Optimal conversion from Rényi Differential Privacy to f f -Differential Privacy Thanks: Preprint. Under review. Anneliese Riess, Juan Felipe Gomez, Flavio du Pin Calmon, Julia Anne Schnabel, Georgios Kaissis Abstract We prove the conjecture stated in Appendix F.3 of 13 : among all conversion rules that map a Rényi Differential Privacy (RDP) profile τ ↦ ρ ⁡ ( τ ) \tau\mapsto\rho(\tau) to a valid hypothesis-testing trade-off f f (or equivalently, an ( ε , δ ) (\varepsilon,\delta) -Differential Privacy curve), the rule based on the intersection of single-order RDP privacy regions is optimal. This optimality holds simultaneously for all valid RDP profiles and for all Type I error levels α \alpha . Concretely, we show that in the space of trade-off functions, the tightest possible bound is f ρ ⁡ ( ⋅ ) ​ ( α ) = sup τ ≥ 0.5 f τ , ρ ⁡ ( τ ) ​ ( α ) f_{\rho(\cdot)}(\alpha)=\sup_{\tau\geq 0.5}f_{\tau,\rho(\tau)}(\alpha) : the pointwise maximum of the single-order bounds for each RDP privacy region. Our proof unifies and sharpens the insights of 2 , 1 , and 13 . Our analysis relies on a precise geometric characterization of the RDP privacy region, leveraging its convexity and the fact that its boundary is determined exclusively by Bernoulli mechanisms. Our results establish that the “intersection-of-RDP-privacy-regions” rule is not only valid, but optimal: no other black-box conversion can uniformly dominate it in the Blackwell sense, marking the fundamental limit of what can be inferred about a mechanism’s privacy solely from its RDP guarantees. 1 Introduction The hypothesis testing interpretation of Differential Privacy (DP), often formalized as f -DP ( 3 ) , has emerged as a rigorous standard for privacy analysis, grounding privacy guarantees in the operational framework of binary hypothesis testing. By characterizing the trade-off between Type I and Type II errors that an adversary faces when distinguishing between adjacent datasets; f -DP provides a complete and geometrically interpretable picture of privacy loss. However, despite the rise of numerical accounting tools, Rényi Differential Privacy (RDP) remains indispensable for many tasks due to its analytical tractability, particularly in complex settings such as graph learning ( 12 ) . Unlike f -DP or Total Variation privacy ( 6 ) , which admit direct variational representations involving a single rejection region, the Rényi divergence does not admit a direct hypothesis-testing interpretation for general distributions ( 2 ) . To bridge the gap between the calculable moments of RDP and the interpretable error trade-offs of f -DP, one must rely on the “two-cut” reduction ( 2 ) to analyze what a divergence constraint implies for the induced binary distributions of a hypothesis test. This necessity drives the construction of the RDP privacy regions : the set of all possible error pairs (\alpha,\beta) compatible with a given RDP guarantee. In this work, we address the problem of optimally converting an RDP profile into f -DP. While prior works, such as 1 , have solved the variational problem for a single Rényi order \tau , a mechanism typically satisfies a continuum of constraints defined by a profile \tau\mapsto\rho(\tau) (also known as functional RDP ( 10 ) ). We prove the conjecture stated in Appendix F.3 of 13 : among all conversion rules that map an RDP profile to a valid hypothesis-testing trade-off f , the lower boundary of the intersection of \tau -RDP privacy regions over all \tau\in[0.5,\infty) is optimal. Any tighter conversion rule would necessarily require information about the mechanism beyond its RDP profile. We solve the functional optimization problem over this entire trajectory, seeking the tightest possible envelope that holds for all mechanisms satisfying a given RDP profile. Our proof unifies and sharpens the insights of 2 , 1 , and 13 . Our main contribution is establishing the fundamental limit of black-box privacy conversion. We prove that the trade-off function derived from the intersection of privacy regions for all valid Rényi orders is the pointwise optimal one. By constructing witness mechanisms, namely, specific instances of Randomized Response that exactly saturate this bound, we demonstrate that no tighter conversion rule can exist without inspecting other properties of the mechanism. This result elevates the proposed conversion from a technical improvement to a definitive conclusion for RDP-to-DP conversion research: we have reached the theoretical ceiling of what can be inferred solely from RDP parameters. Figure 1 : The fundamental limit of RDP-to- f -DP conversion. The blue curve is our optimal conversion bound derived for the RDP profile of a Gaussian mechanism with unit sensitivity and noise scale \sigma=1 , defined by \rho(\tau)=\frac{\tau}{2\sigma^{2}} , \tau\in[0.5,\infty) . The black curve indicates the true trade-off for the Gaussian mechanism f(\alpha)=\Phi(\Phi^{-1}(1-\alpha)-1/\sigma) . The other piecewise linear functions represent witness mechanisms (Randomized Response) that saturate the RDP conversion at various orders. The shaded gray region illustrates the “optimality gap” : the inevitable loss of tightness when converting a mechanism solely via its RDP profile. Our bound (blue) is the tightest possible envelope for the class of all mechanisms sharing this RDP profile, as evidenced by the witness mechanisms touching the converted trade-off. 2 Preliminaries We assume familiarity with the fundamental concepts of differential privacy; however, to ensure this work is self-contained, we review the necessary definitions and notation in this section. We denote by \mathbb{D} the universe of all possible datasets. Let (\mathcal{Y},\Sigma) be a measurable space, and let \mathcal{P}(\mathcal{Y}) denote the set of all probability measures on \mathcal{Y} . 2.1 Privacy Definitions via Divergences A randomized mechanism \mathcal{M}:\mathbb{D}\to\mathcal{P}(\mathcal{Y}) maps a dataset \mathcal{D} to a probability distribution \mathcal{M}(\mathcal{D}) . We define the adjacency relation \mathcal{D}\sim\mathcal{D}^{\prime} for datasets differing by a single record. Definition 2.1 ( (\varepsilon,\delta) -Differential Privacy) . A mechanism \mathcal{M} satisfies (\varepsilon,\delta) -DP if and only if for all \mathcal{D}\sim\mathcal{D}^{\prime} : E_{e^{\varepsilon}}(\mathcal{M}(\mathcal{D})\|\mathcal{M}(\mathcal{D}^{\prime}))\leq\delta, (1) where E_{\gamma}(P\|Q):=\int_{\mathcal{Y}}\max(0,p(y)-\gamma q(y))d\mu(y) is the hockey-stick divergence, with p and q denoting the densities of P and Q with respect to a dominating measure \mu . Definition 2.2 (Rényi Differential Privacy ( 7 ) ) . A mechanism \mathcal{M} satisfies (\tau,\rho) -RDP if and only if for all \mathcal{D}\sim\mathcal{D}^{\prime} : D_{\tau}(\mathcal{M}(\mathcal{D})\|\mathcal{M}(\mathcal{D}^{\prime}))\leq\rho, (2) where D_{\tau}(P\|Q):=\frac{1}{\tau-1}\log\mathbb{E}_{y\sim Q}[(p(y)/q(y))^{\tau}] is the Rényi divergence of order \tau>1 . The function \rho:(1,\infty)\to[0,\infty] satisfying D_{\tau}(\mathcal{M}(\mathcal{D})\|\mathcal{M}(\mathcal{D}^{\prime}))\leq\rho(\tau) for all adjacent \mathcal{D}\sim\mathcal{D}^{\prime} and all \tau>1 is called the RDP profile . In such a case, we say \mathcal{M} satisfies (\tau,\rho(\tau)) -RDP. Extended Domain. While the standard definition of RDP focuses on orders \tau>1 ( 7 ) , 2 demonstrated that the constraints corresponding to \tau\in(0,1) are essential for a tight geometric characterization of the privacy region. Although the privacy parameter \rho is not typically accounted for in these regimes, the underlying Rényi divergence constraints D_{\tau}(P\|Q) are well-defined for all \tau\in\mathbb{R} (assuming P and Q have common support). Consequently, we implicitly view the profile \rho(\cdot) as being defined over the extended domain. However, due to the symmetry of the adjacency relation (implying bounds on D_{\tau}(P\|Q) equate to bounds on D_{1-\tau}(Q\|P) ), we restrict our attention to \tau\in(0.5,\infty) ( 9 ) . Throughout this manuscript, we do not specify the functional form of \rho(\tau) . However, we assume that \rho(\cdot) is a valid RDP profile, meaning that the set of mechanisms satisfying the (\tau,\rho(\tau)) -RDP constraints is non-empty. 2.2 Hypothesis Testing The operational interpretation of DP is best understood through binary hypothesis testing. Consider the task of distinguishing between two adjacent datasets, \mathcal{D} and \mathcal{D}^{\prime} , based on the output Y\in\mathcal{Y} of a mechanism \mathcal{M} . This task corresponds to testing the null hypothesis H_{0}:Y\sim P against the alternative H_{1}:Y\sim Q , where P=\mathcal{M}(\mathcal{D}) and Q=\mathcal{M}(\mathcal{D}^{\prime}) . A decision rule is defined by a rejection region S\subseteq\mathcal{Y} (where we reject H_{0} if Y\in S ). This induces two types of errors: Type I Error: \displaystyle P(S)=\mathbb{P}_{Y\sim P}(Y\in S), Type II Error: \displaystyle Q(S^{c})=\mathbb{P}_{Y\sim Q}(Y\notin S). The difficulty of distinguishing P from Q is fully characterized by the trade-off function T(P,Q) , which maps a Type I error level \alpha\in[0,1] to the minimal possible Type II error: T(P,Q)(\alpha):=\inf_{S\in\Sigma}\{Q(S^{c}):P(S)\leq\alpha\}. (3) Let f be a convex, non-increasing function on the unit square. Then, f -DP ( 3 ) ensures that for any adjacent datasets \mathcal{D}\sim\mathcal{D}^{\prime} , with P=\mathcal{M}(\mathcal{D}) and Q=\mathcal{M}(\mathcal{D}^{\prime}) : T(P,Q)(\alpha)\geq f(\alpha),\quad\forall\alpha\in[0,1]. This effectively limits the power of any adversary to identify the source dataset. Lastly, we say that two DP mechanisms are Blackwell equivalent if and only if their trade-off functions coincide everywhere ( 5 ) . 2.3 Conversion Rules and RDP Classes In many settings, such as private deep learning, the specific mechanism \mathcal{M} and datasets \mathcal{D},\mathcal{D}^{\prime} are unknown or effectively black-box; only the privacy accountant’s output, i.e. the RDP profile \rho(\tau) , is available. Therefore, we must bound the trade-off function for the entire class of distributions permissible under \rho . Let \mathcal{S}_{\rho} be the set of all pairs of probability distributions (P,Q) that satisfy the RDP constraints: \mathcal{S}_{\rho}:=\left\{(P,Q)\in\mathcal{P}(\mathcal{Y})^{2}\;\bigg|\;\begin{aligned} D_{\tau}(P\|Q)&\leq\rho(\tau)\\ D_{\tau}(Q\|P)&\leq\rho(\tau)\end{aligned}\;\forall\tau\geq 0.5\right\}. Definition 2.3 (Admissible Conversion Rule) . Let \Pi be the space of valid RDP profiles and \mathcal{T} the space of trade-off functions. A transformation C:\Pi\to\mathcal{T} is an admissible conversion rule if for any profile \rho\in\Pi , it lower-bounds the trade-off of every pair in \mathcal{S}_{\rho} : C(\rho)(\alpha)\leq\inf_{(P,Q)\in\mathcal{S}_{\rho}}T(P,Q)(\alpha),\quad\forall\alpha\in[0,1]. (4) This definition ensures that C(\rho) is a valid lower bound on the hypothesis testing difficulty. In the simplified case where we possess only a single point-wise guarantee (\tau^{*},\epsilon^{*})\in[0.5,\infty)\times\mathbb{R}_{\geq 0} rather than a full functional profile, we define the conversion rule analogously by considering the profile \rho where \rho(\tau^{*})=\epsilon^{*} and \rho(\tau)=\infty for all \tau\neq\tau^{*} . 2.4 Privacy Region We define the privacy region of a mechanism as the set of all attainable error pairs (\alpha,\beta) for the binary hypothesis testing problem. 11 demonstrated that a mechanism is (\varepsilon,\delta) -DP if and only if all attainable error pairs lie within the region R_{\text{DP}}(\varepsilon,\delta) : R_{\text{DP}}(\varepsilon,\delta)=\left\{(\alpha,\beta)\in[0,1]^{2}\;\bigg|\;\begin{aligned} 1-\alpha&\leq e^{\varepsilon}\beta+\delta\\ 1-\beta&\leq e^{\varepsilon}\alpha+\delta\end{aligned}\right\}. (5) The lower boundary of this region is the tightest lower-bounding piecewise linear trade-off function f_{\varepsilon,\delta}:[0,1]\to[0,1] , given by: f_{\varepsilon,\delta}(\alpha)=\max\left\{0,1-\delta-e^{\varepsilon}\alpha,e^{-\varepsilon}(1-\delta-\alpha)\right\}. The set R_{\text{DP}}(\varepsilon,\delta) represents the collection of permissible error rates. For a mechanism to satisfy (\varepsilon,\delta) -DP, no adversary can construct a test S with error rates (\alpha,\beta) falling outside this region (i.e., closer to the origin (0,0) than the boundary allows). Using the (\varepsilon,\delta) -DP definition via the hockey-stick divergence (see Definition 2.1 ), we can extend the definition of privacy regions to classes of distributions. Consider the set of all distribution pairs consistent with the privacy parameters: \mathcal{S}_{\varepsilon,\delta}:=\left\{(P,Q)\in\mathcal{P}(\mathcal{Y})^{2}\;\bigg|\;\begin{aligned} E_{e^{\varepsilon}}(P\|Q)\leq\delta\\ E_{e^{\varepsilon}}(Q\|P)\leq\delta\end{aligned}\right\}. The privacy region R_{\text{DP}}(\varepsilon,\delta) is precisely the set of all error pairs (\alpha,\beta) attainable by any binary hypothesis test trying to distinguish between any pair P and Q in \mathcal{S}_{\varepsilon,\delta} . 2.4.1 The RDP Privacy Region and 2-Cuts Privacy regions can also be constructed using divergences such as the Rényi divergence ( 2 ) . However, to do so, the concept of k -cuts is required, specifically the 2-cut, which relates high-dimensional distributions to binary hypothesis testing. Unlike the Total Variation distance, the Rényi divergence lacks a direct variational representation involving a single rejection region. To bridge this gap, we use the 2-cut reduction. Intuitively, the 2-cut reduction projects the distinguishability (measured by some divergence) of complex high-dimensional distributions onto the simplest possible space, namely binary outcomes, thereby capturing the worst-case discrimination capability of an adversary. Consider a mechanism satisfying (\tau,\rho) -RDP. For any decision rule defined by rejection region S\subseteq\mathcal{Y} , we can define a randomized binary test with Type I error \alpha=P(S) and Type II error \beta=1-Q(S) . These errors induce two Bernoulli distributions: \mathcal{B}_{P}\sim\text{Bern}(\alpha)\quad\text{and}\quad\mathcal{B}_{Q}\sim\text{Bern}(1-\beta). (6) The data processing inequality (DPI) guarantees that post-processing (in this case, mapping the mechanism’s output to a binary decision) cannot increase the divergence. The “2-cut” of the Rényi divergence, denoted \bar{D}_{\tau}^{2}(P\|Q) , can be interpreted as the worst-case divergence achievable by any such binary reduction: \bar{D}_{\tau}^{2}(P\|Q):=\sup_{S\subseteq\mathcal{Y}}D_{\tau}(\mathcal{B}_{P}\|\mathcal{B}_{Q}). By the DPI, this quantity is upper-bounded by the divergence of the original distributions, which is in turn bounded by the privacy budget \rho : D_{\tau}(\mathcal{B}_{P}\|\mathcal{B}_{Q})\leq\bar{D}_{\tau}^{2}(P\|Q)\leq D_{\tau}(P\|Q)\leq\rho. (7) Consequently, the error rates (\alpha,\beta)\in[0,1]^{2} of every possible binary test must satisfy D_{\tau}(\mathcal{B}_{P}\|\mathcal{B}_{Q})\leq\rho . We formally define the \tau -order RDP privacy region R_{D_{\tau}}(\rho) as the set of all error pairs (\alpha,\beta) attainable by any binary hypothesis test trying to distinguish between any pair P and Q that satisfy D_{\tau}(P\|Q)\leq\rho : \begin{split}&R_{D_{\tau}}(\rho)=\\ &\left\{(\alpha,\beta)\in[0,1]^{2}\;\bigg|\;\begin{aligned} D_{\tau}(\text{Bern}(\alpha)\|\text{Bern}(1-\beta))&\leq\rho\\ D_{\tau}(\text{Bern}(1-\beta)\|\text{Bern}(\alpha))&\leq\rho\end{aligned}\right\}.\end{split} (8) Depending on the value of \tau , these constraints take the following explicit forms derived from ( 7 ): Case 1 ( \tau>1 ): For \tau>1 , the region is the set of (\alpha,\beta) satisfying: \displaystyle\alpha^{\tau}(1-\beta)^{1-\tau}+(1-\alpha)^{\tau}\beta^{1-\tau} \displaystyle\leq e^{(\tau-1)\rho} (9) \displaystyle(1-\beta)^{\tau}\alpha^{1-\tau}+\beta^{\tau}(1-\alpha)^{1-\tau} \displaystyle\leq e^{(\tau-1)\rho}. Case 2 ( \tau=1 : As \tau\to 1 , the RDP constraint converges to the KL-divergence: \displaystyle\alpha\log\frac{\alpha}{1-\beta}+(1-\alpha)\log\frac{1-\alpha}{\beta} \displaystyle\leq\rho (10) \displaystyle(1-\beta)\log\frac{1-\beta}{\alpha}+\beta\log\frac{\beta}{1-\alpha} \displaystyle\leq\rho. Case 3 ( 0<\tau<1 ): For \tau<1 , the definition of Rényi divergence involves a factor of \frac{1}{\tau-1}<0 , which reverses the inequality direction: \displaystyle\alpha^{\tau}(1-\beta)^{1-\tau}+(1-\alpha)^{\tau}\beta^{1-\tau} \displaystyle\geq e^{(\tau-1)\rho} (11) \displaystyle(1-\beta)^{\tau}\alpha^{1-\tau}+\beta^{\tau}(1-\alpha)^{1-\tau} \displaystyle\geq e^{(\tau-1)\rho}. The constraints ( 9 )–( 11 ) have already been derived in 13 ; however, for the self-containedness of this manuscript, we have included their derivation in Appendix A . These inequalities define the \tau -order RDP privacy region R_{D_{\tau}}(\rho) for \tau>1 , \tau=1 , and 0<\tau<1 , respectively. Note that the \tau -order RDP privacy regions for \tau^{\prime}\in[0.5,1) , are sufficient to characterize the \tau -order RDP privacy regions for all 0<\tau<1 (see Apendix A ). 2.4.2 Properties of the RDP Privacy Region Proposition 2.4 (Convexity and Symmetry of the RDP Privacy Region) . For any \tau\in[0.5,\infty) and \rho\geq 0 , the RDP privacy region R_{D_{\tau}}(\rho) is a convex set and is symmetric about \alpha=\beta . Proof. It is an established property that the Rényi divergence D_{\tau}(P\|Q) is jointly quasi-convex in the pair of distributions (P,Q) for all \tau\in(0,\infty) ( 9 ) . Consequently, its sublevel sets are convex in the space of probability distributions . The privacy region R_{D_{\tau}}(\rho) is defined as the intersection of two sets: S_{1}=\{(\alpha,\beta):D_{\tau}(\text{Bern}(\alpha)\|\text{Bern}(1-\beta))\leq\rho\}, S_{2}=\{(\alpha,\beta):D_{\tau}(\text{Bern}(1-\beta)\|\text{Bern}(\alpha))\leq\rho\}. Consider the mapping M:[0,1]^{2}\to\mathcal{P}(\{0,1\})^{2} defined by M(\alpha,\beta)=(\text{Bern}(\alpha),\text{Bern}(1-\beta)) . Identifying the probability measures with vectors in \mathbb{R}^{2} , we have \text{Bern}(\alpha)=[1-\alpha,\alpha]^{\top} and \text{Bern}(1-\beta)=[\beta,1-\beta]^{\top} . This mapping is clearly affine with respect to the parameters \alpha and \beta . Since the mapping from parameters to distributions is affine, the convexity of the sublevel sets in distribution space is preserved when “pulled back” to the parameter space. Thus, both S_{1} and S_{2} are convex sets. Finally, the privacy region R_{D_{\tau}}(\rho)=S_{1}\cap S_{2} is the intersection of two convex sets and is therefore convex. To show symmetry, note that if (\alpha,\beta)\in R_{D_{\tau}}(\rho) , then by definition D_{\tau}(\text{Bern}(\alpha)\|\text{Bern}(1-\beta))\leq\rho and D_{\tau}(\text{Bern}(1-\beta)\|\text{Bern}(\alpha))\leq\rho . Swapping \alpha and \beta in these inequalities yields: \displaystyle D_{\tau}(\text{Bern}(\beta)\|\text{Bern}(1-\alpha))\leq\rho and \displaystyle D_{\tau}(\text{Bern}(1-\alpha)\|\text{Bern}(\beta))\leq\rho. Since D_{\tau}(\text{Bern}(x)\|\text{Bern}(y))=D_{\tau}(\text{Bern}(1-x)\|\text{Bern}(1-y)) , the set of constraints is invariant under the transformation (\alpha,\beta)\mapsto(\beta,\alpha) . ∎ Note that the boundary of the RDP privacy region given by f_{\tau,\rho}(\alpha)=\inf\{\beta:(\alpha,\beta)\in R_{D_{\tau}}(\rho)\} defines a well-behaved trade-off function for all \alpha\in[0,1] : 1. Convexity: f_{\tau,\rho}(\alpha) is a convex function. This follows directly from the convexity of the region R_{D_{\tau}}(\rho) (the lower boundary of a convex set defined on a connected domain is convex). 2. Monotonicity: f_{\tau,\rho}(\alpha) is non-increasing. Increasing the allowable Type I error \alpha relaxes the constraints on the test, allowing for a strictly lower (or equal) Type II error \beta . 3. Symmetry: The RDP region is symmetric about \alpha=\beta . Consequently, (\alpha,\beta)) is in the boundary if and only if (\beta,\alpha) is on the boundary implying that the trade-off function f_{\tau,\rho} is symmetric. 4. Feasibility of Random Guessing: f_{\tau,\rho}(\alpha)\leq 1-\alpha . The trivial “random guessing” test (rejecting H_{0} with probability \alpha independent of the data) yields \beta=1-\alpha . This corresponds to the case where the induced binary distributions are identical ( D_{\tau}=0\leq\rho ). Thus, the diagonal line \beta=1-\alpha lies strictly inside the region, and the lower boundary must lie below it. 3 Optimality of the Single-Order RDP Conversion Figure 2 : Visualization of the \tau -order RDP privacy region R_{D_{\tau}}(\rho) for \tau=1.5 and \rho=0.75 . The bold blue curve depicts the lower boundary f_{\tau,\rho}(\alpha) . The straight lines represent the piecewise linear trade-off functions of specific RR mechanisms that satisfy the (1.5,0.75) -RDP constraint. The red line corresponds to the symmetric RR mechanism, while the other lines correspond to asymmetric RR mechanisms. Note that the envelope of these valid mechanisms exactly reconstructs the RDP boundary, illustrating the optimality result in Proposition 3.1 . We now establish that the boundary of the single-order RDP privacy region is not merely a valid bound, but the optimal one. No alternative conversion rule can extract a tighter trade-off function from a single (\tau,\rho) -RDP guarantee without excluding valid mechanisms. Proposition 3.1 (Optimality of the RDP Boundary) . Let f_{\tau,\rho} be the trade-off function defined by the lower boundary of the RDP privacy region R_{D_{\tau}}(\rho) . For any admissible conversion rule C mapping a single RDP guarantee (\tau,\rho) to a lower bound on trade-off functions, f_{\tau,\rho} is the tightest possible bound. That is, for any other valid lower bound f^{\prime} , f^{\prime}(x)\leq f_{\tau,\rho}(x) for all x\in[0,1] . Proof. The proof proceeds in three steps: (1) Constructing valid Bernoulli mechanisms that achieve every point on the boundary; (2) characterizing their trade-off functions; and (3) proving optimality by contradiction. Step 1: Achievability by Bernoulli Mechanisms. We show that every point on the boundary of R_{D_{\tau}}(\rho) corresponds to the error profile of a specific mechanism with Bernoulli distributions. Consider a mechanism inducing output distributions P=\text{Bern}(q) and Q=\text{Bern}(p) with p>q . The “optimal binary test” for these distributions (the Likelihood Ratio Test at threshold 1) is simply the identity rule: reject H_{0} if the output is 1. This test yields the single error point (\alpha,\beta)=(q,1-p) . We now show that for any point (\alpha^{*},\beta^{*}) on the boundary, we can construct such distributions satisfying the RDP constraint: • Symmetric Case ( p+q=1 ): Consider the intersection of the boundary with the diagonal \alpha=\beta . We assume induced distributions P=\text{Bern}(1-p) and Q=\text{Bern}(p) for p\in[0.5,1) . Define the divergence function g(p):=D_{\tau}(\text{Bern}(1-p)\|\text{Bern}(p)) . The function g(p) is continuous and strictly increasing for p\in[0.5,1) , with g(0.5)=0 and \lim_{p\to 1}g(p)=\infty . By the Intermediate Value Theorem, there exists a unique p^{*} such that the divergence equals \rho . The identity test on these distributions achieves the error profile (\alpha^{*},\beta^{*})=(1-p^{*},1-p^{*}) , exactly satisfying the boundary equation at the diagonal. • Asymmetric Case ( p+q\neq 1 ): Consider any other point (\alpha^{*},\beta^{*}) on the boundary. By definition, the binary Rényi constraint is binding at this point. Without loss of generality, assume D_{\tau}(\text{Bern}(\alpha^{*})\|\text{Bern}(1-\beta^{*}))=\rho . We construct distributions P=\text{Bern}(\alpha^{*}) and Q=\text{Bern}(1-\beta^{*}) . Since (\alpha^{*},\beta^{*})\in R_{D_{\tau}}(\rho) , the reverse divergence (for the symmetric constraint) is also bounded by \rho . Thus, the identity test on this specific pair achieves the error profile (\alpha^{*},\beta^{*}) exactly while satisfying the (\tau,\rho) -RDP constraint. Importantly, these mechanisms, which are instances of Randomized Response (see Appendix B ), satisfy the required RDP constraints. Step 2: Validity of the Trade-off Functions. For any such Bernoulli mechanism with parameters corresponding to a point (\alpha^{*},\beta^{*}) on the boundary, the full trade-off function includes the line segments connecting (0,1) to (\alpha^{*},\beta^{*}) and (\alpha^{*},\beta^{*}) to (1,0) (attainable via randomization). Since the region R_{D_{\tau}}(\rho) is convex and contains these vertices, the entire piecewise linear trade-off curve lies within the region (i.e., it is greater than or equal to the boundary f_{\tau,\rho} pointwise). Step 3: Contradiction. Suppose, for the sake of contradiction, that there exists an admissible conversion rule yielding a trade-off function f_{\text{new}} that strictly improves upon the boundary. That is, there exists some x_{0}\in[0,1] such that f_{\text{new}}(x_{0})>f_{\tau,\rho}(x_{0}) . From Step 1, we know there exists a valid Bernoulli mechanism \mathcal{M}_{\text{Bern}} whose optimal identity test yields exactly the error pair (x_{0},f_{\tau,\rho}(x_{0})) . The true trade-off value of \mathcal{M}_{\text{Bern}} at Type I error x_{0} is exactly f_{\tau,\rho}(x_{0}) . However, the conversion rule asserts that any mechanism satisfying (\tau,\rho) -RDP must have a trade-off at least f_{\text{new}}(x_{0}) . Since f_{\tau,\rho}(x_{0})f_{\rho(\cdot)}(\alpha_{0}). (18) Let \beta^{*}=f_{\rho(\cdot)}(\alpha_{0}) . The point P^{*}=(\alpha_{0},\beta^{*}) lies exactly on the lower boundary of the intersection region \mathcal{R}_{\text{joint}}=\bigcap_{\tau\geq 0.5}R_{D_{\tau}}(\rho(\tau)) . Step 1: Constructing the Witness Mechanism. We invoke the characterization of the boundary (see Step 1 of the proof of Proposition 3.1 ). For the point P^{*}=(\alpha_{0},\beta^{*}) on the boundary, we can construct a specific binary mechanism \mathcal{M}^{*} (an instance of Randomized Response) whose optimal hypothesis test yields the error pair exactly (\alpha,\beta)=(\alpha_{0},\beta^{*}) . Step 2: Verifying Validity. We must establish that \mathcal{M}^{*} is a valid mechanism for the entire profile \rho , i.e., it satisfies D_{\tau}(\mathcal{M}^{*})\leq\rho(\tau) for all \tau\geq 0.5 . By construction, the point P^{*} lies within the global intersection \mathcal{R}_{\text{joint}} , which implies P^{*}\in R_{D_{\tau}}(\rho(\tau)) for every individual \tau . Crucially, because \mathcal{M}^{*} is a binary mechanism (mapping to \{0,1\} ), the DPI for the 2-cut reduction holds with equality. That is, the Rényi divergence of the mechanism is exactly the Rényi divergence of the induced binary distributions defined by its error profile P^{*} : D_{\tau}(\mathcal{M}^{*}(\mathcal{D})\|\mathcal{M}^{*}(\mathcal{D}^{\prime}))=D_{\tau}(\text{Bern}(\alpha_{0})\|\text{Bern}(1-\beta^{*})). Since P^{*} is contained in every region R_{D_{\tau}}(\rho(\tau)) , it satisfies the binary divergence constraints for all \tau . Therefore, D_{\tau}(\mathcal{M}^{*})\leq\rho(\tau) holds for all \tau\geq 0.5 , proving that \mathcal{M}^{*} is a valid instance of the class defined by the profile \rho . Step 3: Contradiction. Since C is defined as an admissible conversion rule, it must provide a valid lower bound for the trade-off function of any mechanism satisfying the profile \rho , including our witness \mathcal{M}^{*} . Evaluating the true trade-off of \mathcal{M}^{*} at \alpha_{0} , we have \beta_{\mathcal{M}^{*}}(\alpha_{0})=\beta^{*} . Admissibility requires: C(\rho)(\alpha_{0})\leq\beta_{\mathcal{M}^{*}}(\alpha_{0})=\beta^{*}. (19) However, our initial hypothesis assumed C(\rho)(\alpha_{0})>\beta^{*} . This is a contradiction. Thus, no such strictly tighter bound exists, and f_{\rho(\cdot)} is universally optimal. ∎ 4.1 Exactness of the RDP Profile for Randomized Response As an interesting auxiliary finding, we show that the entire trade-off curve of the Symmetric Randomized Response mechanism is exactly recovered by our conversion. Proposition 4.5 (Exact Recovery of Randomized Response) . Let \mathcal{M}_{RR} be the Symmetric Randomized Response mechanism with parameter p>0.5 , which satisfies pure \varepsilon -differential privacy with \varepsilon=\ln\frac{p}{1-p} . Let \rho(\tau) be its exact RDP profile for all \tau\geq 0.5 . The joint RDP privacy region \mathcal{R}_{\text{joint}}=\bigcap_{\tau\geq 0.5}R_{D_{\tau}}(\rho(\tau)) coincides exactly with the true attainable privacy region of \mathcal{M}_{RR} . Proof. The proof relies on the behavior of the Rényi divergence as \tau\to\infty . 1. The True Region: The mechanism \mathcal{M}_{RR} satisfies pure (\varepsilon,0) -differential privacy. Its attainable privacy region is the intersection of the two half-planes defined by the standard constraints: 1-\alpha\leq e^{\varepsilon}\beta\quad\text{and}\quad 1-\beta\leq e^{\varepsilon}\alpha. (20) These constraints form the standard piecewise linear trade-off function of \mathcal{M}_{RR} . 2. The RDP Limit: For the Symmetric Randomized Response mechanism with parameter p>0.5 , the exact RDP profile is given by ( 22 ). We examine the asymptotic behavior of the constraint boundary as \tau\to\infty . Recall the constraint inequality: \alpha^{\tau}(1-\beta)^{1-\tau}+(1-\alpha)^{\tau}\beta^{1-\tau}\leq\exp((\tau-1)\rho(\tau)). Substituting \rho(\tau) from ( 22 ), the RHS becomes precisely p^{\tau}(1-p)^{1-\tau}+(1-p)^{\tau}p^{1-\tau} . We take the logarithm of both sides and normalize by \tau to analyze the dominant terms. For the RHS, as \tau\to\infty , the term p^{\tau}(1-p)^{1-\tau} dominates because p>1-p : \displaystyle\lim_{\tau\to\infty}\frac{1}{\tau}\ln\left(\text{RHS}\right) \displaystyle=\lim_{\tau\to\infty}\frac{1}{\tau}\ln\left(p^{\tau}(1-p)^{1-\tau}\left(1+\left(\frac{1-p}{p}\right)^{2\tau-1}\right)\right) \displaystyle=\ln p-\ln(1-p)=\ln\frac{p}{1-p}=\varepsilon. Similarly for the LHS, assuming without loss of generality that \frac{1-\alpha}{\beta}>\frac{\alpha}{1-\beta} , the second term dominates: \displaystyle\lim_{\tau\to\infty}\frac{1}{\tau}\ln\left(\text{LHS}\right)=\ln(1-\alpha)-\ln\beta=\ln\frac{1-\alpha}{\beta}. Thus, the inequality converges to \ln\frac{1-\alpha}{\beta}\leq\varepsilon\implies 1-\alpha\leq e^{\varepsilon}\beta . Applying the same logic to the symmetric constraint yields 1-\beta\leq e^{\varepsilon}\alpha . These recover exactly the linear boundaries of R_{\text{DP}}(\varepsilon,0) . 3. The Intersection: By Lemma 4.2 , the attainable region lies in the intersection \mathcal{R}_{\text{joint}} . Since the limit \tau\to\infty is included in the family of constraints, \mathcal{R}_{\text{joint}} must be contained within the linear region defined by the limit (Pure DP). Conversely, since \mathcal{M}_{RR} satisfies these RDP constraints for all finite \tau , the linear region is contained within every curved RDP region. Therefore, the intersection \bigcap_{\tau\geq 0.5}R_{D_{\tau}} exactly recovers the linear region R_{D_{\infty}} . This proves that the conversion rule that utilizes the intersection of RDP privacy regions is tight for RR. ∎ 5 Conclusion In this work, we have established the fundamental limit of black-box conversions from Rényi DP to f -DP: we have demonstrated that the conversion rule based on the intersection of privacy regions constitutes the “End of the Road” for RDP-to- f -DP conversion research. Our analysis proves that the construction defined by the intersection of single-order regions, equivalent to taking the pointwise maximum across orders in f -DP space. Specifically, we have shown that no admissible black-box conversion rule that takes only the RDP profile \rho(\cdot) as input can yield a strictly tighter trade-off function. Consequently, the intersection bound captures all the information contained in the RDP profile, implying that any further improvement would require further information about mechanism parameters. Furthermore, our analysis reveals the structural simplicity of the “worst-case” mechanisms that saturate this optimal bound. We identified that the boundary is defined by simple Bernoulli processes, with parameters determined by the specific order \tau whose constraint is active at a given query point. This finding mirrors the folklore in pure Differential Privacy that Randomized Response is the least private mechanism for a fixed budget, effectively extending this intuition to the entire RDP spectrum. This structural insight confirms that the theoretical limit is not an abstract artifact, but a tangible boundary realized by concrete, elementary mechanisms. Practically, this result simplifies the implementation of optimal accounting. To compute the optimal f -DP curve from an RDP profile, one does not need solve complex variational problems; it suffices to compute the family of analytic, convex single-order curves and take their pointwise maximum. As shown by 8 , mechanisms with countable support can be converted tightly between RDP and f -DP. Our conversion is optimal for any mechanism described solely by its RDP profile (black box), but it is important to note that it is not guaranteed to be tight for all such mechanisms. For instance, the converted trade-off function for the Gaussian mechanism remains a loose lower bound compared to its analytical form (see Figure 1 ). We regard discovering mechanism classes for which the black-box conversion is near-optimal as a fruitful direction for future work. In conclusion, by definitively closing the gap between the upper bounds derived from RDP constraints and the lower bounds realizable by concrete mechanisms, we have resolved the black-box conversion problem. Appendix A Derivation of the Region Constraints We now formally derive the constraints ( 9 )–( 11 ). Recall that the decision rule induces two Bernoulli distributions: \displaystyle\mathcal{B}_{P} \displaystyle\sim\text{Bern}(\alpha)\implies\mathbb{P}_{\mathcal{B}_{P}}(1)=\alpha,\quad\mathbb{P}_{\mathcal{B}_{P}}(0)=1-\alpha, \displaystyle\mathcal{B}_{Q} \displaystyle\sim\text{Bern}(1-\beta)\implies\mathbb{P}_{\mathcal{B}_{Q}}(1)=1-\beta,\quad\mathbb{P}_{\mathcal{B}_{Q}}(0)=\beta. The RDP condition D_{\tau}(P\|Q)\leq\rho combined with the DPI implies D_{\tau}(\mathcal{B}_{P}\|\mathcal{B}_{Q})\leq\rho . By definition of the Rényi divergence: D_{\tau}(\mathcal{B}_{P}\|\mathcal{B}_{Q})=\frac{1}{\tau-1}\log\left(\sum_{z\in\{0,1\}}\mathbb{P}_{\mathcal{B}_{P}}(z)^{\tau}\mathbb{P}_{\mathcal{B}_{Q}}(z)^{1-\tau}\right)\leq\rho. (21) Case 1: \tau>1 For \tau>1 , the factor \frac{1}{\tau-1} is positive. We can exponentiate both sides without flipping the inequality: \displaystyle\sum_{z\in\{0,1\}}\mathbb{P}_{\mathcal{B}_{P}}(z)^{\tau}\mathbb{P}_{\mathcal{B}_{Q}}(z)^{1-\tau} \displaystyle\leq e^{(\tau-1)\rho} \displaystyle\iff\mathbb{P}_{\mathcal{B}_{P}}(1)^{\tau}\mathbb{P}_{\mathcal{B}_{Q}}(1)^{1-\tau}+\mathbb{P}_{\mathcal{B}_{P}}(0)^{\tau}\mathbb{P}_{\mathcal{B}_{Q}}(0)^{1-\tau} \displaystyle\leq e^{(\tau-1)\rho}. Substituting the Bernoulli parameters \alpha and \beta : \alpha^{\tau}(1-\beta)^{1-\tau}+(1-\alpha)^{\tau}\beta^{1-\tau}\leq e^{(\tau-1)\rho}. The second inequality in ( 9 ) follows from the symmetric privacy constraint D_{\tau}(Q\|P)\leq\rho , which implies D_{\tau}(\mathcal{B}_{Q}\|\mathcal{B}_{P})\leq\rho . This swaps the roles of the distributions: (1-\beta)^{\tau}\alpha^{1-\tau}+\beta^{\tau}(1-\alpha)^{1-\tau}\leq e^{(\tau-1)\rho}. Case 2: \tau=1 (KL-Divergence) By definition of the KL-divergence, ( 21 ) becomes: \displaystyle\sum_{z\in\{0,1\}}\mathbb{P}_{\mathcal{B}_{P}}(z)\log\frac{\mathbb{P}_{\mathcal{B}_{P}}(z)}{\mathbb{P}_{\mathcal{B}_{Q}}(z)} \displaystyle\leq\rho \displaystyle\iff\alpha\log\frac{\alpha}{1-\beta}+(1-\alpha)\log\frac{1-\alpha}{\beta} \displaystyle\leq\rho. Similarly, the symmetric constraint D_{\text{KL}}(\mathcal{B}_{Q}\|\mathcal{B}_{P})\leq\rho yields: \displaystyle(1-\beta)\log\frac{1-\beta}{\alpha}+\beta\log\frac{\beta}{1-\alpha} \displaystyle\leq\rho. Case 3: 0<\tau<1 For \tau<1 , the term \frac{1}{\tau-1} is negative. Isolating the log-sum reverses the inequality direction: \displaystyle\log\left(\sum_{z\in\{0,1\}}\mathbb{P}_{\mathcal{B}_{P}}(z)^{\tau}\mathbb{P}_{\mathcal{B}_{Q}}(z)^{1-\tau}\right)\geq(\tau-1)\rho. Exponentiating both sides yields: \displaystyle\sum_{z\in\{0,1\}}\mathbb{P}_{\mathcal{B}_{P}}(z)^{\tau}\mathbb{P}_{\mathcal{B}_{Q}}(z)^{1-\tau} \displaystyle\geq e^{(\tau-1)\rho} \displaystyle\iff\mathbb{P}_{\mathcal{B}_{P}}(1)^{\tau}\mathbb{P}_{\mathcal{B}_{Q}}(1)^{1-\tau}+\mathbb{P}_{\mathcal{B}_{P}}(0)^{\tau}\mathbb{P}_{\mathcal{B}_{Q}}(0)^{1-\tau} \displaystyle\geq e^{(\tau-1)\rho}. Substituting the Bernoulli parameters \alpha and \beta : \alpha^{\tau}(1-\beta)^{1-\tau}+(1-\alpha)^{\tau}\beta^{1-\tau}\geq e^{(\tau-1)\rho}. As in Case 1, the symmetric constraint D_{\tau}(\mathcal{B}_{Q}\|\mathcal{B}_{P})\leq\rho provides the second inequality required for ( 11 ). Finally, we justify the restriction to \tau\geq 0.5 . Let 0<\tau<0.5 and \tau^{\prime}=1-\tau . Consider the left-hand side of the second inequality in ( 11 ) (derived from D_{\tau}(Q\|P) ): L(\tau):=(1-\beta)^{\tau}\alpha^{1-\tau}+\beta^{\tau}(1-\alpha)^{1-\tau}. Substituting \tau=1-\tau^{\prime} , we observe: L(1-\tau^{\prime})=(1-\beta)^{1-\tau^{\prime}}\alpha^{\tau^{\prime}}+\beta^{1-\tau^{\prime}}(1-\alpha)^{\tau^{\prime}}=\alpha^{\tau^{\prime}}(1-\beta)^{1-\tau^{\prime}}+(1-\alpha)^{\tau^{\prime}}\beta^{1-\tau^{\prime}}. This is exactly the left-hand side of the first inequality for \tau^{\prime} (derived from D_{\tau^{\prime}}(P\|Q) ). Thus, the quantity being constrained is identical due to the symmetry of the divergence pair. Next, we compare the lower bounds. Since \rho>0 and 0<\tau<0.5 , we have \tau-1<-\tau=\tau^{\prime}-1 . The exponential function is strictly increasing, so: e^{(\tau-1)\rho}\hat{p} (which implies the flip dominates), the optimal hypothesis test distinguishing x=0 from x=1 rejects the null when y=0 . The resulting error profile is: \alpha=1-\hat{q}\quad(\text{Type I}),\qquad\beta=\hat{p}\quad(\text{Type II}). Remark B.3 (Mechanical Construction via Mixture Model) . The probability distribution defined in Definition B.2 is not arbitrary; it arises naturally from a two-stage “mixture” process often used to interpret Randomized Response physically (e.g., in survey methodology ( 4 ) or particle processes). Consider the following algorithm: 1. Mixing Step: With probability p , the mechanism flips the input (mapping 0\to 1 and 1\to 0 ) and halts. 2. Noise Step: With probability 1-p , the mechanism ignores the input entirely and outputs a random bit, reporting 0 with probability q and 1 with probability 1-q . This construction provides a structural explanation for the parameters: p controls the correlation with the flipped input (the “deniability” strength), while q controls the bias of the noise.