112. Weighing the Knife's Share

(Epistemic status: The math holds up, and the one real theorem fits on a frosting-stained napkin. The rest is a relaxed palate-cleanser of a post... mostly. CW: notes of buttercream, ash, and... something like iron? For PR, CL, and both AS's.)

 

"I cut, you choose." It's a tale as old as siblings made to share cake - or land. The problem's a deceptively simple one: if you have two (or more) agents that both want the same valuable object, what procedure should they use, in order to ensure that all parties involved feel like they got a fair slice? The solution - for those among today's lucky 10,000 - is as elegant as it is brilliant; perhaps among the first pieces of explicit mechanism design: the cutter, knowing the chooser will then pick, has every incentive to cut two pieces she values equally; the chooser then takes whichever piece he likes better. She can't reasonably envy him, because she was indifferent by construction; he can't envy her, because he chose freely. Neither would trade. For two people and one knife, envy dies in a single cut.

For the more general case, the constraints and problem statement are the natural ones: the cake is of fixed mass/area/stuff-ness, and is rivalrous; each agent must feel, by its own lights, like it got at least \(\frac{1}{n}\) of the value of the cake, where \(n\) is the number of agents; those agents' lights  look like maps from subsets of the cake to numbers, and those functions are continuous. For the moment, we'll also assume that the knives are fixed: there is a single cake-cutter who decides where the cake is to be cut, and that's the end of it; there's some protocols that call for a continuously moving knife and some variety of Dutch auction-type "cry first and get your way" mechanism design, but we'll leave those aside for now. There's also some complicated measure-theoretic things to say here, but they extremely do not matter for our purposes. Probably you should also make use of the whole cake, too, but there's a class of deeply unaesthetic solutions that end up throwing away parts of the cake. Horrid.

For the sake of simplicity, we'll assume that it's an infinitely divisible object, with valuation pretty much linear in the amount of any given natural category within this singular and heterogeneous resource. More concretely: you can cut a cake (unlike gems), you can cut it wherever you like and however you like (unlike some kinds of borders, or indeed the generally more restrictive case of connected pieces), and getting more is smoothly better, with no weird dependencies (unlike other kinds of borders, spectrum auctions, package deals... look, mechanism design just keeps on going, alright?). On the other hand, we need not assume that the cake has been blended smoothly for some godsforsaken reason: it can have frosting, layers, marzipan, jam, bits of candied fruit, a candle, letters, a picture of someone - both face and butt - et cetera, et cetera, et cetera. (Wait, who invited Yul Brynner to this birthday party?) Different people are allowed to like different parts of the cake differently. Indeed, we're counting on that! Because here's the punchline to the novice's version of the problem; two arbitrary agents frictionlessly cutting a spherical cake in a vacuum: even though cake-mass must needs be conserved, both agents can feel like they got more than half the value of the cake. For a simple example, suppose we're just the two of us cutting a cake; I love frosting and jam, but I find the cake proper a little boring and the marzipan would poison me. Meanwhile, you find the sweetness a little cloying, and love marzipan. Any good division algorithm would ensure that you end up getting to crunch the marzipan while I lick jam off my fingers; indeed, under the classic assumption of utility maximization under bounded rationality, whoever cuts should separate the cake into a part with all of the marzipan and a part with most of the jam and frosting, and try to split the cakey part evenly. It's worth noting that there's more than one possible notion of "fairness" at play here: if what we want for our operationalization of what it means for a set of slices to be fair is "proportionality", then we take that to mean that for each cake-sharer, that agent ends up with at least \(\frac{1}{n}\) of the value of the cake by their own lights; the more famous property - "envy-freeness", which is the one we mostly care about here - means that when the crumbs settle, no agent would prefer another agent's piece to their own. Envy-freeness implies proportionality, but the converse fails immediately the moment there's \(n \geq 3\), where one agent could hold a perfectly fair third and still covet another agent's as strictly better, regardless of how well they chose.

There's two problems with this pleasingly simple picture. First off, how does this even work when there's more than two cake-desirous agents in the mix? The two-agent case of a single cut has existed from time out of memory, but the three-agent case (requiring five cuts) had to wait until Selfridge in 1960 (and later, independently, the late lamented Conway), and there wasn't even a bounded algorithm for four or more agents until Aziz and Mackenzie in 2016. Procaccia proved a lower bound for constructive algorithms of \(\Omega(n^2)\) that same year, while Aziz and Mackenzie's construction gives us a cut count of - brace yourself - n^n^n^n^n^n. I can't even make Blogspot's LaTeX accommodate that, except as \(n\uparrow\uparrow 6\). It's an embarrassment, is what it is, if a satisfying one. The worst part of it is, all of this only applies to the algorithmic case: Sperner's lemma tells us - by way of Stromquist and Simmons-Su - that some envy-free and contiguous cake division with the absolute bare minimum of \(n-1\) cuts always exists - but by Stromquist's same token, no finite protocol will ever find such an elegant set of cuts. As ever, Nature mocks us, and all Mathematics can do is play the interpreter.

The second problem with this simplified picture is this: cakes are made of material. Even fudge would give us an easier time; the problem with cake is that it crumbles. Frosting sticks to the knife; crumbs fall from the loaf. Every cut you make ruins some of the cake. With a little thought, we can even model just how much.

Most simply, we can draw on a concept from woodworking: "kerf". When you cut a piece of material - like wood - some of the material is destroyed: turned to chips, or sawdust, or vapor. Woodworkers budget for kerf explicitly when they make their cut lists. The name for the lost material is "kerf". We can thus put in a guess that the ending cake-measure will be \(1 - kw\), where \(k\) is the number of cuts, and \(w\) is the (assumed constant) width around any given cut that a cut will ruin.

But we can already do better, and if you're anything like me, you were nodding along irritably with that last paragraph asking why we're making such stupid approximation choices. How about a perimeter-adhesive model? Now our estimated loss is proportional to the product of [cut surface area] and [film thickness], and for a cake, this makes losses proportional to the total length of all cuts; swap the meaning of \(k\) from "number of cuts" to "total cut length" (or indeed, area, since cakes are three-dimensional), and our expression \(1 - kw\) for final cake-measure gets more precise. Weirder cuts cost us more, as they should; fair division now takes on - as it should - a constraint corresponding to something like constrained isoperimetry. Has anyone asked themself about the price of envy-freeness in terms of perimeter yet? I don't think so, and that's what's been on my mind for a reason to write about this at all.

We can do better yet. Thus far we've had to appeal to the absolute size of a cake and not thought much about the size of any given piece being cut, and we haven't thought at all about the nature of existing envy-free division algorithms. Consider that for a piece \(P\) with bulk \(b(P)\) and girth \(g(P)\), the relative loss per cut is something like \(\frac{wg}{b}\), where by "bulk" I mean something like area or volume or measure or the like, and by girth I mean perimeter or surface or boundary-measure or something like that. The important thing is, all of the exactly envy-free algorithms recurse on trimmings, cutting and trimming away and cutting some more, and they thus hit a wall when the bulk of the trimmings shrinks to about the same as the product of kerf-width and piece-girth - that is, when the blade is pretty close to the entire thing you're cutting - and the blade wrecks some constant (and in such cases in fact large) fraction of what it touches. Take the case of an ordinary 8-inch round cake, with volume ~125 in^3. If we put in a reasonable guess for the kerf width of a tenth of an inch on both sides and recall that the cake's height is something like 2.5 inches, that gives us a budget of ~250 inches worth of cutting, such that if the usual cut is something like 4 inches long, we get a budget of maybe 63 cuts, total, ever. This sounds irrelevantly large until you realize that for the exact-EF algorithm for 4 agents, the best envy-free division procedure candidate - the Aziz-Mackenzie procedure - requires 203 cuts over 584 total queries. Not only will everyone's knife arm get tired, the entire cake will be shredded paste by the end. Let's try not to think too hard about the chaos of a normal-sized birthday party.

That actually brings me to what seems at first blush to be a nitpick: once we start to take kerf into account, it is entirely possible that no fair division exist at all. For a trivial witness, take the case of a perfectly homogeneous cake: any cut at all will ruin some of the cake as kerf, and now however you slice it - pun intended - the slices sum to less than a whole cake. We might patch that by discounting what gets divided by \(n\) to recover a reasonable equivalent. As for envy-freeness, what started off as a more stringent condition than proportionality now becomes a quietly bleak one: no need to scale down, if all you care about is thinking no other agent got the better of you. Indeed, in the kerf-wise setting, there is, equally trivially, always an envy-free solution: a single cut of sufficient length, approximating a space-filling curve. The knife slices on and on, and nothing remains but crumbs, vapor, and bitter recriminations. No one receives anything; no one envies anyone else; everyone goes hungry. (Bad end.) I don't know about you, but that doesn't seem like a successful birthday party to me. Once cutting costs you cake, the objective must correctly value how much cake remains, not just how many cuts or queries it took. But at least the pyrrhic victory that considering kerf brings us illuminates something that envy-free division often leaves tacit: the fact that we generally presuppose some kind of geometry. Indeed, running the same kind of kerf-logic on fair division problems set apart from metric spaces becomes much more difficult - though it's precisely metric spaces where nearly all such problems of interest live, even more abstract divisions on transmission spectra, compute allocation, and inheritance divisions. And I should confess that the cake-wasting solutions I called "unaesthetic" and "horrid" earlier on are found in the literature as a deliberate design choice: Segal-Halevi, Hassidim, and Aumann buy their bounded envy-free protocols precisely by way of throwing cake away on purpose - "Waste Makes Haste", and they mean it. The professionals gave up on the whole cake years ago; they simply called the kerf a feature.

Classic envy-free division algorithms actually tend to focus on not kerf cost but rather query count: how many times do you have to ask "where do you want to cut" or "how much do you like this piece". The Robertson-Webb query model, as it's called, charges primarily for something like how long the cake-cutting takes. This is even a reasonable measure: if you model the cuts as delivered by an idealized machine, costing the quarrelling partygoers no more effort than is required to think about and then specify cuts, then we may as well treat a query and an actual cut as costing the same stroke - the common scale that all algorithms are graded on. Nonetheless, I insist: we live in a world where things are made of stuff and cutting that stuff takes significantly more effort than staring at the stuff and thinking about how much we like it and maybe even marking it gently with a toothpick. 

If you squint a little, even the more classical measure acknowledges as much: if we count deliberation, measurement, and preference-elicitation at all is to admit to the boundedness of the deliberators, and bounded deliberators do not place infinitely precise cuts. Ultimately, the ledger and the blade agree on what exactness costs - they just differ in how they deliver the bill. The appetizingly deep part is, counting in terms of cuts and maybe also cut complexity is close to - if not just plain identical to - one of the kerf metrics described above. The whole assemblage of envy-free cake-division algorithms reorders, if we stop worrying so much about the deliberation and concern ourselves only with the actual effort, the actual cuts, the icing smearing the knife and the crumbs falling away. You can run any protocol in the literature you like in simulate-then-cut mode: start off by deliberating the entire allocation on paper, napkins, or a group chat, and only once done, you execute - physically - the boundaries of the final answer. Under kerf accounting, a protocol's true cost is its final fragmentation, and the whole zoo reorders. Merely-proportional division was always cheap and stays that way: a floor of \(n-1\) cuts, and Even-Paz will walk you there without drama. If you want to push it to \(\epsilon\)-envy-freeness - where maybe someone covets another's piece, just by a crumb - then the cost in cuts stays near that same floor. (Deliberation may still be costly, but at least it doesn't cost us in cake.) It's exact envy-freeness that diverges, of course, and it diverges in your choice of currency, so pick your poison. Choice one: \(n-1\) cuts bought with unbounded deliberation; this is the purview of Stromquist's moving-knife procedure, with four blades hovering over a continuum of infinitesimal decisions and only two cuts ever landing. It's physically optimal but informationally transfinite. Choice two: bounded deliberation, bought with an allocation shattered into astronomical shards as above: recursion-flavored algorithms mean cut-complexity balloons out of control. Time or pieces: something always diverges; something has to give.

But we should take a step back to note that exact envy-freeness is not what's called for here. The physical world has created the problem and dissolves it in the same stroke: a real blade lands with imprecision - that of its own width, let alone the unsteadiness of your hand. Every boundary you might think to cut along thus arrives pre-perturbed; even a nominally exact protocol certifies envy only down to some floor \(\epsilon^\ast\). But we must embrace that every single element of our setup - knife, cake, agents, crumbs - are made of crass matter, and for once, that is a delight: just as we cannot cut a cake to infinite precision, so too does the fine structure of the cake itself remove the possibility of anyone's ever litigating an exact grievance. At the finest scales, the tiny bubbles that baking powder leaves are nothing so much as the sigh of exasperation that the physical world has for the keen-eyed mathematician; if not that, then the grains of sugar of the frosting; if not that, then the wavelength of light; if not that, then the atoms themselves. And overlaying all of it is hedonic indistinguishability; while we may not have measured the just-noticeable-difference of masses of cake, I can assure you that it must be best measured in grams or perhaps milligrams. (Certainly not micrograms.) Admittedly, the unsteadiness of a cake-server far outstrips all of these, and even outstrips the ability of a suspicious child's eye for fairness; people can see fair reason for envy that a knife-wielding server would have trouble eliminating. But the escapes here are well-worn: random selection by way of spinner, more usually a compensatory dab of frosting (a side-payment small enough to fit on a finger tip), and, of course, the birthday celebrant prerogative to pick first and absorb and negate all envy. We brush away the need for exceeding precision on the level of preferences, too: if your partygoers' preferences are the kind of reasonably well-behaved simple desires for frosting over sprinkles, or even for corner pieces over middle, then the exponential tower generally collapses and decent approximations of small exact protocols do you just fine - it's only when truly tangled-up preferences come into play that the tower of exponents looms to be climbed.

On that note, the theorem I promised up front; like I said, the condensed version fit on a frosting-stained napkin. The proof won't, but take it as read that I'd be too busy with cake and celebration to unroll it in full, anyway; this description wouldn't, either, but it's not what I'd be jotting down while I described it. Suppose we take the cake and cut it semi-arbitrarily into \(K\) cells; let's say they're strips, or we're working over the interval cake, or something. (Semi-arbitrary here means: blind to everyone's preferences is fine, but keep the cells roughly even - the bound below requires that no single cell be measure-large, and roughly even-sized cutting delivers that for free.) We discretize on purpose, because it'll let us import all the tools from the indivisible/discontinuous branch of envy-free division we glossed over at the start - most notably, the part about envy-freeness up to one piece. This also means that we need to allocate the cells round-robin, as will allow for the envy to be bounded above by the worth of a single piece. We'll pretend for a moment that kerf losses are total destruction rather than messy wreckage usable for side-payments; the total losses to kerf go like \(Kw\), given that you'll need roughly \(K\) cuts to make up \(K\) pieces, at least for a cake we're treating as linear. (The story is actually better for both cakes and land - there, it'd be closer to \(\sqrt{K}\). If we were divvying up chunks of space - merfolk splitting parts of the ocean depths, say, or miners sussing out mineral rights, or interstellar societies carving up areas of space - then it goes like \(\sqrt[3]{K}\), of course.) Let \(\rho\) be a measure of the most any of the agents assigns any given piece of the cake, measured in utils per absolute measure: \(\rho = 1\) for a perfectly equanimous agent who just wants some cake, any cake, and \(\rho\) is huge if someone really really wants the (small) candle. A cell is then worth at most \(\frac{\rho}{K}\), which is therefore the upper bound on the envy. Let \(\lambda\) be how much you care about any envy that might arise, in terms of how much physical cake you'd throw out if it erased one util of worst-case resentment; we then get an objective function to minimize to figure out how many pieces to cut the cake into: \(L(K) := \lambda \cdot \frac{\rho}{K} + Kw\). (\(K \geq n\), obviously, and this will basically always be true in real life with real cakes unless you're already badly underprovisioned anyway.) But we can use our usual bag of tricks here: the AM-GM inequality says that for any \(a, b \in \mathbb{R}^+\), we always have \(a + b \geq 2 \sqrt{ab}\), with equality when \(a = b\). And delightfully, our product is a constant: \(\lambda \cdot \frac{\rho}{K} \cdot Kw = \lambda \rho w\). Optima are found at extremes only when they can't be found at points of symmetry, and a point of symmetry presents itself: setting \(\lambda \cdot \frac{\rho}{K} = Kw\), we get \(K^\ast = \sqrt{\frac{\lambda \rho}{w}}\), with a total inefficiency of something like \(2\sqrt{\lambda \rho w}\). 

A couple of caveats that I brought up briefly there and moved on from: an atom - a candle loved as a point object, a tanghulu strawberry that can't really be cut - makes \(\rho\) infinite and thus the bound useless, and then lacking a birthday celebrant to hand it to to distribute as they see fit, we're back to random allocation or side payments. This seems reasonable: no knife can ever bound envy over an indivisible. For another, the math changes if you've got more dimensions to work with, but not all that qualitatively: boundary is much cheaper in higher dimensions, shifting the law from square root to cube root in a truer-to-life two-dimensional case, and to a fourth root in that offhanded three-dimensional case, and so on for higher dimensions of Euclidean space. (Now there's a question: what does all this look like in hyperbolic space, with its exponential circumference and area both!)

So there's the answer to the question the title implies: a square root. Not even all that bad. The number of cuts that fairness is worth grows as the square root or cube root of how much you care, divided out by how much the knife wrecks. Exact protocols are the backup plan, if crumbs are worth wasting and your risk aversion times the spinner's variance outweigh that \(2\sqrt{\lambda \rho w}\) and a ruined party from an afternoon of argument. But frankly? For household values of \(n\), household knives, and the kind of people with reasonable maturity and preferences you'd even want at your party... this basically never pays off. Which is, of course, why no birthday party has ever run Aziz-Mackenzie except as a very elaborate mathematician's joke, and why the classic grandmother strategy - cut a bunch of squares, eyeball it, have everyone line up to take the single piece they want with no seconds until everyone's had firsts - was implementing the optimal mechanism the whole time. (To say nothing of a child's lack of reflectivity on the precipitously declining marginal utility of cake.) Best of all, notice what her protocol bills in the Robertson-Webb ledger: zero queries. Nobody asks anyone anything; every scrap of preference information the mechanism uses arrives embodied in the picks themselves. Cut evenly, take turns, and watch your hands: the deliberation cost hasn't been pushed under a rug - it's all but vanished.

And that would be the ending: grandmother vindicated, informal tradition lauded with a touch more rigor and mechanism design, and everyone home by dark with awful stomachaches from too much cake anyway... except that I told you up front what this post was about, and it was never just about cake. "Siblings made to share cake, or land." I've been talking cheerfully and whimsically about cake, but I've been telling you the land version the whole time. Go back and reread all of that in light of the fact that only the nouns change; the strategies won't.

Start with why the expensive-seeming condition was the load-bearing one. Proportionality has never once prevented a war on its own. No state in history overturned the bargaining table because it received less than \(\frac{1}{n}\) by its own valuation; armies get sent to die because of what a neighbor received, and whether it holds up against the balance of force. Envy - irredentist or state viability or simple power politics, pick your favorite flavor - is the casus belli, which is why envy-freeness - not mere fairness - is the condition worth pricing, and why its price matters more than any theorem about dessert.

Then there's the kerf. Oh, the kerf. Lines on maps look so very clean, and just like mathematicians ignore the kerf for cake division, so too is it so very tempting for geopolitical wonks to ignore the kerf that the maps bill only tacitly, in a voice that can always be ignored at the table and never on the ground. But bill it does, annually and forever: garrisons, customs houses and strip-searches and confiscations, walls and their maintenance, the dead strip where nothing is permitted to grow, the restriction of free movement of people and goods and ideas, the exercise of arbitrarily vast state violence inflicted on the many in the name of a terribly lossy abstraction. Crinkliness is the isoperimetric tax written into policy - every salient, every gerrymander, every proud little peninsula of the pen lengthens the invoice forever. Cooch-Behar was a nightmare until it was simplified; if the Schengen area wasn't around, Baarle-Hertog and Baarle-Nassau would be much the same. You want the kerf drawn at 1:1 scale? Four kilometers wide, two-hundred-fifty kilometers long, seventy-three years old a month ago, littered with mines so as to be uncrossable, any number of families torn apart with no hope of reunion any longer. (Don't ask me how I know.) And the other failure mode is speed; speed in the name of expedience and finality and wanting to be done with a rebellious colony and frankly who really cares about them anyway, they're not white. That's how you get a line drawn across a subcontinent in seven weeks by a man who'd never once in his life been east of Paris with partition declared before any of the plans were finished and most of the important questions punted. "The maps at his disposal were out of date / And the Census Returns almost certainly incorrect, / But there was no time to check them, no time to inspect / Contested areas." The kerf there? A million lives ended and ten times as many ruined, at the minimum, and then two decades later, another million dead and who can say how many more ruined, and now and forever one more potential geopolitical flashpoint that no one needs. Quick and dirty: quick is past.

 

It gets worse. I fed you a simplification when the subject was still dessert; I let you believe that kerf dies on the blade, can even become fodder for side payments, and in the kitchen it largely does. The same is not true on maps. The value scraped off of a border is always transferred at a loss: to garrisons, checkpoint economies, smugglers, detention centers, the whole industry of the edge. And an industry, once fed, acquires opinions; people who feed off that kerf will reliably vote for more knife. A border does not merely cost once, nor does it cost some clean annual fee. A border lobbies.

And that brings us right back to that space-filling cut, that silly Peano curve that I let you laugh at. There is always a single partition that ends envy outright: the lossy partition that leaves behind nothing worth envying. You want envy-free division? You have it. We all have it. Nearly everyone is sick to death of it and driven half-mad and has been for the last 70 years with no end in sight. Our leaders have built the knives for such a cut, and we keep them sharpened, for use at moment's notice, and some are day and night trying to figure out how to forge new blades. (Bad end - the real deal, this time.)

And it turns out, this is just plain difficult to solve. You can't simply run the grandmother-protocol above on maps: the cheap fairness we bought in the kitchen pays out in disparate chunks of cake, and you cannot govern or defend disparate chunks, not with a world with borders as hard as this one's are. Land demands contiguous pieces - and that is exactly the constraint that keeps the exact-division tower standing over the map long after the kitchen has escaped it.

What protocol, then, manages to survive contact with land and power? Force the question and the literature answers you, quietly but matter-of-fact: contiguous \(\epsilon\)-envy-freeness. Simmons-Su will hand it to you for any \(\epsilon\) you care to name, in finitely many queries, at the bare minimum of \(n-1\) cuts, with every piece connected, at which point side payments close the \(\epsilon\). Read that twice and you realize that real diplomacy - at its best - already is the map-compatible protocol: a Sperner triangulation walked very slowly by exhausted lawyers, with a trade concession here or there serving for a dab of frosting. Coarse contiguous shares, then swaps, then compensation, then hope. Sometimes it even works.

So that square (or cube) root turns out to be something of an arithmetic of peace. Cut seldom: \(K^\ast\) is small when the accounting is honest about what cutting costs. Cut straight, or in the case of cakes which are varyingly hard to cut, cut with respect for that difficulty. And buy the remainder with the boring machinery your tired kindly grandparents were running all along: eyeballed shares, fair lotteries, compensation for the loudest and especially hungriest at the table. That dab of frosting, reread, is the land swap and the reparations fund and the reconciliation commission and the loosening of border controls: undignified, sticky, and cheaper than the knife by orders of magnitude. That's the whole doctrine right there, and it fits on the same napkin as the theorem: cut seldom, cut respectfully, and pay the difference in frosting. And for goodness' sake, actually end up with cake worth eating. 

Even the frosting has limits, though - there is a residue that no side-payment can ever denominate, and every item in my (miserably nonexhaustive) list is one of my opening assumptions coming due. Sea access: suddenly, preferences stop being additive; connectivity is exactly the package-deal dependency I ruled out for cake, now presenting its bill for land. Mineral rights: the candle, except subterranean - a value-atom that sends \(\rho\) to infinity straight through the middle of somebody's farm, with kerf all the worse for it. Symbolic territory: lexicographic value, for which no \(\lambda\) exists at all - there is no denomination of frosting that buys the holy city or the sacred mountain or the birthplace of significance, and cutting it in half just destroys the value for everyone. Deeper yet, and breaking the frame itself: enclaves of people. The cake contains its eaters; we never get to dodge embodied agency and trying will end badly. Human beings live in the pieces, so cutting the map cuts them - and "making the pieces contiguous" has a twentieth-century name that history politely renders as population transfer. The model has no room for a cake that votes.

And the more you check, the more you think carefully about the problem, the more broken the nice neat frame gets; each of the following is yet another opening assumption presenting its bill and threatening impossibility. (Students of international relations will recognize large parts of this paragraph and the previous one as Fearon's "Rationalist Explanations for War" decorated with frosting and my cake-kerf as the cost to pursue war in treasure, blood, and ruined land.) The assumption of honesty: Sperner's walk runs on honestly reported valuations, but states answer queries about as honestly as poker players do. Worse, my own doctrine loads this particular gun, because compensation for the loudest just incentivizes loudness, and the prospect of side payments just subsidizes the manufacture of grievances. Some fraction of irredentism is just invented \(\rho\). The assumption that the agents are unitary and persistent: that triangulation takes years to walk, at best, and the state that begins it is frequently not the state that finishes it. Governments turn over mid-protocol, successors repudiate signatures and rip up ratifications, and deeper still, whatever cut gets made then manufactures the next generation's valuations, revanchism being a preference baked by the blade, such that the value-landscape shifts underneath the fixed-point theorem while you are still inside it. (And we should know quite well what happens when you nudge around places where non-constructive fixed points live.) Not even a weaker version of the foregoing - the assumption that at least the number of players is fixed - survives: \(n\) is not given but contested and endogenous! Partition mints new claimants overnight, and worse, sometimes the cutters are not the eaters at all - Radcliffe held the knife while other people held the wounds. Who gets a chair is frequently what the war was about. International anarchy puts paid to the assumption of a fair and respected referee: cut-and-choose quietly presumes someone will or indeed can enforce the choosing, and under anarchy the output must be a self-enforcing equilibrium; the solution concept states actually run is not "envy-free" but "defensible", with war as everyone's standing outside option, and a fair division that is not incentive-compatible after the signing is a ceasefire wearing a treaty's clothes. The geometry assumption fails, and fails hard. The map has holes and layers to it; lakes, straits, and pre-existing states are the easiest cases of this, and even those make the contiguous pieces that Sperner promises be potentially trivially unavailable even topologically - just ask Kaliningrad or West Berlin. Worse, sovereignty is stratified: surface and subsurface and airspace and EEZ and sea-lane and half a dozen orbits, each layer with its own adjacency, so that "connected" isn't even a single predicate. The independence assumption fails so hard it makes the cake model look like the lies-to-undergrads it is: pieces keep interacting after allocation. A river gives the upstream party prime access to the downstream party's throat; just ask the states along the Colorado River... or the Nile. Or China and India. Aquifers, fisheries, and monsoons all cross any line you care to draw - Nature need not respect such boundaries - so the value of a piece is a function of who holds the others, circularly ill-defined at precisely the moment that the protocol needs to evaluate it. And last but not least - indeed, most tragically - the fixed-value assumption, deepest of all: cakes are additive and countries are superadditive. A severed railway is worth far less on both sides of the cut; a port without its hinterland is barely half a port. Sometimes the union was most of the value, and that means that every division - however fair, however contiguous - destroys as kerf a surplus that no reallocation recovers. "Cut seldom" has an optimal limiting case, and it is this: don't.

None of that repeals the doctrine, by the way; if anything, it prices it, and it makes me wonder how anything works at all. The boring machinery is not good because it works well, cleanly, fairly, kindly, or even reliably - it is good because every other mechanism on offer is a knife.

So... how does anything work at all? Mostly, it doesn't; it mostly doesn't, and the last half a millennium is an unending testament to the fact that Westphalian sovereignty was a terrible idea whose time had nonetheless come. The map, you see, is not a gallery of solutions, but a gallery of survivors. When a division fails and everyone starts squabbling over the cake, disagreeing over who should get how much and unable to talk to each other, you have the ingredients for a war, and the historical record documents enough wasted cake-kerf to give anyone diabetes. That which holds, holds for two grim foundational reasons. First off, the fact that some learn on hearing, some by watching, and all too many have to piss on the electric fence for themselves. The machinery above demands math, mechanism design, cultural sensitivity, deep care for humanity, and skilled diplomacy and bargaining both, topped with a relative lack of ideology; it's a rare bird who can love all of that for its own sake. No, those powers which are otherwise accustomed to getting their way adopt that machinery - when they are wise, and do - because the history books say what happens when you do not, and often because the kerf-cost of doing otherwise has billed all sides long enough and dearly enough that even a manufactured grievance can no longer pretend the arithmetic away. That envy constant \(\lambda\) is not chosen by the wise, but rather discovered by the bleeding and bereft, and the price for that information is paid in the only currency this post has been counting. Mathematics interprets, as it always will, but the knife instructs. Second, no real division ever terminates. All these exact protocols hold out the promise of a final state, but the world insists on running the anytime version: coarse shares, then maybe some swaps, then occasionally compensation, then hope, with no last step anywhere in the specification. The frosting is not a payment, but a subscription; peace is not achieved, but rather serviced. A treaty that stops being maintained - complete with walls that go back up, guards that get posted again, and tariffs that get abruptly raised - is simply people deciding they're willing to feed more cake into the shredder.

So, once more, with feeling: cut seldom, cut respectfully, and make sure to pay the difference in frosting - and then keep on paying it, on schedule, every year, long after everyone has forgotten why the schedule exists, until you can forget why there needed to be a cut at all. That is how anything works: not well, and not without hard lessons learnt first, and marginally better than the knife.

 

And at last, the tasting notes that I promised up front: the pleasant buttercream of elegant mathematics and mechanism design; the ash of the real-world suffering; and blood: both yours and theirs.