Bayesian inference for a logistic regression model (Part 4)

Part 4: Gradients and the Langevin algorithm


This is the fourth part in a series of posts on MCMC-based Bayesian inference for a logistic regression model. If you are new to this series, please go back to Part 1.

In the previous post we saw how the Metropolis algorithm could be used to generate a Markov chain targeting our posterior distribution. In high dimensions the diffusive nature of the Metropolis random walk proposal becomes increasingly inefficient. It is therefore natural to try and develop algorithms that use additional information about the target distribution. In the case of a differentiable log posterior target, a natural first step in this direction is to try and make use of gradient information.

Gradient of a logistic regression model

There are various ways to derive the gradient of our logistic regression model, but it might be simplest to start from the first form of the log likelihood that we deduced in Part 2:

\displaystyle l(\beta;y) = y^\textsf{T}X\beta - \mathbf{1}^\textsf{T}\log(\mathbf{1}+\exp[X\beta])

We can write this out in component form as

\displaystyle l(\beta;y) = \sum_j\sum_j y_iX_{ij}\beta_j - \sum_i\log\left(1+\exp\left[\sum_jX_{ij}\beta_j\right]\right).

Differentiating wrt \beta_k gives

\displaystyle \frac{\partial l}{\partial \beta_k} = \sum_i y_iX_{ik} - \sum_i \frac{\exp\left[\sum_j X_{ij}\beta_j\right]X_{ik}}{1+\exp\left[\sum_j X_{ij}\beta_j\right]}.

It’s then reasonably clear that stitching all of the partial derivatives together will give the gradient vector

\displaystyle \nabla l = X^\textsf{T}\left[ y - \frac{\mathbf{1}}{\mathbf{1}+\exp[-X\beta]} \right].

This is the gradient of the log likelihood, but we also need the gradient of the log prior. Since we are assuming independent \beta_i \sim N(0,v_i) priors, it is easy to see that the gradient of the log prior is just -\beta\circ v^{-1}. It is the sum of these two terms that gives the gradient of the log posterior.


In R we can implement our gradient function as

glp = function(beta) {
    glpr = -beta/(pscale*pscale)
    gll = as.vector(t(X) %*% (y - 1/(1 + exp(-X %*% beta))))
    glpr + gll


In Python we could use

def glp(beta):
    glpr = -beta/(pscale*pscale)
    gll = (X.T).dot(y - 1/(1 + np.exp(
    return (glpr + gll)

We don’t really need a JAX version, since JAX can auto-diff the log posterior for us.


  def glp(beta: DVD): DVD =
    val glpr = -beta /:/ pvar
    val gll = (X.t)*(y - ones/:/(ones + exp(-X*beta)))
    glpr + gll


Using hmatrix we could use something like

glp :: Matrix Double -> Vector Double -> Vector Double -> Vector Double
glp x y b = let
  glpr = -b / (fromList [100.0, 1, 1, 1, 1, 1, 1, 1])
  gll = (tr x) #> (y - (scalar 1)/((scalar 1) + (cmap exp (-x #> b))))
  in glpr + gll

There’s something interesting to say about Haskell and auto-diff, but getting into this now will be too much of a distraction. I may come back to it in some future post.


Dex is differentiable, so we don’t need a gradient function – we can just use grad lpost. However, for interest and comparison purposes we could nevertheless implement it directly with something like

prscale = map (\ x. 1.0/x) pscale

def glp (b: (Fin 8)=>Float) : (Fin 8)=>Float =
  glpr = -b*prscale*prscale
  gll = (transpose x) **. (y - (map (\eta. 1.0/(1.0 + eta)) (exp (-x **. b))))
  glpr + gll

Langevin diffusions

Now that we have a way of computing the gradient of the log of our target density we need some MCMC algorithms that can make good use of it. In this post we will look at a simple approximate MCMC algorithm derived from an overdamped Langevin diffusion model. In subsequent posts we’ll look at more sophisticated, exact MCMC algorithms.

The multivariate stochastic differential equation (SDE)

\displaystyle dX_t = \frac{1}{2}\nabla\log\pi(X_t)dt + dW_t

has \pi(\cdot) as its equilibrium distribution. Informally, an SDE of this form is a continuous time process with infinitesimal transition kernel

\displaystyle X_{t+dt}|(X_t=x_t) \sim N\left(x_t+\frac{1}{2}\nabla\log\pi(x_t)dt,\mathbf{I}dt\right).

There are various more-or-less formal ways to see that \pi(\cdot) is stationary. A good way is to check it satisfies the Fokker–Planck equation with zero LHS. A less formal approach would be to see that the infinitesimal transition kernel for the process satisfies detailed balance with \pi(\cdot).

Similar arguments show that for any fixed positive definite matrix A, the SDE

\displaystyle dX_t = \frac{1}{2}A\nabla\log\pi(X_t)dt + A^{1/2}dW_t

also has \pi(\cdot) as a stationary distribution. It is quite common to choose a diagonal matrix A to put the components of X_t on a common scale.

The unadjusted Langevin algorithm

Simulating exact sample paths from SDEs such as the overdamped Langevin diffusion model is typically difficult (though not necessarily impossible), so we instead want something simple and tractable as the basis of our MCMC algorithms. Here we will just simulate from the Euler–Maruyama approximation of the process by choosing a small but finite time step \Delta t and using the transition kernel

\displaystyle X_{t+\Delta t}|(X_t=x_t) \sim N\left(x_t+\frac{1}{2}A\nabla\log\pi(x_t)\Delta t, A\Delta t\right)

as the basis of our MCMC method. For sufficiently small \Delta t this should accurately approximate the Langevin dynamics, leading to an equilibrium distribution very close to \pi(\cdot). That said, we would like to choose \Delta t as large as we can get away with, since that will lead to a more rapidly mixing MCMC chain. Below are some implementations of this kernel for a diagonal pre-conditioning matrix.



We can create a kernel for the unadjusted Langevin algorithm in R with the following function.

ulKernel = function(glpi, dt = 1e-4, pre = 1) {
    sdt = sqrt(dt)
    spre = sqrt(pre)
    advance = function(x) x + 0.5*pre*glpi(x)*dt
    function(x, ll) rnorm(p, advance(x), spre*sdt)

Here, we can pass in pre, which is expected to be a vector representing the diagonal of the pre-conditioning matrix, A. We can then use this kernel to generate an MCMC chain as we have seen previously. See the full runnable script for further details.


def ulKernel(glpi, dt = 1e-4, pre = 1):
    p = len(init)
    sdt = np.sqrt(dt)
    spre = np.sqrt(pre)
    advance = lambda x: x + 0.5*pre*glpi(x)*dt
    def kernel(x):
        return advance(x) + np.random.randn(p)*spre*sdt
    return kernel

See the full runnable script for further details.


def ulKernel(lpi, dt = 1e-4, pre = 1):
    p = len(init)
    glpi = jit(grad(lpi))
    sdt = jnp.sqrt(dt)
    spre = jnp.sqrt(pre)
    advance = jit(lambda x: x + 0.5*pre*glpi(x)*dt)
    def kernel(key, x):
        return advance(x) + jax.random.normal(key, [p])*spre*sdt
    return kernel

Note how for JAX we can just pass in the log posterior, and the gradient function can be obtained by automatic differentiation. See the full runnable script for further details.


def ulKernel(glp: DVD => DVD, pre: DVD, dt: Double): DVD => DVD =
  val sdt = math.sqrt(dt)
  val spre = sqrt(pre)
  def advance(beta: DVD): DVD =
    beta + (0.5*dt)*(pre*:*glp(beta))
  beta => advance(beta) + sdt*,_).sample())

See the full runnable script for further details.


ulKernel :: (StatefulGen g m) =>
  (Vector Double -> Vector Double) -> Vector Double -> Double -> g ->
  Vector Double -> m (Vector Double)
ulKernel glpi pre dt g beta = do
  let sdt = sqrt dt
  let spre = cmap sqrt pre
  let p = size pre
  let advance beta = beta + (scalar (0.5*dt))*pre*(glpi beta)
  zl <- (replicateM p . genContVar (normalDistr 0.0 1.0)) g
  let z = fromList zl
  return $latex  advance(beta) + (scalar sdt)*spre*z

See the full runnable script for further details.


In Dex we can write a function that accepts a gradient function

def ulKernel {n} (glpi: (Fin n)=>Float -> (Fin n)=>Float)
    (pre: (Fin n)=>Float) (dt: Float)
    (b: (Fin n)=>Float) (k: Key) : (Fin n)=>Float =
  sdt = sqrt dt
  spre = sqrt pre
  b + (((0.5)*dt) .* (pre*(glpi b))) +
    (sdt .* (spre*(randn_vec k)))

or we can write a function that accepts a log posterior, and uses auto-diff to construct the gradient

def ulKernel {n} (lpi: (Fin n)=>Float -> Float)
    (pre: (Fin n)=>Float) (dt: Float)
    (b: (Fin n)=>Float) (k: Key) : (Fin n)=>Float =
  glpi = grad lpi
  sdt = sqrt dt
  spre = sqrt pre
  b + ((0.5)*dt) .* (pre*(glpi b)) +
    sdt .* (spre*(randn_vec k))

and since Dex is statically typed, we can’t easily mix these functions up.

See the full runnable scripts, without and with auto-diff.

Next steps

In this post we have seen how to construct an MCMC algorithm that makes use of gradient information. But this algorithm is approximate. In the next post we’ll see how to correct for the approximation by using the Langevin updates as proposals within a Metropolis-Hastings algorithm.

Stochastic reaction-diffusion modelling


There is a fairly large literature on reaction-diffusion modelling using partial differential equations (PDEs). There is also a fairly large literature on stochastic modelling of coupled chemical reactions, which account for the discreteness of reacting species at low concentrations. There is some literature on combining the two, to form stochastic reaction-diffusion systems, but much less.

In this post we will look at one approach to the stochastic reaction-diffusion problem, based on an underlying stochastic process often described by the reaction diffusion master equation (RDME). We will start by generating exact realisations from this process using the spatial Gillespie algorithm, before switching to a continuous stochastic approximation often known as the spatial chemical Langevin equation (spatial CLE). For fine discretisations, this spatial CLE is just an explicit numerical scheme for an associated reaction-diffusion stochastic partial differential equation (SPDE), and we can easily contrast such SPDE dynamics with their deterministic PDE approximation. We will investigate using simulation, based on my Scala library, scala-smfsb, which accompanies the third edition of my textbook, Stochastic modelling for systems biology, as discussed in previous posts.

All of the code used to generate the plots and movies in this post is available in my blog repo, and is very simple to build and run.

The Lotka-Volterra reaction network

Exact simulation from the RDME

My favourite toy coupled chemical reaction network is the Lotka-Volterra predator-prey system, presented as the three reactions

X \longrightarrow 2X
X + Y \longrightarrow 2Y
Y \longrightarrow \emptyset

with X representing the prey species and Y the predator. I showed how to simulate realisations from this process using the Scala library in the previous post. Here we will consider simulation of this model in 2d, and simulate exact realisation from the appropriate RDME using the spatial Gillespie algorithm. Full runnable code for this simulation is here, but the key lines are:

val r = 100; val c = 120
val model =[IntState]()
val step = Spatial.gillespie2d(model, DenseVector(0.6, 0.6), maxH=1e12)
val x00 = DenseVector(0, 0)
val x0 = DenseVector(50, 100)
val xx00 = PMatrix(r, c, Vector.fill(r*c)(x00))
val xx0 = xx00.updated(c/2, r/2, x0)
val s = Stream.iterate(xx0)(step(_,0.0,0.1))

which sets up an infinite lazy Stream of states on a 100×120 grid over time steps of 0.1 units with diffusion rates of 0.6 for both species. We can then map this to a stream of images and visualise it using my scala-view library (described in this post). Running gives the following output:


The above image is the final frame of a movie which can be viewed by clicking on the image. In the simulation, blue represents the prey species, X, and red represents the predator, Y. The simulation is initialised with a few prey and predators in the central pixel. At each time step of the simulation, either a reaction or a diffusion event may occur. If diffusion occurs, an individual moves from its current location to one of the four adjacent pixels. This algorithm is extremely computationally intensive, however well it is implemented. The implementation used here (using the function Spatial.gillespie2d in the scala-smfsb library) is quite inefficient. A more efficient implementation would use the next subvolume method or similar algorithm. But since every reaction event is simulated sequentially, this algorithm is always going to be intolerably slow for most interesting problems.

The spatial CLE

The spatial CLE effectively approximates the true RDME dynamics with a set of coupled stochastic differential equations (SDEs) on the spatial grid. This can be interpreted as an explicit scheme for numerically integrating an SPDE. But this numerical scheme is much more efficient, allowing sensible time-stepping of the process, and vectorises and parallelises nicely. The details are in my book, but the Scala implementation is here. Diffusion is implemented efficiently and in parallel using the comonadic approach that I’ve described previously. We can quickly and easily generate large simulations using the spatial CLE. Here is a movie generated on a 250×300 grid.


Again, clicking on the frame should give the movie. We see that although the quantitative details are slightly different to the exact algorithm, the essential qualitative behaviour of the system is captured well by the spatial CLE. Full code for this simulation is here.

Reaction-diffusion PDE

If we remove all of the noise terms from the spatial CLE, we get a set of coupled ODEs, which again, may be interpreted as a numerical scheme for integrating a reaction-diffusion PDE model. Below are the dynamics on the same 250×300 grid.


It seems a bit harsh to describe a reaction-diffusion PDE as “boring”, but it certainly isn’t as interesting as the stochastic dynamics. Also, it has qualitatively quite different behaviour to the stochastic models, with wavefronts being less pronounced and less well separated. The code for this one is here.

Other initialisations

Instead of just seeding the simulation with some individuals in the central pixel, we can initialise 3 pixels. We can look first at a spatial CLE simulation.


Code here.

We can look at the same problem, but now using a PDE.


Code here.

Alternatively, we can initialise every pixel independently with random numbers of predator and prey. A movie for this is given below, following a short warm-up.


Code here.

Again, we can look at the corresponding deterministic integration.


Code here.

The SIR model

Let’s now turn attention to a spatial epidemic process model, the spatial susceptible-infectious-recovered model. Again, we’ll start from the discrete reaction formulation.

S + I \longrightarrow 2I
I \longrightarrow R

I’ll add this model to the next release of scala-smfsb, but in the meantime we can easily define it ourselves with:

def sir[S: State](p: DenseVector[Double] = DenseVector(0.1, 0.5)): Spn[S] =
    List("S", "I", "R"),
    DenseMatrix((1, 1, 0), (0, 1, 0)),
    DenseMatrix((0, 2, 0), (0, 0, 1)),
    (x, t) => {
      val xd = x.toDvd
        xd(0) * xd(1) * p(0), xd(1) * p(1)

We can seed a simulation with a few infectious individuals in the centre of a roughly homogeneous population of susceptibles.

Spatial CLE

This time we’ll skip the exact simulation, since it’s very slow, and go straight to the spatial CLE. A simulation on a 250×300 grid is given below.


Here, green represents S, red I and blue R. In this simulation, I diffuses more slowly than S, and R doesn’t diffuse at all.
Code here.

PDE model

If we ditch the noise to get a reaction-diffusion PDE model, the dynamics are as follows.


Again, we see that the deterministic model is quite different to the stochastic version, and kind-of boring. Code here.

Further information

All of the code used to generate the plots and movies in this post is available in an easily runnable form in my blog repo. It is very easy to adapt the examples to vary parameters and initial conditions, and to study other reaction systems. Further details relating to stochastic reaction-diffusion modelling based on the RDME can be found in Chapter 9 of my textbook, Stochastic modelling for systems biology, third edition.