分数规划是将某个求解最优性问题转化为判定性问题,一般的形式为f(x)=a(x)/b=(x),求解f(x)的最优值,其中a,b,x为连续实数函数。
我们这里讨论f的最小值,即设w=min(f(x)=a(x)/b(x)),那么w=f(xmin)=a(xmin)/b(xmin)=>a(xmin)-w*b(xmin)=0
那么我们设一个新的函数g(w),g(w)=min(a(x)-w*b(x)),首先我们可以得到这个函数的一些性质。
性质:
单减性:即对于任意w1<w2都有g(w1)>g(w2),证明比较容易,假设x1为w1的最优解,x2为w2的最优解,那么有
g(w1)=a(x1)-w1*b(x1)>a(x1)-w2*b(x1)>=a(x2)-w2*b(x2)=g(w2)
唯一性:当且仅当g(w)=0时,w为原问题的最优解。这一性质又被称作Dinkelbach定理。首先我们要证明这一性质,需要证明g(w)=0是w为最优解的充分必要条件。
首先证明必要性,即对于w为最优解时,g(w)=0。设最优解为wmin,那么我们有
wmin=a(xmin)/b(xmin)<=a(x)/b(x),也就是a(x)-wmin*b(x)>=0,因为g(w)是取最小值,所以g(wmin)=0
然后证明充分性,即对于g(w)=0时,w为原文题的最优解。
采用反证法,假设w’比w更优,那么w’=a(x')/b(x')<w,那么a(x')-w*b(x')<0,也就是g(w)<0,与题设g(w)=0不符。
这时候,我们可以通过求出来g(w)的值判断w是否为最优解,这样就将原问题的最优性问题转换为了判定性问题。
比如bzoj3232 http://61.187.179.132/JudgeOnline/problem.php?id=3232
这道题要求求v/c的最优值,这样不容易求解,但是我们可以用网络流最小割求出v-w*c的最小值,然后判断与0的关系,这样二分求解。
//By BLADEVIL const lim =1e-5; var n, m :longint; pre, other :array[0..100010] of longint; len :array[0..100010] of extended; last :array[0..3010] of longint; tot :longint; num :array[0..60,0..60] of longint; key, heng, shu :array[0..60,0..60] of longint; sum :longint; print :extended; que, d :array[0..3010] of longint; source, sink :longint; function min(a,b:extended):extended; begin if a>b then min:=b else min:=a; end; function judge(x:extended):extended; begin if abs(x)<lim then exit(0); if x<0 then exit(-1) else exit(1); end; procedure connect(x,y:longint;z:extended); begin inc(tot); pre[tot]:=last[x]; last[x]:=tot; other[tot]:=y; len[tot]:=z; end; procedure init; var i, j :longint; begin read(n,m); for i:=1 to n do for j:=1 to m do num[i,j]:=(i-1)*m+j; for i:=1 to n do for j:=1 to m do begin read(key[i,j]); sum:=sum+key[i,j]; end; for i:=1 to n+1 do for j:=1 to m do read(heng[i,j]); for i:=1 to n do for j:=1 to m+1 do read(shu[i,j]); source:=num[n,m]+2; sink:=source+1; end; function bfs:boolean; var q, p :longint; h, t, cur :longint; begin fillchar(d,sizeof(d),0); d[source]:=1; h:=0; t:=1; que[1]:=source; while h<t do begin inc(h); cur:=que[h]; q:=last[cur]; while q<>0 do begin p:=other[q]; if (judge(len[q])>0) and (d[p]=0) then begin inc(t); que[t]:=p; d[p]:=d[cur]+1; if p=sink then exit(true); end; q:=pre[q]; end; end; exit(false); end; function dinic(x:longint;flow:extended):extended; var rest, tmp :extended; q, p :longint; begin if x=sink then exit(flow); rest:=flow; q:=last[x]; while q<>0 do begin p:=other[q]; if (judge(len[q])>0) and (d[p]=d[x]+1) and (rest>0) then begin tmp:=dinic(p,min(rest,len[q])); rest:=rest-tmp; len[q]:=len[q]-tmp; len[q xor 1]:=len[q xor 1]+tmp; end; q:=pre[q]; end; exit(flow-rest); end; procedure main; var l, r, mid :extended; cur :longint; ans :extended; i, j :longint; begin l:=0; r:=90; while r-l>lim do begin mid:=(l+r)/2; fillchar(last,sizeof(last),0); tot:=1; for i:=1 to n do for j:=1 to m do begin connect(source,num[i,j],key[i,j]); connect(num[i,j],source,0); end; for i:=1 to n do for j:=1 to m do begin cur:=0; if i=1 then inc(cur,heng[i,j]); if i=n then inc(cur,heng[i+1,j]); if j=1 then inc(cur,shu[i,j]); if j=m then inc(cur,shu[i,j+1]); if cur>0 then begin connect(num[i,j],sink,cur*mid); connect(sink,num[i,j],0); end; end; for i:=1 to n-1 do for j:=1 to m do begin connect(num[i,j],num[i+1,j],heng[i+1,j]*mid); connect(num[i+1,j],num[i,j],heng[i+1,j]*mid); end; for i:=1 to n do for j:=1 to m-1 do begin connect(num[i,j],num[i,j+1],shu[i,j+1]*mid); connect(num[i,j+1],num[i,j],shu[i,j+1]*mid); end; ans:=0; while bfs do ans:=ans+dinic(source,maxlongint); if judge(sum-ans)>0 then l:=mid else r:=mid; end; writeln(l:0:3); end; begin init; main; end.