subdivide each edge into a path of length 3

Max-Cut-Multi ≤p Max-Cut

Max-Cut-Multi≤pMax-Cut\mathrm{Max\text{-}Cut\text{-}Multi}\le_p\mathrm{Max\text{-}Cut} — the hardness of max cut does not come from parallel edges; it survives on simple graphs.

The construction

Let G=(V,E)G=(V,E) be a multigraph. Define G′=(V′,E′)G'=(V',E'):

  • Vertices: V′=V∪V′′V'=V\cup V'', where each edge {u,v}∈E\{u,v\}\in E contributes two vertices u′,v′u',v' unique to that edge.
  • Edges: for each original edge {u,v}\{u,v\}, the three edges {u,u′},{u′,v′},{v′,v}.\{u,u'\},\quad\{u',v'\},\quad\{v',v\}.

So each edge is replaced by a path of length 33 through two private new vertices. The output is ⟨G′, 2∣E∣+k⟩\langle G',\ 2|E|+k\rangle.

G′G' is simple. Two parallel edges {u,v}\{u,v\} receive four distinct intermediate vertices, so their paths are internally disjoint and no parallel edges arise; a loop {u,u}\{u,u\} likewise becomes a path through two new vertices and ceases to be a loop.

Visualization

Guided construction

Max-Cut on multigraphs → simple graphs

Step 1 of 7
|E| = 4k = 3
Max-Cut on a multigraph reduced to Max-Cut on a simple graphOn the left a multigraph with two parallel edges and a loop; on the right the simple graph in which every edge became a path of length three through two private vertices, with the cut sides, the crossing edges and the parity cases marked.G — multigraphe₁e₂e₃e₄abcG′ — simple graphV′ starts as V
original vertex of V

Correctness — both directions

Max-Cut-Multi ∈ ⟹ Max-Cut ∈

Suppose GG has a cut (U,Uˉ)(U,\bar U) of size at least kk. Build a cut in G′G' with at least 2∣E∣+k2|E|+k edges.

Start from U′⊇UU'\supseteq U (old vertices keep their sides) and fix the intermediate vertices edge by edge. Let {u,v}∈E\{u,v\}\in E with intermediates u′,v′u',v'.

  • If the edge crosses (u,vu,v on different sides): put u′u' on vv's side and v′v' on uu's side. Then all three edges {u,u′},{u′,v′},{v′,v}\{u,u'\},\{u',v'\},\{v',v\} cross — contribution 33.
  • If the edge does not cross (u,vu,v on the same side): put u′u' on the opposite side and v′v' on u,vu,v's side. Then {u,u′}\{u,u'\} crosses, {u′,v′}\{u',v'\} crosses, and {v′,v}\{v',v\} does not — contribution 22.

So every edge contributes at least 22, and every crossing edge contributes 33. If the original cut had c≥kc\ge k crossing edges, the total is 2(∣E∣−c)+3c=2∣E∣+c  ≥  2∣E∣+k.2(|E|-c)+3c=2|E|+c\;\ge\;2|E|+k . Hence ⟨G′,2∣E∣+k⟩∈Max-Cut\langle G',2|E|+k\rangle\in\mathrm{Max\text{-}Cut}.

Max-Cut ∈ ⟹ Max-Cut-Multi ∈

Suppose G′G' has a cut (W,Wˉ)(W,\bar W) of size at least 2∣E∣+k2|E|+k. Define U=W∩VU=W\cap V — restrict the cut to the old vertices — and show the induced cut in GG has at least kk edges.

The local bound. Let {u,v}∈E\{u,v\}\in E with its intermediate path. Count how many of the path's three edges can cross, depending only on the sides of uu and vv:

  • if u,vu,v are on the same side of WW: the path u→u′→v′→vu\to u'\to v'\to v starts and ends on one side, so the number of crossings along it is even — that is 00 or 22. In particular at most 22.
  • if u,vu,v are on different sides: the number of crossings along the path is odd — 11 or 33. In particular at most 33.

(The argument is just tracking side changes along a path: the number of changes is even iff the two endpoints share a side.)

The count. Let cc be the number of edges of GG whose endpoints lie on different sides of WW — exactly the crossing edges of (U,Uˉ)(U,\bar U) in GG. The total number of crossing edges in G′G' is bounded by 2(∣E∣−c)+3c=2∣E∣+c.2(|E|-c)+3c=2|E|+c . On the other hand the cut in G′G' was assumed to have at least 2∣E∣+k2|E|+k edges. Hence 2∣E∣+k  ≤  2∣E∣+c⟹c  ≥  k.2|E|+k\;\le\;2|E|+c\quad\Longrightarrow\quad c\;\ge\;k .

So GG has a cut of size at least kk, i.e. ⟨G,k⟩∈Max-Cut-Multi\langle G,k\rangle\in\mathrm{Max\text{-}Cut\text{-}Multi}.

Cost of the reduction

TimeO(|V| + |E|)
SpaceO(|V| + |E|)
Output size|V| + 2|E| vertices, 3|E| edges

A single pass over the edges: two new vertices and three edges each. The threshold 2∣E∣+k2|E|+k is one addition. The output is linear in the input size.

What it buys

With the trivial reduction the other way (Max-Cut≤pMax-Cut-Multi\mathrm{Max\text{-}Cut}\le_p\mathrm{Max\text{-}Cut\text{-}Multi} — a simple graph is a multigraph), the two problems are polynomially equivalent. Hence Max-Cut\mathrm{Max\text{-}Cut} on simple graphs is NP-complete, even though the reduction from 3NAE\mathrm{3NAE} produces a multigraph.

Pitfalls

  • The intermediate vertices must be private to their edge. Were two parallel edges to share intermediates, G′G' would not be simple — which is the entire point of the construction.

  • The threshold is not kk but 2∣E∣+k2|E|+k. Every edge pays a "tax" of 22 regardless, and only the third crossing carries information. Forgetting the tax breaks both directions.

The chain

Where this was read from

  • 2026"Prove G′G' is simple, and that GG has a cut of size at least kk iff G′G' has one of size at least 2∣E∣+k2|E|+k."
  • 2024The same construction, phrased as "adding two new vertices per edge".
  • 2025The same construction on a multigraph with parallel edges and loops.
  • 2026Max-Cut-Simple\mathrm{Max\text{-}Cut\text{-}Simple} against the multigraph version.
  • This reduction does not appear in the 2026B lessons — lesson 10 stops at the multigraph and flags the question as a class exercise. It has been examined in at least four sittings, and the construction above is copied from the phrasing the papers themselves give.