Use Tikhonov Regularization to Solve Fredholm Integral Equation

Apr 26, 2019

Background #

Fredholm in­te­gral equa­tion of the first kind is writ­ten as,

f(x)=abK(x,s)p(s)ds.f(x)=\int_{a}^{b}K(x,s)p(s)\mathrm{d}s.

The prob­lem is to find p(s)p(s), given that f(x)f(x) and K(x,s)K(x,s) are known. This equa­tion oc­curs quite of­ten in many dif­fer­ent ar­eas.

Discretization-based Method #

Here we de­scribe a dis­cretiza­tion-based method to solve the Fredholm in­te­gral equa­tion. The in­te­gral equa­tion is ap­prox­i­mately re­placed by a Riemann sum­ma­tion over grids,

f(xi)=jΔsK(xi,sj)p(sj)f(x_i)=\sum_j \Delta_s K(x_i, s_j) p(s_j)

where Δs\Delta_s is the grid size along the di­men­sion ss and xix_i, sjs_j are the grid points with ii and jj in­di­cat­ing their in­dices. When grid size Δs0\Delta_s\to0, the sum­ma­tion con­verges to the true in­te­gral. It is more con­ve­nient to write it in the ma­trix form,

f=ΔsKp\boldsymbol{f} = \boldsymbol{\Delta}_s \boldsymbol{K} \boldsymbol{p}

where

f=(f(x1),f(x2),,f(xn))T,\boldsymbol{f}=(f(x_1), f(x_2),\cdots,f(x_n))^{\mathrm{T}},

K=(K(x1,s1)K(x1,s2)K(x1,sm)K(x2,s1)K(x2,s2)K(x2,sm)K(xn,s1)K(xn,s2)K(xn,sm))\boldsymbol{K}= \begin{pmatrix} K(x_1,s_1) & K(x_1,s_2) & \cdots & K(x_1,s_m) \\ K(x_2,s_1) & K(x_2,s_2) & \cdots & K(x_2,s_m) \\ \vdots & \vdots & \ddots & \vdots \\ K(x_n,s_1) & K(x_n,s_2) & \cdots & K(x_n,s_m) \end{pmatrix}

p=(p(s1),p(s2),,p(sm))T\boldsymbol{p} = (p(s_1),p(s_2),\cdots,p(s_m))^{\mathrm{T}}

and Δs=ΔsI\boldsymbol{\Delta}_s = \Delta_s \boldsymbol{I} with I\boldsymbol{I} be­ing the iden­tity ma­trix of di­men­sion n×nn \times n.

Now solv­ing the Fredholm in­te­gral equa­tion is equiv­a­lent to solv­ing a sys­tem of lin­ear equa­tions. The stan­dard ap­proach or­di­nary least squares lin­ear re­gres­sion, which is to find the vec­tor p\boldsymbol{p} min­i­miz­ing the norm ΔsKpf22\vert\vert \boldsymbol{\Delta}_s \boldsymbol{K} \boldsymbol{p}-\boldsymbol{f}\vert\vert_2^2. In prin­ci­ple, the Fredholm in­te­gral equa­tion may have non-unique so­lu­tions, thus the cor­re­spond­ing lin­ear equa­tions are also ill-posed. The most com­monly used method for ill-posed prob­lem is Tikhonov reg­u­lar­iza­tion which is to min­i­mize

ΔsKpf22+α2p22\vert\vert\boldsymbol{\Delta}_s \boldsymbol{K} \boldsymbol{p}-\boldsymbol{f}\vert\vert_2^2+\alpha^2\vert\vert\boldsymbol{p}\vert\vert_2^2

Note that this is ac­tu­ally a sub­set of Tikhonov reg­u­lar­iza­tion (also called Ridge reg­u­lar­iza­tion) with α\alpha be­ing a pa­ra­me­ter.

When p(s)p(s) is a prob­a­bil­ity den­sity func­tion #

In many cases, both f(x)f(x) and g(s)g(s) are prob­a­bil­ity den­sity func­tion (PDF), and K(x,s)K(x,s) is a con­di­tional PDF, equiv­a­lent to K(xs)K(x\vert s). Thus, there are two con­straints on the so­lu­tion p(s)p(s), that is p(s)0p(s)\geq 0 and p(s)ds=1\int p(s)\mathrm{d}s = 1. These two con­straints trans­late to p(si)0p(s_i)\geq 0 for any sis_i and Δsip(si)=1\Delta_s\sum_i p(s_i)=1. Hence, we need to solve the Tikhonov reg­u­lar­iza­tion prob­lem sub­ject to these two con­straints.

In the fol­low­ing, I will show how to solve the Tikhonov reg­u­lar­iza­tion prob­lem with both equal­ity and in­equal­ity con­straints. First, I will show that the Tikhonov reg­u­lar­iza­tion prob­lem with non-neg­a­tive con­straint can be eas­ily trans­lated to a reg­u­lar non-neg­a­tive least square prob­lem (NNLS) which can be solved us­ing ac­tive set al­go­rithm.

Let us con­struct the ma­trix,

A=(ΔsKαI)\boldsymbol{A}= \begin{pmatrix} \Delta_s \boldsymbol{K} \\ \alpha \boldsymbol{I} \end{pmatrix}

and the vec­tor,

b=(f0)\boldsymbol{b}= \begin{pmatrix} \boldsymbol{f}\\ \boldsymbol{0} \end{pmatrix}

where I\boldsymbol{I} is the m×mm\times m iden­tity ma­trix and 0\boldsymbol{0} is the zero vec­tor of size mm. It is easy to show that the Tikhonov reg­u­lar­iza­tion prob­lem min(ΔsKpf22+α2p22)\mathrm{min}(\vert\vert\boldsymbol{\Delta}_{s} \boldsymbol{K} \boldsymbol{p} - \boldsymbol{f}\vert\vert_{2}^{2}+\alpha^2\vert\vert\boldsymbol{p}\vert\vert_{2}^{2}) sub­ject to p0\boldsymbol{p}\geq 0 is equiv­a­lent to the reg­u­lar NNLS prob­lem,

min(Apb22), subject to p0\mathrm{min}(\vert\vert\boldsymbol{A}\boldsymbol{p}-\boldsymbol{b}\vert\vert_2^2),\mathrm{\ sub­ject\ to\ }\boldsymbol{p}\geq 0

Now we add the equal­ity con­straint, Δsip(si)=1\Delta_s\sum_i p(s_i)=1 or 1p=1/Δs\boldsymbol{1}\boldsymbol{p}=1/\Delta_s writ­ten in ma­trix form. My im­ple­men­ta­tion of solv­ing such prob­lem fol­lows the al­go­rithm de­scribed in Haskell and Hanson [1]. According to their method, the prob­lem be­comes an­other NNLS prob­lem,

min(1p1/Δs22+ϵ2Apb22), subject to p0\mathrm{min}(\vert\vert\boldsymbol{1}\boldsymbol{p}-1/\Delta_s\vert\vert_2^2+\epsilon^2\vert\vert\boldsymbol{A}\boldsymbol{p}-\boldsymbol{b}\vert\vert_2^2),\mathrm{\ sub­ject\ to\ }\boldsymbol{p}\geq 0

The so­lu­tion to the above equa­tion con­verges to the true so­lu­tion when ϵ0+\epsilon\to0^+. Now I have de­scribed the al­go­rithm to solve the Fredholm equa­tion of the first kind when p(s)p(s) is a prob­a­bil­ity den­sity func­tion. I call the al­go­rithm de­scribed above as non-neg­a­tive Tikhonov reg­u­lar­iza­tion with equal­ity con­straint (NNETR).

Code #

Here I show the core code of the al­go­rithm de­scribed above.

# core algorithm of non-negative equality Tikhonov regularization (NNETR)
def NNETR(K, f, Delta, epsilon, alpha):
    # the first step
    A_nn = np.vstack((K, alpha * np.identity(K.shape[1])))
    b_nn = np.hstack((f, np.zeros(K.shape[1])))

    # the second step
    A_nne = np.vstack((epsilon * A_nn, np.full(A_nn.shape[1], 1.0)))
    b_nne = np.hstack((epsilon * b_nn, 1.0))

    # Use NNLS solver provided by scipy
    sol, residue = scipy.optimize.nnls(A_nne, b_nne)

    # solution should be divided by Delta (grid size)
    sol = sol/Delta
    return sol, residue

Examples #

  • Compounding an ex­po­nen­tial dis­tri­b­u­tion with its rate pa­ra­me­ter dis­trib­uted ac­cord­ing to a gamma dis­tri­b­u­tion yields a Lomax dis­tri­b­u­tion f(x)=a(x+1)(a+1)f(x)=a(x+1)^{-(a+1)}, sup­ported on (0,)(0,\infty), with a>0a>0. k(x,θ)=θeθxk(x,\theta)=\theta e^{-\theta x} is an ex­po­nen­tial den­sity and p(θ)=Γ(a)1θa1eθp(\theta) = \Gamma(a)^{-1}\theta^{a-1}e^{-\theta} is a gamma den­sity.
Example #1: comparison between true and NNETR solution
  • Compounding a Gaussian dis­tri­b­u­tion with mean dis­trib­uted ac­cord­ing to an­other Gaussian dis­tri­b­u­tion yields (again) a Gaussian dis­tri­b­u­tion f(x)=N(a,b2+σ2)f(x)=\math­cal{N}(a,b^2+\sigma^2). k(xμ)=N(μ,σ2)k(x\vert\mu)=\math­cal{N}(\mu,\sigma^2) and p(μ)=N(a,b2)p(\mu)=\math­cal{N}(a,b^2)
Example #2: comparison between true and NNETR solution
  • Compounding an ex­po­nen­tial dis­tri­b­u­tion with its rate pa­ra­me­ter dis­trib­uted ac­cord­ing to a mix­ture dis­tri­b­u­tion of two gamma dis­tri­b­u­tions. Similar to the first ex­am­ple, we use k(x,θ)=θeθxk(x,\theta)=\theta e^{-\theta x}. But here we use p(θ)=qp(θa1)+(1q)p(θa2)p(\theta)=q p(\theta\vert a_1)+(1-q)p(\theta\vert a_2) where qq, a1a_1 and a2a_2 are pa­ra­me­ters. It is clear that p(θ)p(\theta) is a mix­ture be­tween two dif­fer­ent gamma dis­tri­b­u­tions such as it is a bi­modal dis­tri­b­u­tion. Following the first ex­am­ple, we have f(x)=qf(xa1)+(1q)f(xa2)f(x)=qf(x\vert a_1)+(1-q)f(x\vert a_2) where f(xa)=a(x+1)(a+1)f(x\vert a)=a(x+1)^{-(a+1)}.
Example #3: comparison between true and NNETR solution
  • Compounding an ex­po­nen­tial dis­tri­b­u­tion with its rate pa­ra­me­ter dis­trib­uted as a dis­crete dis­tri­b­u­tion.
Example #4: comparison between true and NNETR solution
  1. Haskell, Karen H., and Richard J. Hanson. An al­go­rithm for lin­ear least squares prob­lems with equal­ity and non­neg­a­tiv­ity con­straints.” Mathematical Programming 21.1 (1981): 98-118. ↩︎