Differential Fields
Table of Contents
1. Overview
If we want to try to symbolically solve differential equations, it is tempting to think of analogous problems (like finding the roots of polynomials). We just work over a differential ring, like \(\mathbb{k}=\RR[{[x]}]\) the ring of formal power series with real coefficients (or rational coefficients, or whatever), we can define a formal derivative operator:
\begin{equation} \D (bx^{m} + cx^{n}) = \D(bx^{m}) +\D(cx^{n}) = mbx^{m-1} + ncx^{n-1}. \end{equation}Then we just translate the differential equation we're trying to solve
\begin{equation} y' = f(x,y) \end{equation}into finding an element \(y\in\mathbb{k}\) such that
\begin{equation} \D y = f(x,y). \end{equation}2. Polynomial Methods
Given some representation of polynomials in a single variable, we review a few helper functions that come in handy.
If \(R\) is a principal ideal domain, then \((x,y) = (\gcd(x,y))\) for any \(x,y\in R\).
A Euclidean Domain consists of an integral domain \((R, +, \cdot)\) equipped with a map \(\nu\colon R\setminus\{0\}\to\mathbb{N}\) called the Size function such that
- For all \(a,b\in R\setminus\{0\}\), we have \(\nu(ab)\leq\nu(a)\).
- (Euclidean division) For any \(a,b\in R\) with \(b\neq0\) nonzero, there are \(q,r\in R\) such that \(a=bq+r\) and either \(r=0\) or \(\nu(r)\lt\nu(b)\).
2.1. Long Division
We can use a naive long division algorithm.
(in-package :polynomial) (defun divide (a b &aux q r s) "When A and B are polynomials over a field, division uses this method." (declare (type polynomial a b)) (assert (not (zero? b))) (setf q (zero-polynomial)) (setf r a) (loop while (and (not (zero? r)) (>= (degree r) (degree b))) do (progn (setf s (* (/ (leading-coef r) (leading-coef b)) (expt 'x (- (degree r) (degree b))))) (setf q (+ q s)) (setf r (- r (* b s))))) ;; (assert (equal? a (+ (* q b) r))) (values q r))
Arguably we could implement the quotient and remainder methods using this.
(in-package :polynomial) (defmethod quotient ((a polynomial) (b polynomial)) (multiple-values-bind (q r) (divide a b) q)) (defmethod rem ((a polynomial) (b polynomial)) (multiple-values-bind (q r) (divide a b) r))
2.2. Pseudo-Division
When working over an integral domain \(D\), we can't necessarily compute the quotient and remainder of polynomials. But we can compute the pseudo-quotient and pseudo-remainder (really, they're more like quasi-quotient and quasi-remainder, since "pseudo" implies fraudulent, whereas "quasi" implies "they share similar properties").
The pseudo-quotient is denoted \(\mathrm{pquo}(A,B)\) and pseudo-remainder denoted \(\mathrm{prem}(A,B)\) by Bronstein.
(in-package :polynomial) (defun pseudo-divide (a b &aux c n q r s) (declare (type polynomial a b)) (assert (not (zero? b))) (setf c (leading-coef b)) (setf n (1+ (- (degree a) (degree b)))) (setf q 0) (setf r a) (loop while (and (not (zero? r)) (>= (- (degree r) (degree b)) 0)) do (progn (setf s (* (leading-coef r) (expt 'x (- (degree r) (degree b))))) (decf n) (setf q (+ s (* c q))) (setf r (- (* c r) (* s b))))) (let ((factor (expt c n))) (values (* factor q) (* factor r))))
2.3. Euclidean Algorithm
There are several variations of Euclid's algorithm for GCD with polynomials.
(in-package :polynomial) (defun euclid (a b &key (divide #'poly-divide) &aux x y) (declare (type polynomial a b)) (setf x a) (setf y b) (loop while (not (zero? y)) do (multiple-values-bind (q r) (funcall #'divide x y) (setf x y) (setf y r))) x)
2.3.1. Extended Euclidean Algorithm
The "follow your nose" implementation of the extended Euclidean algorithm could be implemented quite quickly.
(in-package :polynomial) (defun naive-extended-euclid (a b &key (divide #'poly-divide) &aux r1 r2) (declare (type polynomial a b)) (let ((a0 a) (a1 1) (a2 0) (b0 b) (b1 0) (b2 1)) (loop while (not (zero? b0)) do (multiple-values-bind (q r) (funcall #'divide a0 b0) ;; (= a0 (+ r (* b0 q))) (setf a0 b0) (setf b0 r) (setf r1 (- a1 (* q b1))) (setf r2 (- a2 (* q b2))) (setf a1 b1) (setf a2 b2) (setf b1 r1) (setf b2 r2))) ;; (assert (equal? (+ (* a1 a) (* a2 b)) a0)) (values a1 a2 a)))
2.3.2. Half-Extended Euclidean
Alas, if we need only one of the coefficients s or t, we can use a
variant of the Euclidean algorithm which Bronstein calls the
"half-extended Euclidean algorithm" (over some Euclidean domain D):
(in-package :polynomial) (defun half-extended-euclid (a b &key (divide #'poly-divide) &aux r1) (let ((a0 a) (a1 1) (b0 b) (b1 0)) (loop while (not (zero? b0)) do (multiple-values-bind (q r) (funcall #'divide a0 b0) ;; (= a0 (+ r (* b0 q))) (setf a0 b0) (setf b0 r) (setf r1 (- a1 (* q b1))) (setf a1 b1) (setf b1 r1))) (values a1 a)))
2.3.3. Euclidean Algorithm
We can use this half-extended Euclidean algorithm for a more efficient implementation of Euclid's algorithm.
(in-package :polynomial) (defun extended-euclid (a b &key (divide #'poly-divide) &aux u v) (declare (type polynomial a b)) (multiple-values-bind (s g) (half-extended-euclid a b :divide #'divide) (values s (funcall #'divide (- g (* s a)) b) g)))
2.3.4. Diophantine version of Extended Euclidean
We can use the extended Euclidean algorithm to solve Diophantine equations of the form \(sa+tb=c\) where \(a,b,c\in D\) are given and \(s,t\in D\) are the unknowns.
(defun diophantine-naive-extended-euclid (a b c) (multiple-values-bind (s u g) (extended-euclid a b) (multiple-values-bind (q r) (poly-divide c g) (unless (zero? r) (error "~A is not generated in the ideal by ~A and ~A" c a b)) (setf s (* q s)) (setf u (* q u))) (when (and (not (zero? s)) (>= (degree s) (degree b))) (multiple-values-bind (q r) (poly-divide s b) (setf s r) (setf u (+ u (* q a))))) (values s u))
2.3.5. Diophantine Half-Extended Euclidean
When we only want the s in the Euclidean domain such that s*a = c
modulo b and either s = 0 or nu(s) < nu(b). The other solution is
obtained by
where the division is always exact.
(defun diophantine-half-extended-euclidean (a b c) (multiple-values-bind (s g) (half-extended-euclidean a b) (multiple-values-bind (q r) (poly-divide c g) (unless (zero? r) (error "~A is not the ideal generated by ~A and ~A" c a b)) (setf s (* q s))) (when (and (not (zero? s)) (>= (degree s) (degree b))) (multiple-values-bind (q r) (poly-divide s b) ;; (assert (equal? s (+ r (* b q)))) (setf s r))) s)
This gives us a more efficient diophantine extended Euclidean method
(defun diophantine-extended-euclidean (a b c) (let ((s (diophantine-half-extended-euclidean a b c))) (multiple-values-bind (q r) (poly-divide (- c (* s a)) b)) (values s q)))
2.3.6. Partial Fraction Decomposition
We can use Euclidean Algorithm to determine the partial fraction decomposition of a ratio of polynomials. For \(a/d\) we consider the factorization \(d = d_{1}\dots d_{n}\) into distinct coprime (not necessarily irreducible) factors \(\gcd(d_{i},d_{j})=1\) for \(i\neq j\).
(defun distinct-partial-fraction (a &rest denom-factors) "Returns a list of a constant term followed by numerators for the corresponding denominator factors." (if (null denom-factors) a (multiple-values-bind (a0 r) (poly-divide a (apply #'* denom-factors)) (when (singleton? denom-factors) ; n = 1 (return-from distinct-partial-fraction (cons a0 r))) (multiple-values-bind ; n > 1 (a1 s) (extended-euclidean (apply #'* (cdr denom-factors)) r) (assert (< (degree a1) (degree (car denom-factors)))) (let ((numerators (apply #'distinct-partial-fraction s (cdr denom-factors)))) (cons (+ a0 (car numerators)) (cons a1 (cdr numerators))))))))
2.3.7. Partial Fraction Decomposition with Multiplicity
We explicitly assumed each factor \(d_{i}\) was coprime with each other. But what happens if \(d\) has a square factor \(d_{i}^{2}\)? Or more generally a factor \(d_{j}^{e_{j}}\) with nonzero \(e_{j}\in\mathbb{N}\)? How do we do a partial fraction decomposition in this case?
The idea is to use the so-called d-adic expansion of \(a/d^{m}\), writing \(a=dq+a_{m}\) by Euclidean division (where either \(a_{m}=0\) or \(\nu(a_{m})\lt\nu(d)\)). Then
\begin{equation} \frac{a}{d^{m}} = \frac{dq + a_{m}}{d^{m}} = \frac{q}{d^{m-1}} + \frac{a_{m}}{d^{m}}. \end{equation}If \(m=1\), we're done. Otherwise, we recursively find \(a_{0}\), \(a_{1}\), …, \(a_{m-1}\in D\) such that either \(a_{j}=0\) or \(\nu(a_{j})\lt\nu(d)\) for \(j\geq1\), and
\begin{equation} \frac{q}{d^{m-1}} = a_{0} + \sum^{m-1}_{j=1}\frac{a_{j}}{d^{j}}. \end{equation}Thus
\begin{equation} \frac{a}{d^{m}} = \frac{q}{d^{m-1}} + \frac{a_{m}}{d^{m}} = a_{0} + \sum^{m}_{j=1}\frac{a_{j}}{d^{j}}. \end{equation}Now if we have a factorization of \(d=d_{1}^{e_{1}}d_{2}^{e_{2}}\dots d_{n}^{e_{n}}\) (not necessarily into irreducibles) where \(\gcd(d_{i},d_{j})=1\) for any \(i\neq j\) and the \(e_{i}\in\mathbb{N}\). What to do? We take \(b_{i} = d_{i}^{e_{i}}\) for each \(i\), we use the partial fraction decomposition from the previous section
\begin{equation} \frac{a}{d} = a_{0} + \sum^{n}_{j=1}\frac{a_{j}}{b_{j}} = a_{0} + \sum^{n}_{j=1}\frac{a_{j}}{d_{j}^{e_{j}}}. \end{equation}And we use the \(d_{j}\)-adic expansion of each summand to get
\begin{equation} \frac{a}{d} = \widetilde{a} + \sum^{n}_{i=1}\sum^{e_{i}}_{j=1}\frac{a_{ij}}{d_{i}^{j}}. \end{equation}This is the complete partial fraction decomposition of \(a/d\) with respect to the factorization \(d = \prod_{i} d_{i}^{e_{i}}\).
(defun complete-partial-fraction (a &rest denom-expt-pairs) (let* ((all-numerators (apply #'distinct-partial-fraction (cons a (mapcar #'(lambda (d-e) (expt (car d-e) (cadr d-e))) denom-expt-pairs)))) (a0 (car all-numerators)) (numerators (cdr all-numerators))) (do ((d-e-pairs denom-expt-pairs)) () (do (( ))
3. Canonical Representation
Usually we have a field of rational polynomial \(\mathbb{k}=\mathbb{Q}(x)\) with \(D=\D/\D x\), and \(t\) a monomial over \(\mathbb{k}\) such that
\begin{equation} Dt = -t^{2}-\frac{3}{2x}t + \frac{1}{2x} \end{equation}Then we'll want to split a given polynomial
(defun split-factor (p deriv) "Given a derivation `deriv' on `k[t]' and `p' an element of `k[t]', return the pair (normal-polynomial . special-polynomial) such that (= p (* normal-polynomial special-polynomial))." (let ((s (/ (gcd p (deriv p)) (gcd p (d p t))))) (if (zerop (degree s)) (cons p 1) (split-factor (/ p s) deriv))))
4. References
- Manuel Bronstein, Symbolic Integration 1: Transcendental Functions. Springer, second ed., 1996.
4.1. Also See
- Marius van Der Pu, Michael F. Singer, Galois Theory of Linear Differential Equations. Springer, 2003.