in (0,1)$ and $d \ge 2$, there exists a dyadic martingale such that \[ E[\|S_T\|_{V_T^\dagger}^2] \ge (1-\varepsilon^2) T, \] and this transfers to regret lower bounds showing that sublinear doubly-uniform regret is impossible without additional assumptions [2605.01628]. The obstruction is geometric: the adversary can inject energy in directions orthogonal to the current information direction. Under a smoothness condition on the conditional covariate laws, \[ \frac{dP_t(\cdot \mid H_{t-1}^x)}{d\mu}(x) \le S \quad\text{for $\mu$-a.e. }x, \] sublinear regret reappears in $d>1$. The unregularized VAW predictor then satisfies \[ \mathrm{Reg}(T) \lesssim \sqrt{d T \log(T/\delta)} + \log(1/\delta) \] with probability at least $1-\delta$, and the self-normalized concentration inequality becomes \[ \|S_T\|_{V_T^\dagger}^2 \lesssim \sigma^2 \Big( \sqrt{d T \log(2T/\delta)} + \log(2/\delta) \Big) \] without a regularization matrix $\Gamma$ and without boundedness assumptions on $\|X_t\|$ [2605.01628]. This literature also connects directly to an earlier notion of “twice uniform regret.” In sequential linear regression with square loss, uniform regret over $\mathbb{R}^d$ means \[ \sup_{u \in \mathbb{R}^d} \mathrm{Reg}_T(u) = \sum_{t=1}^T (y_t - \hat y_t)^2 - \inf_{u \in \mathbb{R}^d} \sum_{t=1}^T (y_t - u^\top x_t)^2, \] and “twice uniform regret” refers to uniformity over all competitor vectors and worst-case feature and observation sequences [1805.11386]. When features are known beforehand, the adapted metric forecaster achieves \[ \sup_{u \in \mathbb{R}^d} \mathrm{Reg}_T(u) \le d\,B^2 \ln\!\Bigl(1 + \frac{T}{d}\Bigr) + d\,B^2, \] while the minimax lower bound is \[ R^\star_{T,\,[-B,B]} \ge d\,B^2\left(\ln T - (3+\ln d) - \ln\ln T \right). \] For sequentially revealed features, the parameter-free $\lambda=0$ variant satisfies \[ \sup_{u \in \mathbb{R}^d} \mathrm{Reg}_T(u) \le B^2 \sum_{t=1}^T x_t^\top G_t^\dagger x_t, \] which yields asymptotic order $d B^2 \ln T$ for any individual sequence, but a worst-case doubly-uniform bound remains open [1805.11386]. ## 4. Uniformity in accuracy and time in episodic reinforcement learning In episodic finite-horizon reinforcement learning, the closely related concept is Uniform-PAC. For episodic regret, \[ \mathrm{Reg}(T) = \sum_{t=1}^T \big( V_1^*(s_1^{(t)}) - V_1^{\pi_t}(s_1^{(t)}) \big), \] and for PAC-style performance the key count is \[ N_\varepsilon = \sum_{t=1}^{\infty} \mathbb{1}\!\big\{V_1^*(s_1^{(t)}) - V_1^{\pi_t}(s_1^{(t)}) > \varepsilon\big\}. \] An algorithm is Uniform-PAC if, for $\delta>0$, \[ \mathbb{P}\Big(\exists\,\varepsilon>0:\;N_{\varepsilon} > F_{\mathrm{UPAC}}(S,A,H,1/\varepsilon,\log(1/\delta))\Big)\leq\delta, \] with one event controlling all $\varepsilon \in (0,1]$ [1703.07710]. The UBEV algorithm achieves, with probability at least $1-\delta$, simultaneously for all $\varepsilon>0$, \[ N_\varepsilon(\delta) \le O\!\left( \frac{S A H^4}{\varepsilon^2} \min\!\{1+\varepsilon S^2 A,\; S\} \;\mathrm{polylog}\!\Big(A,S,H,\tfrac{1}{\varepsilon},\tfrac{1}{\delta}\Big) \right). \] The conversion theorem then shows that if one has a bound of the form \[ N_\varepsilon \le \frac{C_1}{\varepsilon}\Big(\ln\!\frac{C_3}{\varepsilon}\Big)^k + \frac{C_2}{\varepsilon^2}\Big(\ln\!\frac{C_3}{\varepsilon}\Big)^{2k}, \] then on the same high-probability event one also has, for all $T \ge 1$ simultaneously, \[ R(T) \le \big(\sqrt{C_2 T}+C_1\big)\,\mathrm{polylog}(T,C_3,C_1). \] For UBEV, this yields \[ R(T)=O\!\Big(H^2\big(\sqrt{S A T}+S^3 A^2\big)\;\mathrm{polylog}(S,A,H,T)\Big) \] for all $T$ with probability at least $1-\delta$ [1703.07710]. The paper does not use the phrase “doubly-uniform regret,” but the property is explicit: uniformity in $\varepsilon$ through Uniform-PAC and uniformity in $T$ through anytime high-probability regret. The technical mechanism is time-uniform concentration, including finite-time law-of-the-iterated-logarithm style confidence widths such as \[ \phi(s,a,t) = \sqrt{ \frac{ 2\ln\ln(\max\{e, n(s,a,t)\}) + \ln\!\big(\tfrac{18 S A H}{\delta}\big) }{ n(s,a,t) } }. \] This use of “double uniformity” differs from the OCO and regression usages, but it preserves the same structural theme: a single event controls an entire continuum of thresholds and all horizons [1703.07710]. ## 5. Dynamic contextual pricing and two-class uniformity In dynamic contextual pricing under doubly nonparametric random utility models, the data are contexts $x_t \in \mathbb{R}^d$, prices $p_t \in [0,B]$, and binary purchases $y_t \in \{0,1\}$. The model is \[ v_t = \mu(x_t) + \varepsilon_t,\qquad \mathbb{E}[\varepsilon_t]=0,\qquad y_t = \mathbf{1}\{v_t \ge p_t\}, \] with both $\mu$ and the noise CDF $F$ unknown and modeled nonparametrically. Revenue is \[ \operatorname{rev}_t(p) = p\Big(1 - F(p - \mu(x_t))\Big), \] and regret is \[ R_T = \sum_{t=1}^T \Big( r(x_t,p^\star(x_t)) - r(x_t,p_t) \Big) \] with oracle price \[ p_t^\star = \arg\max_{p\in[0,B]} p\Big(1 - F(p-\mu(x_t))\Big) \] [2405.06866]. Identification is based on two population equations under uniform random exploration prices. Writing $g_t := B y_t$, \[ \mathbb{E}[B y_t \mid x_t] = \mu(x_t), \qquad F(z) = 1 - \mathbb{E}[y_t \mid p_t - \mu(x_t)=z]. \] The oracle pricing map is expressed through \[ \phi(z) = z - \frac{1-F(z)}{F^{(1)}(z)}, \qquad h(v)=v+\phi^{-1}(-v), \] so that \[ p_t^\star = h(\mu(x_t)) \] under the regularity assumption $\phi^{(1)}(z)\ge L_\phi>0$ [2405.06866]. The “doubly-uniform” aspect has two layers. First, the estimators achieve uniform sup-norm control over their domains. For DNN, \[ \sup_{x\in\mathcal X} \big| \widehat\mu^{\mathrm{DNN}}(x;s)-\mu(x) \big| \le C s^{-2/d} + B\sqrt{ \frac{2s[\log(\delta^{-1})+\log d + d\log n]}{n} }, \] and for TDNN, \[ \sup_{x\in\mathcal X} \big| \widehat\mu^{\mathrm{TDNN}}(x;s_1,s_2)-\mu(x) \big| \le C s_1^{-\min\{3,4/d\}} + C B\sqrt{ \frac{2s_1[\log(\delta^{-1})+\log d + d\log n]}{n} }. \] The kernel estimators of $F$ and $F^{(1)}$ are also controlled uniformly over $z$ and over $\mu$ in a neighborhood [2405.06866]. Second, the regret bounds are uniform over both nonparametric classes: \[ \sup_{\mu \in \mathcal{M}_2,\; F \in \mathcal{F}_m} \mathbb{E}[R_T] \le C\; T^{\,\frac{2m+1}{4m-1} \;\lor\; \frac{d+4}{d+8}} \big((\log T)^2 + d\log T\big) \] for the DNN-based policy, and \[ \sup_{\mu \in \mathcal{M}_4,\; F \in \mathcal{F}_m} \mathbb{E}[R_T] \le C\; T^{\,\frac{2m+1}{4m-1} \;\lor\; \frac{d+8}{d+16} \;\lor\; \frac{7}{13}} \big((\log T)^2 + d\log T\big) \] for the TDNN-based policy [2405.06866]. The paper explicitly characterizes this as “doubly-uniform” because the bounds are uniform over both the mean utility class and the noise distribution class. The analysis combines uniform convergence, stability of $\phi^{-1}$, and a second-order expansion showing that per-period regret scales quadratically in the price error: \[ \mathrm{rev}_t(p_t) = \mathrm{rev}_t(p_t^\star) + \tfrac{1}{2}\mathrm{rev}_t^{(2)}(\tilde p)(p_t-p_t^\star)^2. \] This makes the interaction between context dimension $d$ and noise smoothness $m$ explicit in the regret exponent [2405.06866]. ## 6. Related formulations: universality, adaptivity, and bounded regret across problem parameters A recent OCO formulation uses “doubly-uniform” to mean simultaneous universality across curvature classes and adaptivity to gradient variation $V_T$ [2511.19937]. In this setting, \[ R_T = \sum_{t=1}^T f_t(x_t) - \min_{x\in X}\sum_{t=1}^T f_t(x), \qquad V_T := \sum_{t=2}^T \sup_{x\in X}\|\nabla f_t(x)-\nabla f_{t-1}(x)\|^2. \] The goal is a single algorithm that simultaneously achieves \[ O(\sqrt{V_T}) \text{ or } O(\sqrt{V_T\log V_T}), \qquad O((d/\alpha)\log V_T), \qquad O((1/\lambda)\log V_T) \] for convex, exp-concave, and strongly convex losses respectively, while also recovering the standard worst-case $T$-based universal rates [2511.19937]. UniGrad.Correct and UniGrad.Bregman realize this by maintaining $N=\Theta(\log T)$ base learners on exponential grids \[ H^{\exp}=H^{\scvx}=\{2^0/T,2^1/T,\ldots,2^{n-1}/T\}, \qquad n=\lfloor \log_2 T \rfloor + 1. \] UniGrad.Correct achieves \[ O((1/\lambda)\log V_T),\quad O((d/\alpha)\log V_T),\quad O(\sqrt{V_T\log V_T}), \] while UniGrad.Bregman achieves the same curvature-adaptive logarithmic bounds and the optimal convex \[ O(\sqrt{V_T}) \] rate [2511.19937]. The authors describe this as “universality + adaptivity” simultaneously. This is another instance of double uniformity, now across curvature families and variation regimes. A different but historically important example appears in the multi-secretary problem. There, a policy has doubly-uniform regret if there exists a constant $C$ independent of both $n$ and $k$ such that \[ \sup_{(n,k):0\le k\le n} R_{n,k}(\pi) \le C. \] With finite support $\mathcal A=\{a_m<\cdots