---
title: Iterative Dual Orthogonal Projections
url: https://www.emergentmind.com/topics/iterative-dual-orthogonal-projections-dop
type: topic
---

# Iterative Dual Orthogonal Projections

Searching arXiv for the cited papers and closely related terminology.
Iterative Dual Orthogonal Projections (DOP) refers to the iterative method of alternating orthogonal projections between two closed subspaces of a Hilbert space. In the terminology used for alternating projections, “dual” refers to the case of two closed subspaces: given \(M,N\subset H\), orthogonal projections \(P_M,P_N\), and \(x_0\in H\), one studies the iteration \(x_{k+1}=P_NP_Mx_k\), or equivalently \(x_1=P_Mx_0\), \(x_2=P_Nx_1\), \(x_3=P_Mx_2\), and so on. Its foundational property is von Neumann’s theorem: repeated alternation converges in norm to the orthogonal projection of the initial point onto \(M\cap N\), making the dual case the mathematically cleanest and most robust component of the method of alternating projections [1809.05858].

## 1. Hilbert-space formulation and basic operators

The ambient setting is a real or complex Hilbert space \(H\) with inner product \(\langle \cdot,\cdot\rangle\) and norm \(\|x\|=\sqrt{\langle x,x\rangle}\). For a closed subspace \(Y\subset H\), the orthogonal projection \(P_Y:H\to H\) is defined by the decomposition
\[
H=Y\oplus Y^\perp,
\]
so every \(x\in H\) can be written uniquely as
\[
x=y+z,\qquad y\in Y,\ z\in Y^\perp,
\]
and
\[
P_Y(x)=y.
\]

Orthogonal projections satisfy the standard identities emphasized in the alternating-projection literature: linearity and idempotence, \(P_Y^2=P_Y\); self-adjointness, \(P_Y=P_Y^*\); norm nonexpansiveness, \(\|P_Yx\|\le \|x\|\), with equality iff \(x\in Y\); and the best-approximation property, namely that \(P_Yx\) is the unique nearest point in \(Y\) to \(x\) [1809.05858].

In the dual case one fixes two closed subspaces \(M,N\subset H\), with orthogonal projections
\[
P_M:H\to M,\qquad P_N:H\to N,
\]
their intersection \(M\cap N\), and the associated orthogonal projection \(P_{M\cap N}\). This two-subspace configuration already contains the central geometric and operator-theoretic phenomena of the general theory.

## 2. Alternating iteration and von Neumann convergence

The general alternating-projection sequence attached to a family \(M_1,\dots,M_J\subset H\) and an index sequence \((j_n)\subset\{1,\dots,J\}\) is defined by
\[
x_n=P_{j_n}x_{n-1},\qquad n\ge 1.
\]
Iterative DOP specializes this to \(J=2\). A standard dual alternating projection scheme is
\[
x_{k+1}=P_NP_Mx_k,\qquad k=0,1,2,\dots
\]
or, equivalently,
\[
x_0\in H,\quad x_1=P_Mx_0,\quad x_2=P_Nx_1,\quad x_3=P_Mx_2,\quad \dots
\]
Because \(P_M^2=P_M\) and \(P_N^2=P_N\), any sequence of \(P_M\) and \(P_N\) can be reduced to a composition where they alternate, so the two-subspace theorem applies to any order of \(P_M,P_N\) [1809.05858].

The core convergence statement is von Neumann’s theorem. For closed subspaces \(M_1,M_2\subset H\), with projections \(P_1,P_2\), and \(M=M_1\cap M_2\), one has
\[
\|(P_2P_1)^n x-P_Mx\|\to 0\qquad \text{as }n\to\infty.
\]
In dual notation,
\[
(P_NP_M)^n x \xrightarrow{\|\cdot\|} P_{M\cap N}x.
\]
Thus the limit is the unique point in \(M\cap N\) closest to the initial datum. Algorithmically, this means that if projections onto \(M\) and \(N\) are easier to compute than the projection onto \(M\cap N\), the latter can be approximated by repeated dual projections.

A technically important proof route passes through the self-adjoint operator
\[
T=P_1P_2P_1.
\]
The argument uses the spectral theorem for self-adjoint operators, the fact that the associated multiplier satisfies \(0\le m(t)\le 1\) almost everywhere, dominated convergence for \(\|T^n x-T^{n+1}x\|\), and the decomposition
\[
H=\overline{\operatorname{Ran}(I-T)}\oplus \ker(I-T^*)
=\overline{\operatorname{Ran}(I-T)}\oplus (M_1\cap M_2).
\]
From this one obtains \(T^n x\to P_{M_1\cap M_2}x\), hence the norm convergence of the alternating-projection iterates. The significance of this proof is that it works directly with orthogonal projections rather than with general contractions [1809.05858].

## 3. Product structure, canonical factorization, and invariant decomposition

A complementary operator-theoretic description studies the product
\[
T=P_1P_2.
\]
In general, \(T\) is not a projection and need not be selfadjoint or normal, but it is always a contraction. The modern structural characterization of such operators is especially useful for iterative DOP. For \(T\in B(\mathcal H)\), the following are equivalent:

\[
T \text{ is a product of two projections;}
\]
\[
T=P_{\operatorname{Ran}T}P_{\operatorname{Ran}T^*};
\]
\[
TT^*=TP_{\operatorname{Ran}T^*};
\]
\[
TT^*T=T^2.
\]

Accordingly, whenever \(T=P_1P_2\), the canonical projections are
\[
P_1=P_{\operatorname{Ran}T},\qquad P_2=P_{\operatorname{Ran}T^*},
\]
and the product is recovered as
\[
T=P_{\operatorname{Ran}T}P_{\operatorname{Ran}T^*}.
\]
This gives a purely algebraic and range-based test for whether a contraction arises from a dual projection pair. A further identity describes the kernel:
\[
\ker(T)=\bigl[\operatorname{Ran}(I-P_1)\cap \operatorname{Ran}(P_2)\bigr]\oplus \operatorname{Ran}(I-P_2).
\]
These formulas identify precisely where the product acts injectively, where it annihilates vectors, and how the geometry of the two ranges controls the operator [2411.01831].

The contraction \(T=P_1P_2\) also admits the Sz.-Nagy–Foiaş–Langer canonical decomposition
\[
\mathcal H=\mathcal H_u\oplus \mathcal H_{\text{cnu}},
\]
where \(\mathcal H_u\) reduces \(T\) and \(T|\mathcal H_u\) is unitary, while \(T|\mathcal H_{\text{cnu}}\) is completely non-unitary. In the special case of a product of two orthogonal projections, these subspaces become explicit:
\[
\mathcal H_u=\ker(I-P_1P_2)=\operatorname{Ran}(P_1)\cap \operatorname{Ran}(P_2),
\]
and
\[
\mathcal H_{\text{cnu}}=\ker(T)\vee \ker(T^*).
\]
Moreover,
\[
\text{SOT-}\lim_{m\to\infty}(P_1P_2)^m=P_{\mathcal H_u},
\]
and the restriction \(T|\mathcal H_{\text{cnu}}\) is a \(C_{00}\)-contraction, so its powers and adjoint powers converge strongly to \(0\). From the DOP perspective, this means that the iterates preserve exactly the component lying in the intersection of the two target subspaces and asymptotically suppress the completely non-unitary component [2411.01831].

## 4. Geometry, angle, and quantitative convergence

The geometry of iterative DOP is governed by the relative position of the two subspaces outside their intersection. For \(M_1,M_2\subset H\), with \(M=M_1\cap M_2\), the Friedrichs angle is the angle \(\theta\in[0,\pi/2]\) whose cosine is
\[
c(M_1,M_2)
=
\sup\left\{|\langle x,y\rangle|:x\in M_1\cap M^\perp\cap B_H,\ y\in M_2\cap M^\perp\cap B_H\right\}.
\]
A sharp estimate due to Aronszajn and Kayalar–Weinert is
\[
\|(P_{M_2}P_{M_1})^n-P_M\|=c(M_1,M_2)^{2n-1},\qquad n\ge 1.
\]
Hence, if \(0\le c(M_1,M_2)<1\), convergence is uniform and geometric; if \(c(M_1,M_2)=1\), convergence still holds by von Neumann’s theorem but can be arbitrarily slow [1809.05858].

The same phenomenon is visible in Halmos’ two-projection theorem. For orthogonal projections \(P,Q\), Halmos’ canonical form decomposes the Hilbert space into trivial blocks
\[
\mathcal M_{00},\ \mathcal M_{01},\ \mathcal M_{10},\ \mathcal M_{11}
\]
and a generic part \(\mathcal M\oplus \mathcal M'\), on which
\[
P=
\begin{pmatrix}
I & 0\\
0 & 0
\end{pmatrix},
\qquad
Q=
\begin{pmatrix}
H & \sqrt{H(I-H)}\\
\sqrt{H(I-H)} & I-H
\end{pmatrix},
\]
with \(H\) selfadjoint and \(\sigma(H)\subset[0,1]\). On this generic part, the product \(T=QP\) acts fiberwise via
\[
T(t)=
\begin{pmatrix}
t & 0\\
\sqrt{t(1-t)} & 0
\end{pmatrix},
\]
so each fiber is governed by eigenvalues \(\{0,t\}\). This provides an explicit spectral model for the contraction factors that appear in alternating projections and, in finite dimensions, identifies the scalar parameter with \(\cos^2\theta\) for the relevant principal angle [1711.11024].

A related refinement concerns the choice of projection pair representing a fixed product \(T=PQ\). Corach and Maestripieri show that the canonical factorization
\[
T=P_{R(T)}P_{N(T)^\perp}
\]
is optimal in the sense that, among all pairs \((P,Q)\) with product \(T\), it minimizes \(\|P-Q\|\). If \(R(T)\) is closed, this is the unique pair with \(\|P-Q\|<1\); if \(R(T)\) is not closed, then \(\|P-Q\|=1\) for all such pairs. This identifies the canonical pair as the geometry with maximal angular separation and therefore the one most favorable for contraction-based iteration [1011.5237].

## 5. Generalizations beyond two subspaces and failure of norm convergence

The dual case is unusually strong because norm convergence is unconditional. For a family of closed subspaces \(M_1,\dots,M_J\subset H\), periodic and quasiperiodic projection orders still yield norm convergence, but arbitrary orders do not.

In the periodic case, with
\[
T=P_JP_{J-1}\dots P_1,
\]
Halperin’s theorem gives
\[
\|T^n x-P_Mx\|\to 0,\qquad
M=\bigcap_{j=1}^J M_j.
\]
In the quasiperiodic case, each index appears infinitely often and the gaps between successive occurrences are uniformly bounded; Sakai’s theorem then implies that
\[
x_0\in H,\qquad x_n=P_{j_n}x_{n-1}
\]
converges in norm to \(P_Mx_0\). A useful analytical criterion is the estimate
\[
\|x_n-x_m\|^2\le A\sum_{k=m}^{n-1}\|x_{k+1}-x_k\|^2,
\]
which forces the sequence to be Cauchy [1809.05858].

For arbitrary projection orders, one always retains weak convergence. The Amemiya–Ando theorem states that for any Hilbert space, any finite family of closed subspaces, and any sequence of indices \((j_n)\), the sequence
\[
x_n=P_{j_n}x_{n-1}
\]
converges weakly in \(H\). This requires no periodicity, no finite-dimensionality, and no special condition on the intersection. However, weak convergence does not imply norm convergence in infinite dimensions [1809.05858].

The limitation is sharp. There exists a sequence \((j_n)\) such that for any infinite-dimensional Hilbert space \(H\) and any non-zero \(x_0\in H\), one can choose three closed subspaces \(M_1,M_2,M_3\subset H\), intersecting only at the origin, for which
\[
x_n=P_{j_n}x_{n-1}
\]
does not converge in norm. The construction builds words in three projections that move an orthonormal vector \(e_i\) arbitrarily close to \(e_{i+1}\), producing a subsequence that remains within a fixed distance of distinct orthonormal vectors and therefore cannot be norm convergent. The same counterexample shows that Sakai’s inequality does not hold uniformly for arbitrary projection sequences. This delineates the special robustness of iterative DOP: the pathology occurs for \(J\ge 3\), not for \(J=2\) [1809.05858].

## 6. Applications, specialized variants, and terminological ambiguity

In its classical Hilbert-space meaning, iterative DOP is a projection algorithm for computing
\[
P_{M\cap N}x_0
\]
by alternating between the easier projections \(P_M\) and \(P_N\). The literature explicitly identifies several application domains: feasibility problems, projection onto intersections, solving systems of linear equations, and the Schwarz alternating method for PDEs. In the PDE example, two overlapping subdomains \(\Omega_1,\Omega_2\) define subspaces in a Sobolev space, projections onto their orthogonal complements encode local solution operators, and alternating these projections converges in norm to the global weak solution [1809.05858].

A recent engineering specialization uses the phrase “iterative dual orthogonal projections” for a linear beamforming algorithm in downlink multiuser mmWave XL-MIMO systems. There, the method alternates between two orthogonal projections: one onto the orthogonal complement of an interference subspace to eliminate multiuser interference, and one onto a desired-signal subspace to refine the receive combiners. The paper proves that, with each iteration, the signal power for each user increases monotonically, the equivalent noise power after receive combining decreases monotonically, and the spectral efficiency improves accordingly and converges; simulations report rapid convergence, with per-user spectral efficiency curves flattening near their maxima within around 10 iterations, and sum spectral efficiency closely approaching the dirty paper coding benchmark [2508.16888].

The acronym is not uniform across the literature. In distributed fixed-point theory, “DOP” stands for “Distributed quasi-averaged Operator Playing,” a block-separable multi-agent algorithm over directed and unbalanced networks; that paper explicitly states that this DOP is not “Dual Orthogonal Projections” [2008.06636]. Accordingly, the phrase “Iterative Dual Orthogonal Projections” is most precise when reserved for the two-subspace alternating-projection method and for later application papers that explicitly adopt that name for algorithms built from alternating orthogonal projections.

Source: https://www.emergentmind.com/topics/iterative-dual-orthogonal-projections-dop