Stochastic gradients is one of the most important tools in optimization and machine learning (especially for Deep Learning – see for example ConvNet). One of it’s advantage is that it behavior is well understood in general case, by application of methods of statistical mechanics.
In general form stochastic gradient descent could be written as
where is a random variable with expectation zero.
To apply methods of statistical mechanics we can rewrite it in continuous form, as stochastic gradient flow
and random variable F(t) we assume to be white noise for simplicity.
In that moment most of textbooks and papers refer to “methods of statistical mechanics” to show that
stochastic gradient flow has invariant probability distribution, which is called Gibbs distribution
and from here derive some interesting things like temperature and free energy.
The question is – how Gibbs distribution derived from stochastic gradient flow?
First we have to understand what stochastic gradient flow really means.
It’s not a partial differential equation (PDE), because it include random variable, which is not a function . In fact it’s a stochastic differential equation (SDE) . SDE use some complex mathematical machinery and relate to partial differential equations, probability/statistics, measure theory and ergodic theory. However they used a lot in finance, so there is quite a number of textbooks on SDE for non-mathematicians. For the short, minimalistic and accessible book on stochastic differential equation I can’t recommend highly enough introduction to SDE by L.C. Evans
SDE in question is called Ito diffusion. Solution of that equation is a stochastic process – collection of random variables parametrized by time. Sample path of stochastic process in question as a function of time is nowhere differentiable – it’s difficult to talk about it in term of derivatives, so it is defined through it’s integral form.
First I’ll notice that integral of white noise is actually Brownian motion, or Wiener process.
Assume that we have stochastic differential equation written in informal manner
with X -stochastic process and F(t) – white noise
It’s integral form is
where W(t) is a Wiener process
This equation is usually written in the form
This is only a notation for integral equation, d here is not a differential.
Returning to (1)
is an integral along sample path, it’s meaning is obvious, or it can be defined as limit of Riemann sums with respect to time.
The most notable thing here is
– integral with respect to Wiener process (3)
It’s a stochastic integral, and it’s defined in the courses of stochastic differential equation as the limit of Riemann sums of random variables, in the manner similar to definition of ordinary integral.
Curiously, stochastic integral is not quite well defined. Depending on the form of the sum it produce different results, like Ito integral:
Different Riemann sums produce different integral – Stratonovich integral:
Ito integral used more often in statistics because it use – it don’t “look forward”, and Stratonovich more used in theoretical physics.
Returning to Ito integral – Ito integral is stochastic process itself, and it has expectation zero for each t.
From definition of Ito integral follow one of the most important tools of stochastic calculus – Ito Lemma (or Ito formula)
Ito lemma states that for solution of SDE (2)
X, b, W – vectors, g – matrix
were W is Wiener process (actually some more general process) and b and g are good enough
where is the gradient.
From Ito lemma follow Ito product rule for scalar processes: applying Ito formula to process combined from two processes X and Y to function u(V) = XY
Using Ito formula and Ito product rule it is possible to get Feynman–Kac formula (derivation could be found in the wikipedia, it use only Ito formula, Ito product rule and the fact that expectation of Ito integral (3) is zero):
for partial differential equation (PDE)
with terminal condition
solution can be written as conditional expectation:
Feynman–Kac formula establish connection between PDE and stochastic process.
From Feynman–Kac formula taking and we get Kolmogorov backward equation :
with terminal condition (4) have solution as conditional expectation
From Kolmogorov backward equation we can obtain Kolmogorov forward equation, which describe evolution of probability density for random process X (2)
In SDE courses it’s established that (2) is a Markov process and has transitional probability P and transitional density p:
p(x, s, y, t) = probability density at being at y in time t, on condition that it started at x in time s
taking u – solution of (5) with terminal condition (6)
From Markov property
Now we introduce dual operator
By integration by part we can get
and from (7)
This is true for any , wich is independent from p
And we get Kolmogorov forward equation for p. Integrating by x we get the same equation for probability density at any moment T
Now we return to Gibbs invariant distribution for gradient flow
Stochastic gradient flow in SDE notation
– integral of white noise is Wiener process
We want to find invariant probability density . Invariant – means it doesn’t change with time,
so from Kolmogorov forward equation
C = 0 because we want integrable
and at last we get Gibbs distribution
Recalling again the chain of reasoning:
Wiener process →
Kolmogorov forward equation for probability density →
Conditional probability was always baffling me. Empirical, frequentists meaning is clear, but the abstract definition, originating from Kolmogorov – what was its mathematical meaning? How it can be derived? It’s a nontrivial definition and is appearing in the textbooks out the air, without measure theory intuition behind it.
Here I mostly follow Chang&Pollard paper Conditioning as disintegartion. Beware that the paper use non-standard notation, but this post follow more common notation, same as in wikipedia.
Here is example form Chang&Pollard paper:
Suppose we have distribution on concentrated in two straight lines and with respective density and angles with X axis . Observation (X,Y) taken, giving , what is probability that point lies on the line ?
Standard approach would be approximate with and take limit with
Not only taking this limit is kind of cumbersome, it’s also not totally obvious that it’s the same conditional probability that defined in the abstract definition – we are replacing ratio with limit here.
Now what is “correct way” to define conditional probabilities, especially for distributions?
For simplicity we will first talk about single scalar random variable, defined on probability space. We will think of random variable X as function on the sample space. Now condition define fiber – inverse image of .
Disintegration theorem say that probability measure on the sample space can be decomposed into two measures – parametric family of measures induced by original probability on each fiber and “orthogonal” measure on – on the parameter space of the first measure. Here is the space of values of X and serve as parameter space for measures on fibers. Second measure induced by the inverse image of the function (random variable) for each measurable set on . This second measure is called Pushforward measure. Pushforward measure is just for measurable set on (in our case) taking its X inverse image on sample space and measuring it with μ.
Fiber is in fact sample space for conditional event, and measure on fiber is our conditional distribution.
Full statement of the theorem require some term form measure theory. Following wikipedia
* Then there exists a ν-almost everywhere uniquely determined family of probability measures ⊆ P(Y) such that
* the function is Borel measurable, in the sense that is a Borel-measurable function for each Borel-measurable set B ⊆ Y;
* “lives on” the fiber : for ν-almost all x ∈ X,
* for every Borel-measurable function : Y → [0, ∞],
From here for any event E form Y
This was complete statement of the disintegration theorem.
Now returning to Chang&Pollard example. For formal derivation I refer you to the original paper, here we will just “guess” , and by uniqueness it give us disintegration. Our conditional distribution for will be just point masses on the intersections of lines and with axis
Here is delta function – point mass.
Our conditional probability that event lies on with condition , form conditional density thus
Another example from Chen&Pollard. It relate to sufficient statistics. Term sufficient statistic used if we have probability distribution depending on some parameter, like in maximum likelihood estimation. Sufficient statistic is some function of sample, if it’s possible to estimate parameter of distribution from only values of that function in the best possible way – adding more data form the sample will not give more information about parameter of distribution.
Let be uniform distribution on the square . In that case M = max(x, y) is sufficient statistics for . How to show it?
Let take our function , and make disintegration.
is a uniform distribution on edges where x and y equal m and
is density of conditional probability and it doesn’t depend on
For any – means that M is sufficient.
It seems that in most cases disintegration is not a tool for finding conditional distribution. Instead it can help to guess it and form uniqueness prove that the guess is correct. That correctness could be nontrivial – there are some paradoxes similar to Borel Paradox in Chang&Pollard paper.