Response Orthogonality: Theory & Applications

• Response orthogonality is a family of criteria defining system responses via residual moment conditions, fidelity decay, and norm-derivative functionals.
• In convex regression, it enables identification through KKT-implied moment relations and hybrid control functions to address endogeneity.
• In quantum and complex normed spaces, it quantifies many-body state changes and geometric properties using fidelity measures and angular norm derivatives.

Searching arXiv for the primary paper and closely related orthogonality uses of the term. {"query":"arXiv (Dai et al., 26 Jun 2025) Orthogonality conditions for convex regression response orthogonality", "max_results": 5} {"query":"Response orthogonality orthogonality catastrophe localized disordered lattices arXiv (Cosco et al., 2018) multifractal orthogonality catastrophe arXiv (Vasseur et al., 2015)", "max_results": 10} Response orthogonality denotes a family of response-based orthogonality notions whose precise mathematical content depends on context. In convex regression, it is a set of sample moment relations between residuals and regressors, or scaled regressors, implied by the Karush–Kuhn–Tucker system of the estimator and used for identification. In localized quantum systems, it refers to the orthogonal response of many-body states to a local perturbation, typically diagnosed by ground-state fidelity, Loschmidt amplitudes, or threshold power laws. In complex normed spaces, it is defined by the vanishing of an angularly averaged norm derivative $\rho_\infty(x,y)$ and serves as an orthogonality relation derived from first-order norm response rather than from an inner product (Dai et al., 26 Jun 2025, Cosco et al., 2018, Enderami et al., 2022).

1. Convex-regression orthogonality as an identification condition

In convex regression, response orthogonality is the sample analogue of exogeneity moments familiar from linear regression. The additive model is

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$

with $f:\mathbb{R}_+^k \to \mathbb{R}$ concave or convex, and the infinite-dimensional estimator is represented as a constrained nonlinear least-squares program with Afriat-type inequalities,

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & y_i = \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h, \end{aligned}$$

optionally augmented by monotonicity $\boldsymbol{\beta}_i \ge \mathbf{0}$ and linear homogeneity $\alpha_i=0$. The multiplicative model is

$$y_i = f(\mathbf{x}_i)\exp(\varepsilon_i), \qquad \ln y_i = \ln f(\mathbf{x}_i) + \varepsilon_i,$$

estimated through

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & \ln y_i = \ln(\alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i) + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h. \end{aligned}$$

For the additive concavity-only case, the population moment restriction is

$\mathbb{E}[\varepsilon x_k]=0 \quad \forall k,$

and the sample orthogonality condition is

$\sum_{i=1}^n e_i x_{ik}=0 \quad \forall k,$

together with $y_i = f(\mathbf{x}_i) + \varepsilon_i,$0. For the multiplicative concavity-only case, orthogonality is scaled by the fitted value: $y_i = f(\mathbf{x}_i) + \varepsilon_i,$1 and

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$2

These equalities are derived from Lagrangian duality and KKT stationarity. In the additive case, with multipliers $y_i = f(\mathbf{x}_i) + \varepsilon_i,$3 for the regression equations and $y_i = f(\mathbf{x}_i) + \varepsilon_i,$4 for Afriat inequalities,

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$5

stationarity yields $y_i = f(\mathbf{x}_i) + \varepsilon_i,$6, as well as linear relations in $y_i = f(\mathbf{x}_i) + \varepsilon_i,$7 and $y_i = f(\mathbf{x}_i) + \varepsilon_i,$8. Summing and relabeling indices produces $y_i = f(\mathbf{x}_i) + \varepsilon_i,$9. In the multiplicative case, the corresponding stationarity equations introduce the scale factor $f:\mathbb{R}_+^k \to \mathbb{R}$0, which is why the orthogonality moments involve $f:\mathbb{R}_+^k \to \mathbb{R}$1 rather than $f:\mathbb{R}_+^k \to \mathbb{R}$2 alone (Dai et al., 26 Jun 2025).

2. Shape constraints, endogeneity, and hybrid control functions

Monotonicity and homogeneity affect these orthogonality relations differently. When monotonicity is imposed through $f:\mathbb{R}_+^k \to \mathbb{R}$3, the Lagrangian acquires an additional term

$f:\mathbb{R}_+^k \to \mathbb{R}$4

with $f:\mathbb{R}_+^k \to \mathbb{R}$5. This alters the slope stationarity condition and yields strict negative sample correlations. In the additive case,

$f:\mathbb{R}_+^k \to \mathbb{R}$6

hence the population implication becomes

$f:\mathbb{R}_+^k \to \mathbb{R}$7

In the multiplicative case,

$f:\mathbb{R}_+^k \to \mathbb{R}$8

The paper’s intuition is that binding monotonicity constraints push some fitted hyperplanes upward in directions of increasing inputs, so residuals become negatively correlated with inputs. By contrast, linear homogeneity, imposed as $f:\mathbb{R}_+^k \to \mathbb{R}$9, does not change the orthogonality conditions, although $$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & y_i = \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h, \end{aligned}$$0 need not vanish, analogously to OLS through the origin.

The same paper introduces a hybrid instrumental variable control function approach for endogeneity in convex regression. With an endogenous regressor $$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & y_i = \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h, \end{aligned}$$1 and instruments $$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & y_i = \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h, \end{aligned}$$2,

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & y_i = \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h, \end{aligned}$$3

and with control function linearity

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & y_i = \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h, \end{aligned}$$4

estimation proceeds in two stages. The first stage uses OLS to obtain $$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & y_i = \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h, \end{aligned}$$5. The second stage estimates

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & y_i = \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h, \end{aligned}$$6

by convex regression without monotonicity, preserving equality orthogonality: $$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & y_i = \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h, \end{aligned}$$7 Identification is then tied to

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & y_i = \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h, \end{aligned}$$8

The empirical and simulation results reported for this framework are specific. In Monte Carlo experiments based on Cobb–Douglas data-generating processes, OLS in logs performs best without endogeneity, but under endogeneity its RMSE and bias deteriorate quickly. Convex regression is more robust to endogeneity than misspecified OLS in levels. When monotonicity is absent, both IV-based convex regression and the hybrid IV control function mitigate endogeneity, but the hybrid control function consistently yields lower RMSE and bias than 2SLS, especially in larger samples. With monotonicity, equality orthogonality does not hold, yet IV and IV–control function methods still improve performance. In the Chilean manufacturing application, capital is found to be endogenous across industries, and instrumenting capital with lagged investment via the hybrid control function increases estimated capital elasticities relative to convex regression without endogeneity correction (Dai et al., 26 Jun 2025).

3. Fidelity-based response orthogonality in localized insulators

In localized lattice systems, response orthogonality is expressed through ground-state fidelity after an adiabatically switched local impurity. The many-body response is quantified by

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & y_i = \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h, \end{aligned}$$9

instantiated in the paper as

$\boldsymbol{\beta}_i \ge \mathbf{0}$0

where an impurity of strength $\boldsymbol{\beta}_i \ge \mathbf{0}$1 is placed at site $\boldsymbol{\beta}_i \ge \mathbf{0}$2. An orthogonality event is declared when this overlap falls below a threshold $\boldsymbol{\beta}_i \ge \mathbf{0}$3, with $\boldsymbol{\beta}_i \ge \mathbf{0}$4, and the event probability is

$\boldsymbol{\beta}_i \ge \mathbf{0}$5

The underlying single-particle Hamiltonian is

$\boldsymbol{\beta}_i \ge \mathbf{0}$6

Two choices of $\boldsymbol{\beta}_i \ge \mathbf{0}$7 are central. In the Anderson insulator, $\boldsymbol{\beta}_i \ge \mathbf{0}$8 is i.i.d. uniform in $\boldsymbol{\beta}_i \ge \mathbf{0}$9. In the Aubry–André model,

$\alpha_i=0$0

with localization for $\alpha_i=0$1 when $\alpha_i=0$2 is irrational. The many-body ground state is a Slater determinant, and the overlap is computed as

$\alpha_i=0$3

using the occupied single-particle eigenstates before and after the perturbation.

Several response diagnostics accompany fidelity. The radius of disturbance is

$\alpha_i=0$4

and the density witnesses are

$\alpha_i=0$5

For Aubry–André experiments, the odd-even imbalance

$\alpha_i=0$6

is proposed as a directly accessible observable.

The statistical behavior depends sharply on the localization mechanism. In both Anderson and Aubry–André insulators, $\alpha_i=0$7 rises with $\alpha_i=0$8 and saturates at $\alpha_i=0$9 at half filling when $$y_i = f(\mathbf{x}_i)\exp(\varepsilon_i), \qquad \ln y_i = \ln f(\mathbf{x}_i) + \varepsilon_i,$$0. In the Anderson case the increase is smooth and monotonic. In the Aubry–André case, $$y_i = f(\mathbf{x}_i)\exp(\varepsilon_i), \qquad \ln y_i = \ln f(\mathbf{x}_i) + \varepsilon_i,$$1 develops pronounced plateaux, with the principal plateau centered at

$$y_i = f(\mathbf{x}_i)\exp(\varepsilon_i), \qquad \ln y_i = \ln f(\mathbf{x}_i) + \varepsilon_i,$$2

and spanning approximately

$$y_i = f(\mathbf{x}_i)\exp(\varepsilon_i), \qquad \ln y_i = \ln f(\mathbf{x}_i) + \varepsilon_i,$$3

In the strong-localization regime $$y_i = f(\mathbf{x}_i)\exp(\varepsilon_i), \qquad \ln y_i = \ln f(\mathbf{x}_i) + \varepsilon_i,$$4, the site energies are approximated by

$$y_i = f(\mathbf{x}_i)\exp(\varepsilon_i), \qquad \ln y_i = \ln f(\mathbf{x}_i) + \varepsilon_i,$$5

and the main gap centers are

$$y_i = f(\mathbf{x}_i)\exp(\varepsilon_i), \qquad \ln y_i = \ln f(\mathbf{x}_i) + \varepsilon_i,$$6

The mechanism is controlled by quasi-resonant neighbors satisfying

$$y_i = f(\mathbf{x}_i)\exp(\varepsilon_i), \qquad \ln y_i = \ln f(\mathbf{x}_i) + \varepsilon_i,$$7

A two-site effective Hamiltonian,

$$y_i = f(\mathbf{x}_i)\exp(\varepsilon_i), \qquad \ln y_i = \ln f(\mathbf{x}_i) + \varepsilon_i,$$8

and its inverse participation ratio

$$y_i = f(\mathbf{x}_i)\exp(\varepsilon_i), \qquad \ln y_i = \ln f(\mathbf{x}_i) + \varepsilon_i,$$9

quantitatively account for the main plateau. Higher-order gaps, such as

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & \ln y_i = \ln(\alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i) + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h. \end{aligned}$$0

with width of order $$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & \ln y_i = \ln(\alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i) + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h. \end{aligned}$$1, produce a hierarchical devil’s-staircase structure in principle. The contrast with the Anderson model is also visible in the correlation function

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & \ln y_i = \ln(\alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i) + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h. \end{aligned}$$2

which exhibits ordered spectral-spatial patterns in Aubry–André systems and random off-diagonal structure in Anderson systems (Cosco et al., 2018).

4. Orthogonality catastrophes in random critical chains and local impurity quenches

A distinct quantum use of response orthogonality appears in orthogonality catastrophes generated by local perturbations. At one-dimensional random singlet quantum critical points, the relevant quantity is the overlap

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & \ln y_i = \ln(\alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i) + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h. \end{aligned}$$3

between the ground states with and without a local cut. Although these systems are insulating, the overlap vanishes algebraically with system size. In the random antiferromagnetic spin-$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & \ln y_i = \ln(\alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i) + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h. \end{aligned}$$4 Heisenberg and XXZ chains,

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & \ln y_i = \ln(\alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i) + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h. \end{aligned}$$5

while in the random transverse-field Ising critical chain,

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & \ln y_i = \ln(\alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i) + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h. \end{aligned}$$6

The disorder-averaged moments are multifractal,

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & \ln y_i = \ln(\alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i) + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h. \end{aligned}$$7

with

$$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & \ln y_i = \ln(\alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i) + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h. \end{aligned}$$8

satisfying $$\begin{aligned} \min_{\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\varepsilon}} \quad & \frac{1}{2}\sum_{i=1}^n \varepsilon_i^2 \ \text{s.t.}\quad & \ln y_i = \ln(\alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i) + \varepsilon_i \quad \forall i \ & \alpha_i + \boldsymbol{\beta}_i^\prime \mathbf{x}_i \le \alpha_h + \boldsymbol{\beta}_h^\prime \mathbf{x}_i \quad \forall i,h. \end{aligned}$$9 as $\mathbb{E}[\varepsilon x_k]=0 \quad \forall k,$0 and $\mathbb{E}[\varepsilon x_k]=0 \quad \forall k,$1. In the strong-disorder renormalization group description, the overlap is expressed through singlet counts: $\mathbb{E}[\varepsilon x_k]=0 \quad \forall k,$2 with $\mathbb{E}[\varepsilon x_k]=0 \quad \forall k,$3 for Heisenberg/XXZ and $\mathbb{E}[\varepsilon x_k]=0 \quad \forall k,$4 for Ising. The average singlet counts satisfy

$\mathbb{E}[\varepsilon x_k]=0 \quad \forall k,$5

The same framework predicts ultra-slow local-quench dynamics,

$\mathbb{E}[\varepsilon x_k]=0 \quad \forall k,$6

and a low-frequency work-distribution singularity

$\mathbb{E}[\varepsilon x_k]=0 \quad \forall k,$7

A central limitation is that the algebraic catastrophe occurs only at the strict cut; weakening a single bond by a factor $\mathbb{E}[\varepsilon x_k]=0 \quad \forall k,$8 leads instead to a crossover scale

$\mathbb{E}[\varepsilon x_k]=0 \quad \forall k,$9

and eventual saturation of $\sum_{i=1}^n e_i x_{ik}=0 \quad \forall k,$0 (Vasseur et al., 2015).

In trapped Fermi gases, a sudden impurity quench produces the same broad orthogonality-catastrophe phenomenology in a discrete-spectrum setting. The central dynamical quantity is the Loschmidt, or vacuum-persistence, amplitude

$\sum_{i=1}^n e_i x_{ik}=0 \quad \forall k,$1

whose linked-cluster expansion separates into energy shifts, a Gaussian thermal term, and a periodic contribution: $\sum_{i=1}^n e_i x_{ik}=0 \quad \forall k,$2 The Gaussian envelope is

$\sum_{i=1}^n e_i x_{ik}=0 \quad \forall k,$3

while the trap-induced periodic term is

$\sum_{i=1}^n e_i x_{ik}=0 \quad \forall k,$4

At zero temperature this becomes

$\sum_{i=1}^n e_i x_{ik}=0 \quad \forall k,$5

and in the continuum limit $\sum_{i=1}^n e_i x_{ik}=0 \quad \forall k,$6,

$\sum_{i=1}^n e_i x_{ik}=0 \quad \forall k,$7

The trap therefore turns the continuum power law into a periodic structure with revivals at $\sum_{i=1}^n e_i x_{ik}=0 \quad \forall k,$8 and a ladder of spectral subpeaks separated by $\sum_{i=1}^n e_i x_{ik}=0 \quad \forall k,$9 (Sindona et al., 2012).

In quantum impurity models with Fermi-liquid leads, Anderson orthogonality also governs threshold exponents and post-quench dynamics. For a local operator $y_i = f(\mathbf{x}_i) + \varepsilon_i,$00, the generalized Hopfield rule gives

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$01

where $y_i = f(\mathbf{x}_i) + \varepsilon_i,$02 is the phase-shift change in channel $y_i = f(\mathbf{x}_i) + \varepsilon_i,$03, and $y_i = f(\mathbf{x}_i) + \varepsilon_i,$04 is the locally created or annihilated charge in that channel. For a type-1 quench,

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$05

In the interacting resonant level model, this framework identifies intermediate-frequency AO windows distinct from low-frequency Fermi-liquid regimes. In population-switching problems, an added charge-sensor lead contributes an extra $y_i = f(\mathbf{x}_i) + \varepsilon_i,$06 to the pseudospin-flip scaling dimension; once the total dimension exceeds $y_i = f(\mathbf{x}_i) + \varepsilon_i,$07, the perturbation becomes irrelevant and the population switch becomes abrupt (Münder et al., 2011).

5. Norm-derivative response orthogonality in complex normed spaces

In Banach-space geometry, response orthogonality is defined through directional norm derivatives rather than through state overlaps or moment equations. For a complex normed space $y_i = f(\mathbf{x}_i) + \varepsilon_i,$08, the right and left norm derivatives are

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$09

and the angularly averaged functional is

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$10

This averaging removes the real-part bias of $y_i = f(\mathbf{x}_i) + \varepsilon_i,$11 and recovers the full complex behavior in Hilbertian settings.

The basic structural identities are direct. One has

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$12

and

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$13

A quantitative bound is

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$14

where $y_i = f(\mathbf{x}_i) + \varepsilon_i,$15 is the rotundity index of the dual space. A cruder bound,

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$16

also follows from the one-sided derivative estimates.

The associated orthogonality relation is

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$17

For every $y_i = f(\mathbf{x}_i) + \varepsilon_i,$18, there exists $y_i = f(\mathbf{x}_i) + \varepsilon_i,$19 such that

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$20

namely with $y_i = f(\mathbf{x}_i) + \varepsilon_i,$21. The relation interacts sharply with standard Banach-space orthogonalities. The equivalence

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$22

connects $y_i = f(\mathbf{x}_i) + \varepsilon_i,$23-orthogonality to Birkhoff–James orthogonality. In general, $y_i = f(\mathbf{x}_i) + \varepsilon_i,$24 and $y_i = f(\mathbf{x}_i) + \varepsilon_i,$25 are not comparable.

The principal characterization theorem states that

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$26

If the norm is induced by $y_i = f(\mathbf{x}_i) + \varepsilon_i,$27, then

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$28

Thus symmetry of the response functional is exceptional rather than generic. The same functional also yields a linear-preserver theorem: a nonzero bounded linear map $y_i = f(\mathbf{x}_i) + \varepsilon_i,$29 preserves $y_i = f(\mathbf{x}_i) + \varepsilon_i,$30-orthogonality if and only if it is a scalar multiple of an isometry, equivalently

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$31

or

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$32

The paper computes $y_i = f(\mathbf{x}_i) + \varepsilon_i,$33 explicitly in standard spaces. In $y_i = f(\mathbf{x}_i) + \varepsilon_i,$34,

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$35

which already shows non-symmetry: for $y_i = f(\mathbf{x}_i) + \varepsilon_i,$36 and $y_i = f(\mathbf{x}_i) + \varepsilon_i,$37,

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$38

For $y_i = f(\mathbf{x}_i) + \varepsilon_i,$39,

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$40

and in $y_i = f(\mathbf{x}_i) + \varepsilon_i,$41,

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$42

Only when $y_i = f(\mathbf{x}_i) + \varepsilon_i,$43 do these coincide with the Hilbert inner product (Enderami et al., 2022).

6. Comparative interpretation and recurrent points of confusion

Across these literatures, response orthogonality is not a single invariant construction but a family of response-defined orthogonality criteria. In convex regression, it is a KKT-implied residual moment relation: $y_i = f(\mathbf{x}_i) + \varepsilon_i,$44 with identification tied to corresponding population moments. In localized quantum systems, it is encoded by fidelity decay, overlap exponents, or long-time quench responses such as

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$45

In complex normed spaces, it is the vanishing of a first-order norm-response functional,

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$46

Several contrasts are especially important. First, orthogonality need not be symmetric. In Banach-space geometry, symmetry of $y_i = f(\mathbf{x}_i) + \varepsilon_i,$47 is equivalent to the norm’s being induced by an inner product; outside Hilbert spaces, symmetry generally fails. Second, orthogonality need not mean equality moments. In convex regression, equality orthogonality is broken systematically by binding monotonicity constraints and replaced by inequality moments. Third, orthogonality catastrophe is not confined to metals. Localized disordered lattices, random singlet critical points, and trapped insulating systems all exhibit orthogonal many-body responses to local perturbations, though the scaling laws differ: exponential decay of typical overlaps in localized insulators, algebraic decay at random singlet infinite-randomness fixed points, and trap-modulated MND behavior in harmonic Fermi gases (Dai et al., 26 Jun 2025, Cosco et al., 2018, Vasseur et al., 2015, Sindona et al., 2012, Enderami et al., 2022).

A related misconception is that every local perturbation generates the same orthogonality phenomenon. The cited work shows otherwise. In random singlet systems, the algebraic catastrophe occurs for a strict cut, whereas softened bond perturbations cross over to a non-universal constant overlap. In convex regression, orthogonality-based identification generally requires concavity-only specifications, because monotonicity constraints alter the moment structure. In Banach spaces, the conjectured sharp bound

$y_i = f(\mathbf{x}_i) + \varepsilon_i,$48

remains an open problem in full generality, even though it holds in several important classes of spaces. These distinctions indicate that the unifying feature is not a common formula but a common logic: orthogonality is extracted from how a system responds, at first order or asymptotically, to a local perturbation, constraint, or directional variation.