Continuous Optimization Algorithms

Suppose we have a function \(l: \mathbb{R}^n \rightarrow \mathbb{R}\) that we want to minimize. A popular algorithm for accomplishing this is gradient descent, which is an iterative algorithm in which we pick a step size \(\alpha\) and a starting point \(x_0 \in \mathbb{R}^n\) and repeatedly iterate \(x_{t+\alpha} = x_{t} - \alpha * \nabla l(x_t)\). If we take the \(lim_{\alpha \rightarrow 0}\) of this iteration step, we get the differential equation \(\frac{dx}{dt} = -\nabla l(x)\), which we will refer to as the continuous limit of gradient descent. In my last post, I demonstrated that we can derive gradient descent as the Euler’s method discretization of this differential equation.

Many other iterative optimization algorithms have continuous limits. For example, the continuous limit of Newton’s method, in which we iterate \(x_{t+\alpha} = x_{t} - \alpha (\nabla^2 l (x_t))^{-1}\nabla l(x_t)\), is \(\frac{dx}{dt} = -(\nabla^2 l (x))^{-1}\nabla l(x)\).

Some iterative optimization algorithms use placeholder variables to store historical update values or parameter values. For example, in the momentum algorithm we introduce a placeholder variable \(y\) and iterate:

\[x_{t + \alpha} = x_t + \alpha y_{t} \qquad y_{t + \alpha} = y_t - \alpha y_t - \alpha \nabla l(x_t)))\]

The continuous limit of momentum is:

\[dx/dt = y \qquad dy/dt = - y - \nabla l(x)\]

Similarly, in the Adagrad algorithm we introduce a placeholder variable \(y\) and iterate:

\[x_{t + \alpha} = x_t + \frac{\alpha \nabla l(x_t)}{\sqrt{y_t}} \qquad y_{t + \alpha} = y_t + \alpha \nabla l(x_t)^2\]

where the division in the first equality is performed elementwise. The continuous limit of Adagrad is:

\[dx/dt = -\nabla l(x_t) / \sqrt{y_t} \qquad dy/dt = \nabla l(x_t)^2\]

More generally, if we assume that the dimensionality of the placeholder vector is a multiple of the parameter vector dimensionality we can represent the continuous limit of an iterative optimization algorithm for the objective function \(l: \mathbb{R}^{n} \rightarrow \mathbb{R}\) with a map \(d: \mathbb{R}^{n} \times \mathbb{R}^{kn} \rightarrow \mathbb{R}^{kn} \times \mathbb{R}^{kn}\), which we will refer to as a continuous optimizer. Intuitively, the function \(d(x,y) = (dx / dt, dy / dt)\) describes the instantaneous state of the optimization process at a given parameter vector \(x\) and optional placeholder vector \(y\). For example, the Newton’s method continuous optimizer is \(d(x) = (\nabla^2 l (x))^{-1}\nabla l(x)\) and the momentum continuous optimizer is \(d((x,y)) = (y, - y - \nabla l(x))\).

A continuous optimization algorithm is then a dependently typed function \(u: [n] \rightarrow (\mathbb{R}^{n} \rightarrow \mathbb{R}) \rightarrow (\mathbb{R}^{n} \times \mathbb{R}^{kn} \rightarrow \mathbb{R}^n \times \mathbb{R}^{kn})\) that maps smooth objective functions \(l: \mathbb{R}^{n} \rightarrow \mathbb{R}\) to continuous optimizers \(d_l: \mathbb{R}^{n} \times \mathbb{R}^{kn} \rightarrow \mathbb{R}^{n} \times \mathbb{R}^{kn}\). For example, the gradient descent continuous optimization algorithm maps \(l: \mathbb{R}^{n} \rightarrow \mathbb{R}\) to \(-\nabla l: \mathbb{R}^{n} \rightarrow \mathbb{R}^{n}\). Note that given a continuous optimization algorithm \(u\) and an objective \(l\), we can recover the corresponding discrete optimization algorithm with Euler’s method as \(x_{t + \alpha} = x_0 + \alpha u(l)(x_t)\)

Transformation Invariant Continuous Optimization Algorithms

In some situations, we may be able to improve the efficiency of an optimization algorithm by transforming our data first with an invertible function \(f: \mathbb{R}^m \rightarrow \mathbb{R}^n\). That is, rather than optimize the function \(l: \mathbb{R}^{n} \rightarrow \mathbb{R}\), we may be able to reduce the time to convergence by optimizing \(l \circ f: \mathbb{R}^{m} \rightarrow \mathbb{R}\). However, for many optimization algorithms there are classes of transformations to which they are invariant: applying any such transformation to the data before running the optimization algorithm cannot change the results. Formally, consider the invertible transformation \(f: \mathbb{R}^{m} \rightarrow \mathbb{R}^n\) and write \(f_{k+1}\) for the map \((f \otimes f \otimes f ...): \mathbb{R}^m \times \mathbb{R}^{km} \rightarrow \mathbb{R}^n \times \mathbb{R}^{kn}\). We will say that a continuous optimization algorithm \(u\) is invariant to \(f\) if \(u(l \circ f_{k+1}) = f_{k+1}^{-1} \circ u(l) \circ f_{k+1}\). We provide some concrete examples below:

Newton’s Method

Newton’s method is invariant to all linear transformations: \(\begin{aligned} d_{l \circ f}(x) = \\ -(\nabla^{2} (l \circ f)(x))^{-1} \nabla (l \circ f)(x) =\\ -A^{-1}(\nabla^{2}l(Ax))^{-1} A^{-T}A^{T}\nabla l(Ax) = \\ -A^{-1}(\nabla^{2}l(Ax))^{-1}\nabla l(Ax) = \\ -f^{-1}((\nabla^{2}l(f(x)))^{-1} \nabla l(f(x))) = f^{-1}d_l(f(x)) \end{aligned}\)

Gradient Descent

Gradient descent is invariant to orthogonal linear transformations, but not to linear transformations in general. Consider any function of the form \(f(x) = Ax\) where \(A\) is an orthogonal matrix. Then since \(A^{T} = A^{-1}\) we have that:

\[\begin{aligned} d_{l \circ f}(x) = -\nabla (l \circ f)(x) = -A^{T} (\nabla l (Ax)) = -A^{-1} (\nabla l (Ax)) = f^{-1}d_l(f(x)) \end{aligned}\]

Momentum

Momentum is also invariant to orthogonal linear transformations, but not to linear transformations in general:

\[\begin{aligned} d_{l \circ f}(x, y)_x = y = A^{T}Ay = f^{-1}(d_{l}(f(x), f(y)))_x \\ d_{l \circ f}(x, y)_y = y - \nabla (l \circ f)(x)) = - A^{-1}Ay - A^{-1} \nabla l(Ax)) = f^{-1}(d_{l}(f(x), f(y)))_y \end{aligned}\]

Adagrad

Adagrad is not invariant to linear transformations due to the fact that it tracks a nonlinear function of past gradients (sum of squares). However, Adagrad is invariant to permutation maps. If \(f(x) = Px\) is a permutation map then we have:

\[\begin{aligned} d_{l \circ f}(x, y)_x = -\nabla (l\circ f)(x) / \sqrt{y} = -P^{-1}\left(\nabla l(Px) / \sqrt{Py}\right) = f^{-1}(d_{l}(f(x), f(y)))_x \\ d_{l \circ f}(x, y)_y = (\nabla (l\circ f)(x))^2 = (P^{T}\nabla l(Px))^2 = P^{-1}(\nabla l(Px))^2 = f^{-1}(d_{l}(f(x), f(y)))_y \end{aligned}\]

Invariance and Euler’s Method

In the case that there is no placeholder vector, \(f(x) = Ax\) is linear and \(u\) is invariant to \(f\), then the trajectory of the Euler’s method discretization of \(u\) cannot be made to converge faster or better by first applying \(f\) to the data. That is, consider starting at some point \(x_0 \in \mathbb{R}^n\) and repeatedly taking Euler steps \(x_{n+1} = x_n + \alpha u(l)(x_n)\). Now suppose instead that we start at the point \(y_0 = A^{-1}x_0\) and take Euler steps \(y_{n+1} = y_n + \alpha u(l \circ f)(y_n)\). By induction we can see that:

\[y_{n+1} = y_n + \alpha u(l \circ f)(y_n) = y_n + \alpha A^{-1}(u(l)(Ay_n)) = A^{-1} x_n + \alpha A^{-1}(u(l)(x_n)) = A^{-1} x_{n+1}\]

Transformation Invariance is Functoriality

If we characterize continuous optimization algorithms as maps between two categories, we can demonstrate that their invariance properties are equivalent to their functoriality.

In the category \(Fun\), objects are smooth objective functions \(l: \mathbb{R}^{n} \rightarrow \mathbb{R}\) and the morphisms between \(l: \mathbb{R}^{n} \rightarrow \mathbb{R}\) and \(l': \mathbb{R}^{m} \rightarrow \mathbb{R}\) are functions \(f: \mathbb{R}^{m} \rightarrow \mathbb{R}^{n}\) such that \(l' = l \circ f\). In the category \(Opt\), objects are continuous optimizers and the morphisms from \(d: \mathbb{R}^{n} \times \mathbb{R}^{kn} \rightarrow \mathbb{R}^{n} \times \mathbb{R}^{kn}\) to \(d': \mathbb{R}^{m} \times \mathbb{R}^{km} \rightarrow \mathbb{R}^{m} \times \mathbb{R}^{km}\) are functions \(f: \mathbb{R}^{m} \rightarrow \mathbb{R}^{n}\) such that \(d' = f_{k+1}^{-1} \circ d \circ f_{k+1}\).

Given a subcategory of \(Fun\), a continuous optimization algorithm \(u\) defines a functor from that subcategory to \(Opt\) if \(u\) is invariant to any morphism \(f\) in that subcategory. We will call such a functor a continuous optimization functor. That is, we can characterize the invariance of an optimization algorithm in terms of the subcategories of \(Fun\) over which it is functorial. For example:

  • \(Fun_I\) is the subcategory of \(Fun\) in which morphisms are limited to identities. Any continuous optimization algorithm is a functor from \(Fun_I\).
  • \(Fun_P\) is the subcategory of \(Fun\) in which morphisms are limited to permutations. The Adagrad continuous optimization algorithm is a functor from \(Fun_P\).
  • \(Fun_O\) is the subcategory of \(Fun\) in which morphisms are limited to orthogonal linear maps. Both the gradient descent and momentum continuous optimization algorithms are functors from \(Fun_O\).
  • \(Fun_L\) is the subcategory of \(Fun\) in which morphisms are limited to invertible linear maps. The Newton’s method continuous optimization algorithm is a functor from \(Fun_L\).

References