Let C C C be a compact interval of geometry parameters, and let k c ( t ) > 0 k_c(t)>0 k c ( t ) > 0 be a curvature family. Assume its score
J c ( t ) = ∂ c log k c ( t ) = j c ( ζ ( t ) ) J_c(t)=\partial_c\log k_c(t)=j_c(\zeta(t)) J c ( t ) = ∂ c log k c ( t ) = j c ( ζ ( t ))
is jointly continuous and strictly increasing in the state feature ζ \zeta ζ . Let A ( t ) ≥ 0 A(t)\ge0 A ( t ) ≥ 0 and B ( t ) > 0 B(t)>0 B ( t ) > 0 , and set
E c , p ( t ) = p k c ( t ) − A ( t ) p k c ( t ) + B ( t ) , ρ ( c , p ) = sup t ∣ E c , p ( t ) ∣ . E_{c,p}(t)=\frac{pk_c(t)-A(t)}{pk_c(t)+B(t)},\qquad
\rho(c,p)=\sup_t|E_{c,p}(t)|. E c , p ( t ) = p k c ( t ) + B ( t ) p k c ( t ) − A ( t ) , ρ ( c , p ) = t sup ∣ E c , p ( t ) ∣.
Fix a feasible candidate ( c ^ , p ^ ) ∈ C × ( 0 , ∞ ) (\widehat c,\widehat p)\in C\times(0,\infty) ( c , p ) ∈ C × ( 0 , ∞ ) . Suppose three states have strictly ordered feature values, and for σ ∈ { − 1 , 1 } \sigma\in\{-1,1\} σ ∈ { − 1 , 1 } write s i = σ ( − 1 ) i − 1 s_i=\sigma(-1)^{i-1} s i = σ ( − 1 ) i − 1 . Assume
s i E c ^ , p ^ ( t i ) ≥ L > 0 ( i = 1 , 2 , 3 ) , ρ ( c ^ , p ^ ) ≤ U < 1 , L ≤ U . s_iE_{\widehat c,\widehat p}(t_i)\ge L>0\quad(i=1,2,3),
\qquad \rho(\widehat c,\widehat p)\le U<1,\qquad L\le U. s i E c , p ( t i ) ≥ L > 0 ( i = 1 , 2 , 3 ) , ρ ( c , p ) ≤ U < 1 , L ≤ U .
Write A i = A ( t i ) A_i=A(t_i) A i = A ( t i ) , B i = B ( t i ) B_i=B(t_i) B i = B ( t i ) , and require A i > U B i A_i>UB_i A i > U B i whenever s i = − 1 s_i=-1 s i = − 1 . For each trial level r r r , define the error bands
L t ( r ) = A ( t ) − r B ( t ) 1 + r , U t ( r ) = A ( t ) + r B ( t ) 1 − r . L_t(r)=\frac{A(t)-rB(t)}{1+r},\qquad
U_t(r)=\frac{A(t)+rB(t)}{1-r}. L t ( r ) = 1 + r A ( t ) − r B ( t ) , U t ( r ) = 1 − r A ( t ) + r B ( t ) .
Set
β i = { log U t i ( U ) − log U t i ( L ) , s i = 1 , log L t i ( L ) − log L t i ( U ) , s i = − 1 , \beta_i=
\begin{cases}
\log U_{t_i}(U)-\log U_{t_i}(L),&s_i=1,\\
\log L_{t_i}(L)-\log L_{t_i}(U),&s_i=-1,
\end{cases} β i = { log U t i ( U ) − log U t i ( L ) , log L t i ( L ) − log L t i ( U ) , s i = 1 , s i = − 1 ,
δ i j = min c ∈ C [ J c ( t j ) − J c ( t i ) ] > 0 , M J = max c ∈ C , i ∣ J c ( t i ) ∣ \delta_{ij}=\min_{c\in C}[J_c(t_j)-J_c(t_i)]>0,\qquad
M_J=\max_{c\in C,\ i}|J_c(t_i)| δ ij = c ∈ C min [ J c ( t j ) − J c ( t i )] > 0 , M J = c ∈ C , i max ∣ J c ( t i ) ∣
and
( d − , d + ) = { ( ( β 1 + β 2 ) / δ 12 , ( β 2 + β 3 ) / δ 23 ) , σ = 1 , ( ( β 2 + β 3 ) / δ 23 , ( β 1 + β 2 ) / δ 12 ) , σ = − 1. (d_-,d_+)=
\begin{cases}
((\beta_1+\beta_2)/\delta_{12},
(\beta_2+\beta_3)/\delta_{23}),&\sigma=1,\\
((\beta_2+\beta_3)/\delta_{23},
(\beta_1+\beta_2)/\delta_{12}),&\sigma=-1.
\end{cases} ( d − , d + ) = { (( β 1 + β 2 ) / δ 12 , ( β 2 + β 3 ) / δ 23 ) , (( β 2 + β 3 ) / δ 23 , ( β 1 + β 2 ) / δ 12 ) , σ = 1 , σ = − 1.
Every pair ( c , p ) ∈ C × ( 0 , ∞ ) (c,p)\in C\times(0,\infty) ( c , p ) ∈ C × ( 0 , ∞ ) with ρ ( c , p ) ≤ U \rho(c,p)\le U ρ ( c , p ) ≤ U satisfies
c ^ − d − ≤ c ≤ c ^ + d + , ∣ log ( p / p ^ ) ∣ ≤ max i β i + M J max { d − , d + } . \boxed{
\widehat c-d_-\le c\le\widehat c+d_+,\qquad
|\log(p/\widehat p)|
\le\max_i\beta_i+M_J\max\{d_-,d_+\}.} c − d − ≤ c ≤ c + d + , ∣ log ( p / p ) ∣ ≤ i max β i + M J max { d − , d + } .
The minimum over C × ( 0 , ∞ ) C\times(0,\infty) C × ( 0 , ∞ ) exists and belongs to [ L , U ] [L,U] [ L , U ] . If C C C contains the full-family optimum from the geometry classification theorem, that minimizing pair is unique.
This result converts approximate alternating contacts and a certified global error envelope into an enclosure of every improving parameter pair. The upper bound U U U must hold on the entire state domain.
Source: Exact Residual Bounds and Optimal Geometry Selection for Proximal Gradient Methods, revised reading edition, 13 November 2026, Theorem 7.1, page 18; the ordered-family and feasibility assumptions at the start of Section 7.
Related results. Exact profile performance (Theorem 4.1) · Geometry classification (Theorem 5.1) · Error-band equivalence
Further examples and certificates. entropy certified parameter enclosure · polynomial certified parameter enclosure · robust contact localization · rational scale enclosure · localization linear shrinkage