---
title: 'Ortho-derivative: Theory & Applications'
url: https://www.emergentmind.com/topics/ortho-derivative
type: topic
---

# Ortho-derivative: Theory & Applications

Searching arXiv for recent and directly relevant papers on the term and its main technical usages.
“Ortho-derivative” is used in several mathematically distinct senses. In approximation theory and fractional analysis, it denotes an integral construction in which an ordinary derivative is recovered from an orthogonal-polynomial kernel in a small-scale limit, with extensions to Weyl or Riemann–Liouville fractional differentiation, multidimensional domains, and filter-theoretic transfer functions [1406.3972]. In Hilbert \(C^*\)-module theory, the same label is used for the norm-derivative \(\rho_\pm\), which generalizes the real part of an inner product and yields state-based characterizations of orthogonality [2111.14918]. In algebraic topology, “orthogonal derivative” denotes Weiss’s derivative in orthogonal calculus, formulated as a right Quillen functor and classified by spectra with \(O(n)\)-action [1406.0424].

## 1. Classical orthogonal derivative

The classical orthogonal derivative is built from a family \(\{p_n\}\) of orthogonal polynomials. In one formulation, if \(\{p_n\}\) is a family of monic orthogonal polynomials on \([-1,1]\) with weight \(w(t)\),
\[
\int_{-1}^1 p_n(t)\,p_m(t)\,w(t)\,dt = h_n\,\delta_{n,m},
\]
and leading coefficient \(k_n=p_n(t)/t^n\), then the \(n\)th-order orthogonal derivative of a sufficiently smooth \(f\) is
\[
D_n[f](x)
=\lim_{h\to0}\frac{k_n\,n!}{h_n\,h^n}
\int_{-1}^1 p_n(t)\,f\bigl(x+h\,t\bigr)\,w(t)\,dt.
\]
For Legendre polynomials \(P_n\), with \(w(t)=1\) and \(h_n=2/(2n+1)\), this specializes to the familiar integral formula with \(k_n=\frac{2^n\,(n!)^2}{(2n)!}\) [2105.13019].

A closely related general formulation uses a positive Borel measure \(d\mu\) and real polynomials \(p_n\) satisfying
\[
\int_{-\infty}^{\infty} p_n(u)\,p_m(u)\,d\mu(u)=h_n\,\delta_{n,m},
\qquad
p_n(u)=k_n\,u^n+\cdots.
\]
Then, under mild growth conditions on \(f\), the approximate orthogonal derivative
\[
D_n^{(\delta)}[f](x)
:=\frac{k_n\,n!}{h_n}\int_{-\infty}^{\infty} f(x+\delta u)\,p_n(u)\,d\mu(u)
\]
converges absolutely and reproduces the ordinary derivative as \(\delta\to0^+\) [1406.3972].

If the limit is not taken, the construction becomes a finite-scale differentiation formula. For the Legendre case,
\[
G_n[f](x;h)
=\frac{k_n\,n!}{h_n\,h^n}\int_{-1}^1 P_n(t)\,f(x+h\,t)\,dt
\]
satisfies
\[
G_n[f](x;h)=f^{(n)}(x)+C_n\,h^2\,f^{(n+2)}(x)+o(h^2),
\]
so the leading error is \(O(h^2)\) and depends on \(f^{(n+2)}(x)\). For \(n=1\),
\[
G_1[f](x;h)
=\frac32\,\frac1h\int_{-1}^1 t\,f(x+h\,t)\,dt
=f'(x)+\frac{h^2}{10}\,f^{(3)}(x)+o(h^2)
\]
[2105.13019]. This makes clear that the orthogonal derivative is simultaneously a limit definition and a practical approximation scheme.

## 2. Fractional orthogonal derivative and filter interpretation

The fractional orthogonal derivative is obtained by inserting the orthogonal approximation of the ordinary \(n\)th derivative into the Weyl or Riemann–Liouville fractional derivative. For the Weyl case, with \(\Re \alpha<n\),
\[
W^{(\alpha)}[f](x)=(-1)^n\frac{d^n}{dx^n}\Bigl\{W^{(n-\alpha)}[f]\Bigr\}(x),
\qquad
W^{(u)}[f](x)=\frac1{\Gamma(u)}\int_x^{+\infty}(y-x)^{u-1}f(y)\,dy.
\]
Replacing the ordinary derivative by its orthogonal approximation yields an approximate fractional orthogonal derivative, which can be written as a single-integral convolution
\[
D_{\mathrm{ortho}}^{\alpha;(\delta)}[f](x)
=\int_{-\infty}^{\infty}K_n^{\alpha;(\delta)}(y)\,f(x+y)\,dy,
\]
with explicitly computable kernel once the orthogonal-polynomial system is specified [1406.3972].

For Jacobi polynomials \(P_n^{(\alpha,\beta)}\) with weight \((1-u)^\alpha(1+u)^\beta\,du\) on \([-1,1]\), the resulting kernel admits an explicit hypergeometric form. In the notation of the paper, with fractional order renamed \(v\),
\[
D_{v;\alpha,\beta}^{(\delta)}[f](x)
=
\delta^{-v}\,T(v)\,\int_{-1}^1 f(x+\delta y)\,\mathscr K_n^{v;(\alpha,\beta)}(y)\,dy,
\]
where \(\mathscr K_n^{v;(\alpha,\beta)}\) is expressed through \({}_2F_1\) and elementary Jacobi-type factors. Continuity in \(v\) recovers the ordinary Jacobi orthogonal derivative for \(0\le v\le n\) [1406.3972].

Viewed as a linear time-invariant filter, the continuous-time Jacobi-based operator has transfer function
\[
H_{\alpha,\beta}^v(\omega)
=(i\omega)^v e^{-i\omega\delta}
\,{}_1F_1\!\Bigl(n+\alpha+1\,;\,2n+\alpha+\beta+2\,;\,2i\omega\delta\Bigr),
\qquad n=\lfloor v\rfloor+1.
\]
The confluent hypergeometric factor determines the deviation from the ideal fractional law \((i\omega)^v\). In the Jacobi case, \(|H(\omega)|=O(\omega^{v-n-\alpha-\beta-1})\) for large \(\omega\), hence \(|H(\omega)|\to0\) whenever \(v<n+\alpha+\beta+1\). This supplies a built-in high-frequency roll-off absent from the ideal fractional differentiator [1406.3972].

The discrete analogue uses Hahn polynomials \(Q_n(x;a,b,N)\) on \(x=0,1,\dots,N\). The corresponding transfer function is
\[
H_{\rm Hahn}^v(\omega)
=(1-e^{i\omega\delta})^{\,v-n}\,k_n\,n!\,
\sum_{x=0}^N Q_n(x;a,b,N)\,w(x)\,e^{-ix\omega\delta},
\]
which can be reduced to a finite \({}_2F_1\)-representation in \(e^{-i\omega\delta}\) [1406.3972].

A recurrent practical point in this literature is that filter behavior is diagnosed in the frequency domain rather than from the time-domain impulse response. For an ideal derivative of integer order \(n\), \(|H(\omega)|=\omega^n\), so a log–log plot is a straight line of slope \(n\); for a fractional derivative, the low-frequency slope is \(v\). The paper emphasizes that for fractional differentiating filters the informative representation is the log–log plot of \(|H(\omega)|\), which reveals pass-band slope, transition, and high-frequency roll-off [1406.3972].

## 3. Adjustable-precision extensions

A limitation of the classical finite-\(h\) orthogonal derivative is that its leading truncation error is \(O(h^2)\). Two extensions described by Diekema and Koornwinder increase the algebraic order of accuracy to \(O(h^{2m+2})\), with an integer parameter \(m\ge0\) chosen by the user [2105.13019].

In Liptaj’s construction, the Legendre kernel is replaced by
\[
k(t)=\frac{d^n}{dt^n}\,\omega(t),
\qquad
\omega(t)=K\,(1-t^2)^n\sum_{k=0}^m a_{2k}\,t^{2k},
\]
with \(a_0=1\) and the coefficients chosen so that terms up to order \(h^{2m+n+1}\) vanish in the expansion of
\(\int_{-1}^1 k(t)\,f(x+ht)\,dt\). The resulting approximation satisfies
\[
G_{n,m}[f](x;h)=f^{(n)}(x)+O(h^{2m+2}),
\]
and the first neglected term involves \(f^{(n+2m+2)}(x)\). For \(n=1\), the kernel simplifies to
\[
k_m(t)=(-1)^{m+1}\,\frac{2\,\Gamma\!\bigl(m+\tfrac52\bigr)}{\sqrt\pi\,\Gamma(m+1)}
\,t\,P_m^{(0,\tfrac32)}\!\bigl(2t^2-1\bigr),
\]
yielding a first-derivative scheme accurate to \(O(h^{2m+2})\) [2105.13019].

A second extension starts directly from the orthogonal-derivative framework and approximates the \(n\)th derivative by a Legendre sum,
\[
\bigl(-\tfrac1h\bigr)^n\int_{-1}^1
\underbrace{
(-1)^n\sum_{j=0}^m\frac{P_{n+2j}(t)\,P_{n+2j}^{(n)}(0)}{h_{n+2j}}
}_{k(t)}
\,f(x+ht)\,dt
=
f^{(n)}(x)+O(h^{2m+2}).
\]
Here each \(P_{n+2j}\) is an ordinary Legendre polynomial and \(h_k=\int_{-1}^1 P_k^2\). The paper notes explicitly that when \(n>1\) this \(k(t)\) is no longer an orthogonal polynomial [2105.13019].

In the frequency domain, the associated transfer function is
\[
H(h,\omega)
=\bigl(-\tfrac1h\bigr)^n\,i^n\,
\frac{2^n}{\sqrt\pi}
\sum_{j=0}^m
(2n+4j+1)\,
\Gamma\!\bigl(n+j+\tfrac12\bigr)\,\frac{(-1)^j}{j!}\,
j_{n+2j}(h\omega),
\]
where \(j_\nu\) is the spherical Bessel function. The reported plots of \(|H(\omega)|\) show low-pass behavior with successive zeros at higher frequencies, and increasing \(m\) pushes the passband to larger \(\omega\) [2105.13019].

These extensions clarify a common misconception. The label “orthogonal derivative” does not imply that every higher-accuracy extension remains orthogonal in the polynomial-theoretic sense. A precise statement from the literature is that the new kernel is not orthogonal for order greater than one, or, in the Legendre-sum formulation, that it is no longer an orthogonal polynomial when \(n>1\) [2105.13019].

## 4. Two-dimensional orthogonal and fractional constructions

The one-dimensional construction extends to mixed partial derivatives in two variables by integrating against a two-variable polynomial \(p_{n,k}(u,v)\). For a positive measure \(d\mu(u,v)\), the mixed partial \(\partial^n f/\partial x^{n-k}\partial y^k\) is represented by
\[
\frac{\partial^n}{\partial x^{n-k}\partial y^k}f(x,y)
=
(n-k)!\,k!\,\lim_{\delta\to0}
\frac1{\delta^n}
\frac{\int_{\mathbb R^2} f(x+\delta u,y+\delta v)\,p_{n,k}(u,v)\,d\mu(u,v)}
{\int_{\mathbb R^2} p_{n,k}(u,v)\,u^{n-k}v^k\,d\mu(u,v)}.
\]
For product measures on a square region, the construction factorizes. In particular, with Jacobi measures on \([-1,1]^2\), one obtains a product of one-variable Jacobi kernels for \(\partial^{m+n}f/\partial x^m\partial y^n\) [2012.14189].

On the standard triangle
\[
T^2=\{(u,v):u\ge0,\ v\ge0,\ u+v\le1\},
\]
the relevant polynomial system is biorthogonal rather than product-orthogonal. With normalized Jacobi-type weight
\[
W_{\alpha,\beta,\gamma}(u,v)
=\frac{u^\alpha v^\beta(1-u-v)^\gamma}{B(\alpha+1,\beta+1,\gamma+1)},
\qquad
\alpha,\beta,\gamma>-1,
\]
one uses the Rodrigues-type basis
\[
U_{k,n}^{\alpha,\beta,\gamma}(u,v)
=
[u^\alpha v^\beta(1-u-v)^\gamma]^{-1}
\frac{\partial^n}{\partial u^k\partial v^{n-k}}
\bigl[u^{k+\alpha}v^{n-k+\beta}(1-u-v)^{n+\gamma}\bigr].
\]
These functions also admit an Appell \(F_2\) representation and satisfy explicit biorthogonality relations against the monic basis \(q_{n-k,k}(u,v)\) [2012.14189].

The two-dimensional fractional orthogonal derivative combines these constructions with Weyl-type fractional integrals in two variables. For \(\Re\mu,\Re\nu>0\),
\[
W^{-\mu,-\nu}[f](x,y)
=
\frac1{\Gamma(\mu)\Gamma(\nu)}
\int_x^\infty\int_y^\infty
f(U,V)\,(U-x)^{\mu-1}(V-y)^{\nu-1}\,dV\,dU,
\]
and the partial fractional derivative of order \((\mu,\nu)\) is defined using ordinary partial derivatives exactly as in the one-variable Weyl–Riemann–Liouville theory. Replacing the ordinary mixed derivative by its orthogonal approximation produces a kernel representation \(W_\delta^{\mu,\nu,m,l}[f]\) whose inner kernel is obtained by integrating \(p_{m,l}(u,v)\) against fractional powers of \(s-u\) and \(t-v\) [2012.14189].

For the square with product Jacobi weights, the fractional kernel factorizes into products of one-variable \({}_2F_1\)-kernels. For the triangle, the kernel becomes piecewise-defined on five regions of the \((s,t)\)-plane, and each piece is expressed by an elementary prefactor times one of the Appell or Horn functions \(F_2\), \(F_3\), \(H_2\), or Olsson’s \(F_P\) [2012.14189].

The paper records several structural properties: linearity, a semigroup relation
\[
W^{\mu_1,\nu_1}W^{\mu_2,\nu_2}=W^{\mu_1+\mu_2,\nu_1+\nu_2}
\]
under suitable decay conditions, and recovery of ordinary mixed partials when the fractional orders approach integers. A plausible implication is that these multidimensional operators are naturally aligned with spectral and approximation frameworks on non-rectangular domains, particularly where explicit special-function kernels are available [2012.14189].

## 5. Ortho-derivative as norm derivative in Hilbert \(C^*\)-modules

In Banach-space and Hilbert \(C^*\)-module literature, “ortho-derivative” denotes the norm-derivative in the direction of a second vector. For a complex normed space \((X,\|\cdot\|)\) and \(x,y\in X\),
\[
\rho_+(x,y)
=
\lim_{t\to0^+}\frac{\|x+ty\|^2-\|x\|^2}{2t}
=
\|x\|\lim_{t\to0^+}\frac{\|x+ty\|-\|x\|}{t},
\]
and similarly
\[
\rho_-(x,y)
=
\lim_{t\to0^-}\frac{\|x+ty\|^2-\|x\|^2}{2t}.
\]
These one-sided limits exist by convexity and satisfy
\[
\rho_-(x,y)\le\rho_+(x,y),\qquad
|\rho_+(x,y)|\le\|x\|\,\|y\|,\qquad
\rho_+(x,x)=\|x\|^2.
\]
When the norm comes from an inner product, \(\rho_+(x,y)=\rho_-(x,y)=\Re\langle x,y\rangle\), so \(\rho_\pm\) generalize the real part of the inner product [2111.14918].

Let \(\mathscr X\) be a Hilbert \(C^*\)-module over a \(C^*\)-algebra \(\mathscr A\), with \(\mathscr A\)-valued inner product \(\langle\cdot,\cdot\rangle\), and let
\[
R_x=\{\varphi\in\mathcal S(\mathscr A):\varphi(\langle x,x\rangle)=\|x\|^2\}.
\]
The key theorem gives the exact state-space formula
\[
\rho_+(x,y)
=
\max\Bigl\{\Re\,\varphi(\langle x,y\rangle):\varphi\in R_x\Bigr\},
\]
and likewise
\[
\rho_-(x,y)
=
\min\Bigl\{\Re\,\varphi(\langle x,y\rangle):\varphi\in R_x\Bigr\}.
\]
This formula drives the subsequent orthogonality characterizations [2111.14918].

For Birkhoff–James orthogonality, \(x\perp_B y\) means \(\|x+\lambda y\|\ge\|x\|\) for all \(\lambda\in\mathbb C\). The norm-derivative criterion is
\[
x\perp_B y
\Longleftrightarrow
\rho_+(x,y)\ge0
\ \text{and}\ 
\rho_-(x,y)\le0,
\]
and in the Hilbert \(C^*\)-module setting this becomes the state characterization
\[
x\perp_B y
\iff
\exists\,\varphi\in\mathcal S(\mathscr A)
\ \text{with}\
\varphi(\langle x,x\rangle)=\|x\|^2
\ \text{and}\
\varphi(\langle x,y\rangle)=0.
\]
The same framework yields corresponding formulations for strong Birkhoff–James orthogonality \(x\perp_s y\) and \(p\)-orthogonality \(x\perp_p y\) [2111.14918].

The operator-theoretic application given in the paper concerns a generalized Daugavet-type equation in \(\mathbb B(\mathcal H)\). Taking \(x=T\) and \(y=TT^*T\),
\[
\rho_+(T,TT^*T)=\|T\|\,\|TT^*T\|=\|T\|+\|T\|^3,
\]
which recovers
\[
\|T+TT^*T\|=\|T\|+\|TT^*T\|.
\]
In finite dimensions, with \(\mathscr A=M_n(\mathbb C)\), the state description reduces to vector-state criteria such as
\[
T\perp_B S
\iff
\exists\,\xi:\ \| \xi\|=1,\ T\xi\perp S\xi,\ \|T\xi\|=\|T\|.
\]
This usage of “ortho-derivative” is therefore conceptually unrelated to orthogonal-polynomial differentiation, despite the similarity of terminology [2111.14918].

## 6. Orthogonal derivative in Weiss orthogonal calculus

A third usage arises in homotopy theory. In Weiss’s orthogonal calculus, the orthogonal derivative is a functorial construction on enriched functors \(F:J_0\to\mathrm{Top}\), where \(J_0\) is the Top-enriched category of finite-dimensional real inner-product spaces and linear isometric embeddings. The \(n\)th orthogonal derivative is defined by
\[
D_n^{(\mathrm{orth})}(F)(V)
:=
\mathrm{Nat}_{J_0}\bigl(J_0(V,-)^n,F(-)\bigr),
\]
viewed as an object of the intermediate category \(O(n)\times(J_n\mathrm{Top})\), where \(J_n(V,W)=J_0(V,W)^n\) carries the coordinate-permutation action [1406.0424].

Precomposition with the diagonal functor \(J_0\to J_n\), \(f\mapsto(f,\dots,f)\), gives \(D_n^{(\mathrm{orth})}\) the structure of a right adjoint. Its left adjoint sends an \(O(n)\)-equivariant \(J_n\)-diagram \(X\) to the coend
\[
V\mapsto X(V)/O(n)\,\odot\,J_0(V,-).
\]
With the \(n\)-homogeneous model structure on \(\mathrm{Fun}(J_0,\mathrm{Top})\) and the stable model structure on \(O(n)\times(J_n\mathrm{Top})\), this adjunction is Quillen [1406.0424].

The classification theorem states that there are Quillen equivalences
\[
n\text{-homog }\mathrm{Fun}(J_0,\mathrm{Top})
\;\;\longleftrightarrow\;\;
O(n)\times(J_n\mathrm{Top})_{\mathrm{stable}}
\;\;\longleftrightarrow\;\;
O(n)\backslash h\omega\mathrm{Sp},
\]
the category of orthogonal or symmetric spectra with \(O(n)\)-action. On homotopy categories,
\[
\mathrm{Ho}[n\text{-homog }\mathrm{Fun}(J_0,\mathrm{Top})]
\simeq
\mathrm{Ho}[O(n)\backslash h\omega\mathrm{Sp}],
\]
and for an \(n\)-homogeneous functor \(F\),
\[
D_nF(V)\simeq \Omega^\infty\bigl(\Sigma^\infty D_n^{(\mathrm{orth})}F(V)\bigr)^{hO(n)}.
\]
The paper emphasizes that this organization is directly analogous to the model-categorical treatment of Goodwillie’s derivative [1406.0424].

This topological usage shares only the word “orthogonal” with the analytic and \(C^*\)-module usages. The derivative here is neither an integral operator nor a norm directional derivative; it is a categorical construction encoding the \(n\)-homogeneous layer of a functor. The coexistence of these meanings explains why “ortho-derivative” is best treated as a context-dependent term rather than a single universally standardized object.

Source: https://www.emergentmind.com/topics/ortho-derivative