Let $f:X \to Y$ be a proper morphism between smooth algebraic varieties (say over $\mathbb{C}$), let me write $A_X$ for the constant sheaf on $X$ with coefficients of the appropriate type. Then the decomposition theorem of [Beilinson-Bernstein-Deligne] tells me that $f_* A_X$ splits as a direct sum of shifted semisimple perverse sheaves.

So suppose I have collected some of the summands, say into $\mathcal{F}$, and want to know if I have found everything. Invoking proper base change, it is enough to ask, for every point $y \in Y$, whether

$\dim H^*(y,\mathcal{F}_y) = \dim H^*(X_y)$,

since any missing summands would have to contribute something somewhere. But computing the RHS requires actually thinking about the topology of $X_y$. Something which is easier to compute are the virtual Betti numbers, defined for an arbitrary variety in terms of the weight filtration as

$b^j_{vir}(Z) = \sum_i (-1)^{i+j} \dim Gr^j_W H^i_c(Z)$

At first glance these may not look easier, but the point is that they are motivic, i.e., additive under cutting into constructible pieces (and multiplicative under Zariski locally trivial fibrations); evidently they agree with the ordinary Betti numbers for smooth proper varieties. Thus for instance since a rational curve with a single node can be reassembled into $\mathbb{A}^1$, it has a single nonvanishing virtual Betti number, $b^2_{vir} = 1$.

To determine whether $f_* A_X = \mathcal{F}$, the latter known to be a summand of the former, is it enough to compare virtual Betti numbers? To be precise, does it suffice to know, for every $y \in Y$ and all $j$, that $b^j_{vir}(X_y) = \sum_i (-1)^{i+j} \dim Gr^j_W H^i(y,\mathcal{F}_y)$?

I sketch an argument in the affirmative which unfortunately due to my lack of knowledge about what weights are, I cannot seem to make precise. It is enough to show that for any nonzero summand $G$ of $f_* A_X$, there is some $y \in Y$ where not all the virtual Betti numbers of $G_y$ vanish. Let me take $y$ which has a neighborhood where $G$ is a sum of shifted local systems supported on some smooth subvariety $Y'$, and here it seems to me that restricting to $y$ must act as some global shift on the weights. Since $G$ was pure to begin with, being a summand of $f_* A_X$, so is $G_y$ (up to a shift) and hence its virtual Betti numbers are its actual Betti numbers, which see $G$.

I have put the "reference request" tag in the hopes that if the answer to the question is yes, it is some well known thing, maybe somewhere in [BBD], to which someone can point me.

  • $\begingroup$ Maybe I'm missing something here, but haven't you reduced to the case where $G$ is a sum of (pure) variations of Hodge structure (shifted into appropriate degree) on some neighborhood of $y$ in $Y^\prime$. Taking the stalk of such a VHS at $y$ is a pure Hodge structure. I don't see where a change in weights comes in. $\endgroup$ Commented Jul 9, 2011 at 3:01
  • $\begingroup$ Ah, I see now you wanted to work in the BBD setting... $\endgroup$ Commented Jul 9, 2011 at 3:02
  • $\begingroup$ I'm happy to work in any setting where the answer to the question is yes. If it's true in the MHM setting that's certainly good enough for me... I just have even less idea how those work than the BBD formalism. $\endgroup$ Commented Jul 9, 2011 at 4:15

1 Answer 1


(I'll just note beforehand that is this is a "I don't know the answer to your actual question, perhaps you'd rather know the answer to a slightly different one" answer). The answer is yes if you replace virtual Betti number with Frobenius trace; these are also motivic, if admittedly slightly harder to calculate, as they involve counting points. This is covered, for example, in Section 3.4 of my second paper with Geordie Williamson.

  • $\begingroup$ Thanks Ben, that's indeed something I wanted to know as well. $\endgroup$ Commented Jul 9, 2011 at 4:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.