--- title: 'Zero-Deviation: Theory and Applications' url: https://www.emergentmind.com/topics/zero-deviation type: topic --- # Zero-Deviation: Theory and Applications In the cited literature, “zero-deviation” does not denote a single invariant notion. It instead labels several exact conditions in which a deviation quantity vanishes, or a polynomial minimizes deviation from the zero function. Díaz-González, Pijeira-Cabrera and Quintero-Roba study monic polynomials of least deviation from zero in Sobolev \(p\)-norms [2106.06290]. Ghosal et al. define zero fidelity deviation in two-qubit teleportation as the dispersion-free, or universal, regime in which fidelity is independent of the input state [1906.01394]. Schopfer and Poirier use “zero-deviation” for the zero-dissipation extrapolation of the discrepancy between nominally identical quantum Hall resistance standards [1301.5241]. DEeR defines zero aggregation deviation for federated low-rank adaptation as the exact equality between naïve LoRA averaging and the true average update [2410.12926]. ## 1. Cross-domain meaning and formal structure Across these works, the central object is always a deviation quantity with a precise operational interpretation. In approximation theory, the deviation is from the zero function under a Sobolev norm. In teleportation, it is the standard deviation of fidelity over all input states. In quantum Hall metrology, it is the relative discrepancy \(\Delta R/R\) between Hall resistors on the \(\nu=2\) plateau. In federated LoRA, it is the matrix difference between two aggregation rules. | Domain | Quantity | Defining condition | |---|---|---| | Sobolev extremal polynomials | \(\|P_n\|=\inf_{Q\in\mathbb P_n^1}\|Q\|\) | \(P_n\) is monic of exact degree \(n\) | | Two-qubit teleportation | \(\sigma=\sqrt{\langle f^2\rangle-\langle f\rangle^2}\) | \(\sigma=0\) | | Quantum Hall metrology | \(\Delta R/R=(R_H-R_K/2)/(R_K/2)\) | zero-dissipation extrapolation | | Federated LoRA | \(\mathcal O=\bar B\bar A-\frac1K\sum_k B_kA_k\) | \(\mathcal O=0\) | A common technical feature is that zero-deviation is stronger than small average error. In the teleportation setting, \(\sigma=0\) is stronger than merely having \(\bar F>2/3\). In quantum Hall metrology, a flat plateau is not sufficient if \(\overline{R_{xx}}\) remains finite. In federated LoRA, unbiased-looking factor averaging is not sufficient unless the factorization constraint eliminates \(\mathcal O\). In the Sobolev setting, the issue is different: the relevant object is not a variance-like deviation but extremality with respect to zero. ## 2. Least deviation from zero in Sobolev \(p\)-norm For real-coefficient polynomials, let \(\mathbb P\) be the polynomial space and let \(\mathbb P_n^1\) denote the monic polynomials of exact degree \(n\). An \(n\)th minimal or extremal polynomial is a monic \(P_n\in\mathbb P_n^1\) satisfying \[ \|P_n\|=\inf_{Q\in\mathbb P_n^1}\|Q\|. \] In this framework, “least deviation from zero” means minimization of the Sobolev \(p\)-norm of a monic polynomial [2106.06290]. For a vector of finite positive Borel measures \(\vec\mu=(\mu_0,\mu_1,\dots,\mu_m)\) and \(1\le p<\infty\), the continuous Sobolev \(p\)-norm is \[ \|f\|_{W^{m,p}(\vec\mu)} :=\Bigl(\sum_{k=0}^m\int_{\mathbb R}\bigl|f^{(k)}(x)\bigr|^p\,d\mu_k(x)\Bigr)^{1/p}, \] when each \(\mu_k\) is supported on a compact subset of \(\mathbb R\) and has no atoms. If some \(\mu_k\) has positive mass at finitely many points, the same formula defines a discrete Sobolev \(p\)-norm, often written \[ \|f\|^p =\int |f|^p\,du +\sum_{j=1}^N\sum_{k=0}^{m_j} A_{j,k}\,\bigl|f^{(k)}(c_j)\bigr|^p. \] For \(p=1\), \[ \|f\|_{W^{m,1}(\vec\mu)} =\sum_{k=0}^m\int\bigl|f^{(k)}(x)\bigr|\,d\mu_k(x), \] and a point-mass contribution reduces to \(\sum A_{j,k}\lvert f^{(k)}(c_j)\rvert\). For \(10\}\bigr|, \] the \(n\)th extremal polynomial has at least \(n-d^*\) simple real zeros in the convex hull of \(\operatorname{supp}\mu\). In this context, the zero-related statement concerns approximation to the zero function rather than the vanishing of a fluctuation metric. ## 3. Zero fidelity deviation and universality in two-qubit teleportation For a shared two-qubit resource \(\rho\) and an arbitrary pure input qubit \(|\psi\rangle\), the teleportation fidelity is \[ f(\psi)=\langle\psi|\rho_{\mathrm{out}}(\psi)|\psi\rangle. \] The average fidelity is \[ \bar F\equiv\langle f\rangle\equiv\int d\mu(\psi)\,f(\psi), \] where \(d\mu(\psi)\) is the normalized Haar measure on the Bloch sphere, and the fidelity deviation is \[ \sigma\equiv \sqrt{\langle f^2\rangle-\langle f\rangle^2}. \] Ghosal et al. identify \(\sigma\) as the measure of fidelity fluctuations over all input states [1906.01394]. In the Hilbert-Schmidt decomposition, \[ \rho=\frac14\Bigl[I\otimes I + R\cdot\sigma\otimes I + I\otimes S\cdot\sigma + \sum_{i,j=1}^3 T_{ij}\,\sigma_i\otimes\sigma_j\Bigr]. \] In an optimal local-unitary-preprocessed protocol one may take \(R=S=0\) and \(T\) diagonal. If the real eigenvalues of \(T\) are \(t_1,t_2,t_3\), then \[ \sigma=\frac1{\sqrt5}\sqrt{\operatorname{Tr}(TT^T)-\frac13[\operatorname{Tr}T]^2} \] or, since \(T\) is diagonal, \[ \sigma=\frac1{\sqrt5}\sqrt{t_1^2+t_2^2+t_3^2-\frac13(t_1+t_2+t_3)^2}. \] The universality, or dispersion-free, condition is \(\sigma=0\). Requiring \(\sigma=0\) gives \[ \operatorname{Tr}(T^2)=\frac13[\operatorname{Tr}T]^2, \] which for three real numbers holds if and only if \(t_1=t_2=t_3\equiv t\). Hence the necessary and sufficient condition for zero fidelity deviation is \[ T=t\,I_3. \] In that case the teleportation fidelity is independent of the input \(|\psi\rangle\). The relation to quantum advantage is explicit. For useful teleportation, the relevant case is \(\det T\le 0\), and the maximal average fidelity is \[ \bar F_{\max}=\frac14[1+|t_1|+|t_2|+|t_3|]. \] If \(t_1=t_2=t_3\equiv t<0\), then \[ \bar F_{\max}=\frac{1+3|t|}{4},\qquad \sigma=0. \] Since the classical teleportation bound is \(\bar F=2/3\), universal states beating the classical bound satisfy \[ \frac{1+3|t|}{4}> \frac23 \iff |t|>\frac13. \] These are precisely the canonical Werner-like states with \(T=-tI\), \(t\in(1/3,1]\). The examples sharpen the classification. For pure entangled states in Schmidt form, \[ |\psi\rangle=a|00\rangle+b|11\rangle,\qquad a^2+b^2=1,\quad a\ge b>0, \] one has \[ t_1=2ab,\quad t_2=-2ab,\quad t_3=1, \] \[ \bar F_{\max}=\frac{1+2ab}{2},\qquad \sigma=\frac{2ab}{\sqrt5}\sqrt{1-4a^2b^2}. \] In terms of the concurrence \(C=2ab\), \[ \bar F_{\max}=\frac{1+C}{2},\qquad \sigma=\frac{C}{\sqrt5}\sqrt{1-C^2}. \] Thus \(\sigma=0\) only if \(C=1\), namely the maximally entangled case. For Bell-diagonal states \(\rho_{BD}=\sum_{i=0}^3 p_i|\Psi_i\rangle\langle\Psi_i|\), with \(p_0\ge p_1\ge p_2\ge p_3\), one has \(\det T<0\) as soon as \(p_0>1/2\), and \[ \sigma=\frac1{\sqrt5}\sqrt{(p_1-p_2)^2+(p_2-p_3)^2+(p_1-p_3)^2}. \] Zero deviation implies \(p_1=p_2=p_3\), hence a Werner state. Ghosal et al. also exhibit two one-parameter families of \(X\)-states for which \(T=-pI\), hence \(\sigma=0\), while the marginals are not maximally mixed; in each case \(\bar F_{\max}=(1+p)/2\), so these states are universal for \(p>1/3\) and useless for \(p\le 1/3\). ## 4. Zero-deviation near the zero-dissipation state in quantum Hall metrology In the quantum Hall effect, “zero-deviation” refers to the experimental demonstration that four nominally identical GaAs/AlGaAs quantum-Hall-resistance standards, each realizing \(R_{\rm H}\approx R_{\rm K}/2\) at filling factor \(\nu=2\), agree within a few \(10^{-12}\) of the nominal quantized value [1301.5241]. Schopfer and Poirier obtain this result with a modified Wheatstone-bridge technique. The four Hall bars have width \(W=400\,\mu\mathrm m\), identical GaAs/AlGaAs two-dimensional electron gases, \(\mu\approx310{,}000\,\mathrm{cm}^2/\mathrm{Vs}\), and \(n_s\approx5.2\times10^{11}\,\mathrm{cm}^{-2}\). The “triple-connection” of Delahaye makes contact and lead resistances enter only in third order, \[ O[(R_C/R_H)^3]\ll10^{-15}, \] so wire or contact errors are negligible. An AC excitation \(I\) at low frequency (\(0.1\,\mathrm{Hz}\)–\(20\,\mathrm{Hz}\)) is fed into one diagonal of the bridge, and the bridge unbalance current \(I_{ub}(f)\) is detected in the other diagonal by a cryogenic current comparator and SQUID with ultimate resolution \(\sim400\,\mathrm{fA}/\mathrm{Hz}^{1/2}\). In the DC limit, \[ \alpha_j\equiv\frac{2R_{\rm H}^{(j)}-R_{\rm K}}{R_{\rm K}}\qquad (j=1,\ldots,4), \] and \[ [I_{ub}/I]_{C_A}=-[I_{ub}/I]_{C_B} =\frac{(\alpha_2+\alpha_4)-(\alpha_1+\alpha_3)}{2}, \qquad \Delta R/R\equiv\frac{R_H-R_K/2}{R_K/2}=2[I_{ub}/I]_{C_A}. \] By measuring \(I_{ub}(f)\) versus \(f\) and extrapolating the in-phase bridge unbalance to \(f\to0\), the method removes reactive or frequency-dependent error. At \(\nu=2\), the ideal quantized value is \[ R_{\rm H}=R_K/\nu=\frac{h}{e^2\,2}\approx 12{,}906.4035\;\Omega. \] At the central magnetic field \(B_P=10.86\,\mathrm T\), where \(\nu\approx1.98\approx2\) for all four bars and longitudinal dissipation is minimal, linear extrapolation to zero mean longitudinal resistance yields \[ \Delta R/R\bigl(\overline{R_{xx}}=0\bigr)=(-1.9\pm31.8)\times10^{-12}\quad(1\sigma). \] This demonstrates reproducibility between the four standards with a relative uncertainty of \(32\times10^{-12}\). The paper further shows that the mean macroscopic longitudinal resistance \[ \overline{R_{xx}(B)}=\tfrac12\bigl(r_{xx}^1+r_{xx}^2+r_{xx}^3+r_{xx}^4\bigr) \] follows the phenomenological resistivity rule \[ \overline{R_{xx}(B)} =\alpha\,B\,\frac{d}{dB}\Bigl[\Delta R(B)/R\Bigr]. \] From the slope of \(\overline{R_{xx}(B)}\) versus \(B\,d(\Delta R/R)/dB\), one extracts \[ \alpha=2.7\times10^{-2}, \] and this \(\alpha\) remains constant for currents \(I=40\)–\(120\,\mu\mathrm A\). Physically, \(\alpha\approx C\,\delta n_s/n_s\) with \(C\sim1\), because spatial carrier-density fluctuations \(\delta n_s\) generate microscopic \(\rho_{xy}\)-variations and macroscopic dissipation \(\overline{R_{xx}}\) is dominated by these \(\delta\rho_{xy}\) fluctuations rather than the intrinsic \(\rho_{xx}\ll\delta\rho_{xy}\). The vanishing of the plateau slope therefore requires \(\overline{R_{xx}}\to0\). | \(I\) (\(\mu\)A) | \(\overline{R_{xx}}\) (\(\mu\Omega\)) | \(\Delta R/R\) (\(10^{-12}\), combined) | |---|---:|---:| | 40 | \(6.2\pm1.6\) | \(2.9\pm40.2\) | | 80 | \(8.5\pm0.8\) | \(57.8\pm14.2\) | | 120 | \(17.1\pm0.9\) | \(104.5\pm17.2\) | Even a few-\(\mu\Omega\) longitudinal resistance produces nonzero \(\Delta R/R\) on the plateau. Plotting \(\Delta R/R\) versus \(\overline{R_{xx}}\) yields a straight line of slope \(\approx0.08\), and extrapolation to \(\overline{R_{xx}}=0\) recovers the zero-deviation intercept. Below about \(60\,\mu\mathrm A\), the longitudinal resistance is purely temperature-limited and independent of \(I\). Sweeping \(B\) through the \(\nu=2\) plateau to identify the pivot field \(B_P\), and ensuring density homogeneity \(\delta n_s/n_s\lesssim1\%\), are the stated conditions for driving all four bars simultaneously to the zero-dissipation state. ## 5. Zero aggregation deviation in privacy-preserving federated LoRA DEeR studies the combination of low-rank adaptation and federated learning, focusing on aggregation deviation and differential-privacy noise amplification [2410.12926]. With \(K\) clients, each client \(k\) holds LoRA factors \(A_k\in\mathbb R^{r\times n}\) and \(B_k\in\mathbb R^{m\times r}\), and the true model update is \[ \Delta W_k=B_kA_k. \] A naïve FedAvg on LoRA separately averages the factors, \[ \bar A=\frac1K\sum_{k=1}^K A_k,\qquad \bar B=\frac1K\sum_{k=1}^K B_k, \] and uses the global update \[ \widetilde{\Delta W}=\bar B\bar A. \] The correct average of true updates is \[ \Delta W^*=\frac1K\sum_{k=1}^K B_kA_k. \] The deviation is therefore \[ \mathcal O=\widetilde{\Delta W}-\Delta W^* =\bar B\bar A-\frac1K\sum_k B_kA_k. \] An algebraic rearrangement yields \[ \mathcal O =\frac1{K^2}\sum_{k,k'=1}^K (B_k-B_{k'})A_k =\frac1{K^2}\sum_{k,k'=1}^K B_k(A_{k'}-A_k). \] The paper states that \[ \mathcal O=0 \quad\text{if and only if}\quad \forall\,k,k',\ \text{either } B_k=B_{k'} \text{ or } A_k=A_{k'}. \] Hence the necessary and sufficient condition for zero-deviation is that, at aggregation time, one of the two factor sets must be identical across all clients. DEeR enforces this condition every round by alternating minimization. In the \(B\)-step, the server broadcasts \(A_g^{(t)}\), each client solves \[ B_k^{(t+1)}\leftarrow \arg\min_{B_k} f_k\bigl(W_0,A_g^{(t)},B_k\bigr), \] and the server aggregates \[ B_g^{(t+1)}\leftarrow \frac1K\sum_{k=1}^K B_k^{(t+1)}. \] In the \(A\)-step, the server broadcasts \(B_g^{(t+1)}\), each client solves \[ A_k^{(t+1)}\leftarrow \arg\min_{A_k} f_k\bigl(W_0,A_k,B_g^{(t+1)}\bigr), \] and the server aggregates \[ A_g^{(t+1)}\leftarrow \frac1K\sum_{k=1}^K A_k^{(t+1)}. \] During the \(B\)-step, all clients use the same \(A\); during the \(A\)-step, all clients use the same \(B\). The paper therefore concludes that every mini-aggregation satisfies \(\mathcal O=0\). The interaction with differential privacy is handled separately. If Gaussian noise is injected independently into \(A\) and \(B\), then \[ (B_k+\xi_k^B)(A_k+\xi_k^A) = B_kA_k + B_k\xi_k^A + \xi_k^B A_k + \xi_k^B\xi_k^A. \] The linear terms scale with \(\|A_k\|\) and \(\|B_k\|\), so the noise is amplified over rounds. DEeR’s Noise Regulator instead samples a single matrix noise \(\xi^W\sim N(0,\sigma^2C^2I)\) and solves \[ \min_{\xi^B}\|\xi^B A_k-\xi^W\|_F^2 \quad\Rightarrow\quad \xi^B=\xi^W A_k^\top(A_kA_k^\top)^{-1}, \] or \[ \min_{\xi^A}\|B_k\xi^A-\xi^W\|_F^2 \quad\Rightarrow\quad \xi^A=(B_k^\top B_k)^{-1}B_k^\top\xi^W. \] This yields \[ (B_k+\xi^B)A_k=B_kA_k+\xi^W, \qquad B_k(A_k+\xi^A)=B_kA_k+\xi^W. \] The entire mechanism therefore adds exactly one Gaussian noise matrix to each update, with no dependence on the growing norms of \(A_k\) or \(B_k\). ## 6. Recurrent distinctions and common misconceptions Several distinctions recur across these four literatures. In the Sobolev setting, extremality does not imply uniqueness when \(p=1\); the examples \(P_{a,2}(x)=(x+1)(x-a)\) for \(a\in[0,1]\) and \(P_{b,2}(x)=(x+1)(x-b)\) for \(b\in[0,1)\) show explicit non-uniqueness [2106.06290]. In two-qubit teleportation, usefulness and universality are not the same: \(\bar F_{\max}>2/3\) does not by itself imply \(\sigma=0\), and the pure-state formula \[ \sigma=\frac{C}{\sqrt5}\sqrt{1-C^2} \] shows that zero fidelity deviation occurs only at maximal entanglement for that family [1906.01394]. In quantum Hall metrology, the presence of a Hall plateau does not by itself establish zero-deviation. Schopfer and Poirier show that even a few-\(\mu\Omega\) mean longitudinal resistance can produce finite \(\Delta R/R\), and that the correct zero-deviation statement is the extrapolated intercept \[ \Delta R/R\bigl(\overline{R_{xx}}=0\bigr)=(-1.9\pm31.8)\times10^{-12} \] rather than a visual judgment about plateau flatness [1301.5241]. In federated LoRA, averaging factor matrices is not automatically equivalent to averaging their products; DEeR isolates the exact obstruction in \(\mathcal O\) and removes it by forcing one factor set to be identical across clients at aggregation time [2410.12926]. This suggests a common interpretive pattern: zero-deviation is typically a structural guarantee, not merely a small-error regime. In the teleportation and federated-learning cases, the zero condition is an exact algebraic constraint on the correlation matrix or on the LoRA factors. In the quantum Hall case, it is an experimentally defined zero-dissipation limit. In the Sobolev case, the relevant phrase identifies optimal approximation to the zero function and is therefore conceptually adjacent, but not identical, to a vanishing fluctuation or discrepancy metric. Source: https://www.emergentmind.com/topics/zero-deviation