---
title: Projection-Type Gradient Flow
url: https://www.emergentmind.com/topics/projection-type-gradient-flow
type: topic
---

# Projection-Type Gradient Flow

Projection-type gradient flow denotes a family of constrained gradient evolutions in which the ambient descent or ascent direction is not used in its raw form, but is modified by a projection mechanism tied to a feasible set, a tangent space, a moving halfspace, or a geometry-induced admissible cone. In the literature this includes orthogonal projection onto Tangential Cone Condition halfspaces for nonlinear ill-posed equations, Wasserstein crowd motion with actual velocity \( \mathbf u=P_{C_\rho}\mathbf U \), intrinsic projected-gradient iterations on manifolds and hyperbolic space forms, and inexact variable-metric flows of the form \(u'(t)+u(t)-\widetilde P_{\Pi(t)}(u(t)-\alpha(t)M(t)^{-1}\nabla f(u(t)))=0\) [2011.06302] [1002.0686] [2504.11815] [2506.03671]. The term is therefore plural rather than singular: it covers exact projection, relaxed projection, obstacle-type projection in dual variables, and projection-free tangent-space substitutes; it also has nearby but distinct contrast cases where a flow preserves symmetry or concentrates near lower-dimensional structures without containing an explicit projection operator [2109.08362] [2409.14670].

## 1. Core mechanisms and stationarity notions

A first organizing distinction is between projection onto a fixed feasible set and projection onto an auxiliary set generated online. In linearly constrained optimization, the exact \(M\)-orthogonal projection onto \(\ker(B)\) is
\[
P_M = I - M^{-1}B^T S^{-1}B, \qquad S:=BM^{-1}B^T,
\]
and constrained optimality is equivalently
\[
P_M M^{-1}\nabla f(u^\star)=0.
\]
In hyperbolic space forms, the intrinsic projection is instead metric:
\[
P_C^\kappa(p):=\arg\min_{q\in C} d_\kappa(p,q),
\]
and a projected-gradient step takes the form
\[
p_{k+1}=P^\kappa_C\big(\exp^\kappa_{p_k}(-\alpha \grad f(p_k))\big).
\]
In congested transport, the projection acts on velocities rather than states:
\[
\partial_t \rho+\nabla\cdot(\rho \mathbf u)=0, \qquad \mathbf u=P_{C_\rho}\mathbf U.
\]
These are formally different objects, but each replaces an unconstrained direction by an admissible one [2506.03671] [2504.11815] [1002.0686].

Stationarity is correspondingly expressed through the vanishing of a projected residual. For smooth manifolds this is \(P_{T_x}f'(x)=0\); on the sphere,
\[
P_{T_x}=I-xx^T, \qquad P_{T_x}f'(x)=(I-xx^T)f'(x);
\]
for hyperbolic constraints, \(\bar p\in C\) is stationary when
\[
\langle \grad f(\bar p), \Proj^\kappa_{\bar p} p\rangle \ge 0 \qquad \forall p\in C;
\]
and for projected crowd motion the pressure correction \(\nabla p\) lies in the polar cone \(C_\rho^\circ\), so that \(\mathbf U=\mathbf u+\nabla p\) is the orthogonal decomposition associated with projection onto \(C_\rho\) [1906.11580] [2504.11815] [1002.0686].

This suggests a broad but technically coherent definition: projection-type gradient flow is an evolution in which admissibility is enforced through projection onto a feasible geometry, whether that geometry is a convex cone, a tangent space, a manifold-valued metric projection, or a time-dependent approximation of a solution set.

## 2. Dynamically generated projection sets in inverse problems

A particularly explicit projection construction appears in nonlinear ill-posed operator equations
\[
F(x)=y, \qquad F:D(F)\subset X\to Y,
\]
under the Tangential Cone Condition
\[
\|F(\bar x)-F(x)-F'(x)(\bar x-x)\|_Y \le \eta \,\|F(\bar x)-F(x)\|_Y.
\]
For exact data, the halfspace generated at \(x\) is
\[
H_x:=\Big\{z\in X\;\big|\; \langle z-x,\;F'(x)^*F_0(x)\rangle\le -(1-\eta)\|F_0(x)\|^2\Big\},
\]
and the TCC implies
\[
B_\rho(x_0)\cap F^{-1}(y)\subset H_x.
\]
Thus the Landweber direction \(F'(x)^*F_0(x)\) is not merely a descent direction; it is the outward normal of a separating halfspace containing all local exact solutions [2011.06302].

The relaxed projection Landweber family is
\[
x_{k+1}:=x_k-\theta_k\lambda_k\,F'(x_k)^*F_0(x_k),
\]
with
\[
\lambda_k=
\begin{cases}
0, & \text{if }F'(x_k)^*F_0(x_k)=0,\\[1ex]
\dfrac{(1-\eta)\|F_0(x_k)\|^2}{\|F'(x_k)^*F_0(x_k)\|^2}, & \text{otherwise.}
\end{cases}
\]
If \(\hat x\) is the orthogonal projection of \(x_k\) onto \(H_{x_k}\), then
\[
\hat x=x_k-\lambda_k F'(x_k)^*F_0(x_k), \qquad
x_{k+1}=x_k+\theta_k(\hat x-x_k).
\]
The method is therefore exactly a relaxed orthogonal projection onto a dynamically generated halfspace. The monotonicity estimate
\[
\|x^\star-x_k\|^2 \ge \|x^\star-x_{k+1}\|^2 + \theta_k(2-\theta_k) \left((1-\eta)\frac{\|F_0(x_k)\|^2}{\|F'(x_k)^*F_0(x_k)\|}\right)^2
\]
is a Fejér-type decrease with respect to the solution set, induced directly by projection geometry [2011.06302].

The same framework reproduces classical gradient-type methods by specific choices of \(\theta_k\). The paper states that the family encompasses Landweber, the minimal error method, and steepest descent, and that the core PLW scheme remains convergent for \(\eta<1\), whereas Landweber and related gradient analyses require \(\eta<1/2\) [2011.06302]. In this lineage, “projection-type gradient flow” is best read as a discrete geometric dynamics in which first-order information continuously regenerates outer approximations of the solution set.

## 3. Manifold, level-set, and curved-space formulations

On smooth manifolds, projection-type structure is naturally expressed through tangent spaces. For \(Q=\{x:g(x)=0\}\), the tangent projection is
\[
P_{T_x}=I-g'(x)\big(g'(x)^Tg'(x)\big)^{-1}g'(x)^T,
\]
while on the sphere
\[
P_{T_x}=I-xx^T.
\]
The gradient projection algorithms
\[
x_{k+1}=P_Q\left(x_k-\frac{1}{C_1}f'(x_k)\right)
\]
and
\[
z_k=x_k-tP_{T_{x_k}}f'(x_k)
\]
follow the two standard interpretations of constrained descent: ambient metric projection back to \(Q\), or a tangent-space step followed by local feasibility restoration. Under the manifold Lezanski–Polyak–Lojasiewicz condition
\[
\|P_{T_x}f'(x)\|^2\ge \mu(f(x)-f_0),
\]
the paper proves global linear convergence for sphere and smooth-manifold cases [1906.11580].

A distinct but related projection-induced flow appears in density clustering. For regular levels \(\mathcal L_t=\{x:f(x)=t\}\), the metric projection \(P_{t,\eta}\) from \(\mathcal L_t\) onto \(\mathcal L_{t+\eta}\) satisfies
\[
\frac{P_{t,\eta}(x)-x}{\eta} \to \frac{\nabla f(x)}{\|\nabla f(x)\|^2}, \qquad \eta\to 0.
\]
This yields the normalized gradient ascent flow
\[
\dot\zeta_x(t)=
\begin{cases}
\dfrac{\nabla f(\zeta_x(t))}{\|\nabla f(\zeta_x(t))\|^2}, & \nabla f(\zeta_x(t))\neq 0,\\[1.2ex]
0, & \text{otherwise},
\end{cases}
\]
for which
\[
\frac{d}{dt}f(\zeta_x(t))=1.
\]
Between regular levels, \(x\mapsto \zeta_x(s-t)\) is a homeomorphism, and later a diffeomorphism, from \(\mathcal L_t\) to \(\mathcal L_s\) [2109.08362]. Here projection does not constrain the motion to an external feasible set; it induces a reparameterized gradient flow whose time variable is the density level itself.

In negatively curved geometry, hyperbolic projected gradient replaces Euclidean translation by the exponential map and Euclidean projection by intrinsic \(\kappa\)-projection. The update
\[
p_{k+1}=P^\kappa_C\big(\exp^\kappa_{p_k}(-\alpha \grad f(p_k))\big)
\]
is the direct hyperbolic analogue of projected gradient descent. The paper proves that every accumulation point is stationary for both constant and backtracking step sizes, and gives an iteration-complexity bound in terms of \(d_\kappa(p_k,p_{k+1})\) or \(d_\kappa(p_k,z_k)\) [2504.11815]. The geometric infrastructure is unusually explicit: intrinsic projection is linked to Lorentz projection on the cone \(K_C\), and closed-form formulas are given for sets such as
\[
C_+=\{p\in \mathbb H^n_\kappa:\ p\ge 0\}, \qquad
P^\kappa_{C_+}(p)=\frac{p^+}{\sqrt{-\kappa\langle p^+,p^+\rangle}}.
\]

Across these settings, projection-type gradient flow means that curvature, level-set geometry, or manifold constraints are handled intrinsically rather than by an ambient Euclidean correction alone.

## 4. Continuum constrained transport and Wasserstein formulations

One of the clearest continuum realizations is the macroscopic crowd model with hard congestion. The admissible densities are
\[
K = \{ \rho \in  \mathcal{P}_{2}\, , \,\, \hbox{supp}( \rho) \subset \overline{\Omega}\, ,\, \rho = \rho_{out} + \rho_{\Omega}\, ,\, \rho_{\Omega}(x) \leq 1\,\, \hbox{ a.e.} , \,\, \hbox{supp}( \rho_{out}) \subset\Gamma_{out} \},
\]
and the admissible velocity cone is
\[
C_\rho = \{ \mathbf v \in L^2(\Omega)^2\, ,\, \int_\Omega \mathbf v \cdot \nabla q \leq 0\quad \forall q \in   H^1_\rho \}.
\]
The actual velocity is the Hilbert projection
\[
\mathbf u =P_{C_\rho}\mathbf U,
\]
with \(\mathbf U=-\nabla D\). The same model is also the Wasserstein gradient flow of
\[
\Phi(\rho)=J(\rho)+\mathbf I_K(\rho), \qquad
J(\rho)= \int_{\Omega} D(x) \rho(x) \, dx,
\]
constructed by the JKO scheme
\[
\rho_\tau^{k}  = \mathop{\textmd{argmin}\limits_{\mathcal{P}_{2}(\mathbb{R}^d)} \left\{ J(\rho) + \mathbf{I}_{K} (\rho)+ \dfrac{1}{2 \tau} W_{2}^2 (\rho, \rho_\tau^{k-1}) \right\}.
\]
Here projection and gradient flow coincide in a particularly strong sense: the projected-velocity PDE and the Wasserstein minimizing-movement formulation describe the same evolution [1002.0686].

Droplet motion on rough or inclined surfaces yields a different continuum geometry. The state manifold is
\[
M:= \{(\Gamma,u);\ \Gamma=\partial D\in C^1,\ u\in H_0^1(D),\ u\ge 0 \text{ on } D\},
\]
with tangent constraint
\[
T_\eta M:=\{(v_{cl},v);\ v-|\nabla u|\,v_{cl}=0 \text{ on }\Gamma,\ \ v+u\ge 0 \text{ in } D\}.
\]
The Riemannian metric is
\[
g_\eta(q_1,q_2) = \mathcal R\int_\Gamma v_{cl1}v_{cl2}\,ds + \zeta\int_D v_1v_2\frac{dx\,dy}{\sqrt{1+|\nabla u|^2}},
\]
and the free-energy dissipation law is
\[
\frac{d}{dt}F = -\mathcal R\int_{\Gamma(t)}v_{cl}^2\,ds -\zeta\int_{D(t)} \frac{(\partial_t u)^2}{\sqrt{1+|\nabla u|^2}}\,dx\,dy.
\]
When topological changes occur, the formulation becomes a parabolic variational inequality; in the smooth single-droplet regime it reduces to a constrained Hilbert-manifold gradient flow with volume enforced by a Lagrange multiplier [2001.04036].

These two examples occupy different analytical frameworks—Wasserstein space in one case, Hilbert manifolds in the other—but both instantiate the same structural principle: the physical velocity is an admissible projection of a driving force, and the evolution dissipates an energy under a geometry-specific metric.

## 5. Inexact, proximal, and projection-free variants

Projection need not be exact to define a meaningful projected dynamics. For linearly constrained minimization
\[
\min_{u\in \mathbb V} f(u) \qquad \text{subject to } Bu=0,
\]
the inexact projector
\[
\widetilde{P}_{\Pi}= I - M^{-1}B^T\widetilde S^{-1}B
\]
replaces the exact \(P_M\) by a structured Schur-complement approximation. The resulting continuous model is
\[
u'(t) + u(t) - \widetilde{P}_{\Pi(t)} \bigl(u(t)-\alpha(t)M(t)^{-1}\nabla f(u(t))\bigr)=0.
\]
Forward Euler discretization gives
\[
u_{k+1} = (1-\tau_k)u_k + \tau_k\,\widetilde P_{\Pi_k} \bigl(u_k-\alpha_k M_k^{-1}\nabla f(u_k)\bigr),
\]
and \(\tau_k=1\) recovers the original IPPGD iteration. The analysis uses a Lyapunov function
\[
\mathcal E(t)=\lambda \alpha D_f(u(t),u^\star) +\frac12\|(I-\widetilde P_{\Pi_\star})(u(t)-u_\phi^\star)\|_{M_\star}^2,
\]
yielding exponential convergence at the continuous level and linear convergence at the discrete level [2506.03671].

A dual obstacle formulation gives another projection-type mechanism. For the gradient flow
\[
u_t\in -\partial D(u), \qquad D(u)=\int_A |\operatorname{div}u|,
\]
the integrated scalar variable
\[
w(t)=\int_0^t v(s)\,ds
\]
satisfies the obstacle problem
\[
\min\left\{\frac12\int_A |u_0+\nabla w|^2\,dx:\ w\in H_0^1(A),\ |w|\le t\right\}.
\]
Thus the primal subgradient flow is represented through a quadratic minimization over the convex set \(\{|w|\le t\}\). In this setting the projection-type content is not an explicit projector in the primal variable \(u\), but an obstacle projection in the dual cumulative variable \(w\), with weak Hele–Shaw flow emerging from the same structure [1109.6765].

Projection-free methods form the complementary side of the subject. In nonconvex constrained variational minimization, accelerated flows are evolved in tangent spaces
\[
\mathcal F_u= \left\{ w\in U^D: \delta G(u)w =0 \right\}
\]
or, for quadratic constraints, \(\mathcal F_u^B=\{w\in U^D:B(u,w)=0\}\), rather than by exact projection back onto the nonlinear manifold. The resulting BDF\(k\) schemes enforce only the linearized constraint and quantify the defect \(\|B(u^N,u^N)-c\|_Z\). The paper proves, among other estimates, an unconditional \(\mathcal O(s^2)\) bound for BDF2 and an \(\mathcal O(s^4)\) bound for BDF4 under stated discrete-regularity assumptions [2409.14670]. This broadens projection-type gradient flow by showing how exact projection can be replaced by tangent-space restriction while retaining a controlled manifold defect.

## 6. Modern applications, neighboring methods, and common contrasts

Recent generative modeling provides a direct example of projection layered on top of a learned flow. In flow matching, the ideal conditional field is
\[
v_{t,y}^*(x) = -\frac{1}{2t(1-t)} \nabla \left\{ \mathrm{dist}^2_{t\mathcal K_y}(x,1-t) -(1-t)\|x\|^2 \right\},
\]
and classifier-free guidance is interpreted as an approximation of this gradient. The manifold
\[
\mathcal{M}_t := \{z \mid v_\theta(t,z,y)=v_\theta(t,z,\emptyset)\}
\]
leads to the projected update
\[
x_{i+1} \approx \operatorname{Proj}_{\mathcal M_{t_{i+1}}} \left( x_i+\Delta t\cdot v_\theta^{\mathrm{cfg}}(t_i,x_i,y) \right),
\]
implemented by an incremental gradient-descent fixed-point iteration and accelerated by Anderson Acceleration. This is an explicit modern instance of an inexact projected correction attached to a time-dependent flow [2601.21892].

Not every “projection method for flow” is a gradient flow in the strict variational sense. In poroelastic mixture dynamics, trial velocities are corrected by a pressure increment through
\[
\frac{v_s^{n+1}-v_s^*}{\Delta t}=-\nu_s \nabla q^n, \qquad \frac{v_f^{n+1}-v_f^*}{\Delta t}=-\nu_f \nabla q^n,
\]
and the projection equation
\[
\nabla\cdot\left[\Delta t\,\bar\nu(\phi)\nabla q^n\right] = \nabla\cdot \bar{\mathbf V}^*
\]
enforces incompressibility of the total material velocity. The paper explicitly calls this a projection method similar to Chorin projection methods and demonstrates second-order convergence in space and time, but it does not formulate the full system as a gradient flow [2206.14379].

A common misconception is that any gradient flow with smoothing, symmetry preservation, or manifold concentration is therefore projection-type. The lattice/field-theoretic heat flow
\[
\partial_\tau \rho = \partial^2 \rho
\]
is explicitly described as ordinary diffusion or Gaussian smearing and is stated not to be projection-type [1512.00294]. Likewise, the gauge-invariant ERG construction in Yang–Mills theory preserves manifest gauge invariance by construction rather than by an explicit projection operator [2012.03568]. Conversely, some unconstrained flows admit only an interpretive “self-projection”: generic gradient descent trajectories align asymptotically with the talweg tangent space \(T_{x^*}Tal=\mathbb R v_1(x^*)\), but the dynamics remain unconstrained and no operator \(P_{T_xTal}\) appears in the evolution law [2604.11213].

The broader picture is therefore heterogeneous but precise. Projection-type gradient flow is not a single algorithmic recipe. It is a geometric paradigm in which gradient dynamics are constrained, corrected, or approximated by a projection principle: onto convex halfspaces in inverse problems, onto admissible velocity cones in congested transport, onto tangent spaces or intrinsic metric projections on manifolds, onto obstacle sets in dual variables, or onto approximate feasible geometries under inexact and variable-metric computation. Nearby methods that preserve symmetry, smooth fields, or concentrate on lower-dimensional structures may be closely related, but they remain analytically distinct unless an actual projection mechanism—exact, relaxed, or surrogate—is present.

Source: https://www.emergentmind.com/topics/projection-type-gradient-flow