Iterative proportional fitting

IPF has been "re-invented" many times, the earliest by Kruithof in 1937 [6] in relation to telephone traffic ("Kruithof’s double factor method"), Deming and Stephan in 1940[7] for adjusting census crosstabulations, and G.V. Sheleikhovskii for traffic as reported by Bregman.[8] (Deming and Stephan proposed IPFP as an algorithm leading to a minimizer of the Pearson X-squared statistic, which Stephan later reported it does not).[9] Early proofs of uniqueness and convergence came from Sinkhorn (1964),[10] Bacharach (1965),[11] Bishop (1967),[12] and Fienberg (1970).[13] Bishop's proof that IPFP finds the maximum likelihood estimator for any number of dimensions extended a 1959 proof by Brown for 2x2x2... cases. Fienberg's proof by differential geometry exploits the method's constant crossproduct ratios, for strictly positive tables. Csiszár (1975).[14] found necessary and sufficient conditions for general tables having zero entries. Pukelsheim and Simeone (2009) [15] give further results on convergence and error behavior.

An exhaustive treatment of the algorithm and its mathematical foundations can be found in the book of Bishop et al. (1975).[16] Idel (2016)[17] gives a more recent survey.

Other general algorithms can be modified to yield the same limit as the IPFP, for instance the Newton–Raphson method and the EM algorithm. In most cases, IPFP is preferred due to its computational speed, low storage requirements, numerical stability and algebraic simplicity.

Applications of IPFP have grown to include trip distribution models, Fratar or Furness and other applications in transportation planning (Lamond and Stewart), survey weighting, synthesis of cross-classified demographic data, adjusting input–output models in economics, estimating expected quasi-independent contingency tables, biproportional apportionment systems of political representation, and for a preconditioner in linear algebra.[18]

Biproportion, whatever the algorithm used to solve it, is the following concept: ${\displaystyle Z}$, matrix ${\displaystyle Y}$ and matrix ${\displaystyle X}$ are known real nonnegative matrices of dimension ${\displaystyle n,m}$; the interior of ${\displaystyle Y}$ is unknown and ${\displaystyle X}$ is searched such that ${\displaystyle X}$ has the same margins than ${\displaystyle Y}$, i.e. ${\displaystyle Xs=Ys}$ and ${\displaystyle s'X=s'Y}$ (${\displaystyle s}$ being the sum vector, and such that ${\displaystyle X}$ is closed to ${\displaystyle Z}$ following a given criterion, the fitted matrix being of the form ${\displaystyle X=K(Z,Y)=PZQ}$, where ${\displaystyle P}$ and ${\displaystyle Q}$ are diagonal matrices.

${\displaystyle min\sum _{i}\sum _{j}x_{ij}\log(x_{ij}/z_{ij})}$ s.t.${\displaystyle \sum _{j}x_{ij}=y_{i.}}$, ∀${\displaystyle i}$ and ${\displaystyle \sum _{i}x_{ij}=y_{.j}}$, ∀${\displaystyle j}$. The Lagrangian is ${\displaystyle L=\sum _{i}\sum _{j}x_{ij}\log(x_{ij}/z_{ij})-\sum _{i}p_{i}(y_{i.}-\sum _{j}x_{ij})-\sum _{j}q_{j}(y_{.j}-\sum _{i}x_{ij})}$.

Thus ${\displaystyle x_{ij}=z_{ij}\exp -(1+p_{i}+q_{j})}$, for ∀${\displaystyle i,j}$,

which, after posing ${\displaystyle P_{i}=\exp -(1+p_{i})}$ and ${\displaystyle Q_{j}=\exp -q_{j}}$, yields

${\displaystyle x_{ij}=P_{i}z_{ij}Q_{j}}$, ∀${\displaystyle i,j}$, i.e., ${\displaystyle X=PZQ}$,

with ${\displaystyle P_{i}=z_{i}.(\sum _{j}z_{ij}Q_{j})^{-1}}$, ∀${\displaystyle i}$ and ${\displaystyle Q_{j}=z_{.j}(\sum _{i}z_{ij}P_{i})^{-1}}$, ∀${\displaystyle j}$. ${\displaystyle P_{i}}$ and ${\displaystyle Q_{j}}$ form a system that can be solve iteratively:

${\displaystyle P_{i}=z_{i}.^{(t+1)}(\sum _{j}z_{ij}Q_{j}^{(t)})^{-1}}$, ∀${\displaystyle i}$ and ${\displaystyle Q_{j}^{(t+1)}=z_{.j}(\sum _{i}z_{ij}P_{i}^{(t+1)})^{-1}}$, ∀${\displaystyle j}$.

The solution ${\displaystyle X}$ is independent of the initialization chosen (i.e., we can begin by ${\displaystyle q_{j}^{(0)}=1}$, ∀${\displaystyle j}$ or by${\displaystyle p_{i}^{(0)}=1}$, ∀${\displaystyle i}$. If the matrix ${\displaystyle Z}$ is “indecomposable”, then this process has a unique fixed-point because it is deduced from program a program where the function is a convex and continuously derivable function defined on a compact set. In some cases the solution may not exist: see de Mesnard's example cited by Miller and Blair (Miller R.E. & Blair P.D. (2009) Input-output analysis: Foundations and Extensions, Second edition, Cambridge (UK): Cambridge University Press, p. 335-336 (freely available)).

Some properties (see de Mesnard (1994)):

Lack of information: if ${\displaystyle Z}$ brings no information, i.e., ${\displaystyle z_{i}j=z}$, ∀${\displaystyle i,j}$ then ${\displaystyle X=PQ}$.

Idempotency: ${\displaystyle X=K(Z,Y)=Z}$ if ${\displaystyle Y}$ has the same margins than ${\displaystyle Z}$.

Composition of biproportions: ${\displaystyle K(K(Z,Y_{1}),Y_{2}=K(Z,Y_{2})}$; ${\displaystyle K(...K(Z,Y_{1}),Y_{2})...Z_{N})=K(Z,Y_{N})}$.

Zeros: a zero in ${\displaystyle Z}$ is projected as a zero in ${\displaystyle X}$. Thus, a bloc-diagonal matrix is projected as a bloc-diagonal matrix and a triangular matrix is projected as a triangular matrix.

Theorem of separable modifications: if ${\displaystyle Z}$ is premutiplied by a diagonal matrix and/or postmultiplied by a diagonal matrix, then the solution is unchanged.

Theorem of "unicity": If ${\displaystyle K^{q}}$ is any non-specified algorithm, with ${\displaystyle {\hat {X}}=K^{q}(Z,Y)=UZV}$, ${\displaystyle U}$ and ${\displaystyle V}$ being unknown, then ${\displaystyle U}$ and ${\displaystyle V}$ can always be changed into the standard form of ${\displaystyle P}$ and ${\displaystyle Q}$. The demonstrations calls some above properties, particularly the Theorem of separable modifications and the composition of biproportions.

Given a two-way (I × J)-table ${\displaystyle x_{ij}}$, we wish to estimate a new table ${\displaystyle {\hat {m}}_{ij}=a_{i}b_{j}x_{ij}}$ for all i and j such that the marginals satisfy ${\displaystyle \sum _{j}{\hat {m}}_{ij}\ =u_{i},}$ and ${\displaystyle \sum _{i}{\hat {m}}_{ij}\ =v_{j}}$.

Choose initial values ${\displaystyle {\hat {m}}_{ij}^{(0)}:=x_{ij}}$, and for ${\displaystyle \eta \geq 1}$ set

${\displaystyle {\hat {m}}_{ij}^{(2\eta -1)}={\frac {{\hat {m}}_{ij}^{(2\eta -2)}u_{i}}{\sum _{k=1}^{J}{\hat {m}}_{ik}^{(2\eta -2)}}}}$

${\displaystyle {\hat {m}}_{ij}^{(2\eta )}={\frac {{\hat {m}}_{ij}^{(2\eta -1)}v_{j}}{\sum _{k=1}^{I}{\hat {m}}_{kj}^{(2\eta -1)}}}.}$

Repeat these steps until row and column totals are sufficiently close to u and v.

Notes:

• For the RAS form of the algorithm, define the diagonalization operator ${\displaystyle diag:\mathbb {R} ^{k}\longrightarrow \mathbb {R} ^{k\times k}}$, which produces a (diagonal) matrix with its input vector on the main diagonal and zero elsewhere. Then, for each row adjustment, let ${\displaystyle R^{\eta }=diag({\frac {u_{i}}{\sum _{j}m_{ij}^{(2\eta -2)}}})}$, from which ${\displaystyle M^{2\eta -1}=R^{\eta }M^{2\eta -2}}$. Similarly each column adjustment's ${\displaystyle S^{\eta }=diag({\frac {v_{i}}{\sum _{i}m_{ij}^{(2\eta -1)}}})}$, from which ${\displaystyle M^{2\eta }=M^{2\eta -1}S^{\eta }}$. Reducing the operations to the necessary ones, it can easily be seen that RAS does the same as classical IPF. In practice, one would not implement actual matrix multiplication with the whole R and S matrices; the RAS form is more a notational than computational convenience.

Assume the same setting as in the classical IPFP. Alternatively, we can estimate the row and column factors separately: Choose initial values ${\displaystyle {\hat {b}}_{j}^{(0)}:=1}$, and for ${\displaystyle \eta \geq 1}$ set

${\displaystyle {\hat {a}}_{i}^{(\eta )}={\frac {u_{i}}{\sum _{j}\ x_{ij}{\hat {b}}_{j}^{(\eta -1)}}},}$

${\displaystyle {\hat {b}}_{j}^{(\eta )}={\frac {v_{j}}{\sum _{i}\ x_{ij}{\hat {a}}_{i}^{(\eta )}}}}$

Repeat these steps until successive changes of a and b are sufficiently negligible (indicating the resulting row- and column-sums are close to u and v).

Finally, the result matrix is ${\displaystyle {\hat {m}}_{ij}={\hat {a}}_{i}^{(\eta )}{\hat {b}}_{j}^{(\eta )}x_{ij}}$

Notes:

• The two variants of the algorithm are mathematically equivalent, as can be seen by formal induction. With factor estimation, it is not necessary to actually compute each cycle's ${\displaystyle {\hat {m}}_{ij}^{(\eta )}}$.
• The factorization is not unique, since it is ${\displaystyle m_{ij}=a_{i}b_{j}x_{ij}=(\gamma a_{i})({\frac {1}{\gamma }}b_{j})x_{ij}}$ for all ${\displaystyle \gamma >0}$.

The vaguely demanded 'similarity' between M and X can be explained as follows: IPFP (and thus RAS) maintains the crossproduct ratios, i.e.

${\displaystyle {\frac {m_{ij}^{(\eta )}m_{hk}^{(\eta )}}{m_{ik}^{(\eta )}m_{hj}^{(\eta )}}}={\frac {x_{ij}x_{hk}}{x_{ik}x_{hj}}}\ \forall \ \eta \geq 0{\text{ and }}i\neq h,\quad j\neq k}$

since ${\displaystyle m_{ij}^{(\eta )}=a_{i}^{(\eta )}b_{j}^{(\eta )}x_{ij}.}$

This property is sometimes called structure conservation and directly leads to the geometrical interpretation of contingency tables and the proof of convergence in the seminal paper of Fienberg (1970).

Direct factor estimation (algorithm 2) is generally the more efficient way to solve IPF: Whereas a form of the classical IPFP needs

${\displaystyle IJ(2+J)+IJ(2+I)=I^{2}J+IJ^{2}+4IJ\,}$

elementary operations in each iteration step (including a row and a column fitting step), factor estimation needs only

${\displaystyle I(1+J)+J(1+I)=2IJ+I+J\,}$

operations being at least one order in magnitude faster than classical IPFP.

IPFP can be used to estimate expected quasi-independent (incomplete) contingency tables, with ${\displaystyle u_{i}=x_{i+},v_{j}=x_{+j}}$, and ${\displaystyle m_{ij}^{0}=1}$ for included cells and ${\displaystyle m_{ij}^{0}=0}$ for excluded cells. For fully independent (complete) contingency tables, estimation with IPFP concludes exactly in one cycle.

Necessary and sufficient conditions for the existence and uniqueness of MLEs are complicated in the general case (see[22]), but sufficient conditions for 2-dimensional tables are simple:

• the marginals of the observed table do not vanish (that is, ${\displaystyle x_{i+}>0,\ x_{+j}>0}$) and
• the observed table is inseparable (i.e. the table does not permute to a block-diagonal shape).

If unique MLEs exist, IPFP exhibits linear convergence in the worst case (Fienberg 1970), but exponential convergence has also been observed (Pukelsheim and Simeone 2009). If a direct estimator (i.e. a closed form of ${\displaystyle ({\hat {m}}_{ij})}$) exists, IPFP converges after 2 iterations. If unique MLEs do not exist, IPFP converges toward the so-called extended MLEs by design (Haberman 1974), but convergence may be arbitrarily slow and often computationally infeasible.

If all observed values are strictly positive, existence and uniqueness of MLEs and therefore convergence is ensured.

Consider the following table, given with the row- and column-sums and targets.

For executing the classical IPFP, we first adjust the rows:

The first step exactly matched row sums, but not the column sums. Next we adjust the columns:

Now the column sums exactly match their targets, but the row sums no longer match theirs. After completing three cycles, each with a row adjustment and a column adjustment, we get a closer approximation:

The R package mipfp (currently in version 3.1) provides a multi-dimensional implementation of the traditional iterative proportional fitting procedure.[23] The package allows the updating of a N-dimensional array with respect to given target marginal distributions (which, in turn can be multi-dimensional).

Python has an equivalent package, ipfn[24][25] that can be installed via pip. The package supports numpy and pandas input objects.

• Data cleansing
• Data editing
• NM-method
• Triangulation (social science) for quantitative and qualitative study data enhancement.