📊 statistics

Incremental Learning in Mirror Flows

This paper demonstrates that mirror flows initialized near the boundary of a convex domain, when generated by a convex quadratic loss, converge to a limiting process that minimizes the loss over a time-dependent hypothesis set, thereby establishing a general mechanism for incremental learning.

Original authors: Raphaël Berthier, Loucas Pillaud-Vivien

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Raphaël Berthier, Loucas Pillaud-Vivien

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find the perfect spot to park a car in a crowded lot (the "loss" you want to minimize). Usually, you just drive straight toward the best spot. But sometimes, the rules of the lot are weird, or the car has a strange steering mechanism that forces you to take a winding, indirect path.

This paper studies a specific type of "winding path" used in machine learning called a Mirror Flow. Think of a Mirror Flow as a car that doesn't just drive on the road; it drives on a warped, curved map where the rules of distance change depending on where you are.

Here is the core discovery of the paper, explained simply:

1. The "Slow Start" Phenomenon

The authors looked at what happens when you start this "car" right at the very edge of the allowed parking area (the boundary of the domain).

The Setup: Imagine you are trying to learn a pattern. You start with a tiny, almost invisible signal (like a whisper).
The Observation: Instead of slowly growing louder, the system behaves in a very specific way. It stays silent for a while, then suddenly "snaps" into a new, louder state, stays there for a bit, and then snaps into an even louder state.
The Analogy: Think of it like a staircase. You don't slide up a ramp; you stand on one step for a long time, then suddenly jump to the next step. You don't learn everything at once; you learn one simple thing, master it, and then move on to the next, slightly more complex thing. This is called Incremental Learning.

2. The "Limiting" View (The Magic Lens)

The paper proves that if you start your car extremely close to the edge of the map and then zoom out (rescale time and space), the complex, wiggly path of the car turns into a very simple, rigid path.

The Old Way: The car follows a smooth, curved road defined by a complex potential (like a bumpy hill).
The New Way (The Limit): When you zoom out, that bumpy hill disappears. It becomes a flat floor with a sharp, vertical wall (an "indicator function").
What this means: In this simplified, zoomed-out world, the car is forced to move along the wall. It can only move forward if it stays in a specific "hypothesis set" (a specific group of allowed solutions). As time goes on, the wall shifts, allowing the car to access new, more complex solutions it couldn't reach before.

3. Two Different Types of "Parking Lots"

The authors tested this idea in two different geometric worlds to show it works generally:

A. The Non-Negative Orthant (The "Positive Numbers" Lot)

The Setting: Imagine a parking lot where you can only park in the positive quadrant (numbers greater than zero).
The Result: This is mathematically equivalent to a specific type of neural network called a Diagonal Linear Network.
The Behavior: The system learns by turning on features one by one. It starts with zero features, then activates one, then another. It's like building a house brick by brick, where you can't add the second brick until the first one is perfectly settled.

B. The Positive Semidefinite Cone (The "Matrix" Lot)

The Setting: Imagine a parking lot for matrices (grids of numbers) that must be "positive" in a specific mathematical sense.
The Result: This relates to Matrix Factorization (breaking a big picture into smaller pieces).
The Behavior: Here, the learning is different. Instead of just jumping between static points, the system slowly increases its rank (complexity).
- Analogy: Imagine a sculpture. First, it's a flat 2D drawing (Rank 1). Then, it slowly gains depth to become a 3D object (Rank 2), then a more complex 3D object (Rank 3).
- Crucial Difference: Unlike the first example where the system jumps instantly between states, here the system slowly evolves within a fixed complexity level before suddenly gaining a new dimension of complexity. It's a mix of slow, steady growth and sudden leaps.

4. Why This Matters

The paper provides a rigorous mathematical proof that this "step-by-step" learning isn't a bug; it's a fundamental feature of how these systems work when started with very small initial values.

The Takeaway: If you initialize a machine learning model very close to zero (or the edge of its domain), it will naturally learn in stages. It will first capture the simplest, most obvious patterns. Only after it has "mastered" those will it start to learn the finer, more complex details.
The Warning: The authors caution that you cannot assume all complex neural networks behave exactly like the simple "Diagonal Linear Networks" they studied. While the idea of incremental learning holds, the way it happens (jumping vs. slowly evolving) depends heavily on the specific geometry of the problem (like the difference between the "Positive Numbers" lot and the "Matrix" lot).

In summary: The paper shows that starting a learning algorithm with a tiny nudge forces it to learn in a "staircase" fashion. It builds complexity layer by layer, revealing a hidden mechanism where simple structures are learned first, and complexity is added incrementally over time.

Technical Summary: Incremental Learning in Mirror Flows

Problem Statement
The paper investigates the phenomenon of incremental learning in continuous-time optimization dynamics, specifically within the framework of mirror flows. Incremental learning refers to the observation that neural networks trained with gradient descent often learn solutions of increasing complexity, capturing simple structures before progressively incorporating finer details. While this behavior has been rigorously analyzed for diagonal linear networks (DLNs) and matrix factorization, a unified theoretical understanding for a broader class of mirror flows, particularly those initialized near the boundary of their domain, has been lacking. The authors aim to characterize the limiting dynamics of mirror flows generated by a convex quadratic loss and a general convex lower semicontinuous mirror potential when initialized near the domain boundary.

Methodology
The authors employ tools from convex analysis and the theory of subgradient flows to analyze the asymptotic behavior of mirror flows.

Generalized Mirror Flows: The paper defines mirror flows for non-smooth mirror potentials $h \in \Gamma_0(\mathbb{R}^d)$ via a differential inclusion in the dual variable $w$ :
$\frac{dw}{dt} = -\nabla \ell(x), \quad w \in \partial h(x)$
where $\ell$ is a convex quadratic loss. This formulation generalizes the standard smooth mirror flow.
Rescaling and Limiting Regime: The core analytical technique involves a time and space rescaling of the dynamics. The authors consider a sequence of mirror flows initialized at $w^\varepsilon_0$ such that the norm $\mu_\varepsilon = \|w^\varepsilon_0\| \to +\infty$ as $\varepsilon \to 0$ . This corresponds to initializing the primal variable near the boundary of the domain of $h$ . The variables are rescaled as $w^\varepsilon(s) = \frac{1}{\mu_\varepsilon} w^\varepsilon(\mu_\varepsilon s)$ and $x^\varepsilon(s) = x^\varepsilon(\mu_\varepsilon s)$ .
Mosco Convergence: To establish the convergence of these rescaled flows, the authors utilize the theory of evolutionary convergence of subgradient flows under Mosco convergence of the associated convex functions. They show that the rescaled mirror potential $\frac{1}{\mu_\varepsilon} h$ Mosco-converges to the indicator function $\iota_C$ of the domain $C = \text{dom } h$ . Consequently, the dual potential converges to the support function $\sigma_C$ .
Subgradient Flow Reduction: The authors reduce the mirror flow dynamics to a preconditioned subgradient flow in the dual space, leveraging existence and uniqueness results for subgradient flows (specifically from Brézis [Bré73] and Attouch [Att77]) to prove the convergence of the trajectories.

Key Contributions and Results

Characterization of Limiting Dynamics: The main result (Theorem 3.8) proves that the rescaled trajectories of mirror flows, initialized near the boundary, converge uniformly to a limiting mirror flow with a non-smooth potential $\iota_C$ . In this limit, the primal variable $x(s)$ minimizes the loss $\ell$ over a time-dependent hypothesis set defined by the subdifferential of the support function of the domain:
$x(s) \in \text{Argmin}_{y \in \partial \sigma_C(w(s))} \ell(y)$
This set $\partial \sigma_C(w(s))$ evolves throughout training, providing a mechanism for incremental learning.
Unified Framework for Diagonal Linear Networks: The paper provides a rigorous, unified treatment of incremental learning in diagonal linear networks (DLNs) as a specific instance of mirror flows on the non-negative orthant $\mathbb{R}^d_{\geq 0}$ with an entropic potential. It recovers known results regarding sparse activation patterns and transition times, extending them to overparameterized settings and general Hessians.
New Geometries and Rank-Incremental Learning: The authors analyze the von Neumann entropic potential on the positive semidefinite cone $S^d_+$ . They demonstrate that when initialized near the zero matrix, the mirror flow exhibits incremental learning through successive rank increases. Crucially, they distinguish this regime from the non-negative orthant case: on the PSD cone, the primal variable evolves continuously within fixed-rank phases rather than jumping between stationary points, illustrating that incremental learning can arise from an alternation of slow and rapid dynamics.
Comparison with Matrix Factorization: The paper contrasts the mirror flow on the PSD cone with matrix factorization dynamics. While both exhibit rank-incremental learning, the paper notes that the mirror flow offers a sharper framework for analysis, suggesting caution when extrapolating results from diagonal linear networks to matrix factorization or other architectures.
Probability Simplex: The results are extended to optimization on the probability simplex $\Delta_d$ , showing that the limiting dynamics involve piecewise constant trajectories where the active set of coordinates grows over time.

Significance and Claims
The paper claims to provide a general mechanism for incremental learning in mirror flows, moving beyond specific cases like DLNs. By characterizing the limiting dynamics as a non-smooth mirror flow where the hypothesis set evolves via the subdifferential of the support function, the authors offer a rigorous explanation for the "saddle-to-saddle" transitions and plateau phenomena observed in training.

The authors emphasize that their work generalizes previous analyses of diagonal linear networks and suggests that the mirror-flow viewpoint is a more precise framework for understanding incremental learning than direct analysis of matrix factorization or other neural architectures. They explicitly state that their results should be used with caution when extrapolating from DLNs to other models, as the underlying geometries (e.g., orthant vs. PSD cone) can lead to qualitatively different dynamics (e.g., discrete jumps vs. continuous evolution within ranks).

The paper does not propose new experimental protocols or future applications beyond the theoretical analysis and the specific simulations provided to illustrate the convergence of the rescaled flows to the limiting dynamics in the three geometric settings (non-negative orthant, PSD cone, and probability simplex).

1. The "Slow Start" Phenomenon

2. The "Limiting" View (The Magic Lens)

3. Two Different Types of "Parking Lots"

4. Why This Matters

Technical Summary: Incremental Learning in Mirror Flows

More like this